|
|
|
// 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).
|
|
|
|
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
#include <cstring>
|
|
|
|
#include <iomanip>
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
#include <memory>
|
|
|
|
#include <sstream>
|
|
|
|
#include <string>
|
|
|
|
|
|
|
|
#include "db/version_edit.h"
|
|
|
|
#include "db/version_set.h"
|
|
|
|
#include "rocksdb/advanced_options.h"
|
|
|
|
#include "table/unique_id_impl.h"
|
|
|
|
#include "test_util/testharness.h"
|
|
|
|
#include "test_util/testutil.h"
|
|
|
|
#include "util/string_util.h"
|
|
|
|
|
|
|
|
namespace ROCKSDB_NAMESPACE {
|
|
|
|
|
|
|
|
class VersionBuilderTest : public testing::Test {
|
|
|
|
public:
|
|
|
|
const Comparator* ucmp_;
|
|
|
|
InternalKeyComparator icmp_;
|
|
|
|
Options options_;
|
|
|
|
ImmutableOptions ioptions_;
|
|
|
|
MutableCFOptions mutable_cf_options_;
|
|
|
|
VersionStorageInfo vstorage_;
|
|
|
|
uint32_t file_num_;
|
|
|
|
CompactionOptionsFIFO fifo_options_;
|
|
|
|
std::vector<uint64_t> size_being_compacted_;
|
|
|
|
|
|
|
|
VersionBuilderTest()
|
|
|
|
: ucmp_(BytewiseComparator()),
|
|
|
|
icmp_(ucmp_),
|
|
|
|
ioptions_(options_),
|
|
|
|
mutable_cf_options_(options_),
|
|
|
|
vstorage_(&icmp_, ucmp_, options_.num_levels, kCompactionStyleLevel,
|
|
|
|
nullptr, false),
|
|
|
|
file_num_(1) {
|
|
|
|
mutable_cf_options_.RefreshDerivedOptions(ioptions_);
|
|
|
|
size_being_compacted_.resize(options_.num_levels);
|
|
|
|
}
|
|
|
|
|
|
|
|
~VersionBuilderTest() override {
|
|
|
|
for (int i = 0; i < vstorage_.num_levels(); i++) {
|
|
|
|
for (auto* f : vstorage_.LevelFiles(i)) {
|
|
|
|
if (--f->refs == 0) {
|
|
|
|
delete f;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
InternalKey GetInternalKey(const char* ukey,
|
|
|
|
SequenceNumber smallest_seq = 100) {
|
|
|
|
return InternalKey(ukey, smallest_seq, kTypeValue);
|
|
|
|
}
|
|
|
|
|
|
|
|
void Add(int level, uint64_t file_number, const char* smallest,
|
|
|
|
const char* largest, uint64_t file_size = 0, uint32_t path_id = 0,
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
SequenceNumber smallest_seq = 100, SequenceNumber largest_seq = 100,
|
|
|
|
uint64_t num_entries = 0, uint64_t num_deletions = 0,
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
bool sampled = false, SequenceNumber smallest_seqno = 0,
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
SequenceNumber largest_seqno = 0,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
uint64_t oldest_blob_file_number = kInvalidBlobFileNumber,
|
|
|
|
uint64_t epoch_number = kUnknownEpochNumber) {
|
|
|
|
assert(level < vstorage_.num_levels());
|
|
|
|
FileMetaData* f = new FileMetaData(
|
|
|
|
file_number, path_id, file_size, GetInternalKey(smallest, smallest_seq),
|
|
|
|
GetInternalKey(largest, largest_seq), smallest_seqno, largest_seqno,
|
|
|
|
/* marked_for_compact */ false, Temperature::kUnknown,
|
|
|
|
oldest_blob_file_number, kUnknownOldestAncesterTime,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownFileCreationTime, epoch_number, kUnknownFileChecksum,
|
|
|
|
kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
|
|
|
f->compensated_file_size = file_size;
|
|
|
|
f->num_entries = num_entries;
|
|
|
|
f->num_deletions = num_deletions;
|
|
|
|
vstorage_.AddFile(level, f);
|
|
|
|
if (sampled) {
|
|
|
|
f->init_stats_from_file = true;
|
|
|
|
vstorage_.UpdateAccumulatedStats(f);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
void AddBlob(uint64_t blob_file_number, uint64_t total_blob_count,
|
|
|
|
uint64_t total_blob_bytes, std::string checksum_method,
|
|
|
|
std::string checksum_value,
|
|
|
|
BlobFileMetaData::LinkedSsts linked_ssts,
|
|
|
|
uint64_t garbage_blob_count, uint64_t garbage_blob_bytes) {
|
|
|
|
auto shared_meta = SharedBlobFileMetaData::Create(
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
blob_file_number, total_blob_count, total_blob_bytes,
|
|
|
|
std::move(checksum_method), std::move(checksum_value));
|
|
|
|
auto meta =
|
|
|
|
BlobFileMetaData::Create(std::move(shared_meta), std::move(linked_ssts),
|
|
|
|
garbage_blob_count, garbage_blob_bytes);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
|
|
|
vstorage_.AddBlobFile(std::move(meta));
|
|
|
|
}
|
|
|
|
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
void AddDummyFile(uint64_t table_file_number, uint64_t blob_file_number,
|
|
|
|
uint64_t epoch_number) {
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
constexpr int level = 0;
|
|
|
|
constexpr char smallest[] = "bar";
|
|
|
|
constexpr char largest[] = "foo";
|
|
|
|
constexpr uint64_t file_size = 100;
|
|
|
|
constexpr uint32_t path_id = 0;
|
|
|
|
constexpr SequenceNumber smallest_seq = 0;
|
|
|
|
constexpr SequenceNumber largest_seq = 0;
|
|
|
|
constexpr uint64_t num_entries = 0;
|
|
|
|
constexpr uint64_t num_deletions = 0;
|
|
|
|
constexpr bool sampled = false;
|
|
|
|
|
|
|
|
Add(level, table_file_number, smallest, largest, file_size, path_id,
|
|
|
|
smallest_seq, largest_seq, num_entries, num_deletions, sampled,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
smallest_seq, largest_seq, blob_file_number, epoch_number);
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
}
|
|
|
|
|
|
|
|
void AddDummyFileToEdit(VersionEdit* edit, uint64_t table_file_number,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
