You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
rocksdb/utilities/backupable/backupable_db.cc

3017 lines
115 KiB

// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
// This source code is licensed under both the GPLv2 (found in the
// COPYING file in the root directory) and Apache 2.0 License
// (found in the LICENSE.Apache file in the root directory).
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
//
// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file. See the AUTHORS file for names of contributors.
#ifndef ROCKSDB_LITE
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
#include <algorithm>
#include <atomic>
#include <cinttypes>
#include <cstdlib>
#include <functional>
#include <future>
#include <limits>
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
#include <map>
#include <mutex>
#include <sstream>
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
#include <string>
#include <thread>
#include <unordered_map>
#include <unordered_set>
#include <vector>
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
Introduce a new storage specific Env API (#5761) Summary: The current Env API encompasses both storage/file operations, as well as OS related operations. Most of the APIs return a Status, which does not have enough metadata about an error, such as whether its retry-able or not, scope (i.e fault domain) of the error etc., that may be required in order to properly handle a storage error. The file APIs also do not provide enough control over the IO SLA, such as timeout, prioritization, hinting about placement and redundancy etc. This PR separates out the file/storage APIs from Env into a new FileSystem class. The APIs are updated to return an IOStatus with metadata about the error, as well as to take an IOOptions structure as input in order to allow more control over the IO. The user can set both ```options.env``` and ```options.file_system``` to specify that RocksDB should use the former for OS related operations and the latter for storage operations. Internally, a ```CompositeEnvWrapper``` has been introduced that inherits from ```Env``` and redirects individual methods to either an ```Env``` implementation or the ```FileSystem``` as appropriate. When options are sanitized during ```DB::Open```, ```options.env``` is replaced with a newly allocated ```CompositeEnvWrapper``` instance if both env and file_system have been specified. This way, the rest of the RocksDB code can continue to function as before. This PR also ports PosixEnv to the new API by splitting it into two - PosixEnv and PosixFileSystem. PosixEnv is defined as a sub-class of CompositeEnvWrapper, and threading/time functions are overridden with Posix specific implementations in order to avoid an extra level of indirection. The ```CompositeEnvWrapper``` translates ```IOStatus``` return code to ```Status```, and sets the severity to ```kSoftError``` if the io_status is retryable. The error handling code in RocksDB can then recover the DB automatically. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5761 Differential Revision: D18868376 Pulled By: anand1976 fbshipit-source-id: 39efe18a162ea746fabac6360ff529baba48486f
5 years ago
#include "env/composite_env_wrapper.h"
Make backups openable as read-only DBs (#8142) Summary: A current limitation of backups is that you don't know the exact database state of when the backup was taken. With this new feature, you can at least inspect the backup's DB state without restoring it by opening it as a read-only DB. Rather than add something like OpenAsReadOnlyDB to the BackupEngine API, which would inhibit opening stackable DB implementations read-only (if/when their APIs support it), we instead provide a DB name and Env that can be used to open as a read-only DB. Possible follow-up work: * Add a version of GetBackupInfo for a single backup. * Let CreateNewBackup return the BackupID of the newly-created backup. Implementation details: Refactored ChrootFileSystem to split off new base class RemapFileSystem, which allows more general remapping of files. We use this base class to implement BackupEngineImpl::RemapSharedFileSystem. To minimize API impact, I decided to just add these fields `name_for_open` and `env_for_open` to those set by GetBackupInfo when include_file_details=true. Creating the RemapSharedFileSystem adds a bit to the memory consumption, perhaps unnecessarily in some cases, but this has been mitigated by (a) only initialize the RemapSharedFileSystem lazily when GetBackupInfo with include_file_details=true is called, and (b) using the existing `shared_ptr<FileInfo>` objects to hold most of the mapping data. To enhance API safety, RemapSharedFileSystem is wrapped by new ReadOnlyFileSystem which rejects any attempts to write. This uncovered a couple of places in which DB::OpenForReadOnly would write to the filesystem, so I fixed these. Added a release note because this affects logging. Additional minor refactoring in backupable_db.cc to support the new functionality. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8142 Test Plan: new test (run with ASAN and UBSAN), added to stress test and ran it for a while with amplified backup_one_in Reviewed By: ajkr Differential Revision: D27535408 Pulled By: pdillinger fbshipit-source-id: 04666d310aa0261ef6b2385c43ca793ce1dfd148
4 years ago
#include "env/fs_readonly.h"
#include "env/fs_remap.h"
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
#include "file/filename.h"
#include "file/line_file_reader.h"
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
#include "file/sequence_file_reader.h"
#include "file/writable_file_writer.h"
#include "logging/logging.h"
#include "monitoring/iostats_context_imp.h"
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
#include "port/port.h"
#include "rocksdb/rate_limiter.h"
#include "rocksdb/transaction_log.h"
#include "table/sst_file_dumper.h"
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
#include "test_util/sync_point.h"
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
#include "util/cast_util.h"
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
#include "util/channel.h"
#include "util/coding.h"
#include "util/crc32c.h"
#include "util/string_util.h"
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
#include "utilities/backupable/backupable_db_impl.h"
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
#include "utilities/checkpoint/checkpoint_impl.h"
namespace ROCKSDB_NAMESPACE {
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
namespace {
using ShareFilesNaming = BackupEngineOptions::ShareFilesNaming;
Restore file size in backup table file names (and other cleanup) (#7400) Summary: Prior to 6.12, backup files using share_files_with_checksum had the file size encoded in the file name, after the last '\_' and before the last '.'. We considered this an implementation detail subject to change, and indeed removed this information from the file name (with an option to use old behavior) because it was considered ineffective/inefficient for file name uniqueness. However, some downstream RocksDB users were relying on this information since the file size is not explicitly in the backup manifest file. This primary purpose of this change is "retrofitting" the 6.12 release (not yet a public release) to simultaneously support the benefits of the new naming scheme (I/O performance and data correctness at scale) and preserve the file size information, both as default behaviors. With this change, we are essentially making the file size information encoded in the file name an official, though obscure, extension of the backup meta file format. We preserve an option (kLegacyCrc32cAndFileSize) to use the original "legacy" naming scheme, with its caveats, and make it easy to omit the file size information (no kFlagIncludeFileSize), for more compact file names. But note that changing the naming scheme used on an existing db and backup directory can lead to transient space amplification, as some files will be stored under two names in the shared_checksum directory. Because some backups were saved using the original 6.12 naming scheme, we offer two ways of dealing with those files: SST files generated by older 6.12 versions can either use the default naming scheme in effect when the SST files were generated (kFlagMatchInterimNaming, default, no transient space amplification) or can use a new naming scheme (no kFlagMatchInterimNaming, potential space amplification because some already stored files getting a new name). We don't have a natural way to detect which files were generated by previous 6.12 versions, but this change hacks one in by changing DB session ids to now use a more concise encoding, reducing file name length, saving ~dozen bytes from SST files, and making them visually distinct from DB ids so that they are less likely to be mixed up. Two final auxiliary notes: Recognizing that the backup file names have become a de facto part of the backup meta schema, this change makes them easier to parse and extend by putting a distinct marker, 's', before DB session ids embedded in the name. When we extend this to allow custom checksums in the name, they can get their own marker to ensure safe parsing. For backward compatibility, file size does not get a marker but is assumed for `_[0-9]+[.]` Another change from initial 6.12 default behavior is never including file custom checksum in the file name. Looking ahead to 6.13, we do not want the default behavior to cause backup space amplification for someone turning on file custom checksum checking in BackupEngine; we want that to be an easy decision. When implemented, including file custom checksums in backup file names will be a non-default option. Actual file name patterns and priorities, as regexes: kLegacyCrc32cAndFileSize OR pre-6.12 SST file -> [0-9]+_[0-9]+_[0-9]+[.]sst kFlagMatchInterimNaming set (default) AND early 6.12 SST file -> [0-9]+_[0-9a-fA-F-]+[.]sst kUseDbSessionId AND NOT kFlagIncludeFileSize -> [0-9]+_s[0-9A-Z]{20}[.]sst kUseDbSessionId AND kFlagIncludeFileSize (default) -> [0-9]+_s[0-9A-Z]{20}_[0-9]+[.]sst We might add opt-in options for more '\_' separated data in the name, but embedded file size, if present, will always be after last '\_' and before '.sst'. This change was originally applied to version 6.12. (See https://github.com/facebook/rocksdb/issues/7390) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7400 Test Plan: unit tests included. Sync point callbacks are used to mimic previous version SST files. Reviewed By: ajkr Differential Revision: D23759587 Pulled By: pdillinger fbshipit-source-id: f62d8af4e0978de0a34f26288cfbe66049b70025
4 years ago
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
constexpr BackupID kLatestBackupIDMarker = static_cast<BackupID>(-2);
inline uint32_t ChecksumHexToInt32(const std::string& checksum_hex) {
std::string checksum_str;
Slice(checksum_hex).DecodeHex(&checksum_str);
return EndianSwapValue(DecodeFixed32(checksum_str.c_str()));
}
inline std::string ChecksumStrToHex(const std::string& checksum_str) {
return Slice(checksum_str).ToString(true);
}
inline std::string ChecksumInt32ToHex(const uint32_t& checksum_value) {
std::string checksum_str;
PutFixed32(&checksum_str, EndianSwapValue(checksum_value));
return ChecksumStrToHex(checksum_str);
}
Make backups openable as read-only DBs (#8142) Summary: A current limitation of backups is that you don't know the exact database state of when the backup was taken. With this new feature, you can at least inspect the backup's DB state without restoring it by opening it as a read-only DB. Rather than add something like OpenAsReadOnlyDB to the BackupEngine API, which would inhibit opening stackable DB implementations read-only (if/when their APIs support it), we instead provide a DB name and Env that can be used to open as a read-only DB. Possible follow-up work: * Add a version of GetBackupInfo for a single backup. * Let CreateNewBackup return the BackupID of the newly-created backup. Implementation details: Refactored ChrootFileSystem to split off new base class RemapFileSystem, which allows more general remapping of files. We use this base class to implement BackupEngineImpl::RemapSharedFileSystem. To minimize API impact, I decided to just add these fields `name_for_open` and `env_for_open` to those set by GetBackupInfo when include_file_details=true. Creating the RemapSharedFileSystem adds a bit to the memory consumption, perhaps unnecessarily in some cases, but this has been mitigated by (a) only initialize the RemapSharedFileSystem lazily when GetBackupInfo with include_file_details=true is called, and (b) using the existing `shared_ptr<FileInfo>` objects to hold most of the mapping data. To enhance API safety, RemapSharedFileSystem is wrapped by new ReadOnlyFileSystem which rejects any attempts to write. This uncovered a couple of places in which DB::OpenForReadOnly would write to the filesystem, so I fixed these. Added a release note because this affects logging. Additional minor refactoring in backupable_db.cc to support the new functionality. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8142 Test Plan: new test (run with ASAN and UBSAN), added to stress test and ran it for a while with amplified backup_one_in Reviewed By: ajkr Differential Revision: D27535408 Pulled By: pdillinger fbshipit-source-id: 04666d310aa0261ef6b2385c43ca793ce1dfd148
4 years ago
const std::string kPrivateDirName = "private";
const std::string kMetaDirName = "meta";
const std::string kSharedDirName = "shared";
const std::string kSharedChecksumDirName = "shared_checksum";
const std::string kPrivateDirSlash = kPrivateDirName + "/";
const std::string kMetaDirSlash = kMetaDirName + "/";
const std::string kSharedDirSlash = kSharedDirName + "/";
const std::string kSharedChecksumDirSlash = kSharedChecksumDirName + "/";
} // namespace
void BackupStatistics::IncrementNumberSuccessBackup() {
number_success_backup++;
}
void BackupStatistics::IncrementNumberFailBackup() {
number_fail_backup++;
}
uint32_t BackupStatistics::GetNumberSuccessBackup() const {
return number_success_backup;
}
uint32_t BackupStatistics::GetNumberFailBackup() const {
return number_fail_backup;
}
std::string BackupStatistics::ToString() const {
char result[50];
snprintf(result, sizeof(result), "# success backup: %u, # fail backup: %u",
GetNumberSuccessBackup(), GetNumberFailBackup());
return result;
}
void BackupEngineOptions::Dump(Logger* logger) const {
ROCKS_LOG_INFO(logger, " Options.backup_dir: %s",
backup_dir.c_str());
ROCKS_LOG_INFO(logger, " Options.backup_env: %p", backup_env);
ROCKS_LOG_INFO(logger, " Options.share_table_files: %d",
static_cast<int>(share_table_files));
ROCKS_LOG_INFO(logger, " Options.info_log: %p", info_log);
ROCKS_LOG_INFO(logger, " Options.sync: %d",
static_cast<int>(sync));
ROCKS_LOG_INFO(logger, " Options.destroy_old_data: %d",
static_cast<int>(destroy_old_data));
ROCKS_LOG_INFO(logger, " Options.backup_log_files: %d",
static_cast<int>(backup_log_files));
ROCKS_LOG_INFO(logger, " Options.backup_rate_limit: %" PRIu64,
backup_rate_limit);
ROCKS_LOG_INFO(logger, " Options.restore_rate_limit: %" PRIu64,
restore_rate_limit);
ROCKS_LOG_INFO(logger, "Options.max_background_operations: %d",
max_background_operations);
}
// -------- BackupEngineImpl class ---------
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
class BackupEngineImpl {
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
public:
BackupEngineImpl(const BackupEngineOptions& options, Env* db_env,
bool read_only = false);
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
~BackupEngineImpl();
IOStatus CreateNewBackupWithMetadata(const CreateBackupOptions& options,
DB* db, const std::string& app_metadata,
BackupID* new_backup_id_ptr);
IOStatus PurgeOldBackups(uint32_t num_backups_to_keep);
IOStatus DeleteBackup(BackupID backup_id);
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
void StopBackup() { stop_backup_.store(true, std::memory_order_release); }
IOStatus GarbageCollect();
// The returned BackupInfos are in chronological order, which means the
// latest backup comes last.
void GetBackupInfo(std::vector<BackupInfo>* backup_info,
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
bool include_file_details) const;
Status GetBackupInfo(BackupID backup_id, BackupInfo* backup_info,
bool include_file_details = false) const;
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
void GetCorruptedBackups(std::vector<BackupID>* corrupt_backup_ids) const;
IOStatus RestoreDBFromBackup(const RestoreOptions& options,
BackupID backup_id, const std::string& db_dir,
const std::string& wal_dir) const;
IOStatus RestoreDBFromLatestBackup(const RestoreOptions& options,
const std::string& db_dir,
const std::string& wal_dir) const {
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
// Note: don't read latest_valid_backup_id_ outside of lock
return RestoreDBFromBackup(options, kLatestBackupIDMarker, db_dir, wal_dir);
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
IOStatus VerifyBackup(BackupID backup_id,
bool verify_with_checksum = false) const;
IOStatus Initialize();
Restore file size in backup table file names (and other cleanup) (#7400) Summary: Prior to 6.12, backup files using share_files_with_checksum had the file size encoded in the file name, after the last '\_' and before the last '.'. We considered this an implementation detail subject to change, and indeed removed this information from the file name (with an option to use old behavior) because it was considered ineffective/inefficient for file name uniqueness. However, some downstream RocksDB users were relying on this information since the file size is not explicitly in the backup manifest file. This primary purpose of this change is "retrofitting" the 6.12 release (not yet a public release) to simultaneously support the benefits of the new naming scheme (I/O performance and data correctness at scale) and preserve the file size information, both as default behaviors. With this change, we are essentially making the file size information encoded in the file name an official, though obscure, extension of the backup meta file format. We preserve an option (kLegacyCrc32cAndFileSize) to use the original "legacy" naming scheme, with its caveats, and make it easy to omit the file size information (no kFlagIncludeFileSize), for more compact file names. But note that changing the naming scheme used on an existing db and backup directory can lead to transient space amplification, as some files will be stored under two names in the shared_checksum directory. Because some backups were saved using the original 6.12 naming scheme, we offer two ways of dealing with those files: SST files generated by older 6.12 versions can either use the default naming scheme in effect when the SST files were generated (kFlagMatchInterimNaming, default, no transient space amplification) or can use a new naming scheme (no kFlagMatchInterimNaming, potential space amplification because some already stored files getting a new name). We don't have a natural way to detect which files were generated by previous 6.12 versions, but this change hacks one in by changing DB session ids to now use a more concise encoding, reducing file name length, saving ~dozen bytes from SST files, and making them visually distinct from DB ids so that they are less likely to be mixed up. Two final auxiliary notes: Recognizing that the backup file names have become a de facto part of the backup meta schema, this change makes them easier to parse and extend by putting a distinct marker, 's', before DB session ids embedded in the name. When we extend this to allow custom checksums in the name, they can get their own marker to ensure safe parsing. For backward compatibility, file size does not get a marker but is assumed for `_[0-9]+[.]` Another change from initial 6.12 default behavior is never including file custom checksum in the file name. Looking ahead to 6.13, we do not want the default behavior to cause backup space amplification for someone turning on file custom checksum checking in BackupEngine; we want that to be an easy decision. When implemented, including file custom checksums in backup file names will be a non-default option. Actual file name patterns and priorities, as regexes: kLegacyCrc32cAndFileSize OR pre-6.12 SST file -> [0-9]+_[0-9]+_[0-9]+[.]sst kFlagMatchInterimNaming set (default) AND early 6.12 SST file -> [0-9]+_[0-9a-fA-F-]+[.]sst kUseDbSessionId AND NOT kFlagIncludeFileSize -> [0-9]+_s[0-9A-Z]{20}[.]sst kUseDbSessionId AND kFlagIncludeFileSize (default) -> [0-9]+_s[0-9A-Z]{20}_[0-9]+[.]sst We might add opt-in options for more '\_' separated data in the name, but embedded file size, if present, will always be after last '\_' and before '.sst'. This change was originally applied to version 6.12. (See https://github.com/facebook/rocksdb/issues/7390) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7400 Test Plan: unit tests included. Sync point callbacks are used to mimic previous version SST files. Reviewed By: ajkr Differential Revision: D23759587 Pulled By: pdillinger fbshipit-source-id: f62d8af4e0978de0a34f26288cfbe66049b70025
4 years ago
ShareFilesNaming GetNamingNoFlags() const {
return options_.share_files_with_checksum_naming &
BackupEngineOptions::kMaskNoNamingFlags;
Restore file size in backup table file names (and other cleanup) (#7400) Summary: Prior to 6.12, backup files using share_files_with_checksum had the file size encoded in the file name, after the last '\_' and before the last '.'. We considered this an implementation detail subject to change, and indeed removed this information from the file name (with an option to use old behavior) because it was considered ineffective/inefficient for file name uniqueness. However, some downstream RocksDB users were relying on this information since the file size is not explicitly in the backup manifest file. This primary purpose of this change is "retrofitting" the 6.12 release (not yet a public release) to simultaneously support the benefits of the new naming scheme (I/O performance and data correctness at scale) and preserve the file size information, both as default behaviors. With this change, we are essentially making the file size information encoded in the file name an official, though obscure, extension of the backup meta file format. We preserve an option (kLegacyCrc32cAndFileSize) to use the original "legacy" naming scheme, with its caveats, and make it easy to omit the file size information (no kFlagIncludeFileSize), for more compact file names. But note that changing the naming scheme used on an existing db and backup directory can lead to transient space amplification, as some files will be stored under two names in the shared_checksum directory. Because some backups were saved using the original 6.12 naming scheme, we offer two ways of dealing with those files: SST files generated by older 6.12 versions can either use the default naming scheme in effect when the SST files were generated (kFlagMatchInterimNaming, default, no transient space amplification) or can use a new naming scheme (no kFlagMatchInterimNaming, potential space amplification because some already stored files getting a new name). We don't have a natural way to detect which files were generated by previous 6.12 versions, but this change hacks one in by changing DB session ids to now use a more concise encoding, reducing file name length, saving ~dozen bytes from SST files, and making them visually distinct from DB ids so that they are less likely to be mixed up. Two final auxiliary notes: Recognizing that the backup file names have become a de facto part of the backup meta schema, this change makes them easier to parse and extend by putting a distinct marker, 's', before DB session ids embedded in the name. When we extend this to allow custom checksums in the name, they can get their own marker to ensure safe parsing. For backward compatibility, file size does not get a marker but is assumed for `_[0-9]+[.]` Another change from initial 6.12 default behavior is never including file custom checksum in the file name. Looking ahead to 6.13, we do not want the default behavior to cause backup space amplification for someone turning on file custom checksum checking in BackupEngine; we want that to be an easy decision. When implemented, including file custom checksums in backup file names will be a non-default option. Actual file name patterns and priorities, as regexes: kLegacyCrc32cAndFileSize OR pre-6.12 SST file -> [0-9]+_[0-9]+_[0-9]+[.]sst kFlagMatchInterimNaming set (default) AND early 6.12 SST file -> [0-9]+_[0-9a-fA-F-]+[.]sst kUseDbSessionId AND NOT kFlagIncludeFileSize -> [0-9]+_s[0-9A-Z]{20}[.]sst kUseDbSessionId AND kFlagIncludeFileSize (default) -> [0-9]+_s[0-9A-Z]{20}_[0-9]+[.]sst We might add opt-in options for more '\_' separated data in the name, but embedded file size, if present, will always be after last '\_' and before '.sst'. This change was originally applied to version 6.12. (See https://github.com/facebook/rocksdb/issues/7390) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7400 Test Plan: unit tests included. Sync point callbacks are used to mimic previous version SST files. Reviewed By: ajkr Differential Revision: D23759587 Pulled By: pdillinger fbshipit-source-id: f62d8af4e0978de0a34f26288cfbe66049b70025
4 years ago
}
ShareFilesNaming GetNamingFlags() const {
return options_.share_files_with_checksum_naming &
BackupEngineOptions::kMaskNamingFlags;
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
private:
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
void DeleteChildren(const std::string& dir,
uint32_t file_type_filter = 0) const;
IOStatus DeleteBackupNoGC(BackupID backup_id);
// Extends the "result" map with pathname->size mappings for the contents of
// "dir" in "env". Pathnames are prefixed with "dir".
IOStatus ReadChildFileCurrentSizes(
const std::string& dir, const std::shared_ptr<FileSystem>&,
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
std::unordered_map<std::string, uint64_t>* result) const;
struct FileInfo {
FileInfo(const std::string& fname, uint64_t sz, const std::string& checksum,
const std::string& id = "", const std::string& sid = "")
: refs(0),
filename(fname),
size(sz),
checksum_hex(checksum),
db_id(id),
db_session_id(sid) {}
FileInfo(const FileInfo&) = delete;
FileInfo& operator=(const FileInfo&) = delete;
int refs;
const std::string filename;
const uint64_t size;
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
// crc32c checksum as hex. empty == unknown / unavailable
std::string checksum_hex;
// DB identities
// db_id is obtained for potential usage in the future but not used
// currently
const std::string db_id;
// db_session_id appears in the backup SST filename if the table naming
Restore file size in backup table file names (and other cleanup) (#7400) Summary: Prior to 6.12, backup files using share_files_with_checksum had the file size encoded in the file name, after the last '\_' and before the last '.'. We considered this an implementation detail subject to change, and indeed removed this information from the file name (with an option to use old behavior) because it was considered ineffective/inefficient for file name uniqueness. However, some downstream RocksDB users were relying on this information since the file size is not explicitly in the backup manifest file. This primary purpose of this change is "retrofitting" the 6.12 release (not yet a public release) to simultaneously support the benefits of the new naming scheme (I/O performance and data correctness at scale) and preserve the file size information, both as default behaviors. With this change, we are essentially making the file size information encoded in the file name an official, though obscure, extension of the backup meta file format. We preserve an option (kLegacyCrc32cAndFileSize) to use the original "legacy" naming scheme, with its caveats, and make it easy to omit the file size information (no kFlagIncludeFileSize), for more compact file names. But note that changing the naming scheme used on an existing db and backup directory can lead to transient space amplification, as some files will be stored under two names in the shared_checksum directory. Because some backups were saved using the original 6.12 naming scheme, we offer two ways of dealing with those files: SST files generated by older 6.12 versions can either use the default naming scheme in effect when the SST files were generated (kFlagMatchInterimNaming, default, no transient space amplification) or can use a new naming scheme (no kFlagMatchInterimNaming, potential space amplification because some already stored files getting a new name). We don't have a natural way to detect which files were generated by previous 6.12 versions, but this change hacks one in by changing DB session ids to now use a more concise encoding, reducing file name length, saving ~dozen bytes from SST files, and making them visually distinct from DB ids so that they are less likely to be mixed up. Two final auxiliary notes: Recognizing that the backup file names have become a de facto part of the backup meta schema, this change makes them easier to parse and extend by putting a distinct marker, 's', before DB session ids embedded in the name. When we extend this to allow custom checksums in the name, they can get their own marker to ensure safe parsing. For backward compatibility, file size does not get a marker but is assumed for `_[0-9]+[.]` Another change from initial 6.12 default behavior is never including file custom checksum in the file name. Looking ahead to 6.13, we do not want the default behavior to cause backup space amplification for someone turning on file custom checksum checking in BackupEngine; we want that to be an easy decision. When implemented, including file custom checksums in backup file names will be a non-default option. Actual file name patterns and priorities, as regexes: kLegacyCrc32cAndFileSize OR pre-6.12 SST file -> [0-9]+_[0-9]+_[0-9]+[.]sst kFlagMatchInterimNaming set (default) AND early 6.12 SST file -> [0-9]+_[0-9a-fA-F-]+[.]sst kUseDbSessionId AND NOT kFlagIncludeFileSize -> [0-9]+_s[0-9A-Z]{20}[.]sst kUseDbSessionId AND kFlagIncludeFileSize (default) -> [0-9]+_s[0-9A-Z]{20}_[0-9]+[.]sst We might add opt-in options for more '\_' separated data in the name, but embedded file size, if present, will always be after last '\_' and before '.sst'. This change was originally applied to version 6.12. (See https://github.com/facebook/rocksdb/issues/7390) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7400 Test Plan: unit tests included. Sync point callbacks are used to mimic previous version SST files. Reviewed By: ajkr Differential Revision: D23759587 Pulled By: pdillinger fbshipit-source-id: f62d8af4e0978de0a34f26288cfbe66049b70025
4 years ago
// option is kUseDbSessionId
const std::string db_session_id;
Make backups openable as read-only DBs (#8142) Summary: A current limitation of backups is that you don't know the exact database state of when the backup was taken. With this new feature, you can at least inspect the backup's DB state without restoring it by opening it as a read-only DB. Rather than add something like OpenAsReadOnlyDB to the BackupEngine API, which would inhibit opening stackable DB implementations read-only (if/when their APIs support it), we instead provide a DB name and Env that can be used to open as a read-only DB. Possible follow-up work: * Add a version of GetBackupInfo for a single backup. * Let CreateNewBackup return the BackupID of the newly-created backup. Implementation details: Refactored ChrootFileSystem to split off new base class RemapFileSystem, which allows more general remapping of files. We use this base class to implement BackupEngineImpl::RemapSharedFileSystem. To minimize API impact, I decided to just add these fields `name_for_open` and `env_for_open` to those set by GetBackupInfo when include_file_details=true. Creating the RemapSharedFileSystem adds a bit to the memory consumption, perhaps unnecessarily in some cases, but this has been mitigated by (a) only initialize the RemapSharedFileSystem lazily when GetBackupInfo with include_file_details=true is called, and (b) using the existing `shared_ptr<FileInfo>` objects to hold most of the mapping data. To enhance API safety, RemapSharedFileSystem is wrapped by new ReadOnlyFileSystem which rejects any attempts to write. This uncovered a couple of places in which DB::OpenForReadOnly would write to the filesystem, so I fixed these. Added a release note because this affects logging. Additional minor refactoring in backupable_db.cc to support the new functionality. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8142 Test Plan: new test (run with ASAN and UBSAN), added to stress test and ran it for a while with amplified backup_one_in Reviewed By: ajkr Differential Revision: D27535408 Pulled By: pdillinger fbshipit-source-id: 04666d310aa0261ef6b2385c43ca793ce1dfd148
4 years ago
std::string GetDbFileName() {
std::string rv;
// extract the filename part
size_t slash = filename.find_last_of('/');
// file will either be shared/<file>, shared_checksum/<file_crc32c_size>,
// shared_checksum/<file_session>, shared_checksum/<file_crc32c_session>,
// or private/<number>/<file>
assert(slash != std::string::npos);
rv = filename.substr(slash + 1);
// if the file was in shared_checksum, extract the real file name
// in this case the file is <number>_<checksum>_<size>.<type>,
// <number>_<session>.<type>, or <number>_<checksum>_<session>.<type>
if (filename.substr(0, slash) == kSharedChecksumDirName) {
rv = GetFileFromChecksumFile(rv);
}
return rv;
}
};
static inline std::string WithoutTrailingSlash(const std::string& path) {
if (path.empty() || path.back() != '/') {
return path;
} else {
return path.substr(path.size() - 1);
}
}
static inline std::string WithTrailingSlash(const std::string& path) {
if (path.empty() || path.back() != '/') {
return path + '/';
} else {
return path;
}
}
// A filesystem wrapper that makes shared backup files appear to be in the
// private backup directory (dst_dir), so that the private backup dir can
// be opened as a read-only DB.