uint64_t blob_file_number, uint64_t epoch_number) {
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
assert(edit);
|
|
|
|
|
|
|
|
constexpr int level = 0;
|
|
|
|
constexpr uint32_t path_id = 0;
|
|
|
|
constexpr uint64_t file_size = 100;
|
|
|
|
constexpr char smallest[] = "bar";
|
|
|
|
constexpr char largest[] = "foo";
|
|
|
|
constexpr SequenceNumber smallest_seqno = 100;
|
|
|
|
constexpr SequenceNumber largest_seqno = 300;
|
|
|
|
constexpr bool marked_for_compaction = false;
|
|
|
|
|
|
|
|
edit->AddFile(level, table_file_number, path_id, file_size,
|
|
|
|
GetInternalKey(smallest), GetInternalKey(largest),
|
|
|
|
smallest_seqno, largest_seqno, marked_for_compaction,
|
|
|
|
Temperature::kUnknown, blob_file_number,
|
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime,
|
|
|
|
epoch_number, kUnknownFileChecksum,
|
|
|
|
kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
}
|
|
|
|
|
|
|
|
void UpdateVersionStorageInfo(VersionStorageInfo* vstorage) {
|
|
|
|
assert(vstorage);
|
|
|
|
|
|
|
|
vstorage->PrepareForVersionAppend(ioptions_, mutable_cf_options_);
|
|
|
|
vstorage->SetFinalized();
|
|
|
|
}
|
|
|
|
|
|
|
|
void UpdateVersionStorageInfo() { UpdateVersionStorageInfo(&vstorage_); }
|
|
|
|
};
|
|
|
|
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
void UnrefFilesInVersion(VersionStorageInfo* new_vstorage) {
|
|
|
|
for (int i = 0; i < new_vstorage->num_levels(); i++) {
|
|
|
|
for (auto* f : new_vstorage->LevelFiles(i)) {
|
|
|
|
if (--f->refs == 0) {
|
|
|
|
delete f;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(VersionBuilderTest, ApplyAndSaveTo) {
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
Add(0, 1U, "150", "200", 100U, /*path_id*/ 0,
|
|
|
|
/*smallest_seq*/ 100, /*largest_seq*/ 100,
|
|
|
|
/*num_entries*/ 0, /*num_deletions*/ 0,
|
|
|
|
/*sampled*/ false, /*smallest_seqno*/ 0,
|
|
|
|
/*largest_seqno*/ 0,
|
|
|
|
/*oldest_blob_file_number*/ kInvalidBlobFileNumber,
|
|
|
|
/*epoch_number*/ 1);
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
|
|
|
|
Add(1, 66U, "150", "200", 100U);
|
|
|
|
Add(1, 88U, "201", "300", 100U);
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
|
|
|
|
Add(2, 6U, "150", "179", 100U);
|
|
|
|
Add(2, 7U, "180", "220", 100U);
|
|
|
|
Add(2, 8U, "221", "300", 100U);
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
|
|
|
|
Add(3, 26U, "150", "170", 100U);
|
|
|
|
Add(3, 27U, "171", "179", 100U);
|
|
|
|
Add(3, 28U, "191", "220", 100U);
|
|
|
|
Add(3, 29U, "221", "300", 100U);
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
|
|
|
VersionEdit version_edit;
|
|
|
|
version_edit.AddFile(
|
|
|
|
2, 666, 0, 100U, GetInternalKey("301"), GetInternalKey("350"), 200, 200,
|
|
|
|
false, Temperature::kUnknown, kInvalidBlobFileNumber,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime, kUnknownEpochNumber,
|
|
|
|
kUnknownFileChecksum, kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
|
|
|
version_edit.DeleteFile(3, 27U);
|
|
|
|
|
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
|
|
|
|
|
|
|
VersionBuilder version_builder(env_options, &ioptions_, table_cache,
|
|
|
|
&vstorage_, version_set);
|
|
|
|
|
|
|
|
VersionStorageInfo new_vstorage(&icmp_, ucmp_, options_.num_levels,
|
|
|
|
kCompactionStyleLevel, nullptr, false);
|
|
|
|
ASSERT_OK(version_builder.Apply(&version_edit));
|
|
|
|
ASSERT_OK(version_builder.SaveTo(&new_vstorage));
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo(&new_vstorage);
|
|
|
|
|
|
|
|
ASSERT_EQ(400U, new_vstorage.NumLevelBytes(2));
|
|
|
|
ASSERT_EQ(300U, new_vstorage.NumLevelBytes(3));
|
|
|
|
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
UnrefFilesInVersion(&new_vstorage);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(VersionBuilderTest, ApplyAndSaveToDynamic) {
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
ioptions_.level_compaction_dynamic_level_bytes = true;
|
|
|
|
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
Add(0, 1U, "150", "200", 100U, 0, 200U, 200U, 0, 0, false, 200U, 200U,
|
|
|
|
/*oldest_blob_file_number*/ kInvalidBlobFileNumber,
|
|
|
|
/*epoch_number*/ 2);
|
|
|
|
Add(0, 88U, "201", "300", 100U, 0, 100U, 100U, 0, 0, false, 100U, 100U,
|
|
|
|
/*oldest_blob_file_number*/ kInvalidBlobFileNumber,
|
|
|
|
/*epoch_number*/ 1);
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
|
|
|
|
Add(4, 6U, "150", "179", 100U);
|
|
|
|
Add(4, 7U, "180", "220", 100U);
|
|
|
|
Add(4, 8U, "221", "300", 100U);
|
|
|
|
|
|
|
|
Add(5, 26U, "150", "170", 100U);
|
|
|
|
Add(5, 27U, "171", "179", 100U);
|
|
|
|
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
|
|
|
VersionEdit version_edit;
|
|
|
|
version_edit.AddFile(
|
|
|
|
3, 666, 0, 100U, GetInternalKey("301"), GetInternalKey("350"), 200, 200,
|
|
|
|
false, Temperature::kUnknown, kInvalidBlobFileNumber,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime, kUnknownEpochNumber,
|
|
|
|
kUnknownFileChecksum, kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
|
|
|
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
version_edit.DeleteFile(0, 1U);
|
|
|
|
version_edit.DeleteFile(0, 88U);
|
|
|
|
|
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
|
|
|
|
VersionBuilder version_builder(env_options, &ioptions_, table_cache,
|