class RemapSharedFileSystem : public RemapFileSystem {
public:
RemapSharedFileSystem(const std::shared_ptr<FileSystem>& base,
const std::string& dst_dir,
const std::string& src_base_dir,
const std::vector<std::shared_ptr<FileInfo>>& files)
: RemapFileSystem(base),
dst_dir_(WithoutTrailingSlash(dst_dir)),
dst_dir_slash_(WithTrailingSlash(dst_dir)),
src_base_dir_(WithTrailingSlash(src_base_dir)) {
for (auto& info : files) {
if (!StartsWith(info->filename, kPrivateDirSlash)) {
assert(StartsWith(info->filename, kSharedDirSlash) ||
StartsWith(info->filename, kSharedChecksumDirSlash));
remaps_[info->GetDbFileName()] = info;
}
}
}
const char* Name() const override {
return "BackupEngineImpl::RemapSharedFileSystem";
}
// Sometimes a directory listing is required in opening a DB
IOStatus GetChildren(const std::string& dir, const IOOptions& options,
std::vector<std::string>* result,
IODebugContext* dbg) override {
IOStatus s = RemapFileSystem::GetChildren(dir, options, result, dbg);
if (s.ok() && (dir == dst_dir_ || dir == dst_dir_slash_)) {
// Assume remapped files exist
for (auto& r : remaps_) {
result->push_back(r.first);
}
}
return s;
}
// Sometimes a directory listing is required in opening a DB
IOStatus GetChildrenFileAttributes(const std::string& dir,
const IOOptions& options,
std::vector<FileAttributes>* result,
IODebugContext* dbg) override {
IOStatus s =
RemapFileSystem::GetChildrenFileAttributes(dir, options, result, dbg);
if (s.ok() && (dir == dst_dir_ || dir == dst_dir_slash_)) {
// Assume remapped files exist with recorded size
for (auto& r : remaps_) {
result->emplace_back(); // clean up with C++20
FileAttributes& attr = result->back();
attr.name = r.first;
attr.size_bytes = r.second->size;
}
}
return s;
}
protected:
// When a file in dst_dir is requested, see if we need to remap to shared
// file path.
std::pair<IOStatus, std::string> EncodePath(
const std::string& path) override {
if (path.empty() || path[0] != '/') {
return {IOStatus::InvalidArgument(path, "Not an absolute path"), ""};
}
std::pair<IOStatus, std::string> rv{IOStatus(), path};
if (StartsWith(path, dst_dir_slash_)) {
std::string relative = path.substr(dst_dir_slash_.size());
auto it = remaps_.find(relative);
if (it != remaps_.end()) {
rv.second = src_base_dir_ + it->second->filename;
}
}
return rv;
}
private:
// Absolute path to a directory that some extra files will be mapped into.
const std::string dst_dir_;
// Includes a trailing slash.
const std::string dst_dir_slash_;
// Absolute path to a directory containing some files to be mapped into
// dst_dir_. Includes a trailing slash.
const std::string src_base_dir_;
// If remaps_[x] exists, attempt to read dst_dir_ / x should instead read
// src_base_dir_ / remaps_[x]->filename. FileInfo is used to maximize
// sharing with other backup data in memory.
std::unordered_map<std::string, std::shared_ptr<FileInfo>> remaps_;
};
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
class BackupMeta {
public:
BackupMeta(
const std::string& meta_filename, const std::string& meta_tmp_filename,
std::unordered_map<std::string, std::shared_ptr<FileInfo>>* file_infos,
Env* env, const std::shared_ptr<FileSystem>& fs)
: timestamp_(0),
sequence_number_(0),
size_(0),
meta_filename_(meta_filename),
meta_tmp_filename_(meta_tmp_filename),
file_infos_(file_infos),
env_(env),
fs_(fs) {}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
BackupMeta(const BackupMeta&) = delete;
BackupMeta& operator=(const BackupMeta&) = delete;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
~BackupMeta() {}
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
void RecordTimestamp() {
// Best effort
Status s = env_->GetCurrentTime(&timestamp_);
if (!s.ok()) {
timestamp_ = /* something clearly fabricated */ 1;
}
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
int64_t GetTimestamp() const {
return timestamp_;
}
uint64_t GetSize() const {
return size_;
}
uint32_t GetNumberFiles() const {
return static_cast<uint32_t>(files_.size());
}
void SetSequenceNumber(uint64_t sequence_number) {
sequence_number_ = sequence_number;
}
uint64_t GetSequenceNumber() const { return sequence_number_; }
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
const std::string& GetAppMetadata() const { return app_metadata_; }
void SetAppMetadata(const std::string& app_metadata) {
app_metadata_ = app_metadata;
}
IOStatus AddFile(std::shared_ptr<FileInfo> file_info);
IOStatus Delete(bool delete_meta = true);
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
bool Empty() const { return files_.empty(); }
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
std::shared_ptr<FileInfo> GetFile(const std::string& filename) const {
auto it = file_infos_->find(filename);
if (it == file_infos_->end())
return nullptr;
return it->second;
}
const std::vector<std::shared_ptr<FileInfo>>& GetFiles() const {
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
return files_;
}
// @param abs_path_to_size Pre-fetched file sizes (bytes).
IOStatus LoadFromFile(
const std::string& backup_dir,
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
const std::unordered_map<std::string, uint64_t>& abs_path_to_size,
RateLimiter* rate_limiter, Logger* info_log,
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
std::unordered_set<std::string>* reported_ignored_fields);
IOStatus StoreToFile(
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
bool sync, const TEST_FutureSchemaVersion2Options* test_future_options);
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
std::string GetInfoString() {
std::ostringstream ss;
ss << "Timestamp: " << timestamp_ << std::endl;
char human_size[16];
AppendHumanBytes(size_, human_size, sizeof(human_size));
ss << "Size: " << human_size << std::endl;
ss << "Files:" << std::endl;
for (const auto& file : files_) {
AppendHumanBytes(file->size, human_size, sizeof(human_size));
ss << file->filename << ", size " << human_size << ", refs "
<< file->refs << std::endl;
}
return ss.str();
}
Make backups openable as read-only DBs (#8142) Summary: A current limitation of backups is that you don't know the exact database state of when the backup was taken. With this new feature, you can at least inspect the backup's DB state without restoring it by opening it as a read-only DB. Rather than add something like OpenAsReadOnlyDB to the BackupEngine API, which would inhibit opening stackable DB implementations read-only (if/when their APIs support it), we instead provide a DB name and Env that can be used to open as a read-only DB. Possible follow-up work: * Add a version of GetBackupInfo for a single backup. * Let CreateNewBackup return the BackupID of the newly-created backup. Implementation details: Refactored ChrootFileSystem to split off new base class RemapFileSystem, which allows more general remapping of files. We use this base class to implement BackupEngineImpl::RemapSharedFileSystem. To minimize API impact, I decided to just add these fields `name_for_open` and `env_for_open` to those set by GetBackupInfo when include_file_details=true. Creating the RemapSharedFileSystem adds a bit to the memory consumption, perhaps unnecessarily in some cases, but this has been mitigated by (a) only initialize the RemapSharedFileSystem lazily when GetBackupInfo with include_file_details=true is called, and (b) using the existing `shared_ptr<FileInfo>` objects to hold most of the mapping data. To enhance API safety, RemapSharedFileSystem is wrapped by new ReadOnlyFileSystem which rejects any attempts to write. This uncovered a couple of places in which DB::OpenForReadOnly would write to the filesystem, so I fixed these. Added a release note because this affects logging. Additional minor refactoring in backupable_db.cc to support the new functionality. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8142 Test Plan: new test (run with ASAN and UBSAN), added to stress test and ran it for a while with amplified backup_one_in Reviewed By: ajkr Differential Revision: D27535408 Pulled By: pdillinger fbshipit-source-id: 04666d310aa0261ef6b2385c43ca793ce1dfd148
4 years ago
const std::shared_ptr<Env>& GetEnvForOpen() const {
if (!env_for_open_) {
// Lazy initialize
// Find directories
std::string dst_dir = meta_filename_;
auto i = dst_dir.rfind(kMetaDirSlash);
assert(i != std::string::npos);
std::string src_base_dir = dst_dir.substr(0, i);
dst_dir.replace(i, kMetaDirSlash.size(), kPrivateDirSlash);
// Make the RemapSharedFileSystem
std::shared_ptr<FileSystem> remap_fs =
std::make_shared<RemapSharedFileSystem>(fs_, dst_dir, src_base_dir,
files_);
Make backups openable as read-only DBs (#8142) Summary: A current limitation of backups is that you don't know the exact database state of when the backup was taken. With this new feature, you can at least inspect the backup's DB state without restoring it by opening it as a read-only DB. Rather than add something like OpenAsReadOnlyDB to the BackupEngine API, which would inhibit opening stackable DB implementations read-only (if/when their APIs support it), we instead provide a DB name and Env that can be used to open as a read-only DB. Possible follow-up work: * Add a version of GetBackupInfo for a single backup. * Let CreateNewBackup return the BackupID of the newly-created backup. Implementation details: Refactored ChrootFileSystem to split off new base class RemapFileSystem, which allows more general remapping of files. We use this base class to implement BackupEngineImpl::RemapSharedFileSystem. To minimize API impact, I decided to just add these fields `name_for_open` and `env_for_open` to those set by GetBackupInfo when include_file_details=true. Creating the RemapSharedFileSystem adds a bit to the memory consumption, perhaps unnecessarily in some cases, but this has been mitigated by (a) only initialize the RemapSharedFileSystem lazily when GetBackupInfo with include_file_details=true is called, and (b) using the existing `shared_ptr<FileInfo>` objects to hold most of the mapping data. To enhance API safety, RemapSharedFileSystem is wrapped by new ReadOnlyFileSystem which rejects any attempts to write. This uncovered a couple of places in which DB::OpenForReadOnly would write to the filesystem, so I fixed these. Added a release note because this affects logging. Additional minor refactoring in backupable_db.cc to support the new functionality. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8142 Test Plan: new test (run with ASAN and UBSAN), added to stress test and ran it for a while with amplified backup_one_in Reviewed By: ajkr Differential Revision: D27535408 Pulled By: pdillinger fbshipit-source-id: 04666d310aa0261ef6b2385c43ca793ce1dfd148
4 years ago
// Make it read-only for safety
remap_fs = std::make_shared<ReadOnlyFileSystem>(remap_fs);
// Make an Env wrapper
env_for_open_ = std::make_shared<CompositeEnvWrapper>(env_, remap_fs);
}
return env_for_open_;
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
private:
int64_t timestamp_;
// sequence number is only approximate, should not be used
// by clients
uint64_t sequence_number_;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
uint64_t size_;
std::string app_metadata_;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
std::string const meta_filename_;
std::string const meta_tmp_filename_;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
// files with relative paths (without "/" prefix!!)
std::vector<std::shared_ptr<FileInfo>> files_;
std::unordered_map<std::string, std::shared_ptr<FileInfo>>* file_infos_;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
Env* env_;
Make backups openable as read-only DBs (#8142) Summary: A current limitation of backups is that you don't know the exact database state of when the backup was taken. With this new feature, you can at least inspect the backup's DB state without restoring it by opening it as a read-only DB. Rather than add something like OpenAsReadOnlyDB to the BackupEngine API, which would inhibit opening stackable DB implementations read-only (if/when their APIs support it), we instead provide a DB name and Env that can be used to open as a read-only DB. Possible follow-up work: * Add a version of GetBackupInfo for a single backup. * Let CreateNewBackup return the BackupID of the newly-created backup. Implementation details: Refactored ChrootFileSystem to split off new base class RemapFileSystem, which allows more general remapping of files. We use this base class to implement BackupEngineImpl::RemapSharedFileSystem. To minimize API impact, I decided to just add these fields `name_for_open` and `env_for_open` to those set by GetBackupInfo when include_file_details=true. Creating the RemapSharedFileSystem adds a bit to the memory consumption, perhaps unnecessarily in some cases, but this has been mitigated by (a) only initialize the RemapSharedFileSystem lazily when GetBackupInfo with include_file_details=true is called, and (b) using the existing `shared_ptr<FileInfo>` objects to hold most of the mapping data. To enhance API safety, RemapSharedFileSystem is wrapped by new ReadOnlyFileSystem which rejects any attempts to write. This uncovered a couple of places in which DB::OpenForReadOnly would write to the filesystem, so I fixed these. Added a release note because this affects logging. Additional minor refactoring in backupable_db.cc to support the new functionality. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8142 Test Plan: new test (run with ASAN and UBSAN), added to stress test and ran it for a while with amplified backup_one_in Reviewed By: ajkr Differential Revision: D27535408 Pulled By: pdillinger fbshipit-source-id: 04666d310aa0261ef6b2385c43ca793ce1dfd148
4 years ago
mutable std::shared_ptr<Env> env_for_open_;
std::shared_ptr<FileSystem> fs_;
IOOptions iooptions_ = IOOptions();
}; // BackupMeta
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
void SetBackupInfoFromBackupMeta(BackupID id, const BackupMeta& meta,
BackupInfo* backup_info,
bool include_file_details) const;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
inline std::string GetAbsolutePath(
const std::string &relative_path = "") const {
assert(relative_path.size() == 0 || relative_path[0] != '/');
return options_.backup_dir + "/" + relative_path;
}
inline std::string GetPrivateFileRel(BackupID backup_id,
bool tmp = false,
const std::string& file = "") const {
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
assert(file.size() == 0 || file[0] != '/');
Make backups openable as read-only DBs (#8142) Summary: A current limitation of backups is that you don't know the exact database state of when the backup was taken. With this new feature, you can at least inspect the backup's DB state without restoring it by opening it as a read-only DB. Rather than add something like OpenAsReadOnlyDB to the BackupEngine API, which would inhibit opening stackable DB implementations read-only (if/when their APIs support it), we instead provide a DB name and Env that can be used to open as a read-only DB. Possible follow-up work: * Add a version of GetBackupInfo for a single backup. * Let CreateNewBackup return the BackupID of the newly-created backup. Implementation details: Refactored ChrootFileSystem to split off new base class RemapFileSystem, which allows more general remapping of files. We use this base class to implement BackupEngineImpl::RemapSharedFileSystem. To minimize API impact, I decided to just add these fields `name_for_open` and `env_for_open` to those set by GetBackupInfo when include_file_details=true. Creating the RemapSharedFileSystem adds a bit to the memory consumption, perhaps unnecessarily in some cases, but this has been mitigated by (a) only initialize the RemapSharedFileSystem lazily when GetBackupInfo with include_file_details=true is called, and (b) using the existing `shared_ptr<FileInfo>` objects to hold most of the mapping data. To enhance API safety, RemapSharedFileSystem is wrapped by new ReadOnlyFileSystem which rejects any attempts to write. This uncovered a couple of places in which DB::OpenForReadOnly would write to the filesystem, so I fixed these. Added a release note because this affects logging. Additional minor refactoring in backupable_db.cc to support the new functionality. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8142 Test Plan: new test (run with ASAN and UBSAN), added to stress test and ran it for a while with amplified backup_one_in Reviewed By: ajkr Differential Revision: D27535408 Pulled By: pdillinger fbshipit-source-id: 04666d310aa0261ef6b2385c43ca793ce1dfd148
4 years ago
return kPrivateDirSlash + ROCKSDB_NAMESPACE::ToString(backup_id) +
(tmp ? ".tmp" : "") + "/" + file;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
inline std::string GetSharedFileRel(const std::string& file = "",
bool tmp = false) const {
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
assert(file.size() == 0 || file[0] != '/');
Make backups openable as read-only DBs (#8142) Summary: A current limitation of backups is that you don't know the exact database state of when the backup was taken. With this new feature, you can at least inspect the backup's DB state without restoring it by opening it as a read-only DB. Rather than add something like OpenAsReadOnlyDB to the BackupEngine API, which would inhibit opening stackable DB implementations read-only (if/when their APIs support it), we instead provide a DB name and Env that can be used to open as a read-only DB. Possible follow-up work: * Add a version of GetBackupInfo for a single backup. * Let CreateNewBackup return the BackupID of the newly-created backup. Implementation details: Refactored ChrootFileSystem to split off new base class RemapFileSystem, which allows more general remapping of files. We use this base class to implement BackupEngineImpl::RemapSharedFileSystem. To minimize API impact, I decided to just add these fields `name_for_open` and `env_for_open` to those set by GetBackupInfo when include_file_details=true. Creating the RemapSharedFileSystem adds a bit to the memory consumption, perhaps unnecessarily in some cases, but this has been mitigated by (a) only initialize the RemapSharedFileSystem lazily when GetBackupInfo with include_file_details=true is called, and (b) using the existing `shared_ptr<FileInfo>` objects to hold most of the mapping data. To enhance API safety, RemapSharedFileSystem is wrapped by new ReadOnlyFileSystem which rejects any attempts to write. This uncovered a couple of places in which DB::OpenForReadOnly would write to the filesystem, so I fixed these. Added a release note because this affects logging. Additional minor refactoring in backupable_db.cc to support the new functionality. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8142 Test Plan: new test (run with ASAN and UBSAN), added to stress test and ran it for a while with amplified backup_one_in Reviewed By: ajkr Differential Revision: D27535408 Pulled By: pdillinger fbshipit-source-id: 04666d310aa0261ef6b2385c43ca793ce1dfd148
4 years ago
return kSharedDirSlash + std::string(tmp ? "." : "") + file +
(tmp ? ".tmp" : "");
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
inline std::string GetSharedFileWithChecksumRel(const std::string& file = "",
bool tmp = false) const {
assert(file.size() == 0 || file[0] != '/');
Make backups openable as read-only DBs (#8142) Summary: A current limitation of backups is that you don't know the exact database state of when the backup was taken. With this new feature, you can at least inspect the backup's DB state without restoring it by opening it as a read-only DB. Rather than add something like OpenAsReadOnlyDB to the BackupEngine API, which would inhibit opening stackable DB implementations read-only (if/when their APIs support it), we instead provide a DB name and Env that can be used to open as a read-only DB. Possible follow-up work: * Add a version of GetBackupInfo for a single backup. * Let CreateNewBackup return the BackupID of the newly-created backup. Implementation details: Refactored ChrootFileSystem to split off new base class RemapFileSystem, which allows more general remapping of files. We use this base class to implement BackupEngineImpl::RemapSharedFileSystem. To minimize API impact, I decided to just add these fields `name_for_open` and `env_for_open` to those set by GetBackupInfo when include_file_details=true. Creating the RemapSharedFileSystem adds a bit to the memory consumption, perhaps unnecessarily in some cases, but this has been mitigated by (a) only initialize the RemapSharedFileSystem lazily when GetBackupInfo with include_file_details=true is called, and (b) using the existing `shared_ptr<FileInfo>` objects to hold most of the mapping data. To enhance API safety, RemapSharedFileSystem is wrapped by new ReadOnlyFileSystem which rejects any attempts to write. This uncovered a couple of places in which DB::OpenForReadOnly would write to the filesystem, so I fixed these. Added a release note because this affects logging. Additional minor refactoring in backupable_db.cc to support the new functionality. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8142 Test Plan: new test (run with ASAN and UBSAN), added to stress test and ran it for a while with amplified backup_one_in Reviewed By: ajkr Differential Revision: D27535408 Pulled By: pdillinger fbshipit-source-id: 04666d310aa0261ef6b2385c43ca793ce1dfd148
4 years ago
return kSharedChecksumDirSlash + std::string(tmp ? "." : "") + file +
(tmp ? ".tmp" : "");
}
Restore file size in backup table file names (and other cleanup) (#7400) Summary: Prior to 6.12, backup files using share_files_with_checksum had the file size encoded in the file name, after the last '\_' and before the last '.'. We considered this an implementation detail subject to change, and indeed removed this information from the file name (with an option to use old behavior) because it was considered ineffective/inefficient for file name uniqueness. However, some downstream RocksDB users were relying on this information since the file size is not explicitly in the backup manifest file. This primary purpose of this change is "retrofitting" the 6.12 release (not yet a public release) to simultaneously support the benefits of the new naming scheme (I/O performance and data correctness at scale) and preserve the file size information, both as default behaviors. With this change, we are essentially making the file size information encoded in the file name an official, though obscure, extension of the backup meta file format. We preserve an option (kLegacyCrc32cAndFileSize) to use the original "legacy" naming scheme, with its caveats, and make it easy to omit the file size information (no kFlagIncludeFileSize), for more compact file names. But note that changing the naming scheme used on an existing db and backup directory can lead to transient space amplification, as some files will be stored under two names in the shared_checksum directory. Because some backups were saved using the original 6.12 naming scheme, we offer two ways of dealing with those files: SST files generated by older 6.12 versions can either use the default naming scheme in effect when the SST files were generated (kFlagMatchInterimNaming, default, no transient space amplification) or can use a new naming scheme (no kFlagMatchInterimNaming, potential space amplification because some already stored files getting a new name). We don't have a natural way to detect which files were generated by previous 6.12 versions, but this change hacks one in by changing DB session ids to now use a more concise encoding, reducing file name length, saving ~dozen bytes from SST files, and making them visually distinct from DB ids so that they are less likely to be mixed up. Two final auxiliary notes: Recognizing that the backup file names have become a de facto part of the backup meta schema, this change makes them easier to parse and extend by putting a distinct marker, 's', before DB session ids embedded in the name. When we extend this to allow custom checksums in the name, they can get their own marker to ensure safe parsing. For backward compatibility, file size does not get a marker but is assumed for `_[0-9]+[.]` Another change from initial 6.12 default behavior is never including file custom checksum in the file name. Looking ahead to 6.13, we do not want the default behavior to cause backup space amplification for someone turning on file custom checksum checking in BackupEngine; we want that to be an easy decision. When implemented, including file custom checksums in backup file names will be a non-default option. Actual file name patterns and priorities, as regexes: kLegacyCrc32cAndFileSize OR pre-6.12 SST file -> [0-9]+_[0-9]+_[0-9]+[.]sst kFlagMatchInterimNaming set (default) AND early 6.12 SST file -> [0-9]+_[0-9a-fA-F-]+[.]sst kUseDbSessionId AND NOT kFlagIncludeFileSize -> [0-9]+_s[0-9A-Z]{20}[.]sst kUseDbSessionId AND kFlagIncludeFileSize (default) -> [0-9]+_s[0-9A-Z]{20}_[0-9]+[.]sst We might add opt-in options for more '\_' separated data in the name, but embedded file size, if present, will always be after last '\_' and before '.sst'. This change was originally applied to version 6.12. (See https://github.com/facebook/rocksdb/issues/7390) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7400 Test Plan: unit tests included. Sync point callbacks are used to mimic previous version SST files. Reviewed By: ajkr Differential Revision: D23759587 Pulled By: pdillinger fbshipit-source-id: f62d8af4e0978de0a34f26288cfbe66049b70025
4 years ago
inline bool UseLegacyNaming(const std::string& sid) const {
return GetNamingNoFlags() ==
BackupEngineOptions::kLegacyCrc32cAndFileSize ||
Restore file size in backup table file names (and other cleanup) (#7400) Summary: Prior to 6.12, backup files using share_files_with_checksum had the file size encoded in the file name, after the last '\_' and before the last '.'. We considered this an implementation detail subject to change, and indeed removed this information from the file name (with an option to use old behavior) because it was considered ineffective/inefficient for file name uniqueness. However, some downstream RocksDB users were relying on this information since the file size is not explicitly in the backup manifest file. This primary purpose of this change is "retrofitting" the 6.12 release (not yet a public release) to simultaneously support the benefits of the new naming scheme (I/O performance and data correctness at scale) and preserve the file size information, both as default behaviors. With this change, we are essentially making the file size information encoded in the file name an official, though obscure, extension of the backup meta file format. We preserve an option (kLegacyCrc32cAndFileSize) to use the original "legacy" naming scheme, with its caveats, and make it easy to omit the file size information (no kFlagIncludeFileSize), for more compact file names. But note that changing the naming scheme used on an existing db and backup directory can lead to transient space amplification, as some files will be stored under two names in the shared_checksum directory. Because some backups were saved using the original 6.12 naming scheme, we offer two ways of dealing with those files: SST files generated by older 6.12 versions can either use the default naming scheme in effect when the SST files were generated (kFlagMatchInterimNaming, default, no transient space amplification) or can use a new naming scheme (no kFlagMatchInterimNaming, potential space amplification because some already stored files getting a new name). We don't have a natural way to detect which files were generated by previous 6.12 versions, but this change hacks one in by changing DB session ids to now use a more concise encoding, reducing file name length, saving ~dozen bytes from SST files, and making them visually distinct from DB ids so that they are less likely to be mixed up. Two final auxiliary notes: Recognizing that the backup file names have become a de facto part of the backup meta schema, this change makes them easier to parse and extend by putting a distinct marker, 's', before DB session ids embedded in the name. When we extend this to allow custom checksums in the name, they can get their own marker to ensure safe parsing. For backward compatibility, file size does not get a marker but is assumed for `_[0-9]+[.]` Another change from initial 6.12 default behavior is never including file custom checksum in the file name. Looking ahead to 6.13, we do not want the default behavior to cause backup space amplification for someone turning on file custom checksum checking in BackupEngine; we want that to be an easy decision. When implemented, including file custom checksums in backup file names will be a non-default option. Actual file name patterns and priorities, as regexes: kLegacyCrc32cAndFileSize OR pre-6.12 SST file -> [0-9]+_[0-9]+_[0-9]+[.]sst kFlagMatchInterimNaming set (default) AND early 6.12 SST file -> [0-9]+_[0-9a-fA-F-]+[.]sst kUseDbSessionId AND NOT kFlagIncludeFileSize -> [0-9]+_s[0-9A-Z]{20}[.]sst kUseDbSessionId AND kFlagIncludeFileSize (default) -> [0-9]+_s[0-9A-Z]{20}_[0-9]+[.]sst We might add opt-in options for more '\_' separated data in the name, but embedded file size, if present, will always be after last '\_' and before '.sst'. This change was originally applied to version 6.12. (See https://github.com/facebook/rocksdb/issues/7390) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7400 Test Plan: unit tests included. Sync point callbacks are used to mimic previous version SST files. Reviewed By: ajkr Differential Revision: D23759587 Pulled By: pdillinger fbshipit-source-id: f62d8af4e0978de0a34f26288cfbe66049b70025