|
|
|
&vstorage_, version_set);
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
|
|
|
|
VersionStorageInfo new_vstorage(&icmp_, ucmp_, options_.num_levels,
|
|
|
|
kCompactionStyleLevel, nullptr, false);
|
|
|
|
ASSERT_OK(version_builder.Apply(&version_edit));
|
|
|
|
ASSERT_OK(version_builder.SaveTo(&new_vstorage));
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
|
|
|
|
UpdateVersionStorageInfo(&new_vstorage);
|
|
|
|
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
ASSERT_EQ(0U, new_vstorage.NumLevelBytes(0));
|
|
|
|
ASSERT_EQ(100U, new_vstorage.NumLevelBytes(3));
|
|
|
|
ASSERT_EQ(300U, new_vstorage.NumLevelBytes(4));
|
|
|
|
ASSERT_EQ(200U, new_vstorage.NumLevelBytes(5));
|
|
|
|
|
|
|
|
UnrefFilesInVersion(&new_vstorage);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(VersionBuilderTest, ApplyAndSaveToDynamic2) {
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
ioptions_.level_compaction_dynamic_level_bytes = true;
|
|
|
|
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
Add(0, 1U, "150", "200", 100U, 0, 200U, 200U, 0, 0, false, 200U, 200U,
|
|
|
|
/*oldest_blob_file_number*/ kInvalidBlobFileNumber,
|
|
|
|
/*epoch_number*/ 2);
|
|
|
|
Add(0, 88U, "201", "300", 100U, 0, 100U, 100U, 0, 0, false, 100U, 100U,
|
|
|
|
/*oldest_blob_file_number*/ kInvalidBlobFileNumber,
|
|
|
|
/*epoch_number*/ 1);
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
|
|
|
|
Add(4, 6U, "150", "179", 100U);
|
|
|
|
Add(4, 7U, "180", "220", 100U);
|
|
|
|
Add(4, 8U, "221", "300", 100U);
|
|
|
|
|
|
|
|
Add(5, 26U, "150", "170", 100U);
|
|
|
|
Add(5, 27U, "171", "179", 100U);
|
|
|
|
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
|
|
|
VersionEdit version_edit;
|
|
|
|
version_edit.AddFile(
|
|
|
|
4, 666, 0, 100U, GetInternalKey("301"), GetInternalKey("350"), 200, 200,
|
|
|
|
false, Temperature::kUnknown, kInvalidBlobFileNumber,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime, kUnknownEpochNumber,
|
|
|
|
kUnknownFileChecksum, kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
version_edit.DeleteFile(0, 1U);
|
|
|
|
version_edit.DeleteFile(0, 88U);
|
|
|
|
version_edit.DeleteFile(4, 6U);
|
|
|
|
version_edit.DeleteFile(4, 7U);
|
|
|
|
version_edit.DeleteFile(4, 8U);
|
|
|
|
|
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
|
|
|
|
VersionBuilder version_builder(env_options, &ioptions_, table_cache,
|
|
|
|
&vstorage_, version_set);
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
|
|
|
|
VersionStorageInfo new_vstorage(&icmp_, ucmp_, options_.num_levels,
|
|
|
|
kCompactionStyleLevel, nullptr, false);
|
|
|
|
ASSERT_OK(version_builder.Apply(&version_edit));
|
|
|
|
ASSERT_OK(version_builder.SaveTo(&new_vstorage));
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
|
|
|
|
UpdateVersionStorageInfo(&new_vstorage);
|
|
|
|
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
ASSERT_EQ(0U, new_vstorage.NumLevelBytes(0));
|
|
|
|
ASSERT_EQ(100U, new_vstorage.NumLevelBytes(4));
|
|
|
|
ASSERT_EQ(200U, new_vstorage.NumLevelBytes(5));
|
|
|
|
|
|
|
|
UnrefFilesInVersion(&new_vstorage);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(VersionBuilderTest, ApplyMultipleAndSaveTo) {
|
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
|
|
|
VersionEdit version_edit;
|
|
|
|
version_edit.AddFile(
|
|
|
|
2, 666, 0, 100U, GetInternalKey("301"), GetInternalKey("350"), 200, 200,
|
|
|
|
false, Temperature::kUnknown, kInvalidBlobFileNumber,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime, kUnknownEpochNumber,
|
|
|
|
kUnknownFileChecksum, kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
|
|
|
version_edit.AddFile(
|
|
|
|
2, 676, 0, 100U, GetInternalKey("401"), GetInternalKey("450"), 200, 200,
|
|
|
|
false, Temperature::kUnknown, kInvalidBlobFileNumber,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime, kUnknownEpochNumber,
|
|
|
|
kUnknownFileChecksum, kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
|
|
|
version_edit.AddFile(
|
|
|
|
2, 636, 0, 100U, GetInternalKey("601"), GetInternalKey("650"), 200, 200,
|
|
|
|
false, Temperature::kUnknown, kInvalidBlobFileNumber,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime, kUnknownEpochNumber,
|
|
|
|
kUnknownFileChecksum, kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
|
|
|
version_edit.AddFile(
|
|
|
|
2, 616, 0, 100U, GetInternalKey("501"), GetInternalKey("550"), 200, 200,
|
|
|
|
false, Temperature::kUnknown, kInvalidBlobFileNumber,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime, kUnknownEpochNumber,
|
|
|
|
kUnknownFileChecksum, kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
|
|
|
version_edit.AddFile(
|
|
|
|
2, 606, 0, 100U, GetInternalKey("701"), GetInternalKey("750"), 200, 200,
|
|
|
|
false, Temperature::kUnknown, kInvalidBlobFileNumber,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime, kUnknownEpochNumber,
|
|
|
|
kUnknownFileChecksum, kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
|
|
|
|
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
|
|
|
|
|
|
|
VersionBuilder version_builder(env_options, &ioptions_, table_cache,
|
|
|
|
&vstorage_, version_set);
|
|
|
|
|
|
|
|
VersionStorageInfo new_vstorage(&icmp_, ucmp_, options_.num_levels,
|
|
|
|
kCompactionStyleLevel, nullptr, false);
|
|
|
|
ASSERT_OK(version_builder.Apply(&version_edit));
|
|
|
|
ASSERT_OK(version_builder.SaveTo(&new_vstorage));
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo(&new_vstorage);
|
|
|
|
|
|
|
|
ASSERT_EQ(500U, new_vstorage.NumLevelBytes(2));
|
|
|
|
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
UnrefFilesInVersion(&new_vstorage);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(VersionBuilderTest, ApplyDeleteAndSaveTo) {
|
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
|
|
|
|
|
|
|
VersionBuilder version_builder(env_options, &ioptions_, table_cache,
|
|
|
|
&vstorage_, version_set);
|
|
|
|
|
|
|
|
VersionStorageInfo new_vstorage(&icmp_, ucmp_, options_.num_levels,
|
|
|
|
kCompactionStyleLevel, nullptr, false);
|
|
|
|
|
|
|
|
VersionEdit version_edit;
|
|
|
|
version_edit.AddFile(
|
|
|
|
2, 666, 0, 100U, GetInternalKey("301"), GetInternalKey("350"), 200, 200,
|
|
|
|
false, Temperature::kUnknown, kInvalidBlobFileNumber,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime, kUnknownEpochNumber,
|
|
|
|
kUnknownFileChecksum, kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
|
|
|
version_edit.AddFile(
|
|
|
|
2, 676, 0, 100U, GetInternalKey("401"), GetInternalKey("450"), 200, 200,
|
|
|
|
false, Temperature::kUnknown, kInvalidBlobFileNumber,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime, kUnknownEpochNumber,
|
|
|
|
kUnknownFileChecksum, kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
|
|
|
version_edit.AddFile(
|
|
|
|
2, 636, 0, 100U, GetInternalKey("601"), GetInternalKey("650"), 200, 200,
|
|
|
|
false, Temperature::kUnknown, kInvalidBlobFileNumber,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime, kUnknownEpochNumber,
|
|
|
|
kUnknownFileChecksum, kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
|
|
|
version_edit.AddFile(
|
|
|
|
2, 616, 0, 100U, GetInternalKey("501"), GetInternalKey("550"), 200, 200,
|
|
|
|
false, Temperature::kUnknown, kInvalidBlobFileNumber,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime, kUnknownEpochNumber,
|
|
|
|
kUnknownFileChecksum, kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
|
|
|
version_edit.AddFile(
|
|
|
|
2, 606, 0, 100U, GetInternalKey("701"), GetInternalKey("750"), 200, 200,
|
|
|
|
false, Temperature::kUnknown, kInvalidBlobFileNumber,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime, kUnknownEpochNumber,
|
|
|
|
kUnknownFileChecksum, kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
|
|
|
ASSERT_OK(version_builder.Apply(&version_edit));
|
|
|
|
|
|
|
|
VersionEdit version_edit2;
|
|
|
|
version_edit.AddFile(
|
|
|
|
2, 808, 0, 100U, GetInternalKey("901"), GetInternalKey("950"), 200, 200,
|
|
|
|
false, Temperature::kUnknown, kInvalidBlobFileNumber,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime, kUnknownEpochNumber,
|
|
|
|
kUnknownFileChecksum, kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
|
|
|
version_edit2.DeleteFile(2, 616);
|
|
|
|
version_edit2.DeleteFile(2, 636);
|
|
|
|
version_edit.AddFile(
|
|
|
|
2, 806, 0, 100U, GetInternalKey("801"), GetInternalKey("850"), 200, 200,
|
|
|
|
false, Temperature::kUnknown, kInvalidBlobFileNumber,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime, kUnknownEpochNumber,
|
|
|
|
kUnknownFileChecksum, kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
|
|
|
|
|
|
|
ASSERT_OK(version_builder.Apply(&version_edit2));
|
|
|
|
ASSERT_OK(version_builder.SaveTo(&new_vstorage));
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo(&new_vstorage);
|
|
|
|
|
|
|
|
ASSERT_EQ(300U, new_vstorage.NumLevelBytes(2));
|
|
|
|
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
|
|
|
UnrefFilesInVersion(&new_vstorage);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(VersionBuilderTest, ApplyFileDeletionIncorrectLevel) {
|
|
|
|
constexpr int level = 1;
|
|
|
|
constexpr uint64_t file_number = 2345;
|
|
|
|
constexpr char smallest[] = "bar";
|
|
|
|
constexpr char largest[] = "foo";
|
|
|
|
constexpr uint64_t file_size = 100;
|
|
|
|
|
|
|
|
Add(level, file_number, smallest, largest, file_size);
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
|
|
|
|
|
|
|
VersionBuilder builder(env_options, &ioptions_, table_cache, &vstorage_,
|
|
|
|
version_set);
|
|
|
|
|
|
|
|
VersionEdit edit;
|
|
|
|
|
|
|
|
constexpr int incorrect_level = 3;
|
|
|
|
|
|
|
|
edit.DeleteFile(incorrect_level, file_number);
|
|
|
|
|
|
|
|
const Status s = builder.Apply(&edit);
|
|
|
|
ASSERT_TRUE(s.IsCorruption());
|
|
|
|
ASSERT_TRUE(std::strstr(s.getState(),
|
|
|
|
"Cannot delete table file #2345 from level 3 since "
|
|
|
|
"it is on level 1"));
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(VersionBuilderTest, ApplyFileDeletionNotInLSMTree) {
|
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
|
|
|
|
|
|
|
VersionBuilder builder(env_options, &ioptions_, table_cache, &vstorage_,
|
|
|
|
version_set);
|
|
|
|
|
|
|
|
VersionEdit edit;
|
|
|
|
|
|
|
|
constexpr int level = 3;
|
|
|
|
constexpr uint64_t file_number = 1234;
|
|
|
|
|
|
|
|
edit.DeleteFile(level, file_number);
|
|
|
|
|
|
|
|
const Status s = builder.Apply(&edit);
|
|
|
|
ASSERT_TRUE(s.IsCorruption());
|
|
|
|
ASSERT_TRUE(std::strstr(s.getState(),
|
|
|
|
"Cannot delete table file #1234 from level 3 since "
|
|
|
|
"it is not in the LSM tree"));
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(VersionBuilderTest, ApplyFileDeletionAndAddition) {