4 years ago
sid.empty();
}
inline std::string GetSharedFileWithChecksum(
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
const std::string& file, const std::string& checksum_hex,
const uint64_t file_size, const std::string& db_session_id) const {
assert(file.size() == 0 || file[0] != '/');
std::string file_copy = file;
Restore file size in backup table file names (and other cleanup) (#7400) Summary: Prior to 6.12, backup files using share_files_with_checksum had the file size encoded in the file name, after the last '\_' and before the last '.'. We considered this an implementation detail subject to change, and indeed removed this information from the file name (with an option to use old behavior) because it was considered ineffective/inefficient for file name uniqueness. However, some downstream RocksDB users were relying on this information since the file size is not explicitly in the backup manifest file. This primary purpose of this change is "retrofitting" the 6.12 release (not yet a public release) to simultaneously support the benefits of the new naming scheme (I/O performance and data correctness at scale) and preserve the file size information, both as default behaviors. With this change, we are essentially making the file size information encoded in the file name an official, though obscure, extension of the backup meta file format. We preserve an option (kLegacyCrc32cAndFileSize) to use the original "legacy" naming scheme, with its caveats, and make it easy to omit the file size information (no kFlagIncludeFileSize), for more compact file names. But note that changing the naming scheme used on an existing db and backup directory can lead to transient space amplification, as some files will be stored under two names in the shared_checksum directory. Because some backups were saved using the original 6.12 naming scheme, we offer two ways of dealing with those files: SST files generated by older 6.12 versions can either use the default naming scheme in effect when the SST files were generated (kFlagMatchInterimNaming, default, no transient space amplification) or can use a new naming scheme (no kFlagMatchInterimNaming, potential space amplification because some already stored files getting a new name). We don't have a natural way to detect which files were generated by previous 6.12 versions, but this change hacks one in by changing DB session ids to now use a more concise encoding, reducing file name length, saving ~dozen bytes from SST files, and making them visually distinct from DB ids so that they are less likely to be mixed up. Two final auxiliary notes: Recognizing that the backup file names have become a de facto part of the backup meta schema, this change makes them easier to parse and extend by putting a distinct marker, 's', before DB session ids embedded in the name. When we extend this to allow custom checksums in the name, they can get their own marker to ensure safe parsing. For backward compatibility, file size does not get a marker but is assumed for `_[0-9]+[.]` Another change from initial 6.12 default behavior is never including file custom checksum in the file name. Looking ahead to 6.13, we do not want the default behavior to cause backup space amplification for someone turning on file custom checksum checking in BackupEngine; we want that to be an easy decision. When implemented, including file custom checksums in backup file names will be a non-default option. Actual file name patterns and priorities, as regexes: kLegacyCrc32cAndFileSize OR pre-6.12 SST file -> [0-9]+_[0-9]+_[0-9]+[.]sst kFlagMatchInterimNaming set (default) AND early 6.12 SST file -> [0-9]+_[0-9a-fA-F-]+[.]sst kUseDbSessionId AND NOT kFlagIncludeFileSize -> [0-9]+_s[0-9A-Z]{20}[.]sst kUseDbSessionId AND kFlagIncludeFileSize (default) -> [0-9]+_s[0-9A-Z]{20}_[0-9]+[.]sst We might add opt-in options for more '\_' separated data in the name, but embedded file size, if present, will always be after last '\_' and before '.sst'. This change was originally applied to version 6.12. (See https://github.com/facebook/rocksdb/issues/7390) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7400 Test Plan: unit tests included. Sync point callbacks are used to mimic previous version SST files. Reviewed By: ajkr Differential Revision: D23759587 Pulled By: pdillinger fbshipit-source-id: f62d8af4e0978de0a34f26288cfbe66049b70025
4 years ago
if (UseLegacyNaming(db_session_id)) {
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
assert(!checksum_hex.empty());
Restore file size in backup table file names (and other cleanup) (#7400) Summary: Prior to 6.12, backup files using share_files_with_checksum had the file size encoded in the file name, after the last '\_' and before the last '.'. We considered this an implementation detail subject to change, and indeed removed this information from the file name (with an option to use old behavior) because it was considered ineffective/inefficient for file name uniqueness. However, some downstream RocksDB users were relying on this information since the file size is not explicitly in the backup manifest file. This primary purpose of this change is "retrofitting" the 6.12 release (not yet a public release) to simultaneously support the benefits of the new naming scheme (I/O performance and data correctness at scale) and preserve the file size information, both as default behaviors. With this change, we are essentially making the file size information encoded in the file name an official, though obscure, extension of the backup meta file format. We preserve an option (kLegacyCrc32cAndFileSize) to use the original "legacy" naming scheme, with its caveats, and make it easy to omit the file size information (no kFlagIncludeFileSize), for more compact file names. But note that changing the naming scheme used on an existing db and backup directory can lead to transient space amplification, as some files will be stored under two names in the shared_checksum directory. Because some backups were saved using the original 6.12 naming scheme, we offer two ways of dealing with those files: SST files generated by older 6.12 versions can either use the default naming scheme in effect when the SST files were generated (kFlagMatchInterimNaming, default, no transient space amplification) or can use a new naming scheme (no kFlagMatchInterimNaming, potential space amplification because some already stored files getting a new name). We don't have a natural way to detect which files were generated by previous 6.12 versions, but this change hacks one in by changing DB session ids to now use a more concise encoding, reducing file name length, saving ~dozen bytes from SST files, and making them visually distinct from DB ids so that they are less likely to be mixed up. Two final auxiliary notes: Recognizing that the backup file names have become a de facto part of the backup meta schema, this change makes them easier to parse and extend by putting a distinct marker, 's', before DB session ids embedded in the name. When we extend this to allow custom checksums in the name, they can get their own marker to ensure safe parsing. For backward compatibility, file size does not get a marker but is assumed for `_[0-9]+[.]` Another change from initial 6.12 default behavior is never including file custom checksum in the file name. Looking ahead to 6.13, we do not want the default behavior to cause backup space amplification for someone turning on file custom checksum checking in BackupEngine; we want that to be an easy decision. When implemented, including file custom checksums in backup file names will be a non-default option. Actual file name patterns and priorities, as regexes: kLegacyCrc32cAndFileSize OR pre-6.12 SST file -> [0-9]+_[0-9]+_[0-9]+[.]sst kFlagMatchInterimNaming set (default) AND early 6.12 SST file -> [0-9]+_[0-9a-fA-F-]+[.]sst kUseDbSessionId AND NOT kFlagIncludeFileSize -> [0-9]+_s[0-9A-Z]{20}[.]sst kUseDbSessionId AND kFlagIncludeFileSize (default) -> [0-9]+_s[0-9A-Z]{20}_[0-9]+[.]sst We might add opt-in options for more '\_' separated data in the name, but embedded file size, if present, will always be after last '\_' and before '.sst'. This change was originally applied to version 6.12. (See https://github.com/facebook/rocksdb/issues/7390) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7400 Test Plan: unit tests included. Sync point callbacks are used to mimic previous version SST files. Reviewed By: ajkr Differential Revision: D23759587 Pulled By: pdillinger fbshipit-source-id: f62d8af4e0978de0a34f26288cfbe66049b70025
4 years ago
file_copy.insert(file_copy.find_last_of('.'),
"_" + ToString(ChecksumHexToInt32(checksum_hex)) + "_" +
ToString(file_size));
} else {
Restore file size in backup table file names (and other cleanup) (#7400) Summary: Prior to 6.12, backup files using share_files_with_checksum had the file size encoded in the file name, after the last '\_' and before the last '.'. We considered this an implementation detail subject to change, and indeed removed this information from the file name (with an option to use old behavior) because it was considered ineffective/inefficient for file name uniqueness. However, some downstream RocksDB users were relying on this information since the file size is not explicitly in the backup manifest file. This primary purpose of this change is "retrofitting" the 6.12 release (not yet a public release) to simultaneously support the benefits of the new naming scheme (I/O performance and data correctness at scale) and preserve the file size information, both as default behaviors. With this change, we are essentially making the file size information encoded in the file name an official, though obscure, extension of the backup meta file format. We preserve an option (kLegacyCrc32cAndFileSize) to use the original "legacy" naming scheme, with its caveats, and make it easy to omit the file size information (no kFlagIncludeFileSize), for more compact file names. But note that changing the naming scheme used on an existing db and backup directory can lead to transient space amplification, as some files will be stored under two names in the shared_checksum directory. Because some backups were saved using the original 6.12 naming scheme, we offer two ways of dealing with those files: SST files generated by older 6.12 versions can either use the default naming scheme in effect when the SST files were generated (kFlagMatchInterimNaming, default, no transient space amplification) or can use a new naming scheme (no kFlagMatchInterimNaming, potential space amplification because some already stored files getting a new name). We don't have a natural way to detect which files were generated by previous 6.12 versions, but this change hacks one in by changing DB session ids to now use a more concise encoding, reducing file name length, saving ~dozen bytes from SST files, and making them visually distinct from DB ids so that they are less likely to be mixed up. Two final auxiliary notes: Recognizing that the backup file names have become a de facto part of the backup meta schema, this change makes them easier to parse and extend by putting a distinct marker, 's', before DB session ids embedded in the name. When we extend this to allow custom checksums in the name, they can get their own marker to ensure safe parsing. For backward compatibility, file size does not get a marker but is assumed for `_[0-9]+[.]` Another change from initial 6.12 default behavior is never including file custom checksum in the file name. Looking ahead to 6.13, we do not want the default behavior to cause backup space amplification for someone turning on file custom checksum checking in BackupEngine; we want that to be an easy decision. When implemented, including file custom checksums in backup file names will be a non-default option. Actual file name patterns and priorities, as regexes: kLegacyCrc32cAndFileSize OR pre-6.12 SST file -> [0-9]+_[0-9]+_[0-9]+[.]sst kFlagMatchInterimNaming set (default) AND early 6.12 SST file -> [0-9]+_[0-9a-fA-F-]+[.]sst kUseDbSessionId AND NOT kFlagIncludeFileSize -> [0-9]+_s[0-9A-Z]{20}[.]sst kUseDbSessionId AND kFlagIncludeFileSize (default) -> [0-9]+_s[0-9A-Z]{20}_[0-9]+[.]sst We might add opt-in options for more '\_' separated data in the name, but embedded file size, if present, will always be after last '\_' and before '.sst'. This change was originally applied to version 6.12. (See https://github.com/facebook/rocksdb/issues/7390) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7400 Test Plan: unit tests included. Sync point callbacks are used to mimic previous version SST files. Reviewed By: ajkr Differential Revision: D23759587 Pulled By: pdillinger fbshipit-source-id: f62d8af4e0978de0a34f26288cfbe66049b70025
4 years ago
file_copy.insert(file_copy.find_last_of('.'), "_s" + db_session_id);
if (GetNamingFlags() & BackupEngineOptions::kFlagIncludeFileSize) {
Restore file size in backup table file names (and other cleanup) (#7400) Summary: Prior to 6.12, backup files using share_files_with_checksum had the file size encoded in the file name, after the last '\_' and before the last '.'. We considered this an implementation detail subject to change, and indeed removed this information from the file name (with an option to use old behavior) because it was considered ineffective/inefficient for file name uniqueness. However, some downstream RocksDB users were relying on this information since the file size is not explicitly in the backup manifest file. This primary purpose of this change is "retrofitting" the 6.12 release (not yet a public release) to simultaneously support the benefits of the new naming scheme (I/O performance and data correctness at scale) and preserve the file size information, both as default behaviors. With this change, we are essentially making the file size information encoded in the file name an official, though obscure, extension of the backup meta file format. We preserve an option (kLegacyCrc32cAndFileSize) to use the original "legacy" naming scheme, with its caveats, and make it easy to omit the file size information (no kFlagIncludeFileSize), for more compact file names. But note that changing the naming scheme used on an existing db and backup directory can lead to transient space amplification, as some files will be stored under two names in the shared_checksum directory. Because some backups were saved using the original 6.12 naming scheme, we offer two ways of dealing with those files: SST files generated by older 6.12 versions can either use the default naming scheme in effect when the SST files were generated (kFlagMatchInterimNaming, default, no transient space amplification) or can use a new naming scheme (no kFlagMatchInterimNaming, potential space amplification because some already stored files getting a new name). We don't have a natural way to detect which files were generated by previous 6.12 versions, but this change hacks one in by changing DB session ids to now use a more concise encoding, reducing file name length, saving ~dozen bytes from SST files, and making them visually distinct from DB ids so that they are less likely to be mixed up. Two final auxiliary notes: Recognizing that the backup file names have become a de facto part of the backup meta schema, this change makes them easier to parse and extend by putting a distinct marker, 's', before DB session ids embedded in the name. When we extend this to allow custom checksums in the name, they can get their own marker to ensure safe parsing. For backward compatibility, file size does not get a marker but is assumed for `_[0-9]+[.]` Another change from initial 6.12 default behavior is never including file custom checksum in the file name. Looking ahead to 6.13, we do not want the default behavior to cause backup space amplification for someone turning on file custom checksum checking in BackupEngine; we want that to be an easy decision. When implemented, including file custom checksums in backup file names will be a non-default option. Actual file name patterns and priorities, as regexes: kLegacyCrc32cAndFileSize OR pre-6.12 SST file -> [0-9]+_[0-9]+_[0-9]+[.]sst kFlagMatchInterimNaming set (default) AND early 6.12 SST file -> [0-9]+_[0-9a-fA-F-]+[.]sst kUseDbSessionId AND NOT kFlagIncludeFileSize -> [0-9]+_s[0-9A-Z]{20}[.]sst kUseDbSessionId AND kFlagIncludeFileSize (default) -> [0-9]+_s[0-9A-Z]{20}_[0-9]+[.]sst We might add opt-in options for more '\_' separated data in the name, but embedded file size, if present, will always be after last '\_' and before '.sst'. This change was originally applied to version 6.12. (See https://github.com/facebook/rocksdb/issues/7390) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7400 Test Plan: unit tests included. Sync point callbacks are used to mimic previous version SST files. Reviewed By: ajkr Differential Revision: D23759587 Pulled By: pdillinger fbshipit-source-id: f62d8af4e0978de0a34f26288cfbe66049b70025
4 years ago
file_copy.insert(file_copy.find_last_of('.'),
"_" + ToString(file_size));
}
}
Restore file size in backup table file names (and other cleanup) (#7400) Summary: Prior to 6.12, backup files using share_files_with_checksum had the file size encoded in the file name, after the last '\_' and before the last '.'. We considered this an implementation detail subject to change, and indeed removed this information from the file name (with an option to use old behavior) because it was considered ineffective/inefficient for file name uniqueness. However, some downstream RocksDB users were relying on this information since the file size is not explicitly in the backup manifest file. This primary purpose of this change is "retrofitting" the 6.12 release (not yet a public release) to simultaneously support the benefits of the new naming scheme (I/O performance and data correctness at scale) and preserve the file size information, both as default behaviors. With this change, we are essentially making the file size information encoded in the file name an official, though obscure, extension of the backup meta file format. We preserve an option (kLegacyCrc32cAndFileSize) to use the original "legacy" naming scheme, with its caveats, and make it easy to omit the file size information (no kFlagIncludeFileSize), for more compact file names. But note that changing the naming scheme used on an existing db and backup directory can lead to transient space amplification, as some files will be stored under two names in the shared_checksum directory. Because some backups were saved using the original 6.12 naming scheme, we offer two ways of dealing with those files: SST files generated by older 6.12 versions can either use the default naming scheme in effect when the SST files were generated (kFlagMatchInterimNaming, default, no transient space amplification) or can use a new naming scheme (no kFlagMatchInterimNaming, potential space amplification because some already stored files getting a new name). We don't have a natural way to detect which files were generated by previous 6.12 versions, but this change hacks one in by changing DB session ids to now use a more concise encoding, reducing file name length, saving ~dozen bytes from SST files, and making them visually distinct from DB ids so that they are less likely to be mixed up. Two final auxiliary notes: Recognizing that the backup file names have become a de facto part of the backup meta schema, this change makes them easier to parse and extend by putting a distinct marker, 's', before DB session ids embedded in the name. When we extend this to allow custom checksums in the name, they can get their own marker to ensure safe parsing. For backward compatibility, file size does not get a marker but is assumed for `_[0-9]+[.]` Another change from initial 6.12 default behavior is never including file custom checksum in the file name. Looking ahead to 6.13, we do not want the default behavior to cause backup space amplification for someone turning on file custom checksum checking in BackupEngine; we want that to be an easy decision. When implemented, including file custom checksums in backup file names will be a non-default option. Actual file name patterns and priorities, as regexes: kLegacyCrc32cAndFileSize OR pre-6.12 SST file -> [0-9]+_[0-9]+_[0-9]+[.]sst kFlagMatchInterimNaming set (default) AND early 6.12 SST file -> [0-9]+_[0-9a-fA-F-]+[.]sst kUseDbSessionId AND NOT kFlagIncludeFileSize -> [0-9]+_s[0-9A-Z]{20}[.]sst kUseDbSessionId AND kFlagIncludeFileSize (default) -> [0-9]+_s[0-9A-Z]{20}_[0-9]+[.]sst We might add opt-in options for more '\_' separated data in the name, but embedded file size, if present, will always be after last '\_' and before '.sst'. This change was originally applied to version 6.12. (See https://github.com/facebook/rocksdb/issues/7390) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7400 Test Plan: unit tests included. Sync point callbacks are used to mimic previous version SST files. Reviewed By: ajkr Differential Revision: D23759587 Pulled By: pdillinger fbshipit-source-id: f62d8af4e0978de0a34f26288cfbe66049b70025
4 years ago
return file_copy;
}
Make backups openable as read-only DBs (#8142) Summary: A current limitation of backups is that you don't know the exact database state of when the backup was taken. With this new feature, you can at least inspect the backup's DB state without restoring it by opening it as a read-only DB. Rather than add something like OpenAsReadOnlyDB to the BackupEngine API, which would inhibit opening stackable DB implementations read-only (if/when their APIs support it), we instead provide a DB name and Env that can be used to open as a read-only DB. Possible follow-up work: * Add a version of GetBackupInfo for a single backup. * Let CreateNewBackup return the BackupID of the newly-created backup. Implementation details: Refactored ChrootFileSystem to split off new base class RemapFileSystem, which allows more general remapping of files. We use this base class to implement BackupEngineImpl::RemapSharedFileSystem. To minimize API impact, I decided to just add these fields `name_for_open` and `env_for_open` to those set by GetBackupInfo when include_file_details=true. Creating the RemapSharedFileSystem adds a bit to the memory consumption, perhaps unnecessarily in some cases, but this has been mitigated by (a) only initialize the RemapSharedFileSystem lazily when GetBackupInfo with include_file_details=true is called, and (b) using the existing `shared_ptr<FileInfo>` objects to hold most of the mapping data. To enhance API safety, RemapSharedFileSystem is wrapped by new ReadOnlyFileSystem which rejects any attempts to write. This uncovered a couple of places in which DB::OpenForReadOnly would write to the filesystem, so I fixed these. Added a release note because this affects logging. Additional minor refactoring in backupable_db.cc to support the new functionality. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8142 Test Plan: new test (run with ASAN and UBSAN), added to stress test and ran it for a while with amplified backup_one_in Reviewed By: ajkr Differential Revision: D27535408 Pulled By: pdillinger fbshipit-source-id: 04666d310aa0261ef6b2385c43ca793ce1dfd148
4 years ago
static inline std::string GetFileFromChecksumFile(const std::string& file) {
assert(file.size() == 0 || file[0] != '/');
std::string file_copy = file;
size_t first_underscore = file_copy.find_first_of('_');
return file_copy.erase(first_underscore,
file_copy.find_last_of('.') - first_underscore);
}
inline std::string GetBackupMetaFile(BackupID backup_id, bool tmp) const {
Make backups openable as read-only DBs (#8142) Summary: A current limitation of backups is that you don't know the exact database state of when the backup was taken. With this new feature, you can at least inspect the backup's DB state without restoring it by opening it as a read-only DB. Rather than add something like OpenAsReadOnlyDB to the BackupEngine API, which would inhibit opening stackable DB implementations read-only (if/when their APIs support it), we instead provide a DB name and Env that can be used to open as a read-only DB. Possible follow-up work: * Add a version of GetBackupInfo for a single backup. * Let CreateNewBackup return the BackupID of the newly-created backup. Implementation details: Refactored ChrootFileSystem to split off new base class RemapFileSystem, which allows more general remapping of files. We use this base class to implement BackupEngineImpl::RemapSharedFileSystem. To minimize API impact, I decided to just add these fields `name_for_open` and `env_for_open` to those set by GetBackupInfo when include_file_details=true. Creating the RemapSharedFileSystem adds a bit to the memory consumption, perhaps unnecessarily in some cases, but this has been mitigated by (a) only initialize the RemapSharedFileSystem lazily when GetBackupInfo with include_file_details=true is called, and (b) using the existing `shared_ptr<FileInfo>` objects to hold most of the mapping data. To enhance API safety, RemapSharedFileSystem is wrapped by new ReadOnlyFileSystem which rejects any attempts to write. This uncovered a couple of places in which DB::OpenForReadOnly would write to the filesystem, so I fixed these. Added a release note because this affects logging. Additional minor refactoring in backupable_db.cc to support the new functionality. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8142 Test Plan: new test (run with ASAN and UBSAN), added to stress test and ran it for a while with amplified backup_one_in Reviewed By: ajkr Differential Revision: D27535408 Pulled By: pdillinger fbshipit-source-id: 04666d310aa0261ef6b2385c43ca793ce1dfd148
4 years ago
return GetAbsolutePath(kMetaDirName) + "/" + (tmp ? "." : "") +
ROCKSDB_NAMESPACE::ToString(backup_id) + (tmp ? ".tmp" : "");
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
// If size_limit == 0, there is no size limit, copy everything.
//
// Exactly one of src and contents must be non-empty.
//
// @param src If non-empty, the file is copied from this pathname.
// @param contents If non-empty, the file will be created with these contents.
IOStatus CopyOrCreateFile(
const std::string& src, const std::string& dst,
const std::string& contents, Env* src_env, Env* dst_env,
const EnvOptions& src_env_options, bool sync, RateLimiter* rate_limiter,
uint64_t* size = nullptr, std::string* checksum_hex = nullptr,
uint64_t size_limit = 0,
std::function<void()> progress_callback = []() {});
IOStatus ReadFileAndComputeChecksum(const std::string& src,
const std::shared_ptr<FileSystem>& src_fs,
const EnvOptions& src_env_options,
uint64_t size_limit,
std::string* checksum_hex) const;
// Obtain db_id and db_session_id from the table properties of file_path
Status GetFileDbIdentities(Env* src_env, const EnvOptions& src_env_options,
const std::string& file_path,
RateLimiter* rate_limiter, std::string* db_id,
std::string* db_session_id);
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
struct CopyOrCreateResult {
~CopyOrCreateResult() {
// The Status needs to be ignored here for two reasons.
// First, if the BackupEngineImpl shuts down with jobs outstanding, then
// it is possible that the Status in the future/promise is never read,
// resulting in an unchecked Status. Second, if there are items in the
// channel when the BackupEngineImpl is shutdown, these will also have
// Status that have not been checked. This
// TODO: Fix those issues so that the Status
io_status.PermitUncheckedError();
}
uint64_t size;
std::string checksum_hex;
std::string db_id;
std::string db_session_id;
IOStatus io_status;
};
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
// Exactly one of src_path and contents must be non-empty. If src_path is
// non-empty, the file is copied from this pathname. Otherwise, if contents is
// non-empty, the file will be created at dst_path with these contents.
struct CopyOrCreateWorkItem {
std::string src_path;
std::string dst_path;
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
std::string contents;
Env* src_env;
Env* dst_env;
EnvOptions src_env_options;
bool sync;
RateLimiter* rate_limiter;
uint64_t size_limit;
Statistics* stats;
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
std::promise<CopyOrCreateResult> result;
std::function<void()> progress_callback;
std::string src_checksum_func_name;
std::string src_checksum_hex;
std::string db_id;
std::string db_session_id;
utilities/backupable : Fix coverity issues Summary: 1. Class BackupMeta ``` 52 : timestamp_(0), size_(0), meta_filename_(meta_filename), CID 1168103 (#1 of 1): Uninitialized scalar field (UNINIT_CTOR) 2. uninit_member: Non-static class member sequence_number_ is not initialized in this constructor nor in any functions that it calls. 153 file_infos_(file_infos), env_(env) {} ``` 2. class BackupEngineImpl ``` 513 } 7. uninit_member: Non-static class member latest_backup_id_ is not initialized in this constructor nor in any functions that it calls. CID 1322803 (#1 of 1): Uninitialized scalar field (UNINIT_CTOR) 9. uninit_member: Non-static class member latest_valid_backup_id_ is not initialized in this constructor nor in any functions that it calls. 514} ``` 3. struct BackupAfterCopyOrCreateWorkItem ``` 368 struct BackupAfterCopyOrCreateWorkItem { 369 std::future<CopyOrCreateResult> result; 1. member_decl: Class member declaration for shared. 370 bool shared; 3. member_decl: Class member declaration for needed_to_copy. 371 bool needed_to_copy; 5. member_decl: Class member declaration for backup_env. 372 Env* backup_env; 373 std::string dst_path_tmp; 374 std::string dst_path; 375 std::string dst_relative; 2. uninit_member: Non-static class member shared is not initialized in this constructor nor in any functions that it calls. 4. uninit_member: Non-static class member needed_to_copy is not initialized in this constructor nor in any functions that it calls. CID 1396122 (#1 of 1): Uninitialized pointer field (UNINIT_CTOR) 6. uninit_member: Non-static class member backup_env is not initialized in this constructor nor in any functions that it calls. 376 BackupAfterCopyOrCreateWorkItem() {} ``` 4. struct CopyOrCreateWorkItem ``` 318 struct CopyOrCreateWorkItem { 319 std::string src_path; 320 std::string dst_path; 321 std::string contents; 1. member_decl: Class member declaration for src_env. 322 Env* src_env; 3. member_decl: Class member declaration for dst_env. 323 Env* dst_env; 5. member_decl: Class member declaration for sync. 324 bool sync; 7. member_decl: Class member declaration for rate_limiter. 325 RateLimiter* rate_limiter; 9. member_decl: Class member declaration for size_limit. 326 uint64_t size_limit; 327 std::promise<CopyOrCreateResult> result; 328 std::function<void()> progress_callback; 329 2. uninit_member: Non-static class member src_env is not initialized in this constructor nor in any functions that it calls. 4. uninit_member: Non-static class member dst_env is not initialized in this constructor nor in any functions that it calls. 6. uninit_member: Non-static class member sync is not initialized in this constructor nor in any functions that it calls. 8. uninit_member: Non-static class member rate_limiter is not initialized in this constructor nor in any functions that it calls. CID 1396123 (#1 of 1): Uninitialized pointer field (UNINIT_CTOR) 10. uninit_member: Non-static class member size_limit is not initialized in this constructor nor in any functions that it calls. 330 CopyOrCreateWorkItem() {} ``` 5. struct RestoreAfterCopyOrCreateWorkItem ``` struct RestoreAfterCopyOrCreateWorkItem { 410 std::future<CopyOrCreateResult> result; 1. member_decl: Class member declaration for checksum_value. 411 uint32_t checksum_value; CID 1396153 (#1 of 1): Uninitialized scalar field (UNINIT_CTOR) 2. uninit_member: Non-static class member checksum_value is not initialized in this constructor nor in any functions that it calls. 412 RestoreAfterCopyOrCreateWorkItem() {} ``` Closes https://github.com/facebook/rocksdb/pull/3131 Differential Revision: D6428556 Pulled By: sagar0 fbshipit-source-id: a86675444543eff028e3cae6942197a143a112c4
7 years ago
CopyOrCreateWorkItem()
: src_path(""),
dst_path(""),
contents(""),
src_env(nullptr),
dst_env(nullptr),
src_env_options(),
sync(false),
rate_limiter(nullptr),
size_limit(0),
stats(nullptr),
src_checksum_func_name(kUnknownFileChecksumFuncName),
src_checksum_hex(""),
db_id(""),
db_session_id("") {}
utilities/backupable : Fix coverity issues Summary: 1. Class BackupMeta ``` 52 : timestamp_(0), size_(0), meta_filename_(meta_filename), CID 1168103 (#1 of 1): Uninitialized scalar field (UNINIT_CTOR) 2. uninit_member: Non-static class member sequence_number_ is not initialized in this constructor nor in any functions that it calls. 153 file_infos_(file_infos), env_(env) {} ``` 2. class BackupEngineImpl ``` 513 } 7. uninit_member: Non-static class member latest_backup_id_ is not initialized in this constructor nor in any functions that it calls. CID 1322803 (#1 of 1): Uninitialized scalar field (UNINIT_CTOR) 9. uninit_member: Non-static class member latest_valid_backup_id_ is not initialized in this constructor nor in any functions that it calls. 514} ``` 3. struct BackupAfterCopyOrCreateWorkItem ``` 368 struct BackupAfterCopyOrCreateWorkItem { 369 std::future<CopyOrCreateResult> result; 1. member_decl: Class member declaration for shared. 370 bool shared; 3. member_decl: Class member declaration for needed_to_copy. 371 bool needed_to_copy; 5. member_decl: Class member declaration for backup_env. 372 Env* backup_env; 373 std::string dst_path_tmp; 374 std::string dst_path; 375 std::string dst_relative; 2. uninit_member: Non-static class member shared is not initialized in this constructor nor in any functions that it calls. 4. uninit_member: Non-static class member needed_to_copy is not initialized in this constructor nor in any functions that it calls. CID 1396122 (#1 of 1): Uninitialized pointer field (UNINIT_CTOR) 6. uninit_member: Non-static class member backup_env is not initialized in this constructor nor in any functions that it calls. 376 BackupAfterCopyOrCreateWorkItem() {} ``` 4. struct CopyOrCreateWorkItem ``` 318 struct CopyOrCreateWorkItem { 319 std::string src_path; 320 std::string dst_path; 321 std::string contents; 1. member_decl: Class member declaration for src_env. 322 Env* src_env; 3. member_decl: Class member declaration for dst_env. 323 Env* dst_env; 5. member_decl: Class member declaration for sync. 324 bool sync; 7. member_decl: Class member declaration for rate_limiter. 325 RateLimiter* rate_limiter; 9. member_decl: Class member declaration for size_limit. 326 uint64_t size_limit; 327 std::promise<CopyOrCreateResult> result; 328 std::function<void()> progress_callback; 329 2. uninit_member: Non-static class member src_env is not initialized in this constructor nor in any functions that it calls. 4. uninit_member: Non-static class member dst_env is not initialized in this constructor nor in any functions that it calls. 6. uninit_member: Non-static class member sync is not initialized in this constructor nor in any functions that it calls. 8. uninit_member: Non-static class member rate_limiter is not initialized in this constructor nor in any functions that it calls. CID 1396123 (#1 of 1): Uninitialized pointer field (UNINIT_CTOR) 10. uninit_member: Non-static class member size_limit is not initialized in this constructor nor in any functions that it calls. 330 CopyOrCreateWorkItem() {} ``` 5. struct RestoreAfterCopyOrCreateWorkItem ``` struct RestoreAfterCopyOrCreateWorkItem { 410 std::future<CopyOrCreateResult> result; 1. member_decl: Class member declaration for checksum_value. 411 uint32_t checksum_value; CID 1396153 (#1 of 1): Uninitialized scalar field (UNINIT_CTOR) 2. uninit_member: Non-static class member checksum_value is not initialized in this constructor nor in any functions that it calls. 412 RestoreAfterCopyOrCreateWorkItem() {} ``` Closes https://github.com/facebook/rocksdb/pull/3131 Differential Revision: D6428556 Pulled By: sagar0 fbshipit-source-id: a86675444543eff028e3cae6942197a143a112c4
7 years ago
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
CopyOrCreateWorkItem(const CopyOrCreateWorkItem&) = delete;
CopyOrCreateWorkItem& operator=(const CopyOrCreateWorkItem&) = delete;
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
CopyOrCreateWorkItem(CopyOrCreateWorkItem&& o) ROCKSDB_NOEXCEPT {
*this = std::move(o);
}
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
CopyOrCreateWorkItem& operator=(CopyOrCreateWorkItem&& o) ROCKSDB_NOEXCEPT {
src_path = std::move(o.src_path);
dst_path = std::move(o.dst_path);
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
contents = std::move(o.contents);
src_env = o.src_env;
dst_env = o.dst_env;
src_env_options = std::move(o.src_env_options);
sync = o.sync;
rate_limiter = o.rate_limiter;
size_limit = o.size_limit;
stats = o.stats;
result = std::move(o.result);
progress_callback = std::move(o.progress_callback);
src_checksum_func_name = std::move(o.src_checksum_func_name);
src_checksum_hex = std::move(o.src_checksum_hex);
db_id = std::move(o.db_id);
db_session_id = std::move(o.db_session_id);
return *this;
}
CopyOrCreateWorkItem(
std::string _src_path, std::string _dst_path, std::string _contents,
Env* _src_env, Env* _dst_env, EnvOptions _src_env_options, bool _sync,
RateLimiter* _rate_limiter, uint64_t _size_limit, Statistics* _stats,
std::function<void()> _progress_callback = []() {},
const std::string& _src_checksum_func_name =
kUnknownFileChecksumFuncName,
const std::string& _src_checksum_hex = "",
const std::string& _db_id = "", const std::string& _db_session_id = "")
: src_path(std::move(_src_path)),
dst_path(std::move(_dst_path)),
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
contents(std::move(_contents)),
src_env(_src_env),
dst_env(_dst_env),
src_env_options(std::move(_src_env_options)),
sync(_sync),
rate_limiter(_rate_limiter),
size_limit(_size_limit),
stats(_stats),
progress_callback(_progress_callback),
src_checksum_func_name(_src_checksum_func_name),
src_checksum_hex(_src_checksum_hex),
db_id(_db_id),
db_session_id(_db_session_id) {}
};
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
struct BackupAfterCopyOrCreateWorkItem {
std::future<CopyOrCreateResult> result;
bool shared;
bool needed_to_copy;
Env* backup_env;
std::string dst_path_tmp;
std::string dst_path;
std::string dst_relative;
utilities/backupable : Fix coverity issues Summary: 1. Class BackupMeta ``` 52 : timestamp_(0), size_(0), meta_filename_(meta_filename), CID 1168103 (#1 of 1): Uninitialized scalar field (UNINIT_CTOR) 2. uninit_member: Non-static class member sequence_number_ is not initialized in this constructor nor in any functions that it calls. 153 file_infos_(file_infos), env_(env) {} ``` 2. class BackupEngineImpl ``` 513 } 7. uninit_member: Non-static class member latest_backup_id_ is not initialized in this constructor nor in any functions that it calls. CID 1322803 (#1 of 1): Uninitialized scalar field (UNINIT_CTOR) 9. uninit_member: Non-static class member latest_valid_backup_id_ is not initialized in this constructor nor in any functions that it calls. 514} ``` 3. struct BackupAfterCopyOrCreateWorkItem ``` 368 struct BackupAfterCopyOrCreateWorkItem { 369 std::future<CopyOrCreateResult> result; 1. member_decl: Class member declaration for shared. 370 bool shared; 3. member_decl: Class member declaration for needed_to_copy. 371 bool needed_to_copy; 5. member_decl: Class member declaration for backup_env. 372 Env* backup_env; 373 std::string dst_path_tmp; 374 std::string dst_path; 375 std::string dst_relative; 2. uninit_member: Non-static class member shared is not initialized in this constructor nor in any functions that it calls. 4. uninit_member: Non-static class member needed_to_copy is not initialized in this constructor nor in any functions that it calls. CID 1396122 (#1 of 1): Uninitialized pointer field (UNINIT_CTOR) 6. uninit_member: Non-static class member backup_env is not initialized in this constructor nor in any functions that it calls. 376 BackupAfterCopyOrCreateWorkItem() {} ``` 4. struct CopyOrCreateWorkItem ``` 318 struct CopyOrCreateWorkItem { 319 std::string src_path; 320 std::string dst_path; 321 std::string contents; 1. member_decl: Class member declaration for src_env. 322 Env* src_env; 3. member_decl: Class member declaration for dst_env. 323 Env* dst_env; 5. member_decl: Class member declaration for sync. 324 bool sync; 7. member_decl: Class member declaration for rate_limiter. 325 RateLimiter* rate_limiter; 9. member_decl: Class member declaration for size_limit. 326 uint64_t size_limit; 327 std::promise<CopyOrCreateResult> result; 328 std::function<void()> progress_callback; 329 2. uninit_member: Non-static class member src_env is not initialized in this constructor nor in any functions that it calls. 4. uninit_member: Non-static class member dst_env is not initialized in this constructor nor in any functions that it calls. 6. uninit_member: Non-static class member sync is not initialized in this constructor nor in any functions that it calls. 8. uninit_member: Non-static class member rate_limiter is not initialized in this constructor nor in any functions that it calls. CID 1396123 (#1 of 1): Uninitialized pointer field (UNINIT_CTOR) 10. uninit_member: Non-static class member size_limit is not initialized in this constructor nor in any functions that it calls. 330 CopyOrCreateWorkItem() {} ``` 5. struct RestoreAfterCopyOrCreateWorkItem ``` struct RestoreAfterCopyOrCreateWorkItem { 410 std::future<CopyOrCreateResult> result; 1. member_decl: Class member declaration for checksum_value. 411 uint32_t checksum_value; CID 1396153 (#1 of 1): Uninitialized scalar field (UNINIT_CTOR) 2. uninit_member: Non-static class member checksum_value is not initialized in this constructor nor in any functions that it calls. 412 RestoreAfterCopyOrCreateWorkItem() {} ``` Closes https://github.com/facebook/rocksdb/pull/3131 Differential Revision: D6428556 Pulled By: sagar0 fbshipit-source-id: a86675444543eff028e3cae6942197a143a112c4
7 years ago
BackupAfterCopyOrCreateWorkItem()
: shared(false),
needed_to_copy(false),
backup_env(nullptr),
dst_path_tmp(""),
dst_path(""),
dst_relative("") {}
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
BackupAfterCopyOrCreateWorkItem(BackupAfterCopyOrCreateWorkItem&& o)
ROCKSDB_NOEXCEPT {
*this = std::move(o);
}
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
BackupAfterCopyOrCreateWorkItem& operator=(
BackupAfterCopyOrCreateWorkItem&& o) ROCKSDB_NOEXCEPT {
result = std::move(o.result);
shared = o.shared;
needed_to_copy = o.needed_to_copy;
backup_env = o.backup_env;
dst_path_tmp = std::move(o.dst_path_tmp);
dst_path = std::move(o.dst_path);
dst_relative = std::move(o.dst_relative);
return *this;
}
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
BackupAfterCopyOrCreateWorkItem(std::future<CopyOrCreateResult>&& _result,
bool _shared, bool _needed_to_copy,
Env* _backup_env, std::string _dst_path_tmp,
std::string _dst_path,
std::string _dst_relative)
: result(std::move(_result)),
shared(_shared),
needed_to_copy(_needed_to_copy),
backup_env(_backup_env),
dst_path_tmp(std::move(_dst_path_tmp)),
dst_path(std::move(_dst_path)),
dst_relative(std::move(_dst_relative)) {}
};
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
struct RestoreAfterCopyOrCreateWorkItem {
std::future<CopyOrCreateResult> result;
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
std::string from_file;
std::string to_file;
std::string checksum_hex;
RestoreAfterCopyOrCreateWorkItem() : checksum_hex("") {}
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
RestoreAfterCopyOrCreateWorkItem(std::future<CopyOrCreateResult>&& _result,
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
const std::string& _from_file,
const std::string& _to_file,
const std::string& _checksum_hex)
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
: result(std::move(_result)),
from_file(_from_file),
to_file(_to_file),
checksum_hex(_checksum_hex) {}
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
RestoreAfterCopyOrCreateWorkItem(RestoreAfterCopyOrCreateWorkItem&& o)
ROCKSDB_NOEXCEPT {
*this = std::move(o);
}
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
RestoreAfterCopyOrCreateWorkItem& operator=(
RestoreAfterCopyOrCreateWorkItem&& o) ROCKSDB_NOEXCEPT {
result = std::move(o.result);
checksum_hex = std::move(o.checksum_hex);
return *this;
}
};
bool initialized_;
std::mutex byte_report_mutex_;
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
mutable channel<CopyOrCreateWorkItem> files_to_copy_or_create_;
std::vector<port::Thread> threads_;
std::atomic<CpuPriority> threads_cpu_priority_;
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
// Certain operations like PurgeOldBackups and DeleteBackup will trigger
// automatic GarbageCollect (true) unless we've already done one in this
// session and have not failed to delete backup files since then (false).
bool might_need_garbage_collect_ = true;
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
// Adds a file to the backup work queue to be copied or created if it doesn't
// already exist.
//
// Exactly one of src_dir and contents must be non-empty.
//
// @param src_dir If non-empty, the file in this directory named fname will be
// copied.
// @param fname Name of destination file and, in case of copy, source file.
// @param contents If non-empty, the file will be created with these contents.
IOStatus AddBackupFileWorkItem(
std::unordered_set<std::string>& live_dst_paths,
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
std::vector<BackupAfterCopyOrCreateWorkItem>& backup_items_to_finish,
BackupID backup_id, bool shared, const std::string& src_dir,
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
const std::string& fname, // starts with "/"
const EnvOptions& src_env_options, RateLimiter* rate_limiter,
FileType file_type, uint64_t size_bytes, Statistics* stats,
uint64_t size_limit = 0, bool shared_checksum = false,
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
std::function<void()> progress_callback = []() {},
const std::string& contents = std::string(),
const std::string& src_checksum_func_name = kUnknownFileChecksumFuncName,
const std::string& src_checksum_str = kUnknownFileChecksum);
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
// backup state data
BackupID latest_backup_id_;
BackupID latest_valid_backup_id_;
std::map<BackupID, std::unique_ptr<BackupMeta>> backups_;
std::map<BackupID, std::pair<IOStatus, std::unique_ptr<BackupMeta>>>
corrupt_backups_;
std::unordered_map<std::string,
std::shared_ptr<FileInfo>> backuped_file_infos_;
std::atomic<bool> stop_backup_;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
// options data
BackupEngineOptions options_;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
Env* db_env_;
Env* backup_env_;
// directories
std::unique_ptr<FSDirectory> backup_directory_;
std::unique_ptr<FSDirectory> shared_directory_;
std::unique_ptr<FSDirectory> meta_directory_;
std::unique_ptr<FSDirectory> private_directory_;
static const size_t kDefaultCopyFileBufferSize = 5 * 1024 * 1024LL; // 5MB
bool read_only_;
BackupStatistics backup_statistics_;
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
std::unordered_set<std::string> reported_ignored_fields_;
static const size_t kMaxAppMetaSize = 1024 * 1024; // 1MB
std::shared_ptr<FileSystem> db_fs_;
std::shared_ptr<FileSystem> backup_fs_;
IOOptions io_options_ = IOOptions();
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
public:
std::unique_ptr<TEST_FutureSchemaVersion2Options> test_future_options_;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
};
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
// -------- BackupEngineImplThreadSafe class ---------
// This locking layer for thread safety in the public API is layered on
// top to prevent accidental recursive locking with RWMutex, which is UB.
// Note: BackupEngineReadOnlyBase inherited twice, but has no fields
class BackupEngineImplThreadSafe : public BackupEngine,
public BackupEngineReadOnly {
public:
BackupEngineImplThreadSafe(const BackupEngineOptions& options, Env* db_env,
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
bool read_only = false)
: impl_(options, db_env, read_only) {}
~BackupEngineImplThreadSafe() override {}
using BackupEngine::CreateNewBackupWithMetadata;
IOStatus CreateNewBackupWithMetadata(const CreateBackupOptions& options,
DB* db, const std::string& app_metadata,
BackupID* new_backup_id) override {
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
WriteLock lock(&mutex_);
return impl_.CreateNewBackupWithMetadata(options, db, app_metadata,
new_backup_id);
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
}
IOStatus PurgeOldBackups(uint32_t num_backups_to_keep) override {
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
WriteLock lock(&mutex_);
return impl_.PurgeOldBackups(num_backups_to_keep);
}
IOStatus DeleteBackup(BackupID backup_id) override {
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
WriteLock lock(&mutex_);
return impl_.DeleteBackup(backup_id);
}
void StopBackup() override {
// No locking needed
impl_.StopBackup();
}
IOStatus GarbageCollect() override {
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
WriteLock lock(&mutex_);
return impl_.GarbageCollect();
}
Status GetLatestBackupInfo(BackupInfo* backup_info,
bool include_file_details = false) const override {
ReadLock lock(&mutex_);
return impl_.GetBackupInfo(kLatestBackupIDMarker, backup_info,
include_file_details);
}
Status GetBackupInfo(BackupID backup_id, BackupInfo* backup_info,
bool include_file_details = false) const override {
ReadLock lock(&mutex_);
return impl_.GetBackupInfo(backup_id, backup_info, include_file_details);
}
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
void GetBackupInfo(std::vector<BackupInfo>* backup_info,
bool include_file_details) const override {
ReadLock lock(&mutex_);
impl_.GetBackupInfo(backup_info, include_file_details);
}
void GetCorruptedBackups(
std::vector<BackupID>* corrupt_backup_ids) const override {
ReadLock lock(&mutex_);
impl_.GetCorruptedBackups(corrupt_backup_ids);
}
using BackupEngine::RestoreDBFromBackup;
IOStatus RestoreDBFromBackup(const RestoreOptions& options,
BackupID backup_id, const std::string& db_dir,
const std::string& wal_dir) const override {
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
ReadLock lock(&mutex_);
return impl_.RestoreDBFromBackup(options, backup_id, db_dir, wal_dir);
}
using BackupEngine::RestoreDBFromLatestBackup;
IOStatus RestoreDBFromLatestBackup(
const RestoreOptions& options, const std::string& db_dir,
const std::string& wal_dir) const override {
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
// Defer to above function, which locks
return RestoreDBFromBackup(options, kLatestBackupIDMarker, db_dir, wal_dir);
}
IOStatus VerifyBackup(BackupID backup_id,
bool verify_with_checksum = false) const override {
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
ReadLock lock(&mutex_);
return impl_.VerifyBackup(backup_id, verify_with_checksum);
}
// Not public API but needed
IOStatus Initialize() {
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
// No locking needed
return impl_.Initialize();
}
// Not public API but used in testing
void TEST_EnableWriteFutureSchemaVersion2(
const TEST_FutureSchemaVersion2Options& options) {
impl_.test_future_options_.reset(
new TEST_FutureSchemaVersion2Options(options));
}
private:
mutable port::RWMutex mutex_;
BackupEngineImpl impl_;
};
IOStatus BackupEngine::Open(const BackupEngineOptions& options, Env* env,
BackupEngine** backup_engine_ptr) {
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
std::unique_ptr<BackupEngineImplThreadSafe> backup_engine(
new BackupEngineImplThreadSafe(options, env));
auto s = backup_engine->Initialize();
if (!s.ok()) {
*backup_engine_ptr = nullptr;
return s;
}
*backup_engine_ptr = backup_engine.release();
return IOStatus::OK();
}
BackupEngineImpl::BackupEngineImpl(const BackupEngineOptions& options,
Env* db_env, bool read_only)
: initialized_(false),
threads_cpu_priority_(),
utilities/backupable : Fix coverity issues Summary: 1. Class BackupMeta ``` 52 : timestamp_(0), size_(0), meta_filename_(meta_filename), CID 1168103 (#1 of 1): Uninitialized scalar field (UNINIT_CTOR) 2. uninit_member: Non-static class member sequence_number_ is not initialized in this constructor nor in any functions that it calls. 153 file_infos_(file_infos), env_(env) {} ``` 2. class BackupEngineImpl ``` 513 } 7. uninit_member: Non-static class member latest_backup_id_ is not initialized in this constructor nor in any functions that it calls. CID 1322803 (#1 of 1): Uninitialized scalar field (UNINIT_CTOR) 9. uninit_member: Non-static class member latest_valid_backup_id_ is not initialized in this constructor nor in any functions that it calls. 514} ``` 3. struct BackupAfterCopyOrCreateWorkItem ``` 368 struct BackupAfterCopyOrCreateWorkItem { 369 std::future<CopyOrCreateResult> result; 1. member_decl: Class member declaration for shared. 370 bool shared; 3. member_decl: Class member declaration for needed_to_copy. 371 bool needed_to_copy; 5. member_decl: Class member declaration for backup_env. 372 Env* backup_env; 373 std::string dst_path_tmp; 374 std::string dst_path; 375 std::string dst_relative; 2. uninit_member: Non-static class member shared is not initialized in this constructor nor in any functions that it calls. 4. uninit_member: Non-static class member needed_to_copy is not initialized in this constructor nor in any functions that it calls. CID 1396122 (#1 of 1): Uninitialized pointer field (UNINIT_CTOR) 6. uninit_member: Non-static class member backup_env is not initialized in this constructor nor in any functions that it calls. 376 BackupAfterCopyOrCreateWorkItem() {} ``` 4. struct CopyOrCreateWorkItem ``` 318 struct CopyOrCreateWorkItem { 319 std::string src_path; 320 std::string dst_path; 321 std::string contents; 1. member_decl: Class member declaration for src_env. 322 Env* src_env; 3. member_decl: Class member declaration for dst_env. 323 Env* dst_env; 5. member_decl: Class member declaration for sync. 324 bool sync; 7. member_decl: Class member declaration for rate_limiter. 325 RateLimiter* rate_limiter; 9. member_decl: Class member declaration for size_limit. 326 uint64_t size_limit; 327 std::promise<CopyOrCreateResult> result; 328 std::function<void()> progress_callback; 329 2. uninit_member: Non-static class member src_env is not initialized in this constructor nor in any functions that it calls. 4. uninit_member: Non-static class member dst_env is not initialized in this constructor nor in any functions that it calls. 6. uninit_member: Non-static class member sync is not initialized in this constructor nor in any functions that it calls. 8. uninit_member: Non-static class member rate_limiter is not initialized in this constructor nor in any functions that it calls. CID 1396123 (#1 of 1): Uninitialized pointer field (UNINIT_CTOR) 10. uninit_member: Non-static class member size_limit is not initialized in this constructor nor in any functions that it calls. 330 CopyOrCreateWorkItem() {} ``` 5. struct RestoreAfterCopyOrCreateWorkItem ``` struct RestoreAfterCopyOrCreateWorkItem { 410 std::future<CopyOrCreateResult> result; 1. member_decl: Class member declaration for checksum_value. 411 uint32_t checksum_value; CID 1396153 (#1 of 1): Uninitialized scalar field (UNINIT_CTOR) 2. uninit_member: Non-static class member checksum_value is not initialized in this constructor nor in any functions that it calls. 412 RestoreAfterCopyOrCreateWorkItem() {} ``` Closes https://github.com/facebook/rocksdb/pull/3131 Differential Revision: D6428556 Pulled By: sagar0 fbshipit-source-id: a86675444543eff028e3cae6942197a143a112c4
7 years ago
latest_backup_id_(0),
latest_valid_backup_id_(0),
stop_backup_(false),
options_(options),
db_env_(db_env),
backup_env_(options.backup_env != nullptr ? options.backup_env : db_env_),
read_only_(read_only) {
if (options_.backup_rate_limiter == nullptr &&
options_.backup_rate_limit > 0) {
options_.backup_rate_limiter.reset(
NewGenericRateLimiter(options_.backup_rate_limit));
}
if (options_.restore_rate_limiter == nullptr &&
options_.restore_rate_limit > 0) {
options_.restore_rate_limiter.reset(
NewGenericRateLimiter(options_.restore_rate_limit));
}
db_fs_ = db_env_->GetFileSystem();
backup_fs_ = backup_env_->GetFileSystem();
}
BackupEngineImpl::~BackupEngineImpl() {
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
files_to_copy_or_create_.sendEof();
for (auto& t : threads_) {
t.join();
}
LogFlush(options_.info_log);
for (const auto& it : corrupt_backups_) {
it.second.first.PermitUncheckedError();
}
}
IOStatus BackupEngineImpl::Initialize() {
assert(!initialized_);
initialized_ = true;
if (read_only_) {
ROCKS_LOG_INFO(options_.info_log, "Starting read_only backup engine");
}
options_.Dump(options_.info_log);
Make backups openable as read-only DBs (#8142) Summary: A current limitation of backups is that you don't know the exact database state of when the backup was taken. With this new feature, you can at least inspect the backup's DB state without restoring it by opening it as a read-only DB. Rather than add something like OpenAsReadOnlyDB to the BackupEngine API, which would inhibit opening stackable DB implementations read-only (if/when their APIs support it), we instead provide a DB name and Env that can be used to open as a read-only DB. Possible follow-up work: * Add a version of GetBackupInfo for a single backup. * Let CreateNewBackup return the BackupID of the newly-created backup. Implementation details: Refactored ChrootFileSystem to split off new base class RemapFileSystem, which allows more general remapping of files. We use this base class to implement BackupEngineImpl::RemapSharedFileSystem. To minimize API impact, I decided to just add these fields `name_for_open` and `env_for_open` to those set by GetBackupInfo when include_file_details=true. Creating the RemapSharedFileSystem adds a bit to the memory consumption, perhaps unnecessarily in some cases, but this has been mitigated by (a) only initialize the RemapSharedFileSystem lazily when GetBackupInfo with include_file_details=true is called, and (b) using the existing `shared_ptr<FileInfo>` objects to hold most of the mapping data. To enhance API safety, RemapSharedFileSystem is wrapped by new ReadOnlyFileSystem which rejects any attempts to write. This uncovered a couple of places in which DB::OpenForReadOnly would write to the filesystem, so I fixed these. Added a release note because this affects logging. Additional minor refactoring in backupable_db.cc to support the new functionality. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8142 Test Plan: new test (run with ASAN and UBSAN), added to stress test and ran it for a while with amplified backup_one_in Reviewed By: ajkr Differential Revision: D27535408 Pulled By: pdillinger fbshipit-source-id: 04666d310aa0261ef6b2385c43ca793ce1dfd148
4 years ago
auto meta_path = GetAbsolutePath(kMetaDirName);
if (!read_only_) {
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
// we might need to clean up from previous crash or I/O errors
might_need_garbage_collect_ = true;
if (options_.max_valid_backups_to_open != port::kMaxInt32) {
options_.max_valid_backups_to_open = port::kMaxInt32;
ROCKS_LOG_WARN(
options_.info_log,
"`max_valid_backups_to_open` is not set to the default value. Ignoring "
"its value since BackupEngine is not read-only.");
}
// gather the list of directories that we need to create
std::vector<std::pair<std::string, std::unique_ptr<FSDirectory>*>>
directories;
directories.emplace_back(GetAbsolutePath(), &backup_directory_);
if (options_.share_table_files) {
if (options_.share_files_with_checksum) {
directories.emplace_back(
GetAbsolutePath(GetSharedFileWithChecksumRel()),
&shared_directory_);
} else {
directories.emplace_back(GetAbsolutePath(GetSharedFileRel()),
&shared_directory_);
}
}
Make backups openable as read-only DBs (#8142) Summary: A current limitation of backups is that you don't know the exact database state of when the backup was taken. With this new feature, you can at least inspect the backup's DB state without restoring it by opening it as a read-only DB. Rather than add something like OpenAsReadOnlyDB to the BackupEngine API, which would inhibit opening stackable DB implementations read-only (if/when their APIs support it), we instead provide a DB name and Env that can be used to open as a read-only DB. Possible follow-up work: * Add a version of GetBackupInfo for a single backup. * Let CreateNewBackup return the BackupID of the newly-created backup. Implementation details: Refactored ChrootFileSystem to split off new base class RemapFileSystem, which allows more general remapping of files. We use this base class to implement BackupEngineImpl::RemapSharedFileSystem. To minimize API impact, I decided to just add these fields `name_for_open` and `env_for_open` to those set by GetBackupInfo when include_file_details=true. Creating the RemapSharedFileSystem adds a bit to the memory consumption, perhaps unnecessarily in some cases, but this has been mitigated by (a) only initialize the RemapSharedFileSystem lazily when GetBackupInfo with include_file_details=true is called, and (b) using the existing `shared_ptr<FileInfo>` objects to hold most of the mapping data. To enhance API safety, RemapSharedFileSystem is wrapped by new ReadOnlyFileSystem which rejects any attempts to write. This uncovered a couple of places in which DB::OpenForReadOnly would write to the filesystem, so I fixed these. Added a release note because this affects logging. Additional minor refactoring in backupable_db.cc to support the new functionality. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8142 Test Plan: new test (run with ASAN and UBSAN), added to stress test and ran it for a while with amplified backup_one_in Reviewed By: ajkr Differential Revision: D27535408 Pulled By: pdillinger fbshipit-source-id: 04666d310aa0261ef6b2385c43ca793ce1dfd148
4 years ago
directories.emplace_back(GetAbsolutePath(kPrivateDirName),
&private_directory_);
Make backups openable as read-only DBs (#8142) Summary: A current limitation of backups is that you don't know the exact database state of when the backup was taken. With this new feature, you can at least inspect the backup's DB state without restoring it by opening it as a read-only DB. Rather than add something like OpenAsReadOnlyDB to the BackupEngine API, which would inhibit opening stackable DB implementations read-only (if/when their APIs support it), we instead provide a DB name and Env that can be used to open as a read-only DB. Possible follow-up work: * Add a version of GetBackupInfo for a single backup. * Let CreateNewBackup return the BackupID of the newly-created backup. Implementation details: Refactored ChrootFileSystem to split off new base class RemapFileSystem, which allows more general remapping of files. We use this base class to implement BackupEngineImpl::RemapSharedFileSystem. To minimize API impact, I decided to just add these fields `name_for_open` and `env_for_open` to those set by GetBackupInfo when include_file_details=true. Creating the RemapSharedFileSystem adds a bit to the memory consumption, perhaps unnecessarily in some cases, but this has been mitigated by (a) only initialize the RemapSharedFileSystem lazily when GetBackupInfo with include_file_details=true is called, and (b) using the existing `shared_ptr<FileInfo>` objects to hold most of the mapping data. To enhance API safety, RemapSharedFileSystem is wrapped by new ReadOnlyFileSystem which rejects any attempts to write. This uncovered a couple of places in which DB::OpenForReadOnly would write to the filesystem, so I fixed these. Added a release note because this affects logging. Additional minor refactoring in backupable_db.cc to support the new functionality. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8142 Test Plan: new test (run with ASAN and UBSAN), added to stress test and ran it for a while with amplified backup_one_in Reviewed By: ajkr Differential Revision: D27535408 Pulled By: pdillinger fbshipit-source-id: 04666d310aa0261ef6b2385c43ca793ce1dfd148
4 years ago
directories.emplace_back(meta_path, &meta_directory_);
// create all the dirs we need
for (const auto& d : directories) {
IOStatus io_s =
backup_fs_->CreateDirIfMissing(d.first, io_options_, nullptr);
if (io_s.ok()) {
io_s =
backup_fs_->NewDirectory(d.first, io_options_, d.second, nullptr);
}
if (!io_s.ok()) {
return io_s;
}
}
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
std::vector<std::string> backup_meta_files;
{
IOStatus io_s = backup_fs_->GetChildren(meta_path, io_options_,
&backup_meta_files, nullptr);
if (io_s.IsNotFound()) {
return IOStatus::NotFound(meta_path + " is missing");
} else if (!io_s.ok()) {
return io_s;
}
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
// create backups_ structure
for (auto& file : backup_meta_files) {
ROCKS_LOG_INFO(options_.info_log, "Detected backup %s", file.c_str());
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
BackupID backup_id = 0;
sscanf(file.c_str(), "%u", &backup_id);
if (backup_id == 0 || file != ROCKSDB_NAMESPACE::ToString(backup_id)) {
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
// Invalid file name, will be deleted with auto-GC when user
// initiates an append or write operation. (Behave as read-only until
// then.)