|
|
|
|
constexpr int level = 1;
|
|
|
|
constexpr uint64_t file_number = 2345;
|
|
|
|
constexpr char smallest[] = "bar";
|
|
|
|
constexpr char largest[] = "foo";
|
|
|
|
constexpr uint64_t file_size = 10000;
|
|
|
|
constexpr uint32_t path_id = 0;
|
|
|
|
constexpr SequenceNumber smallest_seq = 100;
|
|
|
|
constexpr SequenceNumber largest_seq = 500;
|
|
|
|
constexpr uint64_t num_entries = 0;
|
|
|
|
constexpr uint64_t num_deletions = 0;
|
|
|
|
constexpr bool sampled = false;
|
|
|
|
constexpr SequenceNumber smallest_seqno = 1;
|
|
|
|
constexpr SequenceNumber largest_seqno = 1000;
|
|
|
|
|
|
|
|
Add(level, file_number, smallest, largest, file_size, path_id, smallest_seq,
|
|
|
|
largest_seq, num_entries, num_deletions, sampled, smallest_seqno,
|
|
|
|
largest_seqno);
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
|
|
|
|
|
|
|
VersionBuilder builder(env_options, &ioptions_, table_cache, &vstorage_,
|
|
|
|
version_set);
|
|
|
|
|
|
|
|
VersionEdit deletion;
|
|
|
|
|
|
|
|
deletion.DeleteFile(level, file_number);
|
|
|
|
|
|
|
|
ASSERT_OK(builder.Apply(&deletion));
|
|
|
|
|
|
|
|
VersionEdit addition;
|
|
|
|
|
|
|
|
constexpr bool marked_for_compaction = false;
|
|
|
|
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
addition.AddFile(
|
|
|
|
level, file_number, path_id, file_size,
|
|
|
|
GetInternalKey(smallest, smallest_seq),
|
|
|
|
GetInternalKey(largest, largest_seq), smallest_seqno, largest_seqno,
|
|
|
|
marked_for_compaction, Temperature::kUnknown, kInvalidBlobFileNumber,
|
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime, kUnknownEpochNumber,
|
|
|
|
kUnknownFileChecksum, kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
|
|
|
|
|
|
|
ASSERT_OK(builder.Apply(&addition));
|
|
|
|
|
|
|
|
constexpr bool force_consistency_checks = false;
|
|
|
|
VersionStorageInfo new_vstorage(&icmp_, ucmp_, options_.num_levels,
|
|
|
|
kCompactionStyleLevel, &vstorage_,
|
|
|
|
force_consistency_checks);
|
|
|
|
|
|
|
|
ASSERT_OK(builder.SaveTo(&new_vstorage));
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo(&new_vstorage);
|
|
|
|
|
|
|
|
ASSERT_EQ(new_vstorage.GetFileLocation(file_number).GetLevel(), level);
|
|
|
|
|
|
|
|
UnrefFilesInVersion(&new_vstorage);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(VersionBuilderTest, ApplyFileAdditionAlreadyInBase) {
|
|
|
|
constexpr int level = 1;
|
|
|
|
constexpr uint64_t file_number = 2345;
|
|
|
|
constexpr char smallest[] = "bar";
|
|
|
|
constexpr char largest[] = "foo";
|
|
|
|
constexpr uint64_t file_size = 10000;
|
|
|
|
|
|
|
|
Add(level, file_number, smallest, largest, file_size);
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
|
|
|
|
|
|
|
VersionBuilder builder(env_options, &ioptions_, table_cache, &vstorage_,
|
|
|
|
version_set);
|
|
|
|
|
|
|
|
VersionEdit edit;
|
|
|
|
|
|
|
|
constexpr int new_level = 2;
|
|
|
|
constexpr uint32_t path_id = 0;
|
|
|
|
constexpr SequenceNumber smallest_seqno = 100;
|
|
|
|
constexpr SequenceNumber largest_seqno = 1000;
|
|
|
|
constexpr bool marked_for_compaction = false;
|
|
|
|
|
|
|
|
edit.AddFile(
|
|
|
|
new_level, file_number, path_id, file_size, GetInternalKey(smallest),
|
|
|
|
GetInternalKey(largest), smallest_seqno, largest_seqno,
|
|
|
|
marked_for_compaction, Temperature::kUnknown, kInvalidBlobFileNumber,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime, kUnknownEpochNumber,
|
|
|
|
kUnknownFileChecksum, kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
|
|
|
|
|
|
|
const Status s = builder.Apply(&edit);
|
|
|
|
ASSERT_TRUE(s.IsCorruption());
|
|
|
|
ASSERT_TRUE(std::strstr(s.getState(),
|
|
|
|
"Cannot add table file #2345 to level 2 since it is "
|
|
|
|
"already in the LSM tree on level 1"));
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(VersionBuilderTest, ApplyFileAdditionAlreadyApplied) {
|
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
|
|
|
|
|
|
|
VersionBuilder builder(env_options, &ioptions_, table_cache, &vstorage_,
|
|
|
|
version_set);
|
|
|
|
|
|
|
|
VersionEdit edit;
|
|
|
|
|
|
|
|
constexpr int level = 3;
|
|
|
|
constexpr uint64_t file_number = 2345;
|
|
|
|
constexpr uint32_t path_id = 0;
|
|
|
|
constexpr uint64_t file_size = 10000;
|
|
|
|
constexpr char smallest[] = "bar";
|
|
|
|
constexpr char largest[] = "foo";
|
|
|
|
constexpr SequenceNumber smallest_seqno = 100;
|
|
|
|
constexpr SequenceNumber largest_seqno = 1000;
|
|
|
|
constexpr bool marked_for_compaction = false;
|
|
|
|
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
edit.AddFile(
|
|
|
|
level, file_number, path_id, file_size, GetInternalKey(smallest),
|
|
|
|
GetInternalKey(largest), smallest_seqno, largest_seqno,
|
|
|
|
marked_for_compaction, Temperature::kUnknown, kInvalidBlobFileNumber,
|
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime, kUnknownEpochNumber,
|
|
|
|
kUnknownFileChecksum, kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
|
|
|
|
|
|
|
ASSERT_OK(builder.Apply(&edit));
|
|
|
|
|
|
|
|
VersionEdit other_edit;
|
|
|
|
|
|
|
|
constexpr int new_level = 2;
|
|
|
|
|
|
|
|
other_edit.AddFile(
|
|
|
|
new_level, file_number, path_id, file_size, GetInternalKey(smallest),
|
|
|
|
GetInternalKey(largest), smallest_seqno, largest_seqno,
|
|
|
|