ROCKS_LOG_INFO(options_.info_log, "Skipping unrecognized meta file %s",
file.c_str());
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
continue;
}
assert(backups_.find(backup_id) == backups_.end());
// Insert all the (backup_id, BackupMeta) that will be loaded later
// The loading performed later will check whether there are corrupt backups
// and move the corrupt backups to corrupt_backups_
backups_.insert(std::make_pair(
backup_id, std::unique_ptr<BackupMeta>(new BackupMeta(
GetBackupMetaFile(backup_id, false /* tmp */),
GetBackupMetaFile(backup_id, true /* tmp */),
&backuped_file_infos_, backup_env_, backup_fs_))));
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
latest_backup_id_ = 0;
latest_valid_backup_id_ = 0;
10 years ago
if (options_.destroy_old_data) { // Destroy old data
assert(!read_only_);
ROCKS_LOG_INFO(
options_.info_log,
"Backup Engine started with destroy_old_data == true, deleting all "
"backups");
IOStatus io_s = PurgeOldBackups(0);
if (io_s.ok()) {
io_s = GarbageCollect();
}
if (!io_s.ok()) {
return io_s;
}
} else { // Load data from storage
// abs_path_to_size: maps absolute paths of files in backup directory to
// their corresponding sizes
std::unordered_map<std::string, uint64_t> abs_path_to_size;
// Insert files and their sizes in backup sub-directories (shared and
// shared_checksum) to abs_path_to_size
for (const auto& rel_dir :
{GetSharedFileRel(), GetSharedFileWithChecksumRel()}) {
const auto abs_dir = GetAbsolutePath(rel_dir);
IOStatus io_s =
ReadChildFileCurrentSizes(abs_dir, backup_fs_, &abs_path_to_size);
if (!io_s.ok()) {
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
// I/O error likely impacting all backups
return io_s;
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
}
}
// load the backups if any, until valid_backups_to_open of the latest
// non-corrupted backups have been successfully opened.
int valid_backups_to_open = options_.max_valid_backups_to_open;
for (auto backup_iter = backups_.rbegin();
backup_iter != backups_.rend();
++backup_iter) {
assert(latest_backup_id_ == 0 || latest_backup_id_ > backup_iter->first);
if (latest_backup_id_ == 0) {
latest_backup_id_ = backup_iter->first;
}
if (valid_backups_to_open == 0) {
break;
}
// Insert files and their sizes in backup sub-directories
// (private/backup_id) to abs_path_to_size
IOStatus io_s = ReadChildFileCurrentSizes(
GetAbsolutePath(GetPrivateFileRel(backup_iter->first)), backup_fs_,
&abs_path_to_size);
if (io_s.ok()) {
io_s = backup_iter->second->LoadFromFile(
options_.backup_dir, abs_path_to_size,
options_.backup_rate_limiter.get(), options_.info_log,
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
&reported_ignored_fields_);
}
if (io_s.IsCorruption() || io_s.IsNotSupported()) {
ROCKS_LOG_INFO(options_.info_log, "Backup %u corrupted -- %s",
backup_iter->first, io_s.ToString().c_str());
corrupt_backups_.insert(std::make_pair(
backup_iter->first,
std::make_pair(io_s, std::move(backup_iter->second))));
} else if (!io_s.ok()) {
// Distinguish corruption errors from errors in the backup Env.
// Errors in the backup Env (i.e., this code path) will cause Open() to
// fail, whereas corruption errors would not cause Open() failures.
return io_s;
} else {
ROCKS_LOG_INFO(options_.info_log, "Loading backup %" PRIu32 " OK:\n%s",
backup_iter->first,
backup_iter->second->GetInfoString().c_str());
assert(latest_valid_backup_id_ == 0 ||
latest_valid_backup_id_ > backup_iter->first);
if (latest_valid_backup_id_ == 0) {
latest_valid_backup_id_ = backup_iter->first;
}
--valid_backups_to_open;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
}
for (const auto& corrupt : corrupt_backups_) {
backups_.erase(backups_.find(corrupt.first));
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
// erase the backups before max_valid_backups_to_open
int num_unopened_backups;
if (options_.max_valid_backups_to_open == 0) {
num_unopened_backups = 0;
} else {
num_unopened_backups =
std::max(0, static_cast<int>(backups_.size()) -
options_.max_valid_backups_to_open);
}
for (int i = 0; i < num_unopened_backups; ++i) {
assert(backups_.begin()->second->Empty());
backups_.erase(backups_.begin());
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
ROCKS_LOG_INFO(options_.info_log, "Latest backup is %u", latest_backup_id_);
ROCKS_LOG_INFO(options_.info_log, "Latest valid backup is %u",
latest_valid_backup_id_);
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
// set up threads perform copies from files_to_copy_or_create_ in the
// background
threads_cpu_priority_ = CpuPriority::kNormal;
threads_.reserve(options_.max_background_operations);
for (int t = 0; t < options_.max_background_operations; t++) {
threads_.emplace_back([this]() {
#if defined(_GNU_SOURCE) && defined(__GLIBC_PREREQ)
#if __GLIBC_PREREQ(2, 12)
pthread_setname_np(pthread_self(), "backup_engine");
#endif
#endif
CpuPriority current_priority = CpuPriority::kNormal;
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
CopyOrCreateWorkItem work_item;
while (files_to_copy_or_create_.read(work_item)) {
CpuPriority priority = threads_cpu_priority_;
if (current_priority != priority) {
TEST_SYNC_POINT_CALLBACK(
"BackupEngineImpl::Initialize:SetCpuPriority", &priority);
port::SetCpuPriority(0, priority);
current_priority = priority;
}
// `bytes_read` and `bytes_written` stats are enabled based on
// compile-time support and cannot be dynamically toggled. So we do not
// need to worry about `PerfLevel` here, unlike many other
// `IOStatsContext` / `PerfContext` stats.
uint64_t prev_bytes_read = IOSTATS(bytes_read);
uint64_t prev_bytes_written = IOSTATS(bytes_written);
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
CopyOrCreateResult result;
result.io_status = CopyOrCreateFile(
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
work_item.src_path, work_item.dst_path, work_item.contents,
work_item.src_env, work_item.dst_env, work_item.src_env_options,
work_item.sync, work_item.rate_limiter, &result.size,
&result.checksum_hex, work_item.size_limit,
work_item.progress_callback);
RecordTick(work_item.stats, BACKUP_READ_BYTES,
IOSTATS(bytes_read) - prev_bytes_read);
RecordTick(work_item.stats, BACKUP_WRITE_BYTES,
IOSTATS(bytes_written) - prev_bytes_written);
result.db_id = work_item.db_id;
result.db_session_id = work_item.db_session_id;
if (result.io_status.ok() && !work_item.src_checksum_hex.empty()) {
// unknown checksum function name implies no db table file checksum in
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
// db manifest; work_item.src_checksum_hex not empty means
// backup engine has calculated its crc32c checksum for the table
// file; therefore, we are able to compare the checksums.
if (work_item.src_checksum_func_name ==
kUnknownFileChecksumFuncName ||
work_item.src_checksum_func_name == kDbFileChecksumFuncName) {
if (work_item.src_checksum_hex != result.checksum_hex) {
std::string checksum_info(
"Expected checksum is " + work_item.src_checksum_hex +
" while computed checksum is " + result.checksum_hex);
result.io_status = IOStatus::Corruption(
"Checksum mismatch after copying to " + work_item.dst_path +
": " + checksum_info);
}
} else {
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
// FIXME(peterd): dead code?
std::string checksum_function_info(
"Existing checksum function is " +
work_item.src_checksum_func_name +
" while provided checksum function is " +
kBackupFileChecksumFuncName);
ROCKS_LOG_INFO(
options_.info_log,
"Unable to verify checksum after copying to %s: %s\n",
work_item.dst_path.c_str(), checksum_function_info.c_str());
}
}
work_item.result.set_value(std::move(result));
}
});
}
ROCKS_LOG_INFO(options_.info_log, "Initialized BackupEngine");
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
return IOStatus::OK();
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
IOStatus BackupEngineImpl::CreateNewBackupWithMetadata(
const CreateBackupOptions& options, DB* db, const std::string& app_metadata,
BackupID* new_backup_id_ptr) {
assert(initialized_);
assert(!read_only_);
if (app_metadata.size() > kMaxAppMetaSize) {
return IOStatus::InvalidArgument("App metadata too large");
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
if (options.decrease_background_thread_cpu_priority) {
if (options.background_thread_cpu_priority < threads_cpu_priority_) {
threads_cpu_priority_.store(options.background_thread_cpu_priority);
}
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
BackupID new_backup_id = latest_backup_id_ + 1;
// `bytes_read` and `bytes_written` stats are enabled based on compile-time
// support and cannot be dynamically toggled. So we do not need to worry about
// `PerfLevel` here, unlike many other `IOStatsContext` / `PerfContext` stats.
uint64_t prev_bytes_read = IOSTATS(bytes_read);
uint64_t prev_bytes_written = IOSTATS(bytes_written);
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
assert(backups_.find(new_backup_id) == backups_.end());
auto private_dir = GetAbsolutePath(GetPrivateFileRel(new_backup_id));
IOStatus io_s = backup_fs_->FileExists(private_dir, io_options_, nullptr);
if (io_s.ok()) {
// maybe last backup failed and left partial state behind, clean it up.
// need to do this before updating backups_ such that a private dir
More fixes to auto-GarbageCollect in BackupEngine (#6023) Summary: Production: * Fixes GarbageCollect (and auto-GC triggered by PurgeOldBackups, DeleteBackup, or CreateNewBackup) to clean up backup directory independent of current settings (except max_valid_backups_to_open; see issue https://github.com/facebook/rocksdb/issues/4997) and prior settings used with same backup directory. * Fixes GarbageCollect (and auto-GC) not to attempt to remove "." and ".." entries from directories. * Clarifies contract with users in modifying BackupEngine operations. In short, leftovers from any incomplete operation are cleaned up by any subsequent call to that same kind of operation (PurgeOldBackups and DeleteBackup considered the same kind of operation). GarbageCollect is available to clean up after all kinds. (NB: right now PurgeOldBackups and DeleteBackup will clean up after incomplete CreateNewBackup, but we aren't promising to continue that behavior.) Pull Request resolved: https://github.com/facebook/rocksdb/pull/6023 Test Plan: * Refactors open parameters to use an option enum, for readability, etc. (Also fixes an unused parameter bug in the redundant OpenDBAndBackupEngineShareWithChecksum.) * Fixes an apparent bug in ShareTableFilesWithChecksumsTransition in which old backup data was destroyed in the transition to be tested. That test is now augmented to ensure GarbageCollect (or auto-GC) does not remove shared files when BackupEngine is opened with share_table_files=false. * Augments DeleteTmpFiles test to ensure that CreateNewBackup does auto-GC when an incompletely created backup is detected. Differential Revision: D18453559 Pulled By: pdillinger fbshipit-source-id: 5e54e7b08d711b161bc9c656181012b69a8feac4
5 years ago
// named after new_backup_id will be cleaned up.
// (If an incomplete new backup is followed by an incomplete delete
// of the latest full backup, then there could be more than one next
// id with a private dir, the last thing to be deleted in delete
// backup, but all will be cleaned up with a GarbageCollect.)
io_s = GarbageCollect();
} else if (io_s.IsNotFound()) {
// normal case, the new backup's private dir doesn't exist yet
io_s = IOStatus::OK();
}
auto ret = backups_.insert(std::make_pair(
new_backup_id, std::unique_ptr<BackupMeta>(new BackupMeta(
GetBackupMetaFile(new_backup_id, false /* tmp */),
GetBackupMetaFile(new_backup_id, true /* tmp */),
&backuped_file_infos_, backup_env_, backup_fs_))));
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
assert(ret.second == true);
auto& new_backup = ret.first->second;
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
new_backup->RecordTimestamp();
new_backup->SetAppMetadata(app_metadata);
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
auto start_backup = backup_env_->NowMicros();
ROCKS_LOG_INFO(options_.info_log,
"Started the backup process -- creating backup %u",
new_backup_id);
if (options_.share_table_files && !options_.share_files_with_checksum) {
ROCKS_LOG_WARN(options_.info_log,
"BackupEngineOptions::share_files_with_checksum=false is "
"DEPRECATED and could lead to data loss.");
}
if (io_s.ok()) {
io_s = backup_fs_->CreateDir(private_dir, io_options_, nullptr);
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
// A set into which we will insert the dst_paths that are calculated for live
// files and live WAL files.
// This is used to check whether a live files shares a dst_path with another
// live file.
std::unordered_set<std::string> live_dst_paths;
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
std::vector<BackupAfterCopyOrCreateWorkItem> backup_items_to_finish;
// Add a CopyOrCreateWorkItem to the channel for each live file
Status disabled = db->DisableFileDeletions();
DBOptions db_options = db->GetDBOptions();
Statistics* stats = db_options.statistics.get();
if (io_s.ok()) {
CheckpointImpl checkpoint(db);
uint64_t sequence_number = 0;
FileChecksumGenFactory* db_checksum_factory =
db_options.file_checksum_gen_factory.get();
const std::string kFileChecksumGenFactoryName =
"FileChecksumGenCrc32cFactory";
bool compare_checksum =
db_checksum_factory != nullptr &&
db_checksum_factory->Name() == kFileChecksumGenFactoryName
? true
: false;
EnvOptions src_raw_env_options(db_options);
RateLimiter* rate_limiter = options_.backup_rate_limiter.get();
io_s = status_to_io_status(checkpoint.CreateCustomCheckpoint(
db_options,
[&](const std::string& /*src_dirname*/, const std::string& /*fname*/,
FileType) {
// custom checkpoint will switch to calling copy_file_cb after it sees
// NotSupported returned from link_file_cb.
return IOStatus::NotSupported();
} /* link_file_cb */,
[&](const std::string& src_dirname, const std::string& fname,
uint64_t size_limit_bytes, FileType type,
const std::string& checksum_func_name,
const std::string& checksum_val) {
if (type == kWalFile && !options_.backup_log_files) {
return IOStatus::OK();
}
Log(options_.info_log, "add file for backup %s", fname.c_str());
uint64_t size_bytes = 0;
IOStatus io_st;
if (type == kTableFile || type == kBlobFile) {
io_st = db_fs_->GetFileSize(src_dirname + fname, io_options_,
&size_bytes, nullptr);
}
EnvOptions src_env_options;
switch (type) {
case kWalFile:
src_env_options =
db_env_->OptimizeForLogRead(src_raw_env_options);
break;
case kTableFile:
src_env_options = db_env_->OptimizeForCompactionTableRead(
src_raw_env_options, ImmutableDBOptions(db_options));
break;
case kDescriptorFile:
src_env_options =
db_env_->OptimizeForManifestRead(src_raw_env_options);
break;
case kBlobFile:
src_env_options = db_env_->OptimizeForBlobFileRead(
src_raw_env_options, ImmutableDBOptions(db_options));
break;
default:
// Other backed up files (like options file) are not read by live
// DB, so don't need to worry about avoiding mixing buffered and
// direct I/O. Just use plain defaults.
src_env_options = src_raw_env_options;
break;
}
if (io_st.ok()) {
io_st = AddBackupFileWorkItem(
live_dst_paths, backup_items_to_finish, new_backup_id,
options_.share_table_files &&
(type == kTableFile || type == kBlobFile),
src_dirname, fname, src_env_options, rate_limiter, type,
size_bytes, db_options.statistics.get(), size_limit_bytes,
options_.share_files_with_checksum &&
(type == kTableFile || type == kBlobFile),
options.progress_callback, "" /* contents */,
checksum_func_name, checksum_val);
}
return io_st;
} /* copy_file_cb */,
[&](const std::string& fname, const std::string& contents,
FileType type) {
Log(options_.info_log, "add file for backup %s", fname.c_str());
return AddBackupFileWorkItem(
live_dst_paths, backup_items_to_finish, new_backup_id,
false /* shared */, "" /* src_dir */, fname,
EnvOptions() /* src_env_options */, rate_limiter, type,
contents.size(), db_options.statistics.get(), 0 /* size_limit */,
false /* shared_checksum */, options.progress_callback, contents);
} /* create_file_cb */,
&sequence_number, options.flush_before_backup ? 0 : port::kMaxUint64,
compare_checksum));
if (io_s.ok()) {
new_backup->SetSequenceNumber(sequence_number);
}
}
ROCKS_LOG_INFO(options_.info_log, "add files for backup done, wait finish.");
IOStatus item_io_status;
for (auto& item : backup_items_to_finish) {
item.result.wait();
auto result = item.result.get();
item_io_status = result.io_status;
if (item_io_status.ok() && item.shared && item.needed_to_copy) {
item_io_status = item.backup_env->GetFileSystem()->RenameFile(
item.dst_path_tmp, item.dst_path, io_options_, nullptr);
}
if (item_io_status.ok()) {
item_io_status = new_backup.get()->AddFile(std::make_shared<FileInfo>(
item.dst_relative, result.size, result.checksum_hex, result.db_id,
result.db_session_id));
}
if (!item_io_status.ok()) {
io_s = item_io_status;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
}
// we copied all the files, enable file deletions
if (disabled.ok()) { // If we successfully disabled file deletions
db->EnableFileDeletions(false).PermitUncheckedError();
}
auto backup_time = backup_env_->NowMicros() - start_backup;
if (io_s.ok()) {
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
// persist the backup metadata on the disk
io_s = new_backup->StoreToFile(options_.sync, test_future_options_.get());
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
if (io_s.ok() && options_.sync) {
std::unique_ptr<FSDirectory> backup_private_directory;
backup_fs_
->NewDirectory(GetAbsolutePath(GetPrivateFileRel(new_backup_id, false)),
io_options_, &backup_private_directory, nullptr)
.PermitUncheckedError();
if (backup_private_directory != nullptr) {
io_s = backup_private_directory->Fsync(io_options_, nullptr);
}
if (io_s.ok() && private_directory_ != nullptr) {
io_s = private_directory_->Fsync(io_options_, nullptr);
}
if (io_s.ok() && meta_directory_ != nullptr) {
io_s = meta_directory_->Fsync(io_options_, nullptr);
}
if (io_s.ok() && shared_directory_ != nullptr) {
io_s = shared_directory_->Fsync(io_options_, nullptr);
}
if (io_s.ok() && backup_directory_ != nullptr) {
io_s = backup_directory_->Fsync(io_options_, nullptr);
}
}
if (io_s.ok()) {
backup_statistics_.IncrementNumberSuccessBackup();
// here we know that we succeeded and installed the new backup
latest_backup_id_ = new_backup_id;
latest_valid_backup_id_ = new_backup_id;
if (new_backup_id_ptr) {
*new_backup_id_ptr = new_backup_id;
}
ROCKS_LOG_INFO(options_.info_log, "Backup DONE. All is good");
// backup_speed is in byte/second
double backup_speed = new_backup->GetSize() / (1.048576 * backup_time);
ROCKS_LOG_INFO(options_.info_log, "Backup number of files: %u",
new_backup->GetNumberFiles());
char human_size[16];
AppendHumanBytes(new_backup->GetSize(), human_size, sizeof(human_size));
ROCKS_LOG_INFO(options_.info_log, "Backup size: %s", human_size);
ROCKS_LOG_INFO(options_.info_log, "Backup time: %" PRIu64 " microseconds",
backup_time);
ROCKS_LOG_INFO(options_.info_log, "Backup speed: %.3f MB/s", backup_speed);
ROCKS_LOG_INFO(options_.info_log, "Backup Statistics %s",
backup_statistics_.ToString().c_str());
} else {
backup_statistics_.IncrementNumberFailBackup();
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
// clean all the files we might have created
ROCKS_LOG_INFO(options_.info_log, "Backup failed -- %s",
io_s.ToString().c_str());
ROCKS_LOG_INFO(options_.info_log, "Backup Statistics %s\n",
backup_statistics_.ToString().c_str());
// delete files that we might have already written
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
might_need_garbage_collect_ = true;
DeleteBackup(new_backup_id).PermitUncheckedError();
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
RecordTick(stats, BACKUP_READ_BYTES, IOSTATS(bytes_read) - prev_bytes_read);
RecordTick(stats, BACKUP_WRITE_BYTES,
IOSTATS(bytes_written) - prev_bytes_written);
return io_s;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
IOStatus BackupEngineImpl::PurgeOldBackups(uint32_t num_backups_to_keep) {
assert(initialized_);
assert(!read_only_);
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
// Best effort deletion even with errors
IOStatus overall_status = IOStatus::OK();
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
ROCKS_LOG_INFO(options_.info_log, "Purging old backups, keeping %u",
num_backups_to_keep);
std::vector<BackupID> to_delete;
auto itr = backups_.begin();
while ((backups_.size() - to_delete.size()) > num_backups_to_keep) {
to_delete.push_back(itr->first);
itr++;
}
for (auto backup_id : to_delete) {
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
// Do not GC until end
IOStatus io_s = DeleteBackupNoGC(backup_id);
if (!io_s.ok()) {
overall_status = io_s;
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
// Clean up after any incomplete backup deletion, potentially from
// earlier session.
if (might_need_garbage_collect_) {
IOStatus io_s = GarbageCollect();
if (!io_s.ok() && overall_status.ok()) {
overall_status = io_s;
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
}
}
return overall_status;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
IOStatus BackupEngineImpl::DeleteBackup(BackupID backup_id) {
IOStatus s1 = DeleteBackupNoGC(backup_id);
IOStatus s2 = IOStatus::OK();
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
// Clean up after any incomplete backup deletion, potentially from
// earlier session.
if (might_need_garbage_collect_) {
s2 = GarbageCollect();
}
if (!s1.ok()) {
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
// Any failure in the primary objective trumps any failure in the
// secondary objective.
s2.PermitUncheckedError();
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
return s1;
} else {
return s2;
}
}
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
// Does not auto-GarbageCollect nor lock
IOStatus BackupEngineImpl::DeleteBackupNoGC(BackupID backup_id) {
assert(initialized_);
assert(!read_only_);
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
ROCKS_LOG_INFO(options_.info_log, "Deleting backup %u", backup_id);
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
auto backup = backups_.find(backup_id);
if (backup != backups_.end()) {
IOStatus io_s = backup->second->Delete();
if (!io_s.ok()) {
return io_s;
}
backups_.erase(backup);
} else {
auto corrupt = corrupt_backups_.find(backup_id);
if (corrupt == corrupt_backups_.end()) {
return IOStatus::NotFound("Backup not found");
}
IOStatus io_s = corrupt->second.second->Delete();
if (!io_s.ok()) {
return io_s;
}
corrupt->second.first.PermitUncheckedError();
corrupt_backups_.erase(corrupt);
}
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
// After removing meta file, best effort deletion even with errors.
// (Don't delete other files if we can't delete the meta file right
// now.)
std::vector<std::string> to_delete;
for (auto& itr : backuped_file_infos_) {
if (itr.second->refs == 0) {
IOStatus io_s = backup_fs_->DeleteFile(GetAbsolutePath(itr.first),
io_options_, nullptr);
ROCKS_LOG_INFO(options_.info_log, "Deleting %s -- %s", itr.first.c_str(),
io_s.ToString().c_str());
to_delete.push_back(itr.first);
if (!io_s.ok()) {
// Trying again later might work
might_need_garbage_collect_ = true;
}
}
}
for (auto& td : to_delete) {
backuped_file_infos_.erase(td);
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
// take care of private dirs -- GarbageCollect() will take care of them
// if they are not empty
std::string private_dir = GetPrivateFileRel(backup_id);
IOStatus io_s =
backup_fs_->DeleteDir(GetAbsolutePath(private_dir), io_options_, nullptr);
ROCKS_LOG_INFO(options_.info_log, "Deleting private dir %s -- %s",
private_dir.c_str(), io_s.ToString().c_str());
if (!io_s.ok()) {
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
// Full gc or trying again later might work
might_need_garbage_collect_ = true;
}
return IOStatus::OK();
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
void BackupEngineImpl::SetBackupInfoFromBackupMeta(
BackupID id, const BackupMeta& meta, BackupInfo* backup_info,
bool include_file_details) const {
*backup_info = BackupInfo(id, meta.GetTimestamp(), meta.GetSize(),
meta.GetNumberFiles(), meta.GetAppMetadata());
if (include_file_details) {
auto& file_details = backup_info->file_details;
file_details.reserve(meta.GetFiles().size());
for (auto& file_ptr : meta.GetFiles()) {
BackupFileInfo& finfo = *file_details.emplace(file_details.end());
finfo.relative_filename = file_ptr->filename;
finfo.size = file_ptr->size;
}
backup_info->name_for_open = GetAbsolutePath(GetPrivateFileRel(id));
backup_info->name_for_open.pop_back(); // remove trailing '/'
backup_info->env_for_open = meta.GetEnvForOpen();
}
}
Status BackupEngineImpl::GetBackupInfo(BackupID backup_id,
BackupInfo* backup_info,
bool include_file_details) const {
assert(initialized_);
if (backup_id == kLatestBackupIDMarker) {
// Note: Read latest_valid_backup_id_ inside of lock
backup_id = latest_valid_backup_id_;
}
auto corrupt_itr = corrupt_backups_.find(backup_id);
if (corrupt_itr != corrupt_backups_.end()) {
return Status::Corruption(corrupt_itr->second.first.ToString());
}
auto backup_itr = backups_.find(backup_id);
if (backup_itr == backups_.end()) {
return Status::NotFound("Backup not found");
}
auto& backup = backup_itr->second;
if (backup->Empty()) {
return Status::NotFound("Backup not found");
}
SetBackupInfoFromBackupMeta(backup_id, *backup, backup_info,
include_file_details);
return Status::OK();
}
void BackupEngineImpl::GetBackupInfo(std::vector<BackupInfo>* backup_info,
bool include_file_details) const {
assert(initialized_);
backup_info->resize(backups_.size());
size_t i = 0;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
for (auto& backup : backups_) {
const BackupMeta& meta = *backup.second;
if (!meta.Empty()) {
SetBackupInfoFromBackupMeta(backup.first, meta, &backup_info->at(i++),
include_file_details);
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
}
}
void BackupEngineImpl::GetCorruptedBackups(
std::vector<BackupID>* corrupt_backup_ids) const {
assert(initialized_);
corrupt_backup_ids->reserve(corrupt_backups_.size());
for (auto& backup : corrupt_backups_) {
corrupt_backup_ids->push_back(backup.first);
}
}
IOStatus BackupEngineImpl::RestoreDBFromBackup(
const RestoreOptions& options, BackupID backup_id,
const std::string& db_dir, const std::string& wal_dir) const {
assert(initialized_);
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
if (backup_id == kLatestBackupIDMarker) {
// Note: Read latest_valid_backup_id_ inside of lock
backup_id = latest_valid_backup_id_;
}
auto corrupt_itr = corrupt_backups_.find(backup_id);
if (corrupt_itr != corrupt_backups_.end()) {
return corrupt_itr->second.first;
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
auto backup_itr = backups_.find(backup_id);
if (backup_itr == backups_.end()) {
return IOStatus::NotFound("Backup not found");
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
auto& backup = backup_itr->second;
if (backup->Empty()) {
return IOStatus::NotFound("Backup not found");
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
ROCKS_LOG_INFO(options_.info_log, "Restoring backup id %u\n", backup_id);
ROCKS_LOG_INFO(options_.info_log, "keep_log_files: %d\n",
static_cast<int>(options.keep_log_files));
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
// just in case. Ignore errors
db_fs_->CreateDirIfMissing(db_dir, io_options_, nullptr)
.PermitUncheckedError();
db_fs_->CreateDirIfMissing(wal_dir, io_options_, nullptr)
.PermitUncheckedError();
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
if (options.keep_log_files) {
// delete files in db_dir, but keep all the log files
DeleteChildren(db_dir, 1 << kWalFile);
// move all the files from archive dir to wal_dir
std::string archive_dir = ArchivalDirectory(wal_dir);
std::vector<std::string> archive_files;
db_fs_->GetChildren(archive_dir, io_options_, &archive_files, nullptr)
.PermitUncheckedError(); // ignore errors
for (const auto& f : archive_files) {
uint64_t number;
FileType type;
bool ok = ParseFileName(f, &number, &type);
if (ok && type == kWalFile) {
ROCKS_LOG_INFO(options_.info_log,
"Moving log file from archive/ to wal_dir: %s",
f.c_str());
IOStatus io_s = db_fs_->RenameFile(
archive_dir + "/" + f, wal_dir + "/" + f, io_options_, nullptr);
if (!io_s.ok()) {
// if we can't move log file from archive_dir to wal_dir,
// we should fail, since it might mean data loss
return io_s;
}
}
}
} else {
DeleteChildren(wal_dir);
DeleteChildren(ArchivalDirectory(wal_dir));
DeleteChildren(db_dir);
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
IOStatus io_s;
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
std::vector<RestoreAfterCopyOrCreateWorkItem> restore_items_to_finish;
std::string temporary_current_file;
std::string final_current_file;
std::unique_ptr<FSDirectory> db_dir_for_fsync;
std::unique_ptr<FSDirectory> wal_dir_for_fsync;
for (const auto& file_info : backup->GetFiles()) {
const std::string& file = file_info->filename;
Make backups openable as read-only DBs (#8142) Summary: A current limitation of backups is that you don't know the exact database state of when the backup was taken. With this new feature, you can at least inspect the backup's DB state without restoring it by opening it as a read-only DB. Rather than add something like OpenAsReadOnlyDB to the BackupEngine API, which would inhibit opening stackable DB implementations read-only (if/when their APIs support it), we instead provide a DB name and Env that can be used to open as a read-only DB. Possible follow-up work: * Add a version of GetBackupInfo for a single backup. * Let CreateNewBackup return the BackupID of the newly-created backup. Implementation details: Refactored ChrootFileSystem to split off new base class RemapFileSystem, which allows more general remapping of files. We use this base class to implement BackupEngineImpl::RemapSharedFileSystem. To minimize API impact, I decided to just add these fields `name_for_open` and `env_for_open` to those set by GetBackupInfo when include_file_details=true. Creating the RemapSharedFileSystem adds a bit to the memory consumption, perhaps unnecessarily in some cases, but this has been mitigated by (a) only initialize the RemapSharedFileSystem lazily when GetBackupInfo with include_file_details=true is called, and (b) using the existing `shared_ptr<FileInfo>` objects to hold most of the mapping data. To enhance API safety, RemapSharedFileSystem is wrapped by new ReadOnlyFileSystem which rejects any attempts to write. This uncovered a couple of places in which DB::OpenForReadOnly would write to the filesystem, so I fixed these. Added a release note because this affects logging. Additional minor refactoring in backupable_db.cc to support the new functionality. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8142 Test Plan: new test (run with ASAN and UBSAN), added to stress test and ran it for a while with amplified backup_one_in Reviewed By: ajkr Differential Revision: D27535408 Pulled By: pdillinger fbshipit-source-id: 04666d310aa0261ef6b2385c43ca793ce1dfd148
4 years ago
// 1. get DB filename
std::string dst = file_info->GetDbFileName();
// 2. find the filetype
uint64_t number;
FileType type;
bool ok = ParseFileName(dst, &number, &type);
if (!ok) {
return IOStatus::Corruption("Backup corrupted: Fail to parse filename " +
dst);
}
// 3. Construct the final path
// kWalFile lives in wal_dir and all the rest live in db_dir
if (type == kWalFile) {
dst = wal_dir + "/" + dst;
if (options_.sync && !wal_dir_for_fsync) {
io_s = db_fs_->NewDirectory(wal_dir, io_options_, &wal_dir_for_fsync,
nullptr);
if (!io_s.ok()) {
return io_s;
}
}
} else {
dst = db_dir + "/" + dst;
if (options_.sync && !db_dir_for_fsync) {
io_s = db_fs_->NewDirectory(db_dir, io_options_, &db_dir_for_fsync,
nullptr);
if (!io_s.ok()) {
return io_s;
}
}
}
// For atomicity, initially restore CURRENT file to a temporary name.
// This is useful even without options_.sync e.g. in case the restore
// process is interrupted.
if (type == kCurrentFile) {
final_current_file = dst;
dst = temporary_current_file = dst + ".tmp";
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
ROCKS_LOG_INFO(options_.info_log, "Restoring %s to %s\n", file.c_str(),
dst.c_str());
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
CopyOrCreateWorkItem copy_or_create_work_item(
GetAbsolutePath(file), dst, "" /* contents */, backup_env_, db_env_,
EnvOptions() /* src_env_options */, options_.sync,
options_.restore_rate_limiter.get(), 0 /* size_limit */,
nullptr /* stats */);
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
RestoreAfterCopyOrCreateWorkItem after_copy_or_create_work_item(
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
copy_or_create_work_item.result.get_future(), file, dst,
file_info->checksum_hex);
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
files_to_copy_or_create_.write(std::move(copy_or_create_work_item));
restore_items_to_finish.push_back(
std::move(after_copy_or_create_work_item));
}
IOStatus item_io_status;
for (auto& item : restore_items_to_finish) {
item.result.wait();
auto result = item.result.get();
item_io_status = result.io_status;
// Note: It is possible that both of the following bad-status cases occur
// during copying. But, we only return one status.
if (!item_io_status.ok()) {
io_s = item_io_status;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
break;
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
} else if (!item.checksum_hex.empty() &&
item.checksum_hex != result.checksum_hex) {
io_s = IOStatus::Corruption(
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
"While restoring " + item.from_file + " -> " + item.to_file +
": expected checksum is " + item.checksum_hex +
" while computed checksum is " + result.checksum_hex);
break;
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
// When enabled, the first Fsync is to ensure all files are fully persisted
// before renaming CURRENT.tmp
if (io_s.ok() && db_dir_for_fsync) {
ROCKS_LOG_INFO(options_.info_log, "Restore: fsync\n");
io_s = db_dir_for_fsync->Fsync(io_options_, nullptr);
}
if (io_s.ok() && wal_dir_for_fsync) {
io_s = wal_dir_for_fsync->Fsync(io_options_, nullptr);
}
if (io_s.ok() && !temporary_current_file.empty()) {
ROCKS_LOG_INFO(options_.info_log, "Restore: atomic rename CURRENT.tmp\n");
assert(!final_current_file.empty());
io_s = db_fs_->RenameFile(temporary_current_file, final_current_file,
io_options_, nullptr);
}
if (io_s.ok() && db_dir_for_fsync && !temporary_current_file.empty()) {
// Second Fsync is to ensure the final atomic rename of DB restore is
// fully persisted even if power goes out right after restore operation
// returns success
assert(db_dir_for_fsync);
io_s = db_dir_for_fsync->Fsync(io_options_, nullptr);
}
ROCKS_LOG_INFO(options_.info_log, "Restoring done -- %s\n",
io_s.ToString().c_str());
return io_s;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
IOStatus BackupEngineImpl::VerifyBackup(BackupID backup_id,
bool verify_with_checksum) const {
assert(initialized_);
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
// Check if backup_id is corrupted, or valid and registered
auto corrupt_itr = corrupt_backups_.find(backup_id);
if (corrupt_itr != corrupt_backups_.end()) {
return corrupt_itr->second.first;
}
auto backup_itr = backups_.find(backup_id);
if (backup_itr == backups_.end()) {
return IOStatus::NotFound();
}
auto& backup = backup_itr->second;
if (backup->Empty()) {
return IOStatus::NotFound();
}
ROCKS_LOG_INFO(options_.info_log, "Verifying backup id %u\n", backup_id);
// Find all existing backup files belong to backup_id
std::unordered_map<std::string, uint64_t> curr_abs_path_to_size;
for (const auto& rel_dir : {GetPrivateFileRel(backup_id), GetSharedFileRel(),
GetSharedFileWithChecksumRel()}) {
const auto abs_dir = GetAbsolutePath(rel_dir);
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
// Shared directories allowed to be missing in some cases. Expected but
// missing files will be reported a few lines down.
ReadChildFileCurrentSizes(abs_dir, backup_fs_, &curr_abs_path_to_size)
.PermitUncheckedError();
}
// For all files registered in backup
for (const auto& file_info : backup->GetFiles()) {
const auto abs_path = GetAbsolutePath(file_info->filename);
// check existence of the file
if (curr_abs_path_to_size.find(abs_path) == curr_abs_path_to_size.end()) {
return IOStatus::NotFound("File missing: " + abs_path);
}
// verify file size
if (file_info->size != curr_abs_path_to_size[abs_path]) {
std::string size_info("Expected file size is " +
ToString(file_info->size) +
" while found file size is " +
ToString(curr_abs_path_to_size[abs_path]));
return IOStatus::Corruption("File corrupted: File size mismatch for " +
abs_path + ": " + size_info);
}
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
if (verify_with_checksum && !file_info->checksum_hex.empty()) {
// verify file checksum
std::string checksum_hex;
ROCKS_LOG_INFO(options_.info_log, "Verifying %s checksum...\n",
abs_path.c_str());
IOStatus io_s =
ReadFileAndComputeChecksum(abs_path, backup_fs_, EnvOptions(),
0 /* size_limit */, &checksum_hex);
if (!io_s.ok()) {
return io_s;
} else if (file_info->checksum_hex != checksum_hex) {
std::string checksum_info(
"Expected checksum is " + file_info->checksum_hex +
" while computed checksum is " + checksum_hex);
return IOStatus::Corruption("File corrupted: Checksum mismatch for " +
abs_path + ": " + checksum_info);
}
}
}
return IOStatus::OK();
}
IOStatus BackupEngineImpl::CopyOrCreateFile(
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
const std::string& src, const std::string& dst, const std::string& contents,
Env* src_env, Env* dst_env, const EnvOptions& src_env_options, bool sync,
RateLimiter* rate_limiter, uint64_t* size, std::string* checksum_hex,
uint64_t size_limit, std::function<void()> progress_callback) {
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
assert(src.empty() != contents.empty());
IOStatus io_s;
std::unique_ptr<FSWritableFile> dst_file;
std::unique_ptr<FSSequentialFile> src_file;
FileOptions dst_file_options;
dst_file_options.use_mmap_writes = false;
// TODO:(gzh) maybe use direct reads/writes here if possible
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
if (size != nullptr) {
*size = 0;
}
uint32_t checksum_value = 0;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
// Check if size limit is set. if not, set it to very big number
if (size_limit == 0) {
size_limit = std::numeric_limits<uint64_t>::max();
}
io_s = dst_env->GetFileSystem()->NewWritableFile(dst, dst_file_options,
&dst_file, nullptr);
if (io_s.ok() && !src.empty()) {
io_s = src_env->GetFileSystem()->NewSequentialFile(
src, FileOptions(src_env_options), &src_file, nullptr);
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
if (!io_s.ok()) {
return io_s;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
size_t buf_size =
rate_limiter ? static_cast<size_t>(rate_limiter->GetSingleBurstBytes())
: kDefaultCopyFileBufferSize;
std::unique_ptr<WritableFileWriter> dest_writer(
new WritableFileWriter(std::move(dst_file), dst, dst_file_options));
std::unique_ptr<SequentialFileReader> src_reader;
std::unique_ptr<char[]> buf;
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
if (!src.empty()) {
src_reader.reset(new SequentialFileReader(std::move(src_file), src));
buf.reset(new char[buf_size]);
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
Slice data;
uint64_t processed_buffer_size = 0;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
do {
if (stop_backup_.load(std::memory_order_acquire)) {
return status_to_io_status(Status::Incomplete("Backup stopped"));
}
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
if (!src.empty()) {
size_t buffer_to_read =
(buf_size < size_limit) ? buf_size : static_cast<size_t>(size_limit);
io_s = src_reader->Read(buffer_to_read, &data, buf.get());
if (rate_limiter != nullptr) {
rate_limiter->Request(data.size(), Env::IO_LOW, nullptr /* stats */,
RateLimiter::OpType::kRead);
}
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
processed_buffer_size += buffer_to_read;
} else {
data = contents;
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
size_limit -= data.size();
TEST_SYNC_POINT_CALLBACK(
"BackupEngineImpl::CopyOrCreateFile:CorruptionDuringBackup",
(src.length() > 4 && src.rfind(".sst") == src.length() - 4) ? &data
: nullptr);
if (!io_s.ok()) {
return io_s;
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
if (size != nullptr) {
*size += data.size();
}
if (checksum_hex != nullptr) {
checksum_value = crc32c::Extend(checksum_value, data.data(), data.size());
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
io_s = dest_writer->Append(data);
if (rate_limiter != nullptr) {
rate_limiter->Request(data.size(), Env::IO_LOW, nullptr /* stats */,
RateLimiter::OpType::kWrite);
}
if (processed_buffer_size > options_.callback_trigger_interval_size) {
processed_buffer_size -= options_.callback_trigger_interval_size;
std::lock_guard<std::mutex> lock(byte_report_mutex_);
progress_callback();
}
} while (io_s.ok() && contents.empty() && data.size() > 0 && size_limit > 0);
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
// Convert uint32_t checksum to hex checksum
if (checksum_hex != nullptr) {
checksum_hex->assign(ChecksumInt32ToHex(checksum_value));
}
if (io_s.ok() && sync) {
io_s = dest_writer->Sync(false);
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
if (io_s.ok()) {
io_s = dest_writer->Close();
}
return io_s;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
// fname will always start with "/"
IOStatus BackupEngineImpl::AddBackupFileWorkItem(
std::unordered_set<std::string>& live_dst_paths,
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
std::vector<BackupAfterCopyOrCreateWorkItem>& backup_items_to_finish,
BackupID backup_id, bool shared, const std::string& src_dir,
const std::string& fname, const EnvOptions& src_env_options,
RateLimiter* rate_limiter, FileType file_type, uint64_t size_bytes,
Statistics* stats, uint64_t size_limit, bool shared_checksum,
std::function<void()> progress_callback, const std::string& contents,
const std::string& src_checksum_func_name,
const std::string& src_checksum_str) {
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
assert(!fname.empty() && fname[0] == '/');
assert(contents.empty() != src_dir.empty());
std::string dst_relative = fname.substr(1);
std::string dst_relative_tmp;
std::string db_id;
std::string db_session_id;
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
// crc32c checksum in hex. empty == unavailable / unknown
std::string checksum_hex;
// Whenever a default checksum function name is passed in, we will compares
// the corresponding checksum values after copying. Note that only table and
// blob files may have a known checksum function name passed in.
//
// If no default checksum function name is passed in and db session id is not
// available, we will calculate the checksum *before* copying in two cases
// (we always calcuate checksums when copying or creating for any file types):
// a) share_files_with_checksum is true and file type is table;
// b) share_table_files is true and the file exists already.
//
// Step 0: Check if default checksum function name is passed in
if (kDbFileChecksumFuncName == src_checksum_func_name) {
if (src_checksum_str == kUnknownFileChecksum) {
return status_to_io_status(
Status::Aborted("Unknown checksum value for " + fname));
}
checksum_hex = ChecksumStrToHex(src_checksum_str);
}
// Step 1: Prepare the relative path to destination
if (shared && shared_checksum) {
if (GetNamingNoFlags() != BackupEngineOptions::kLegacyCrc32cAndFileSize &&
file_type != kBlobFile) {
// Prepare db_session_id to add to the file name
// Ignore the returned status
// In the failed cases, db_id and db_session_id will be empty
GetFileDbIdentities(db_env_, src_env_options, src_dir + fname,
rate_limiter, &db_id, &db_session_id)
.PermitUncheckedError();
}
// Calculate checksum if checksum and db session id are not available.
// If db session id is available, we will not calculate the checksum
// since the session id should suffice to avoid file name collision in
// the shared_checksum directory.
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
if (checksum_hex.empty() && db_session_id.empty()) {
IOStatus io_s = ReadFileAndComputeChecksum(
src_dir + fname, db_fs_, src_env_options, size_limit, &checksum_hex);
if (!io_s.ok()) {
return io_s;
}
}
if (size_bytes == port::kMaxUint64) {
return IOStatus::NotFound("File missing: " + src_dir + fname);
}
// dst_relative depends on the following conditions:
Restore file size in backup table file names (and other cleanup) (#7400) Summary: Prior to 6.12, backup files using share_files_with_checksum had the file size encoded in the file name, after the last '\_' and before the last '.'. We considered this an implementation detail subject to change, and indeed removed this information from the file name (with an option to use old behavior) because it was considered ineffective/inefficient for file name uniqueness. However, some downstream RocksDB users were relying on this information since the file size is not explicitly in the backup manifest file. This primary purpose of this change is "retrofitting" the 6.12 release (not yet a public release) to simultaneously support the benefits of the new naming scheme (I/O performance and data correctness at scale) and preserve the file size information, both as default behaviors. With this change, we are essentially making the file size information encoded in the file name an official, though obscure, extension of the backup meta file format. We preserve an option (kLegacyCrc32cAndFileSize) to use the original "legacy" naming scheme, with its caveats, and make it easy to omit the file size information (no kFlagIncludeFileSize), for more compact file names. But note that changing the naming scheme used on an existing db and backup directory can lead to transient space amplification, as some files will be stored under two names in the shared_checksum directory. Because some backups were saved using the original 6.12 naming scheme, we offer two ways of dealing with those files: SST files generated by older 6.12 versions can either use the default naming scheme in effect when the SST files were generated (kFlagMatchInterimNaming, default, no transient space amplification) or can use a new naming scheme (no kFlagMatchInterimNaming, potential space amplification because some already stored files getting a new name). We don't have a natural way to detect which files were generated by previous 6.12 versions, but this change hacks one in by changing DB session ids to now use a more concise encoding, reducing file name length, saving ~dozen bytes from SST files, and making them visually distinct from DB ids so that they are less likely to be mixed up. Two final auxiliary notes: Recognizing that the backup file names have become a de facto part of the backup meta schema, this change makes them easier to parse and extend by putting a distinct marker, 's', before DB session ids embedded in the name. When we extend this to allow custom checksums in the name, they can get their own marker to ensure safe parsing. For backward compatibility, file size does not get a marker but is assumed for `_[0-9]+[.]` Another change from initial 6.12 default behavior is never including file custom checksum in the file name. Looking ahead to 6.13, we do not want the default behavior to cause backup space amplification for someone turning on file custom checksum checking in BackupEngine; we want that to be an easy decision. When implemented, including file custom checksums in backup file names will be a non-default option. Actual file name patterns and priorities, as regexes: kLegacyCrc32cAndFileSize OR pre-6.12 SST file -> [0-9]+_[0-9]+_[0-9]+[.]sst kFlagMatchInterimNaming set (default) AND early 6.12 SST file -> [0-9]+_[0-9a-fA-F-]+[.]sst kUseDbSessionId AND NOT kFlagIncludeFileSize -> [0-9]+_s[0-9A-Z]{20}[.]sst kUseDbSessionId AND kFlagIncludeFileSize (default) -> [0-9]+_s[0-9A-Z]{20}_[0-9]+[.]sst We might add opt-in options for more '\_' separated data in the name, but embedded file size, if present, will always be after last '\_' and before '.sst'. This change was originally applied to version 6.12. (See https://github.com/facebook/rocksdb/issues/7390) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7400 Test Plan: unit tests included. Sync point callbacks are used to mimic previous version SST files. Reviewed By: ajkr Differential Revision: D23759587 Pulled By: pdillinger fbshipit-source-id: f62d8af4e0978de0a34f26288cfbe66049b70025
4 years ago
// 1) the naming scheme is kUseDbSessionId,
// 2) db_session_id is not empty,
// 3) checksum is available in the DB manifest.
// If 1,2,3) are satisfied, then dst_relative will be of the form:
// shared_checksum/<file_number>_<checksum>_<db_session_id>.sst
// If 1,2) are satisfied, then dst_relative will be of the form:
// shared_checksum/<file_number>_<db_session_id>.sst
// Otherwise, dst_relative is of the form
// shared_checksum/<file_number>_<checksum>_<size>.sst
//
// For blob files, db_session_id is not supported with the blob file format.
// It uses original/legacy naming scheme.
// dst_relative will be of the form:
// shared_checksum/<file_number>_<checksum>_<size>.blob
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
dst_relative = GetSharedFileWithChecksum(dst_relative, checksum_hex,
size_bytes, db_session_id);
dst_relative_tmp = GetSharedFileWithChecksumRel(dst_relative, true);
dst_relative = GetSharedFileWithChecksumRel(dst_relative, false);
} else if (shared) {
dst_relative_tmp = GetSharedFileRel(dst_relative, true);
dst_relative = GetSharedFileRel(dst_relative, false);
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
} else {
dst_relative = GetPrivateFileRel(backup_id, false, dst_relative);
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
// We copy into `temp_dest_path` and, once finished, rename it to
// `final_dest_path`. This allows files to atomically appear at
// `final_dest_path`. We can copy directly to the final path when atomicity
// is unnecessary, like for files in private backup directories.
const std::string* copy_dest_path;
std::string temp_dest_path;
std::string final_dest_path = GetAbsolutePath(dst_relative);
if (!dst_relative_tmp.empty()) {
temp_dest_path = GetAbsolutePath(dst_relative_tmp);
copy_dest_path = &temp_dest_path;
} else {
copy_dest_path = &final_dest_path;
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
// Step 2: Determine whether to copy or not
// if it's shared, we also need to check if it exists -- if it does, no need
// to copy it again.
bool need_to_copy = true;
// true if final_dest_path is the same path as another live file
const bool same_path =
live_dst_paths.find(final_dest_path) != live_dst_paths.end();
bool file_exists = false;
if (shared && !same_path) {
// Should be in shared directory but not a live path, check existence in
// shared directory
IOStatus exist =
backup_fs_->FileExists(final_dest_path, io_options_, nullptr);
if (exist.ok()) {
file_exists = true;
} else if (exist.IsNotFound()) {
file_exists = false;
} else {
return exist;
}
}
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
if (!contents.empty()) {
need_to_copy = false;
} else if (shared && (same_path || file_exists)) {
need_to_copy = false;
Less I/O for incremental backups, slightly better corruption detection (#7413) Summary: Two relatively simple functional changes to incremental backup behavior, integrated with a minor refactoring to reduce code redundancy and improve error/log message. There are nuances to the impact of these changes, but I believe they are fundamentally good and generally safe. Those functional changes: * Incremental backups no longer read DB table files that are already saved to a shared part of the backup directory, unless `share_files_with_checksum` is used with `kLegacyCrc32cAndFileSize` naming (discouraged) where crc32c full file checksums are needed to determine file naming. * Justification: incremental backups should not need to read the whole DB, especially without rate limiting. (Although other BackupEngine reads are not rate limited either, other non-trivial reads are generally limited by a corresponding write, as in copying files.) Also, the fact that this is not already fixed was arguably a bug/oversight in the implementation of https://github.com/facebook/rocksdb/issues/7110. * When considering whether a table file is already backed up in a shared part of backup directory, BackupEngine would already query the sizes of source (DB) and pre-existing destination (backup) files. BackupEngine now uses these file sizes to detect corruption, as at least one of (a) old backup, (b) backup in progress, or (c) current DB is corrupt if there's a size mismatch. * Justification: a random related fix that also helps to cover a small hole in corruption checking uncovered by the other functional change: * For `share_table_files` without "checksum" (not recommended), the other change regresses in detecting fundamentally unsafe use of this option combination: when you might generate different versions of same SST file number. As demonstrated by `BackupableDBTest.FailOverwritingBackups,` this regression is greatly mitigated by the new file size checking. Nevertheless, almost no reason to use `share_files_with_checksum=false` should remain, and comments are updated appropriately. Also, this change renames internal function `CalculateChecksum` to `ReadFileAndComputeChecksum` to make the performance impact of this function clear in code reviews. It is not clear what 'same_path' is for in backupable_db.cc, and I suspect it cannot be true for a DB with unique file names (like DBImpl). Nevertheless, I've tried to keep its functionality intact when `true` to minimize risk for now, despite having no unit tests for which it is true. Select impact details (much more in unit tests): For `share_files_with_checksum`, I am confident there is no regression (vs. pre-6.12) in detecting DB or backup corruption at backup creation time, mostly because the old design did not leverage this extra checksum computation for detecting inconsistencies at backup creation time. (With computed checksums in names, a recently corrupted file just looked like a different file vs. what was already backed up.) Even in the hypothetical case of DB session id collision (~100 bits entropy collision), file size in name and/or our file size check add an extra layer of protection against false success in creating an accurate new backup. (Unit test included.) `DB::VerifyChecksum` and `BackupEngine::VerifyBackup` with checksum checking are still able to catch corruptions that `CreateNewBackup` does not. Note that when custom file checksum support is added to BackupEngine, that will essentially give the same power as `DB::VerifyChecksum` into `CreateNewBackup`. We could add options for `CreateNewBackup` to cover some of what would be caught by `VerifyBackup` with checksum checking. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7413 Test Plan: Two new unit tests included, both of which fail without these changes. Although we don't test the I/O improvement directly, we test it indirectly in DB corruption detection power that was inadvertently unlocked with new backup file naming PLUS computing current content checksums (now removed). (I don't think that case of DB corruption detection justifies reading the whole DB on incremental backup.) Reviewed By: zhichao-cao Differential Revision: D23818480 Pulled By: pdillinger fbshipit-source-id: 148aff16f001af5b9fd4b22f155311c2461f1bac
4 years ago
auto find_result = backuped_file_infos_.find(dst_relative);
if (find_result == backuped_file_infos_.end() && !same_path) {
// file exists but not referenced
ROCKS_LOG_INFO(
options_.info_log,
"%s already present, but not referenced by any backup. We will "
"overwrite the file.",
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
fname.c_str());
need_to_copy = true;
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
// Defer any failure reporting to when we try to write the file
backup_fs_->DeleteFile(final_dest_path, io_options_, nullptr)
.PermitUncheckedError();
} else {
Less I/O for incremental backups, slightly better corruption detection (#7413) Summary: Two relatively simple functional changes to incremental backup behavior, integrated with a minor refactoring to reduce code redundancy and improve error/log message. There are nuances to the impact of these changes, but I believe they are fundamentally good and generally safe. Those functional changes: * Incremental backups no longer read DB table files that are already saved to a shared part of the backup directory, unless `share_files_with_checksum` is used with `kLegacyCrc32cAndFileSize` naming (discouraged) where crc32c full file checksums are needed to determine file naming. * Justification: incremental backups should not need to read the whole DB, especially without rate limiting. (Although other BackupEngine reads are not rate limited either, other non-trivial reads are generally limited by a corresponding write, as in copying files.) Also, the fact that this is not already fixed was arguably a bug/oversight in the implementation of https://github.com/facebook/rocksdb/issues/7110. * When considering whether a table file is already backed up in a shared part of backup directory, BackupEngine would already query the sizes of source (DB) and pre-existing destination (backup) files. BackupEngine now uses these file sizes to detect corruption, as at least one of (a) old backup, (b) backup in progress, or (c) current DB is corrupt if there's a size mismatch. * Justification: a random related fix that also helps to cover a small hole in corruption checking uncovered by the other functional change: * For `share_table_files` without "checksum" (not recommended), the other change regresses in detecting fundamentally unsafe use of this option combination: when you might generate different versions of same SST file number. As demonstrated by `BackupableDBTest.FailOverwritingBackups,` this regression is greatly mitigated by the new file size checking. Nevertheless, almost no reason to use `share_files_with_checksum=false` should remain, and comments are updated appropriately. Also, this change renames internal function `CalculateChecksum` to `ReadFileAndComputeChecksum` to make the performance impact of this function clear in code reviews. It is not clear what 'same_path' is for in backupable_db.cc, and I suspect it cannot be true for a DB with unique file names (like DBImpl). Nevertheless, I've tried to keep its functionality intact when `true` to minimize risk for now, despite having no unit tests for which it is true. Select impact details (much more in unit tests): For `share_files_with_checksum`, I am confident there is no regression (vs. pre-6.12) in detecting DB or backup corruption at backup creation time, mostly because the old design did not leverage this extra checksum computation for detecting inconsistencies at backup creation time. (With computed checksums in names, a recently corrupted file just looked like a different file vs. what was already backed up.) Even in the hypothetical case of DB session id collision (~100 bits entropy collision), file size in name and/or our file size check add an extra layer of protection against false success in creating an accurate new backup. (Unit test included.) `DB::VerifyChecksum` and `BackupEngine::VerifyBackup` with checksum checking are still able to catch corruptions that `CreateNewBackup` does not. Note that when custom file checksum support is added to BackupEngine, that will essentially give the same power as `DB::VerifyChecksum` into `CreateNewBackup`. We could add options for `CreateNewBackup` to cover some of what would be caught by `VerifyBackup` with checksum checking. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7413 Test Plan: Two new unit tests included, both of which fail without these changes. Although we don't test the I/O improvement directly, we test it indirectly in DB corruption detection power that was inadvertently unlocked with new backup file naming PLUS computing current content checksums (now removed). (I don't think that case of DB corruption detection justifies reading the whole DB on incremental backup.) Reviewed By: zhichao-cao Differential Revision: D23818480 Pulled By: pdillinger fbshipit-source-id: 148aff16f001af5b9fd4b22f155311c2461f1bac
4 years ago
// file exists and referenced
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
if (checksum_hex.empty()) {
// same_path should not happen for a standard DB, so OK to
// read file contents to check for checksum mismatch between
// two files from same DB getting same name.