marked_for_compaction, Temperature::kUnknown, kInvalidBlobFileNumber,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime, kUnknownEpochNumber,
|
|
|
|
kUnknownFileChecksum, kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
|
|
|
|
|
|
|
const Status s = builder.Apply(&other_edit);
|
|
|
|
ASSERT_TRUE(s.IsCorruption());
|
|
|
|
ASSERT_TRUE(std::strstr(s.getState(),
|
|
|
|
"Cannot add table file #2345 to level 2 since it is "
|
|
|
|
"already in the LSM tree on level 3"));
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(VersionBuilderTest, ApplyFileAdditionAndDeletion) {
|
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
|
|
|
constexpr int level = 1;
|
|
|
|
constexpr uint64_t file_number = 2345;
|
|
|
|
constexpr uint32_t path_id = 0;
|
|
|
|
constexpr uint64_t file_size = 10000;
|
|
|
|
constexpr char smallest[] = "bar";
|
|
|
|
constexpr char largest[] = "foo";
|
|
|
|
constexpr SequenceNumber smallest_seqno = 100;
|
|
|
|
constexpr SequenceNumber largest_seqno = 1000;
|
|
|
|
constexpr bool marked_for_compaction = false;
|
|
|
|
|
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
|
|
|
|
|
|
|
VersionBuilder builder(env_options, &ioptions_, table_cache, &vstorage_,
|
|
|
|
version_set);
|
|
|
|
|
|
|
|
VersionEdit addition;
|
|
|
|
|
|
|
|
addition.AddFile(
|
|
|
|
level, file_number, path_id, file_size, GetInternalKey(smallest),
|
|
|
|
GetInternalKey(largest), smallest_seqno, largest_seqno,
|
|
|
|
marked_for_compaction, Temperature::kUnknown, kInvalidBlobFileNumber,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime, kUnknownEpochNumber,
|
|
|
|
kUnknownFileChecksum, kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
|
|
|
|
|
|
|
ASSERT_OK(builder.Apply(&addition));
|
|
|
|
|
|
|
|
VersionEdit deletion;
|
|
|
|
|
|
|
|
deletion.DeleteFile(level, file_number);
|
|
|
|
|
|
|
|
ASSERT_OK(builder.Apply(&deletion));
|
|
|
|
|
|
|
|
constexpr bool force_consistency_checks = false;
|
|
|
|
VersionStorageInfo new_vstorage(&icmp_, ucmp_, options_.num_levels,
|
|
|
|
kCompactionStyleLevel, &vstorage_,
|
|
|
|
force_consistency_checks);
|
|
|
|
|
|
|
|
ASSERT_OK(builder.SaveTo(&new_vstorage));
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo(&new_vstorage);
|
|
|
|
|
|
|
|
ASSERT_FALSE(new_vstorage.GetFileLocation(file_number).IsValid());
|
|
|
|
|
|
|
|
UnrefFilesInVersion(&new_vstorage);
|
|
|
|
}
|
|
|
|
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
TEST_F(VersionBuilderTest, ApplyBlobFileAddition) {
|
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
|
|
|
|
|
|
|
VersionBuilder builder(env_options, &ioptions_, table_cache, &vstorage_,
|
|
|
|
version_set);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
|
|
|
VersionEdit edit;
|
|
|
|
|
|
|
|
constexpr uint64_t blob_file_number = 1234;
|
|
|
|
constexpr uint64_t total_blob_count = 5678;
|
|
|
|
constexpr uint64_t total_blob_bytes = 999999;
|
|
|
|
constexpr char checksum_method[] = "SHA1";
|
|
|
|
constexpr char checksum_value[] =
|
|
|
|
"\xbd\xb7\xf3\x4a\x59\xdf\xa1\x59\x2c\xe7\xf5\x2e\x99\xf9\x8c\x57\x0c\x52"
|
|
|
|
"\x5c\xbd";
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
|
|
|
edit.AddBlobFile(blob_file_number, total_blob_count, total_blob_bytes,
|
|
|
|
checksum_method, checksum_value);
|
|
|
|
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
// Add dummy table file to ensure the blob file is referenced.
|
|
|
|
constexpr uint64_t table_file_number = 1;
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
AddDummyFileToEdit(&edit, table_file_number, blob_file_number,
|
|
|
|
1 /*epoch_number*/);
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
ASSERT_OK(builder.Apply(&edit));
|
|
|
|
|
|
|
|
constexpr bool force_consistency_checks = false;
|
|
|
|
VersionStorageInfo new_vstorage(&icmp_, ucmp_, options_.num_levels,
|
|
|
|
kCompactionStyleLevel, &vstorage_,
|
|
|
|
force_consistency_checks);
|
|
|
|
|
|
|
|
ASSERT_OK(builder.SaveTo(&new_vstorage));
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo(&new_vstorage);
|
|
|
|
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
const auto& new_blob_files = new_vstorage.GetBlobFiles();
|
|
|
|
ASSERT_EQ(new_blob_files.size(), 1);
|
|
|
|
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
const auto new_meta = new_vstorage.GetBlobFileMetaData(blob_file_number);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
|
|
|
ASSERT_NE(new_meta, nullptr);
|
|
|
|
ASSERT_EQ(new_meta->GetBlobFileNumber(), blob_file_number);
|
|
|
|
ASSERT_EQ(new_meta->GetTotalBlobCount(), total_blob_count);
|
|
|
|
ASSERT_EQ(new_meta->GetTotalBlobBytes(), total_blob_bytes);
|
|
|
|
ASSERT_EQ(new_meta->GetChecksumMethod(), checksum_method);
|
|
|
|
ASSERT_EQ(new_meta->GetChecksumValue(), checksum_value);
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
ASSERT_EQ(new_meta->GetLinkedSsts(),
|
|
|
|
BlobFileMetaData::LinkedSsts{table_file_number});
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
ASSERT_EQ(new_meta->GetGarbageBlobCount(), 0);
|
|
|
|
ASSERT_EQ(new_meta->GetGarbageBlobBytes(), 0);
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
|
|
|
|
UnrefFilesInVersion(&new_vstorage);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(VersionBuilderTest, ApplyBlobFileAdditionAlreadyInBase) {
|
|
|
|
// Attempt to add a blob file that is already present in the base version.