// For compatibility with future meta file that might not have
// crc32c checksum available, consider it might be empty, but
// we don't currently generate meta file without crc32c checksum.
// Therefore we have to read & compute it if we don't have it.
if (!same_path && !find_result->second->checksum_hex.empty()) {
Less I/O for incremental backups, slightly better corruption detection (#7413) Summary: Two relatively simple functional changes to incremental backup behavior, integrated with a minor refactoring to reduce code redundancy and improve error/log message. There are nuances to the impact of these changes, but I believe they are fundamentally good and generally safe. Those functional changes: * Incremental backups no longer read DB table files that are already saved to a shared part of the backup directory, unless `share_files_with_checksum` is used with `kLegacyCrc32cAndFileSize` naming (discouraged) where crc32c full file checksums are needed to determine file naming. * Justification: incremental backups should not need to read the whole DB, especially without rate limiting. (Although other BackupEngine reads are not rate limited either, other non-trivial reads are generally limited by a corresponding write, as in copying files.) Also, the fact that this is not already fixed was arguably a bug/oversight in the implementation of https://github.com/facebook/rocksdb/issues/7110. * When considering whether a table file is already backed up in a shared part of backup directory, BackupEngine would already query the sizes of source (DB) and pre-existing destination (backup) files. BackupEngine now uses these file sizes to detect corruption, as at least one of (a) old backup, (b) backup in progress, or (c) current DB is corrupt if there's a size mismatch. * Justification: a random related fix that also helps to cover a small hole in corruption checking uncovered by the other functional change: * For `share_table_files` without "checksum" (not recommended), the other change regresses in detecting fundamentally unsafe use of this option combination: when you might generate different versions of same SST file number. As demonstrated by `BackupableDBTest.FailOverwritingBackups,` this regression is greatly mitigated by the new file size checking. Nevertheless, almost no reason to use `share_files_with_checksum=false` should remain, and comments are updated appropriately. Also, this change renames internal function `CalculateChecksum` to `ReadFileAndComputeChecksum` to make the performance impact of this function clear in code reviews. It is not clear what 'same_path' is for in backupable_db.cc, and I suspect it cannot be true for a DB with unique file names (like DBImpl). Nevertheless, I've tried to keep its functionality intact when `true` to minimize risk for now, despite having no unit tests for which it is true. Select impact details (much more in unit tests): For `share_files_with_checksum`, I am confident there is no regression (vs. pre-6.12) in detecting DB or backup corruption at backup creation time, mostly because the old design did not leverage this extra checksum computation for detecting inconsistencies at backup creation time. (With computed checksums in names, a recently corrupted file just looked like a different file vs. what was already backed up.) Even in the hypothetical case of DB session id collision (~100 bits entropy collision), file size in name and/or our file size check add an extra layer of protection against false success in creating an accurate new backup. (Unit test included.) `DB::VerifyChecksum` and `BackupEngine::VerifyBackup` with checksum checking are still able to catch corruptions that `CreateNewBackup` does not. Note that when custom file checksum support is added to BackupEngine, that will essentially give the same power as `DB::VerifyChecksum` into `CreateNewBackup`. We could add options for `CreateNewBackup` to cover some of what would be caught by `VerifyBackup` with checksum checking. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7413 Test Plan: Two new unit tests included, both of which fail without these changes. Although we don't test the I/O improvement directly, we test it indirectly in DB corruption detection power that was inadvertently unlocked with new backup file naming PLUS computing current content checksums (now removed). (I don't think that case of DB corruption detection justifies reading the whole DB on incremental backup.) Reviewed By: zhichao-cao Differential Revision: D23818480 Pulled By: pdillinger fbshipit-source-id: 148aff16f001af5b9fd4b22f155311c2461f1bac
4 years ago
assert(find_result != backuped_file_infos_.end());
// Note: to save I/O on incremental backups, we copy prior known
// checksum of the file instead of reading entire file contents
// to recompute it.
checksum_hex = find_result->second->checksum_hex;
// Regarding corruption detection, consider:
// (a) the DB file is corrupt (since previous backup) and the backup
// file is OK: we failed to detect, but the backup is safe. DB can
// be repaired/restored once its corruption is detected.
// (b) the backup file is corrupt (since previous backup) and the
// db file is OK: we failed to detect, but the backup is corrupt.
// CreateNewBackup should support fast incremental backups and
// there's no way to support that without reading all the files.
// We might add an option for extra checks on incremental backup,
// but until then, use VerifyBackups to check existing backup data.
// (c) file name collision with legitimately different content.
// This is almost inconceivable with a well-generated DB session
// ID, but even in that case, we double check the file sizes in
// BackupMeta::AddFile.
} else {
IOStatus io_s = ReadFileAndComputeChecksum(src_dir + fname, db_fs_,
src_env_options,
size_limit, &checksum_hex);
if (!io_s.ok()) {
return io_s;
Less I/O for incremental backups, slightly better corruption detection (#7413) Summary: Two relatively simple functional changes to incremental backup behavior, integrated with a minor refactoring to reduce code redundancy and improve error/log message. There are nuances to the impact of these changes, but I believe they are fundamentally good and generally safe. Those functional changes: * Incremental backups no longer read DB table files that are already saved to a shared part of the backup directory, unless `share_files_with_checksum` is used with `kLegacyCrc32cAndFileSize` naming (discouraged) where crc32c full file checksums are needed to determine file naming. * Justification: incremental backups should not need to read the whole DB, especially without rate limiting. (Although other BackupEngine reads are not rate limited either, other non-trivial reads are generally limited by a corresponding write, as in copying files.) Also, the fact that this is not already fixed was arguably a bug/oversight in the implementation of https://github.com/facebook/rocksdb/issues/7110. * When considering whether a table file is already backed up in a shared part of backup directory, BackupEngine would already query the sizes of source (DB) and pre-existing destination (backup) files. BackupEngine now uses these file sizes to detect corruption, as at least one of (a) old backup, (b) backup in progress, or (c) current DB is corrupt if there's a size mismatch. * Justification: a random related fix that also helps to cover a small hole in corruption checking uncovered by the other functional change: * For `share_table_files` without "checksum" (not recommended), the other change regresses in detecting fundamentally unsafe use of this option combination: when you might generate different versions of same SST file number. As demonstrated by `BackupableDBTest.FailOverwritingBackups,` this regression is greatly mitigated by the new file size checking. Nevertheless, almost no reason to use `share_files_with_checksum=false` should remain, and comments are updated appropriately. Also, this change renames internal function `CalculateChecksum` to `ReadFileAndComputeChecksum` to make the performance impact of this function clear in code reviews. It is not clear what 'same_path' is for in backupable_db.cc, and I suspect it cannot be true for a DB with unique file names (like DBImpl). Nevertheless, I've tried to keep its functionality intact when `true` to minimize risk for now, despite having no unit tests for which it is true. Select impact details (much more in unit tests): For `share_files_with_checksum`, I am confident there is no regression (vs. pre-6.12) in detecting DB or backup corruption at backup creation time, mostly because the old design did not leverage this extra checksum computation for detecting inconsistencies at backup creation time. (With computed checksums in names, a recently corrupted file just looked like a different file vs. what was already backed up.) Even in the hypothetical case of DB session id collision (~100 bits entropy collision), file size in name and/or our file size check add an extra layer of protection against false success in creating an accurate new backup. (Unit test included.) `DB::VerifyChecksum` and `BackupEngine::VerifyBackup` with checksum checking are still able to catch corruptions that `CreateNewBackup` does not. Note that when custom file checksum support is added to BackupEngine, that will essentially give the same power as `DB::VerifyChecksum` into `CreateNewBackup`. We could add options for `CreateNewBackup` to cover some of what would be caught by `VerifyBackup` with checksum checking. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7413 Test Plan: Two new unit tests included, both of which fail without these changes. Although we don't test the I/O improvement directly, we test it indirectly in DB corruption detection power that was inadvertently unlocked with new backup file naming PLUS computing current content checksums (now removed). (I don't think that case of DB corruption detection justifies reading the whole DB on incremental backup.) Reviewed By: zhichao-cao Differential Revision: D23818480 Pulled By: pdillinger fbshipit-source-id: 148aff16f001af5b9fd4b22f155311c2461f1bac
4 years ago
}
}
Less I/O for incremental backups, slightly better corruption detection (#7413) Summary: Two relatively simple functional changes to incremental backup behavior, integrated with a minor refactoring to reduce code redundancy and improve error/log message. There are nuances to the impact of these changes, but I believe they are fundamentally good and generally safe. Those functional changes: * Incremental backups no longer read DB table files that are already saved to a shared part of the backup directory, unless `share_files_with_checksum` is used with `kLegacyCrc32cAndFileSize` naming (discouraged) where crc32c full file checksums are needed to determine file naming. * Justification: incremental backups should not need to read the whole DB, especially without rate limiting. (Although other BackupEngine reads are not rate limited either, other non-trivial reads are generally limited by a corresponding write, as in copying files.) Also, the fact that this is not already fixed was arguably a bug/oversight in the implementation of https://github.com/facebook/rocksdb/issues/7110. * When considering whether a table file is already backed up in a shared part of backup directory, BackupEngine would already query the sizes of source (DB) and pre-existing destination (backup) files. BackupEngine now uses these file sizes to detect corruption, as at least one of (a) old backup, (b) backup in progress, or (c) current DB is corrupt if there's a size mismatch. * Justification: a random related fix that also helps to cover a small hole in corruption checking uncovered by the other functional change: * For `share_table_files` without "checksum" (not recommended), the other change regresses in detecting fundamentally unsafe use of this option combination: when you might generate different versions of same SST file number. As demonstrated by `BackupableDBTest.FailOverwritingBackups,` this regression is greatly mitigated by the new file size checking. Nevertheless, almost no reason to use `share_files_with_checksum=false` should remain, and comments are updated appropriately. Also, this change renames internal function `CalculateChecksum` to `ReadFileAndComputeChecksum` to make the performance impact of this function clear in code reviews. It is not clear what 'same_path' is for in backupable_db.cc, and I suspect it cannot be true for a DB with unique file names (like DBImpl). Nevertheless, I've tried to keep its functionality intact when `true` to minimize risk for now, despite having no unit tests for which it is true. Select impact details (much more in unit tests): For `share_files_with_checksum`, I am confident there is no regression (vs. pre-6.12) in detecting DB or backup corruption at backup creation time, mostly because the old design did not leverage this extra checksum computation for detecting inconsistencies at backup creation time. (With computed checksums in names, a recently corrupted file just looked like a different file vs. what was already backed up.) Even in the hypothetical case of DB session id collision (~100 bits entropy collision), file size in name and/or our file size check add an extra layer of protection against false success in creating an accurate new backup. (Unit test included.) `DB::VerifyChecksum` and `BackupEngine::VerifyBackup` with checksum checking are still able to catch corruptions that `CreateNewBackup` does not. Note that when custom file checksum support is added to BackupEngine, that will essentially give the same power as `DB::VerifyChecksum` into `CreateNewBackup`. We could add options for `CreateNewBackup` to cover some of what would be caught by `VerifyBackup` with checksum checking. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7413 Test Plan: Two new unit tests included, both of which fail without these changes. Although we don't test the I/O improvement directly, we test it indirectly in DB corruption detection power that was inadvertently unlocked with new backup file naming PLUS computing current content checksums (now removed). (I don't think that case of DB corruption detection justifies reading the whole DB on incremental backup.) Reviewed By: zhichao-cao Differential Revision: D23818480 Pulled By: pdillinger fbshipit-source-id: 148aff16f001af5b9fd4b22f155311c2461f1bac
4 years ago
}
if (!db_session_id.empty()) {
ROCKS_LOG_INFO(options_.info_log,
"%s already present, with checksum %s, size %" PRIu64
" and DB session identity %s",
fname.c_str(), checksum_hex.c_str(), size_bytes,
db_session_id.c_str());
} else {
ROCKS_LOG_INFO(options_.info_log,
"%s already present, with checksum %s and size %" PRIu64,
fname.c_str(), checksum_hex.c_str(), size_bytes);
}
}
}
live_dst_paths.insert(final_dest_path);
// Step 3: Add work item
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
if (!contents.empty() || need_to_copy) {
ROCKS_LOG_INFO(options_.info_log, "Copying %s to %s", fname.c_str(),
copy_dest_path->c_str());
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
CopyOrCreateWorkItem copy_or_create_work_item(
src_dir.empty() ? "" : src_dir + fname, *copy_dest_path, contents,
db_env_, backup_env_, src_env_options, options_.sync, rate_limiter,
size_limit, stats, progress_callback, src_checksum_func_name,
checksum_hex, db_id, db_session_id);
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
BackupAfterCopyOrCreateWorkItem after_copy_or_create_work_item(
copy_or_create_work_item.result.get_future(), shared, need_to_copy,
backup_env_, temp_dest_path, final_dest_path, dst_relative);
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
files_to_copy_or_create_.write(std::move(copy_or_create_work_item));
backup_items_to_finish.push_back(std::move(after_copy_or_create_work_item));
} else {
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
std::promise<CopyOrCreateResult> promise_result;
BackupAfterCopyOrCreateWorkItem after_copy_or_create_work_item(
promise_result.get_future(), shared, need_to_copy, backup_env_,
temp_dest_path, final_dest_path, dst_relative);
Handle concurrent manifest update and backup creation Summary: Fixed two related race conditions in backup creation. (1) CreateNewBackup() uses DB::DisableFileDeletions() to prevent table files from being deleted while it is copying; however, the MANIFEST file could still rotate during this time. The fix is to stop deleting the old manifest in the rotation logic. It will be deleted safely later when PurgeObsoleteFiles() runs (can only happen when file deletions are enabled). (2) CreateNewBackup() did not account for the CURRENT file being mutable. This is significant because the files returned by GetLiveFiles() contain a particular manifest filename, but the manifest to which CURRENT refers can change at any time. This causes problems when CURRENT changes between the call to GetLiveFiles() and when it's copied to the backup directory. To workaround this, I manually forge a CURRENT file referring to the manifest filename returned in GetLiveFiles(). (2) also applies to the checkpointing code, so let me know if this approach is good and I'll make the same change there. Test Plan: new test for roll manifest during backup creation. running the test before this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... IO error: /tmp/rocksdbtest-9383/backupable_db/MANIFEST-000001: No such file or directory running the test after this change: $ ./backupable_db_test --gtest_filter=BackupableDBTest.ChangeManifestDuringBackupCreation ... [ RUN ] BackupableDBTest.ChangeManifestDuringBackupCreation [ OK ] BackupableDBTest.ChangeManifestDuringBackupCreation (2836 ms) Reviewers: IslamAbdelRahman, anthony, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D54711
9 years ago
backup_items_to_finish.push_back(std::move(after_copy_or_create_work_item));
CopyOrCreateResult result;
result.io_status = IOStatus::OK();
result.size = size_bytes;
result.checksum_hex = std::move(checksum_hex);
result.db_id = std::move(db_id);
result.db_session_id = std::move(db_session_id);
promise_result.set_value(std::move(result));
}
return IOStatus::OK();
}
IOStatus BackupEngineImpl::ReadFileAndComputeChecksum(
const std::string& src, const std::shared_ptr<FileSystem>& src_fs,
const EnvOptions& src_env_options, uint64_t size_limit,
std::string* checksum_hex) const {
if (checksum_hex == nullptr) {
return status_to_io_status(Status::Aborted("Checksum pointer is null"));
}
uint32_t checksum_value = 0;
if (size_limit == 0) {
size_limit = std::numeric_limits<uint64_t>::max();
}
std::unique_ptr<SequentialFileReader> src_reader;
IOStatus io_s = SequentialFileReader::Create(
src_fs, src, FileOptions(src_env_options), &src_reader, nullptr);
if (!io_s.ok()) {
return io_s;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
RateLimiter* rate_limiter = options_.backup_rate_limiter.get();
size_t buf_size =
rate_limiter ? static_cast<size_t>(rate_limiter->GetSingleBurstBytes())
: kDefaultCopyFileBufferSize;
std::unique_ptr<char[]> buf(new char[buf_size]);
Slice data;
do {
if (stop_backup_.load(std::memory_order_acquire)) {
return status_to_io_status(Status::Incomplete("Backup stopped"));
}
size_t buffer_to_read =
(buf_size < size_limit) ? buf_size : static_cast<size_t>(size_limit);
io_s = src_reader->Read(buffer_to_read, &data, buf.get());
if (rate_limiter != nullptr) {
rate_limiter->Request(data.size(), Env::IO_LOW, nullptr /* stats */,
RateLimiter::OpType::kRead);
}
if (!io_s.ok()) {
return io_s;
}
size_limit -= data.size();
checksum_value = crc32c::Extend(checksum_value, data.data(), data.size());
} while (data.size() > 0 && size_limit > 0);
checksum_hex->assign(ChecksumInt32ToHex(checksum_value));
return io_s;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
Status BackupEngineImpl::GetFileDbIdentities(Env* src_env,
const EnvOptions& src_env_options,
const std::string& file_path,
RateLimiter* rate_limiter,
std::string* db_id,
std::string* db_session_id) {
assert(db_id != nullptr || db_session_id != nullptr);
Options options;
options.env = src_env;
SstFileDumper sst_reader(options, file_path,
2 * 1024 * 1024
/* readahead_size */,
false /* verify_checksum */, false /* output_hex */,
false /* decode_blob_index */, src_env_options,
true /* silent */);
const TableProperties* table_properties = nullptr;
std::shared_ptr<const TableProperties> tp;
Status s = sst_reader.getStatus();
if (s.ok()) {
// Try to get table properties from the table reader of sst_reader
if (!sst_reader.ReadTableProperties(&tp).ok()) {
// Try to use table properites from the initialization of sst_reader
table_properties = sst_reader.GetInitTableProperties();
} else {
table_properties = tp.get();
if (table_properties != nullptr && rate_limiter != nullptr) {
// sizeof(*table_properties) is a sufficent but far-from-exact
// approximation of read bytes due to metaindex block, std::string
// properties and varint compression
rate_limiter->Request(sizeof(*table_properties), Env::IO_LOW,
nullptr /* stats */, RateLimiter::OpType::kRead);
}
}
} else {
ROCKS_LOG_INFO(options_.info_log, "Failed to read %s: %s",
file_path.c_str(), s.ToString().c_str());
return s;
}
if (table_properties != nullptr) {
if (db_id != nullptr) {
db_id->assign(table_properties->db_id);
}
if (db_session_id != nullptr) {
db_session_id->assign(table_properties->db_session_id);
if (db_session_id->empty()) {
s = Status::NotFound("DB session identity not found in " + file_path);
ROCKS_LOG_INFO(options_.info_log, "%s", s.ToString().c_str());
return s;
}
}
return Status::OK();
} else {
s = Status::Corruption("Table properties missing in " + file_path);
ROCKS_LOG_INFO(options_.info_log, "%s", s.ToString().c_str());
return s;
}
}
void BackupEngineImpl::DeleteChildren(const std::string& dir,
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
uint32_t file_type_filter) const {
std::vector<std::string> children;
db_fs_->GetChildren(dir, io_options_, &children, nullptr)
.PermitUncheckedError(); // ignore errors
for (const auto& f : children) {
uint64_t number;
FileType type;
bool ok = ParseFileName(f, &number, &type);
if (ok && (file_type_filter & (1 << type))) {
// don't delete this file
continue;
}
db_fs_->DeleteFile(dir + "/" + f, io_options_, nullptr)
.PermitUncheckedError(); // ignore errors
}
}
IOStatus BackupEngineImpl::ReadChildFileCurrentSizes(
const std::string& dir, const std::shared_ptr<FileSystem>& fs,
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
std::unordered_map<std::string, uint64_t>* result) const {
assert(result != nullptr);
std::vector<Env::FileAttributes> files_attrs;
IOStatus io_status = fs->FileExists(dir, io_options_, nullptr);
if (io_status.ok()) {
io_status =
fs->GetChildrenFileAttributes(dir, io_options_, &files_attrs, nullptr);
} else if (io_status.IsNotFound()) {
// Insert no entries can be considered success
io_status = IOStatus::OK();
}
const bool slash_needed = dir.empty() || dir.back() != '/';
for (const auto& file_attrs : files_attrs) {
result->emplace(dir + (slash_needed ? "/" : "") + file_attrs.name,
file_attrs.size_bytes);
}
return io_status;
}
IOStatus BackupEngineImpl::GarbageCollect() {
assert(!read_only_);
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
// We will make a best effort to remove all garbage even in the presence
// of inconsistencies or I/O failures that inhibit finding garbage.
IOStatus overall_status = IOStatus::OK();
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
// If all goes well, we don't need another auto-GC this session
might_need_garbage_collect_ = false;
ROCKS_LOG_INFO(options_.info_log, "Starting garbage collection");
// delete obsolete shared files
for (bool with_checksum : {false, true}) {
std::vector<std::string> shared_children;
{
std::string shared_path;
if (with_checksum) {
shared_path = GetAbsolutePath(GetSharedFileWithChecksumRel());
} else {
shared_path = GetAbsolutePath(GetSharedFileRel());
}
IOStatus io_s = backup_fs_->FileExists(shared_path, io_options_, nullptr);
if (io_s.ok()) {
io_s = backup_fs_->GetChildren(shared_path, io_options_,
&shared_children, nullptr);
} else if (io_s.IsNotFound()) {
io_s = IOStatus::OK();
}
if (!io_s.ok()) {
overall_status = io_s;
// Trying again later might work
might_need_garbage_collect_ = true;
}
}
for (auto& child : shared_children) {
std::string rel_fname;
if (with_checksum) {
rel_fname = GetSharedFileWithChecksumRel(child);
} else {
rel_fname = GetSharedFileRel(child);
}
auto child_itr = backuped_file_infos_.find(rel_fname);
// if it's not refcounted, delete it
if (child_itr == backuped_file_infos_.end() ||
child_itr->second->refs == 0) {
// this might be a directory, but DeleteFile will just fail in that
// case, so we're good
IOStatus io_s = backup_fs_->DeleteFile(GetAbsolutePath(rel_fname),
io_options_, nullptr);
ROCKS_LOG_INFO(options_.info_log, "Deleting %s -- %s",
rel_fname.c_str(), io_s.ToString().c_str());
backuped_file_infos_.erase(rel_fname);
if (!io_s.ok()) {
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
// Trying again later might work
might_need_garbage_collect_ = true;
}
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
// delete obsolete private files
std::vector<std::string> private_children;
{
IOStatus io_s =
backup_fs_->GetChildren(GetAbsolutePath(kPrivateDirName), io_options_,
&private_children, nullptr);
if (!io_s.ok()) {
overall_status = io_s;
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
// Trying again later might work
might_need_garbage_collect_ = true;
}
}
for (auto& child : private_children) {
BackupID backup_id = 0;
bool tmp_dir = child.find(".tmp") != std::string::npos;
sscanf(child.c_str(), "%u", &backup_id);
if (!tmp_dir && // if it's tmp_dir, delete it
(backup_id == 0 || backups_.find(backup_id) != backups_.end())) {
// it's either not a number or it's still alive. continue
continue;
}
// here we have to delete the dir and all its children
std::string full_private_path =
GetAbsolutePath(GetPrivateFileRel(backup_id));
std::vector<std::string> subchildren;
if (backup_fs_
->GetChildren(full_private_path, io_options_, &subchildren, nullptr)
.ok()) {
for (auto& subchild : subchildren) {
IOStatus io_s = backup_fs_->DeleteFile(full_private_path + subchild,
io_options_, nullptr);
ROCKS_LOG_INFO(options_.info_log, "Deleting %s -- %s",
(full_private_path + subchild).c_str(),
io_s.ToString().c_str());
if (!io_s.ok()) {
// Trying again later might work
might_need_garbage_collect_ = true;
}
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
// finally delete the private dir
IOStatus io_s =
backup_fs_->DeleteDir(full_private_path, io_options_, nullptr);
ROCKS_LOG_INFO(options_.info_log, "Deleting dir %s -- %s",
full_private_path.c_str(), io_s.ToString().c_str());
if (!io_s.ok()) {
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
// Trying again later might work
might_need_garbage_collect_ = true;
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
Auto-GarbageCollect on PurgeOldBackups and DeleteBackup (#6015) Summary: Only if there is a crash, power failure, or I/O error in DeleteBackup, shared or private files from the backup might be left behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only by GarbageCollect. This makes the BackupEngine API "leaky by default." Even if it means a modest performance hit, I think we should make Delete and Purge do as they say, with ongoing best effort: i.e. future calls will attempt to finish any incomplete work from earlier calls. This change does that by having DeleteBackup and PurgeOldBackups do a GarbageCollect, unless (to minimize performance hit) this BackupEngine has already done a GarbageCollect and there have been no deletion-related I/O errors in that GarbageCollect or since then. Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files. Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015 Test Plan: Updated unit tests Differential Revision: D18401333 Pulled By: pdillinger fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
5 years ago
assert(overall_status.ok() || might_need_garbage_collect_);
return overall_status;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
// ------- BackupMeta class --------
IOStatus BackupEngineImpl::BackupMeta::AddFile(
std::shared_ptr<FileInfo> file_info) {
auto itr = file_infos_->find(file_info->filename);
if (itr == file_infos_->end()) {
auto ret = file_infos_->insert({file_info->filename, file_info});
if (ret.second) {
itr = ret.first;
itr->second->refs = 1;
} else {
// if this happens, something is seriously wrong
return IOStatus::Corruption("In memory metadata insertion error");
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
} else {
Less I/O for incremental backups, slightly better corruption detection (#7413) Summary: Two relatively simple functional changes to incremental backup behavior, integrated with a minor refactoring to reduce code redundancy and improve error/log message. There are nuances to the impact of these changes, but I believe they are fundamentally good and generally safe. Those functional changes: * Incremental backups no longer read DB table files that are already saved to a shared part of the backup directory, unless `share_files_with_checksum` is used with `kLegacyCrc32cAndFileSize` naming (discouraged) where crc32c full file checksums are needed to determine file naming. * Justification: incremental backups should not need to read the whole DB, especially without rate limiting. (Although other BackupEngine reads are not rate limited either, other non-trivial reads are generally limited by a corresponding write, as in copying files.) Also, the fact that this is not already fixed was arguably a bug/oversight in the implementation of https://github.com/facebook/rocksdb/issues/7110. * When considering whether a table file is already backed up in a shared part of backup directory, BackupEngine would already query the sizes of source (DB) and pre-existing destination (backup) files. BackupEngine now uses these file sizes to detect corruption, as at least one of (a) old backup, (b) backup in progress, or (c) current DB is corrupt if there's a size mismatch. * Justification: a random related fix that also helps to cover a small hole in corruption checking uncovered by the other functional change: * For `share_table_files` without "checksum" (not recommended), the other change regresses in detecting fundamentally unsafe use of this option combination: when you might generate different versions of same SST file number. As demonstrated by `BackupableDBTest.FailOverwritingBackups,` this regression is greatly mitigated by the new file size checking. Nevertheless, almost no reason to use `share_files_with_checksum=false` should remain, and comments are updated appropriately. Also, this change renames internal function `CalculateChecksum` to `ReadFileAndComputeChecksum` to make the performance impact of this function clear in code reviews. It is not clear what 'same_path' is for in backupable_db.cc, and I suspect it cannot be true for a DB with unique file names (like DBImpl). Nevertheless, I've tried to keep its functionality intact when `true` to minimize risk for now, despite having no unit tests for which it is true. Select impact details (much more in unit tests): For `share_files_with_checksum`, I am confident there is no regression (vs. pre-6.12) in detecting DB or backup corruption at backup creation time, mostly because the old design did not leverage this extra checksum computation for detecting inconsistencies at backup creation time. (With computed checksums in names, a recently corrupted file just looked like a different file vs. what was already backed up.) Even in the hypothetical case of DB session id collision (~100 bits entropy collision), file size in name and/or our file size check add an extra layer of protection against false success in creating an accurate new backup. (Unit test included.) `DB::VerifyChecksum` and `BackupEngine::VerifyBackup` with checksum checking are still able to catch corruptions that `CreateNewBackup` does not. Note that when custom file checksum support is added to BackupEngine, that will essentially give the same power as `DB::VerifyChecksum` into `CreateNewBackup`. We could add options for `CreateNewBackup` to cover some of what would be caught by `VerifyBackup` with checksum checking. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7413 Test Plan: Two new unit tests included, both of which fail without these changes. Although we don't test the I/O improvement directly, we test it indirectly in DB corruption detection power that was inadvertently unlocked with new backup file naming PLUS computing current content checksums (now removed). (I don't think that case of DB corruption detection justifies reading the whole DB on incremental backup.) Reviewed By: zhichao-cao Differential Revision: D23818480 Pulled By: pdillinger fbshipit-source-id: 148aff16f001af5b9fd4b22f155311c2461f1bac
4 years ago
// Compare sizes, because we scanned that off the filesystem on both
// ends. This is like a check in VerifyBackup.