|
|
|
|
|
|
|
|
constexpr uint64_t blob_file_number = 1234;
|
|
|
|
constexpr uint64_t total_blob_count = 5678;
|
|
|
|
constexpr uint64_t total_blob_bytes = 999999;
|
|
|
|
constexpr char checksum_method[] = "SHA1";
|
|
|
|
constexpr char checksum_value[] =
|
|
|
|
"\xbd\xb7\xf3\x4a\x59\xdf\xa1\x59\x2c\xe7\xf5\x2e\x99\xf9\x8c\x57\x0c\x52"
|
|
|
|
"\x5c\xbd";
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
constexpr uint64_t garbage_blob_count = 123;
|
|
|
|
constexpr uint64_t garbage_blob_bytes = 456789;
|
|
|
|
|
|
|
|
AddBlob(blob_file_number, total_blob_count, total_blob_bytes, checksum_method,
|
|
|
|
checksum_value, BlobFileMetaData::LinkedSsts(), garbage_blob_count,
|
|
|
|
garbage_blob_bytes);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
|
|
|
|
|
|
|
VersionBuilder builder(env_options, &ioptions_, table_cache, &vstorage_,
|
|
|
|
version_set);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
|
|
|
VersionEdit edit;
|
|
|
|
|
|
|
|
edit.AddBlobFile(blob_file_number, total_blob_count, total_blob_bytes,
|
|
|
|
checksum_method, checksum_value);
|
|
|
|
|
|
|
|
const Status s = builder.Apply(&edit);
|
|
|
|
ASSERT_TRUE(s.IsCorruption());
|
|
|
|
ASSERT_TRUE(std::strstr(s.getState(), "Blob file #1234 already added"));
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(VersionBuilderTest, ApplyBlobFileAdditionAlreadyApplied) {
|
|
|
|
// Attempt to add the same blob file twice using version edits.
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
|
|
|
|
|
|
|
VersionBuilder builder(env_options, &ioptions_, table_cache, &vstorage_,
|
|
|
|
version_set);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
|
|
|
VersionEdit edit;
|
|
|
|
|
|
|
|
constexpr uint64_t blob_file_number = 1234;
|
|
|
|
constexpr uint64_t total_blob_count = 5678;
|
|
|
|
constexpr uint64_t total_blob_bytes = 999999;
|
|
|
|
constexpr char checksum_method[] = "SHA1";
|
|
|
|
constexpr char checksum_value[] =
|
|
|
|
"\xbd\xb7\xf3\x4a\x59\xdf\xa1\x59\x2c\xe7\xf5\x2e\x99\xf9\x8c\x57\x0c\x52"
|
|
|
|
"\x5c\xbd";
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
|
|
|
edit.AddBlobFile(blob_file_number, total_blob_count, total_blob_bytes,
|
|
|
|
checksum_method, checksum_value);
|
|
|
|
|
|
|
|
ASSERT_OK(builder.Apply(&edit));
|
|
|
|
|
|
|
|
const Status s = builder.Apply(&edit);
|
|
|
|
ASSERT_TRUE(s.IsCorruption());
|
|
|
|
ASSERT_TRUE(std::strstr(s.getState(), "Blob file #1234 already added"));
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(VersionBuilderTest, ApplyBlobFileGarbageFileInBase) {
|
|
|
|
// Increase the amount of garbage for a blob file present in the base version.
|
|
|
|
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
constexpr uint64_t table_file_number = 1;
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
constexpr uint64_t blob_file_number = 1234;
|
|
|
|
constexpr uint64_t total_blob_count = 5678;
|
|
|
|
constexpr uint64_t total_blob_bytes = 999999;
|
|
|
|
constexpr char checksum_method[] = "SHA1";
|
|
|
|
constexpr char checksum_value[] =
|
|
|
|
"\xbd\xb7\xf3\x4a\x59\xdf\xa1\x59\x2c\xe7\xf5\x2e\x99\xf9\x8c\x57\x0c\x52"
|
|
|
|
"\x5c\xbd";
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
constexpr uint64_t garbage_blob_count = 123;
|
|
|
|
constexpr uint64_t garbage_blob_bytes = 456789;
|
|
|
|
|
|
|
|
AddBlob(blob_file_number, total_blob_count, total_blob_bytes, checksum_method,
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
checksum_value, BlobFileMetaData::LinkedSsts{table_file_number},
|
|
|
|
garbage_blob_count, garbage_blob_bytes);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
const auto meta = vstorage_.GetBlobFileMetaData(blob_file_number);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
ASSERT_NE(meta, nullptr);
|
|
|
|
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
// Add dummy table file to ensure the blob file is referenced.
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
AddDummyFile(table_file_number, blob_file_number, 1 /*epoch_number*/);
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
|
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
|
|
|
|
|
|
|
VersionBuilder builder(env_options, &ioptions_, table_cache, &vstorage_,
|
|
|
|
version_set);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
|
|
|
VersionEdit edit;
|
|
|
|
|
|
|
|
constexpr uint64_t new_garbage_blob_count = 456;
|
|
|
|
constexpr uint64_t new_garbage_blob_bytes = 111111;
|
|
|
|
|
|
|
|
edit.AddBlobFileGarbage(blob_file_number, new_garbage_blob_count,
|
|
|
|
new_garbage_blob_bytes);
|
|
|
|
|
|
|
|
ASSERT_OK(builder.Apply(&edit));
|
|
|
|
|
|
|
|
constexpr bool force_consistency_checks = false;
|
|
|
|
VersionStorageInfo new_vstorage(&icmp_, ucmp_, options_.num_levels,
|
|
|
|
kCompactionStyleLevel, &vstorage_,
|
|
|
|
force_consistency_checks);
|
|
|
|
|
|
|
|
ASSERT_OK(builder.SaveTo(&new_vstorage));
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo(&new_vstorage);
|
|
|
|
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
const auto& new_blob_files = new_vstorage.GetBlobFiles();
|
|
|
|
ASSERT_EQ(new_blob_files.size(), 1);
|
|
|
|
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
const auto new_meta = new_vstorage.GetBlobFileMetaData(blob_file_number);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
|
|
|
ASSERT_NE(new_meta, nullptr);
|
|
|
|
ASSERT_EQ(new_meta->GetSharedMeta(), meta->GetSharedMeta());
|
|
|
|
ASSERT_EQ(new_meta->GetBlobFileNumber(), blob_file_number);
|
|
|
|
ASSERT_EQ(new_meta->GetTotalBlobCount(), total_blob_count);
|
|
|
|
ASSERT_EQ(new_meta->GetTotalBlobBytes(), total_blob_bytes);
|
|
|
|
ASSERT_EQ(new_meta->GetChecksumMethod(), checksum_method);
|
|
|
|
ASSERT_EQ(new_meta->GetChecksumValue(), checksum_value);
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
ASSERT_EQ(new_meta->GetLinkedSsts(),
|
|
|
|
BlobFileMetaData::LinkedSsts{table_file_number});
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
ASSERT_EQ(new_meta->GetGarbageBlobCount(),
|
|
|
|
garbage_blob_count + new_garbage_blob_count);
|
|
|
|
ASSERT_EQ(new_meta->GetGarbageBlobBytes(),
|
|
|
|
garbage_blob_bytes + new_garbage_blob_bytes);
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
|
|
|
|
UnrefFilesInVersion(&new_vstorage);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(VersionBuilderTest, ApplyBlobFileGarbageFileAdditionApplied) {
|
|
|
|
// Increase the amount of garbage for a blob file added using a version edit.