if (itr->second->size != file_info->size) {
std::string msg = "Size mismatch for existing backup file: ";
msg.append(file_info->filename);
msg.append(" Size in backup is " + ToString(itr->second->size) +
" while size in DB is " + ToString(file_info->size));
msg.append(
" If this DB file checks as not corrupt, try deleting old"
" backups or backing up to a different backup directory.");
return IOStatus::Corruption(msg);
Less I/O for incremental backups, slightly better corruption detection (#7413) Summary: Two relatively simple functional changes to incremental backup behavior, integrated with a minor refactoring to reduce code redundancy and improve error/log message. There are nuances to the impact of these changes, but I believe they are fundamentally good and generally safe. Those functional changes: * Incremental backups no longer read DB table files that are already saved to a shared part of the backup directory, unless `share_files_with_checksum` is used with `kLegacyCrc32cAndFileSize` naming (discouraged) where crc32c full file checksums are needed to determine file naming. * Justification: incremental backups should not need to read the whole DB, especially without rate limiting. (Although other BackupEngine reads are not rate limited either, other non-trivial reads are generally limited by a corresponding write, as in copying files.) Also, the fact that this is not already fixed was arguably a bug/oversight in the implementation of https://github.com/facebook/rocksdb/issues/7110. * When considering whether a table file is already backed up in a shared part of backup directory, BackupEngine would already query the sizes of source (DB) and pre-existing destination (backup) files. BackupEngine now uses these file sizes to detect corruption, as at least one of (a) old backup, (b) backup in progress, or (c) current DB is corrupt if there's a size mismatch. * Justification: a random related fix that also helps to cover a small hole in corruption checking uncovered by the other functional change: * For `share_table_files` without "checksum" (not recommended), the other change regresses in detecting fundamentally unsafe use of this option combination: when you might generate different versions of same SST file number. As demonstrated by `BackupableDBTest.FailOverwritingBackups,` this regression is greatly mitigated by the new file size checking. Nevertheless, almost no reason to use `share_files_with_checksum=false` should remain, and comments are updated appropriately. Also, this change renames internal function `CalculateChecksum` to `ReadFileAndComputeChecksum` to make the performance impact of this function clear in code reviews. It is not clear what 'same_path' is for in backupable_db.cc, and I suspect it cannot be true for a DB with unique file names (like DBImpl). Nevertheless, I've tried to keep its functionality intact when `true` to minimize risk for now, despite having no unit tests for which it is true. Select impact details (much more in unit tests): For `share_files_with_checksum`, I am confident there is no regression (vs. pre-6.12) in detecting DB or backup corruption at backup creation time, mostly because the old design did not leverage this extra checksum computation for detecting inconsistencies at backup creation time. (With computed checksums in names, a recently corrupted file just looked like a different file vs. what was already backed up.) Even in the hypothetical case of DB session id collision (~100 bits entropy collision), file size in name and/or our file size check add an extra layer of protection against false success in creating an accurate new backup. (Unit test included.) `DB::VerifyChecksum` and `BackupEngine::VerifyBackup` with checksum checking are still able to catch corruptions that `CreateNewBackup` does not. Note that when custom file checksum support is added to BackupEngine, that will essentially give the same power as `DB::VerifyChecksum` into `CreateNewBackup`. We could add options for `CreateNewBackup` to cover some of what would be caught by `VerifyBackup` with checksum checking. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7413 Test Plan: Two new unit tests included, both of which fail without these changes. Although we don't test the I/O improvement directly, we test it indirectly in DB corruption detection power that was inadvertently unlocked with new backup file naming PLUS computing current content checksums (now removed). (I don't think that case of DB corruption detection justifies reading the whole DB on incremental backup.) Reviewed By: zhichao-cao Differential Revision: D23818480 Pulled By: pdillinger fbshipit-source-id: 148aff16f001af5b9fd4b22f155311c2461f1bac
4 years ago
}
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
if (file_info->checksum_hex.empty()) {
// No checksum available to check
} else if (itr->second->checksum_hex.empty()) {
// Remember checksum if newly acquired
itr->second->checksum_hex = file_info->checksum_hex;
} else if (itr->second->checksum_hex != file_info->checksum_hex) {
// Note: to save I/O, these will be equal trivially on already backed
// up files that don't have the checksum in their name. And it should
// never fail for files that do have checksum in their name.
Less I/O for incremental backups, slightly better corruption detection (#7413) Summary: Two relatively simple functional changes to incremental backup behavior, integrated with a minor refactoring to reduce code redundancy and improve error/log message. There are nuances to the impact of these changes, but I believe they are fundamentally good and generally safe. Those functional changes: * Incremental backups no longer read DB table files that are already saved to a shared part of the backup directory, unless `share_files_with_checksum` is used with `kLegacyCrc32cAndFileSize` naming (discouraged) where crc32c full file checksums are needed to determine file naming. * Justification: incremental backups should not need to read the whole DB, especially without rate limiting. (Although other BackupEngine reads are not rate limited either, other non-trivial reads are generally limited by a corresponding write, as in copying files.) Also, the fact that this is not already fixed was arguably a bug/oversight in the implementation of https://github.com/facebook/rocksdb/issues/7110. * When considering whether a table file is already backed up in a shared part of backup directory, BackupEngine would already query the sizes of source (DB) and pre-existing destination (backup) files. BackupEngine now uses these file sizes to detect corruption, as at least one of (a) old backup, (b) backup in progress, or (c) current DB is corrupt if there's a size mismatch. * Justification: a random related fix that also helps to cover a small hole in corruption checking uncovered by the other functional change: * For `share_table_files` without "checksum" (not recommended), the other change regresses in detecting fundamentally unsafe use of this option combination: when you might generate different versions of same SST file number. As demonstrated by `BackupableDBTest.FailOverwritingBackups,` this regression is greatly mitigated by the new file size checking. Nevertheless, almost no reason to use `share_files_with_checksum=false` should remain, and comments are updated appropriately. Also, this change renames internal function `CalculateChecksum` to `ReadFileAndComputeChecksum` to make the performance impact of this function clear in code reviews. It is not clear what 'same_path' is for in backupable_db.cc, and I suspect it cannot be true for a DB with unique file names (like DBImpl). Nevertheless, I've tried to keep its functionality intact when `true` to minimize risk for now, despite having no unit tests for which it is true. Select impact details (much more in unit tests): For `share_files_with_checksum`, I am confident there is no regression (vs. pre-6.12) in detecting DB or backup corruption at backup creation time, mostly because the old design did not leverage this extra checksum computation for detecting inconsistencies at backup creation time. (With computed checksums in names, a recently corrupted file just looked like a different file vs. what was already backed up.) Even in the hypothetical case of DB session id collision (~100 bits entropy collision), file size in name and/or our file size check add an extra layer of protection against false success in creating an accurate new backup. (Unit test included.) `DB::VerifyChecksum` and `BackupEngine::VerifyBackup` with checksum checking are still able to catch corruptions that `CreateNewBackup` does not. Note that when custom file checksum support is added to BackupEngine, that will essentially give the same power as `DB::VerifyChecksum` into `CreateNewBackup`. We could add options for `CreateNewBackup` to cover some of what would be caught by `VerifyBackup` with checksum checking. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7413 Test Plan: Two new unit tests included, both of which fail without these changes. Although we don't test the I/O improvement directly, we test it indirectly in DB corruption detection power that was inadvertently unlocked with new backup file naming PLUS computing current content checksums (now removed). (I don't think that case of DB corruption detection justifies reading the whole DB on incremental backup.) Reviewed By: zhichao-cao Differential Revision: D23818480 Pulled By: pdillinger fbshipit-source-id: 148aff16f001af5b9fd4b22f155311c2461f1bac
4 years ago
// Should never reach here, but produce an appropriate corruption
// message in case we do in a release build.
assert(false);
std::string msg = "Checksum mismatch for existing backup file: ";
msg.append(file_info->filename);
msg.append(" Expected checksum is " + itr->second->checksum_hex +
" while computed checksum is " + file_info->checksum_hex);
msg.append(
" If this DB file checks as not corrupt, try deleting old"
" backups or backing up to a different backup directory.");
return IOStatus::Corruption(msg);
}
++itr->second->refs; // increase refcount if already present
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
size_ += file_info->size;
files_.push_back(itr->second);
return IOStatus::OK();
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
IOStatus BackupEngineImpl::BackupMeta::Delete(bool delete_meta) {
IOStatus io_s;
for (const auto& file : files_) {
--file->refs; // decrease refcount
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
files_.clear();
// delete meta file
if (delete_meta) {
io_s = fs_->FileExists(meta_filename_, iooptions_, nullptr);
if (io_s.ok()) {
io_s = fs_->DeleteFile(meta_filename_, iooptions_, nullptr);
} else if (io_s.IsNotFound()) {
io_s = IOStatus::OK(); // nothing to delete
}
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
timestamp_ = 0;
return io_s;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
// Constants for backup meta file schema (see LoadFromFile)
namespace {
const std::string kSchemaVersionPrefix{"schema_version "};
const std::string kFooterMarker{"// FOOTER"};
const std::string kAppMetaDataFieldName{"metadata"};
// WART: The checksums are crc32c but named "crc32"
const std::string kFileCrc32cFieldName{"crc32"};
const std::string kFileSizeFieldName{"size"};
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
// Marks a (future) field that should cause failure if not recognized.
// Other fields are assumed to be ignorable. For example, in the future
// we might add
// ni::file_name_escape uri_percent
// to indicate all file names have had spaces and special characters
// escaped using a URI percent encoding.
const std::string kNonIgnorableFieldPrefix{"ni::"};
} // namespace
// Each backup meta file is of the format (schema version 1):
//----------------------------------------------------------
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
// <timestamp>
// <seq number>
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
// metadata <metadata> (optional)
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
// <number of files>
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
// <file1> crc32 <crc32c_as_unsigned_decimal>
// <file2> crc32 <crc32c_as_unsigned_decimal>
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
// ...
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
//----------------------------------------------------------
//
// For schema version 2.x (not in public APIs, but
// forward-compatibility started):
//----------------------------------------------------------
// schema_version <ver>
// <timestamp>
// <seq number>
// [<field name> <field data>]
// ...
// <number of files>
// <file1>( <field name> <field data no spaces>)*
// <file2>( <field name> <field data no spaces>)*
// ...
// [// FOOTER]
// [<field name> <field data>]
// ...
//----------------------------------------------------------
// where
// <ver> ::= [0-9]+([.][0-9]+)
// <field name> ::= [A-Za-z_][A-Za-z_0-9.]+
// <field data> is anything but newline
// <field data no spaces> is anything but space and newline
// Although "// FOOTER" wouldn't strictly be required as a delimiter
// given the number of files is included, it is there for parsing
// sanity in case of corruption. It is only required if followed
// by footer fields, such as a checksum of the meta file (so far).
// Unrecognized fields are ignored, to support schema evolution on
// non-critical features with forward compatibility. Update schema
// major version for breaking changes. Schema minor versions are indicated
// only for diagnostic/debugging purposes.
//
// Fields in schema version 2.0:
// * Top-level meta fields:
// * Only "metadata" as in schema version 1
// * File meta fields:
// * "crc32" - a crc32c checksum as in schema version 1
// * "size" - the size of the file (new)
// * Footer meta fields:
// * None yet (future use for meta file checksum anticipated)
//
IOStatus BackupEngineImpl::BackupMeta::LoadFromFile(
const std::string& backup_dir,
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
const std::unordered_map<std::string, uint64_t>& abs_path_to_size,
RateLimiter* rate_limiter, Logger* info_log,
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
std::unordered_set<std::string>* reported_ignored_fields) {
assert(reported_ignored_fields);
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
assert(Empty());
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
std::unique_ptr<LineFileReader> backup_meta_reader;
{
IOStatus io_s = LineFileReader::Create(fs_, meta_filename_, FileOptions(),
&backup_meta_reader, nullptr);
if (!io_s.ok()) {
return io_s;
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
// If we don't read an explicit schema_version, that implies version 1,
// which is what we call the original backup meta schema.
int schema_major_version = 1;
// Failures handled at the end
std::string line;
if (backup_meta_reader->ReadLine(&line)) {
if (rate_limiter != nullptr) {
rate_limiter->Request(line.size(), Env::IO_LOW, nullptr /* stats */,
RateLimiter::OpType::kRead);
}
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
if (StartsWith(line, kSchemaVersionPrefix)) {
std::string ver = line.substr(kSchemaVersionPrefix.size());
if (ver == "2" || StartsWith(ver, "2.")) {
schema_major_version = 2;
} else {
return IOStatus::NotSupported(
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
"Unsupported/unrecognized schema version: " + ver);
}
line.clear();
} else if (line.empty()) {
return IOStatus::Corruption("Unexpected empty line");
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
}
}
if (!line.empty()) {
timestamp_ = std::strtoull(line.c_str(), nullptr, /*base*/ 10);
} else if (backup_meta_reader->ReadLine(&line)) {
if (rate_limiter != nullptr) {
rate_limiter->Request(line.size(), Env::IO_LOW, nullptr /* stats */,
RateLimiter::OpType::kRead);
}
timestamp_ = std::strtoull(line.c_str(), nullptr, /*base*/ 10);
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
if (backup_meta_reader->ReadLine(&line)) {
if (rate_limiter != nullptr) {
rate_limiter->Request(line.size(), Env::IO_LOW, nullptr /* stats */,
RateLimiter::OpType::kRead);
}
sequence_number_ = std::strtoull(line.c_str(), nullptr, /*base*/ 10);
}
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
uint32_t num_files = UINT32_MAX;
while (backup_meta_reader->ReadLine(&line)) {
if (rate_limiter != nullptr) {
rate_limiter->Request(line.size(), Env::IO_LOW, nullptr /* stats */,
RateLimiter::OpType::kRead);
}
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
if (line.empty()) {
return IOStatus::Corruption("Unexpected empty line");
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
}
// Number -> number of files -> exit loop reading optional meta fields
if (line[0] >= '0' && line[0] <= '9') {
num_files = static_cast<uint32_t>(strtoul(line.c_str(), nullptr, 10));
break;
}
// else, must be a meta field assignment
auto space_pos = line.find_first_of(' ');
if (space_pos == std::string::npos) {
return IOStatus::Corruption("Expected number of files or meta field");
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
}
std::string field_name = line.substr(0, space_pos);
std::string field_data = line.substr(space_pos + 1);
if (field_name == kAppMetaDataFieldName) {
// app metadata present
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
bool decode_success = Slice(field_data).DecodeHex(&app_metadata_);
if (!decode_success) {
return IOStatus::Corruption(
"Failed to decode stored hex encoded app metadata");
}
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
} else if (schema_major_version < 2) {
return IOStatus::Corruption("Expected number of files or \"" +
kAppMetaDataFieldName + "\" field");
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
} else if (StartsWith(field_name, kNonIgnorableFieldPrefix)) {
return IOStatus::NotSupported("Unrecognized non-ignorable meta field " +
field_name + " (from future version?)");
} else {
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
// Warn the first time we see any particular unrecognized meta field
if (reported_ignored_fields->insert("meta:" + field_name).second) {
ROCKS_LOG_WARN(info_log, "Ignoring unrecognized backup meta field %s",
field_name.c_str());
}
}
}
std::vector<std::shared_ptr<FileInfo>> files;
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
bool footer_present = false;
while (backup_meta_reader->ReadLine(&line)) {
if (rate_limiter != nullptr) {
rate_limiter->Request(line.size(), Env::IO_LOW, nullptr /* stats */,
RateLimiter::OpType::kRead);
}
std::vector<std::string> components = StringSplit(line, ' ');
if (components.size() < 1) {
return IOStatus::Corruption("Empty line instead of file entry.");
}
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
if (schema_major_version >= 2 && components.size() == 2 &&
line == kFooterMarker) {
footer_present = true;
break;
}
const std::string& filename = components[0];
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
uint64_t actual_size;
const std::shared_ptr<FileInfo> file_info = GetFile(filename);
if (file_info) {
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
actual_size = file_info->size;
} else {
std::string abs_path = backup_dir + "/" + filename;
auto e = abs_path_to_size.find(abs_path);
if (e == abs_path_to_size.end()) {
return IOStatus::Corruption(
"Pathname in meta file not found on disk: " + abs_path);
}
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
actual_size = e->second;
}
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
if (schema_major_version >= 2) {
if (components.size() % 2 != 1) {
return IOStatus::Corruption(
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
"Bad number of line components for file entry.");
}
} else {
// Check restricted original schema
if (components.size() < 3) {
return IOStatus::Corruption("File checksum is missing for " + filename +
" in " + meta_filename_);
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
}
if (components[1] != kFileCrc32cFieldName) {
return IOStatus::Corruption("Unknown checksum type for " + filename +
" in " + meta_filename_);
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
}
if (components.size() > 3) {
return IOStatus::Corruption("Extra data for entry " + filename +
" in " + meta_filename_);
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
}
}
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
std::string checksum_hex;
for (unsigned i = 1; i < components.size(); i += 2) {
const std::string& field_name = components[i];
const std::string& field_data = components[i + 1];
if (field_name == kFileCrc32cFieldName) {
uint32_t checksum_value =
static_cast<uint32_t>(strtoul(field_data.c_str(), nullptr, 10));
if (field_data != ROCKSDB_NAMESPACE::ToString(checksum_value)) {
return IOStatus::Corruption("Invalid checksum value for " + filename +
" in " + meta_filename_);
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
}
checksum_hex = ChecksumInt32ToHex(checksum_value);
} else if (field_name == kFileSizeFieldName) {
uint64_t ex_size =
std::strtoull(field_data.c_str(), nullptr, /*base*/ 10);
if (ex_size != actual_size) {
return IOStatus::Corruption(
"For file " + filename + " expected size " + ToString(ex_size) +
" but found size" + ToString(actual_size));
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
}
} else if (StartsWith(field_name, kNonIgnorableFieldPrefix)) {
return IOStatus::NotSupported("Unrecognized non-ignorable file field " +
field_name + " (from future version?)");
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
} else {
// Warn the first time we see any particular unrecognized file field
if (reported_ignored_fields->insert("file:" + field_name).second) {
ROCKS_LOG_WARN(info_log, "Ignoring unrecognized backup file field %s",
field_name.c_str());
}
}
}
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
files.emplace_back(new FileInfo(filename, actual_size, checksum_hex));
}
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
if (footer_present) {
assert(schema_major_version >= 2);
while (backup_meta_reader->ReadLine(&line)) {
if (rate_limiter != nullptr) {
rate_limiter->Request(line.size(), Env::IO_LOW, nullptr /* stats */,
RateLimiter::OpType::kRead);
}
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
if (line.empty()) {
return IOStatus::Corruption("Unexpected empty line");
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
}
auto space_pos = line.find_first_of(' ');
if (space_pos == std::string::npos) {
return IOStatus::Corruption("Expected footer field");
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
}
std::string field_name = line.substr(0, space_pos);
std::string field_data = line.substr(space_pos + 1);
if (StartsWith(field_name, kNonIgnorableFieldPrefix)) {
return IOStatus::NotSupported("Unrecognized non-ignorable field " +
field_name + " (from future version?)");
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
} else if (reported_ignored_fields->insert("footer:" + field_name)
.second) {
// Warn the first time we see any particular unrecognized footer field
ROCKS_LOG_WARN(info_log,
"Ignoring unrecognized backup meta footer field %s",
field_name.c_str());
}
}
}
{
IOStatus io_s = backup_meta_reader->GetStatus();
if (!io_s.ok()) {
return io_s;
}
}
if (num_files != files.size()) {
return IOStatus::Corruption(
"Inconsistent number of files or missing/incomplete header in " +
meta_filename_);
}
files_.reserve(files.size());
for (const auto& file_info : files) {
IOStatus io_s = AddFile(file_info);
if (!io_s.ok()) {
return io_s;
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
return IOStatus::OK();
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
IOStatus BackupEngineImpl::BackupMeta::StoreToFile(
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
bool sync, const TEST_FutureSchemaVersion2Options* test_future_options) {
IOStatus io_s;
std::unique_ptr<FSWritableFile> backup_meta_file;
FileOptions file_options;
file_options.use_mmap_writes = false;
file_options.use_direct_writes = false;
io_s = fs_->NewWritableFile(meta_tmp_filename_, file_options,
&backup_meta_file, nullptr);
if (!io_s.ok()) {
return io_s;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
std::ostringstream buf;
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
if (test_future_options) {
buf << kSchemaVersionPrefix << test_future_options->version << "\n";
}
buf << static_cast<unsigned long long>(timestamp_) << "\n";
buf << sequence_number_ << "\n";
if (!app_metadata_.empty()) {
std::string hex_encoded_metadata =
Slice(app_metadata_).ToString(/* hex */ true);
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
buf << kAppMetaDataFieldName << " " << hex_encoded_metadata << "\n";
}
if (test_future_options) {
for (auto& e : test_future_options->meta_fields) {
buf << e.first << " " << e.second << "\n";
}
}
buf << files_.size() << "\n";
for (const auto& file : files_) {
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
buf << file->filename;
if (test_future_options == nullptr ||
test_future_options->crc32c_checksums) {
// use crc32c for now, switch to something else if needed
buf << " " << kFileCrc32cFieldName << " "
<< ChecksumHexToInt32(file->checksum_hex);
}
if (test_future_options && test_future_options->file_sizes) {
buf << " " << kFileSizeFieldName << " " << ToString(file->size);
}
if (test_future_options) {
for (auto& e : test_future_options->file_fields) {
buf << " " << e.first << " " << e.second;
}
}
buf << "\n";
}
if (test_future_options && !test_future_options->footer_fields.empty()) {
buf << kFooterMarker << "\n";
for (auto& e : test_future_options->footer_fields) {
buf << e.first << " " << e.second << "\n";
}
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
io_s = backup_meta_file->Append(Slice(buf.str()), iooptions_, nullptr);
IOSTATS_ADD(bytes_written, buf.str().size());
if (io_s.ok() && sync) {
io_s = backup_meta_file->Sync(iooptions_, nullptr);
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
if (io_s.ok()) {
io_s = backup_meta_file->Close(iooptions_, nullptr);
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
if (io_s.ok()) {
io_s = fs_->RenameFile(meta_tmp_filename_, meta_filename_, iooptions_,
nullptr);
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
return io_s;
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
}
IOStatus BackupEngineReadOnly::Open(const BackupEngineOptions& options,
Env* env,
BackupEngineReadOnly** backup_engine_ptr) {
if (options.destroy_old_data) {
return IOStatus::InvalidArgument(
"Can't destroy old data with ReadOnly BackupEngine");
}
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
std::unique_ptr<BackupEngineImplThreadSafe> backup_engine(
new BackupEngineImplThreadSafe(options, env, true /*read_only*/));
auto s = backup_engine->Initialize();
if (!s.ok()) {
*backup_engine_ptr = nullptr;
return s;
}
*backup_engine_ptr = backup_engine.release();
return IOStatus::OK();
}
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
void TEST_EnableWriteFutureSchemaVersion2(
BackupEngine* engine, const TEST_FutureSchemaVersion2Options& options) {
Add thread safety to BackupEngine, explain more (#8115) Summary: BackupEngine previously had unclear but strict concurrency requirements that the API user must follow for safe use. Now we make that clear, by separating operations into "Read," "Append," and "Write" operations, and specifying which combinations are safe across threads on the same BackupEngine object (previously none; now all, using a read-write lock), and which are safe across different BackupEngine instances open on the same backup_dir. The changes to backupable_db.h should be backward compatible. It is mostly about eliminating copies of what should be the same function and (unsurprisingly) useful documentation comments were often placed on only one of the two copies. With the re-organization, we are also grouping different categories of operations. In the future we might add BackupEngineReadAppendOnly, but that didn't seem necessary. To mark API Read operations 'const', I had to mark some implementation functions 'const' and some fields mutable. Functional changes: * Added RWMutex locking around public API functions to implement thread safety on a single object. To avoid future bugs, this is another internal class layered on top (removing many "override" in BackupEngineImpl). It would be possible to allow more concurrency between operations, rather than mutual exclusion, but IMHO not worth the work. * Fixed a race between Open() (Initialize()) and CreateNewBackup() for different objects on the same backup_dir, where Initialize() could delete the temporary meta file created during CreateNewBackup(). (This was found by the new test.) Also cleaned up a couple of "status checked" TODOs, and improved a checksum mismatch error message to include involved files. Potential follow-up work: * CreateNewBackup has an API wart because it doesn't tell you the BackupID it just created, which makes it of limited use in a multithreaded setting. * We could also consider a Refresh() function to catch up to changes made from another BackupEngine object to the same dir. * Use a lock file to prevent multiple writer BackupEngines, but this won't work on remote filesystems not supporting lock files. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115 Test Plan: new mini-stress test in backup unit tests, run with gcc, clang, ASC, TSAN, and UBSAN, 100 iterations each. Reviewed By: ajkr Differential Revision: D27347589 Pulled By: pdillinger fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
BackupEngineImplThreadSafe* impl =
static_cast_with_check<BackupEngineImplThreadSafe>(engine);
impl->TEST_EnableWriteFutureSchemaVersion2(options);
Begin forward compatibility for new backup meta schema (#8069) Summary: This does not add any new public APIs or published functionality, but adds the ability to read and use (and in tests, write) backups with a new meta file schema, based on the old schema but not forward-compatible (before this change). The new schema enables some capabilities not in the old: * Explicit versioning, so that users get clean error messages the next time we want to break forward compatibility. * Ignoring unrecognized fields (with warning), so that new non-critical features can be added without breaking forward compatibility. * Rejecting future "non-ignorable" fields, so that new features critical to some use-cases could potentially be added outside of linear schema versions, with broken forward compatibility. * Fields at the end of the meta file, such as for checksum of the meta file's contents (up to that point) * New optional 'size' field for each file, which is checked when present * Optionally omitting 'crc32' field, so that we aren't required to have a crc32c checksum for files to take a backup. (E.g. to support backup via hard links and to better support file custom checksums.) Because we do not have a JSON parser and to share code, the new schema is simply derived from the old schema. BackupEngine code is updated to allow missing checksums in some places, and to make that easier, `has_checksum` and `verify_checksum_after_work` are eliminated. Empty `checksum_hex` indicates checksum is unknown. I'm not too afraid of regressing on data integrity, because (a) we have pretty good test coverage of corruption detection in backups, and (b) we are increasingly relying on the DB itself for data integrity rather than it being an exclusive feature of backups. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8069 Test Plan: new unit tests, added to crash test (some local run with boosted backup probability) Reviewed By: ajkr Differential Revision: D27139824 Pulled By: pdillinger fbshipit-source-id: 9e0e4decfb42bb84783d64d2d246456d97e8e8c5
4 years ago
}
} // namespace ROCKSDB_NAMESPACE
#endif // ROCKSDB_LITE