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
|
|
|
|
|
|
|
VersionBuilder builder(env_options, &ioptions_, table_cache, &vstorage_,
|
|
|
|
version_set);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
|
|
|
VersionEdit addition;
|
|
|
|
|
|
|
|
constexpr uint64_t blob_file_number = 1234;
|
|
|
|
constexpr uint64_t total_blob_count = 5678;
|
|
|
|
constexpr uint64_t total_blob_bytes = 999999;
|
|
|
|
constexpr char checksum_method[] = "SHA1";
|
|
|
|
constexpr char checksum_value[] =
|
|
|
|
"\xbd\xb7\xf3\x4a\x59\xdf\xa1\x59\x2c\xe7\xf5\x2e\x99\xf9\x8c\x57\x0c\x52"
|
|
|
|
"\x5c\xbd";
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
|
|
|
addition.AddBlobFile(blob_file_number, total_blob_count, total_blob_bytes,
|
|
|
|
checksum_method, checksum_value);
|
|
|
|
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
// Add dummy table file to ensure the blob file is referenced.
|
|
|
|
constexpr uint64_t table_file_number = 1;
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
AddDummyFileToEdit(&addition, table_file_number, blob_file_number,
|
|
|
|
1 /*epoch_number*/);
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
ASSERT_OK(builder.Apply(&addition));
|
|
|
|
|
|
|
|
constexpr uint64_t garbage_blob_count = 123;
|
|
|
|
constexpr uint64_t garbage_blob_bytes = 456789;
|
|
|
|
|
|
|
|
VersionEdit garbage;
|
|
|
|
|
|
|
|
garbage.AddBlobFileGarbage(blob_file_number, garbage_blob_count,
|
|
|
|
garbage_blob_bytes);
|
|
|
|
|
|
|
|
ASSERT_OK(builder.Apply(&garbage));
|
|
|
|
|
|
|
|
constexpr bool force_consistency_checks = false;
|
|
|
|
VersionStorageInfo new_vstorage(&icmp_, ucmp_, options_.num_levels,
|
|
|
|
kCompactionStyleLevel, &vstorage_,
|
|
|
|
force_consistency_checks);
|
|
|
|
|
|
|
|
ASSERT_OK(builder.SaveTo(&new_vstorage));
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo(&new_vstorage);
|
|
|
|
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
const auto& new_blob_files = new_vstorage.GetBlobFiles();
|
|
|
|
ASSERT_EQ(new_blob_files.size(), 1);
|
|
|
|
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
const auto new_meta = new_vstorage.GetBlobFileMetaData(blob_file_number);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
|
|
|
ASSERT_NE(new_meta, nullptr);
|
|
|
|
ASSERT_EQ(new_meta->GetBlobFileNumber(), blob_file_number);
|
|
|
|
ASSERT_EQ(new_meta->GetTotalBlobCount(), total_blob_count);
|
|
|
|
ASSERT_EQ(new_meta->GetTotalBlobBytes(), total_blob_bytes);
|
|
|
|
ASSERT_EQ(new_meta->GetChecksumMethod(), checksum_method);
|
|
|
|
ASSERT_EQ(new_meta->GetChecksumValue(), checksum_value);
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
ASSERT_EQ(new_meta->GetLinkedSsts(),
|
|
|
|
BlobFileMetaData::LinkedSsts{table_file_number});
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
ASSERT_EQ(new_meta->GetGarbageBlobCount(), garbage_blob_count);
|
|
|
|
ASSERT_EQ(new_meta->GetGarbageBlobBytes(), garbage_blob_bytes);
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
|
|
|
|
UnrefFilesInVersion(&new_vstorage);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(VersionBuilderTest, ApplyBlobFileGarbageFileNotFound) {
|
|
|
|
// Attempt to increase the amount of garbage for a blob file that is
|
|
|
|
// neither in the base version, nor was it added using a version edit.
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
|
|
|
|
|
|
|
VersionBuilder builder(env_options, &ioptions_, table_cache, &vstorage_,
|
|
|
|
version_set);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
|
|
|
VersionEdit edit;
|
|
|
|
|
|
|
|
constexpr uint64_t blob_file_number = 1234;
|
|
|
|
constexpr uint64_t garbage_blob_count = 5678;
|
|
|
|
constexpr uint64_t garbage_blob_bytes = 999999;
|
|
|
|
|
|
|
|
edit.AddBlobFileGarbage(blob_file_number, garbage_blob_count,
|
|
|
|
garbage_blob_bytes);
|
|
|
|
|
|
|
|
const Status s = builder.Apply(&edit);
|
|
|
|
ASSERT_TRUE(s.IsCorruption());
|
|
|
|
ASSERT_TRUE(std::strstr(s.getState(), "Blob file #1234 not found"));
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(VersionBuilderTest, BlobFileGarbageOverflow) {
|
|
|
|
// Test that VersionEdits that would result in the count/total size of garbage
|
|
|
|
// exceeding the count/total size of all blobs are rejected.
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
|
|
|
|
|
|
|
VersionBuilder builder(env_options, &ioptions_, table_cache, &vstorage_,
|
|
|
|
version_set);
|
|
|
|
|
|
|
|
VersionEdit addition;
|
|
|
|
|
|
|
|
constexpr uint64_t blob_file_number = 1234;
|
|
|
|
constexpr uint64_t total_blob_count = 5678;
|
|
|
|
constexpr uint64_t total_blob_bytes = 999999;
|
|
|
|
constexpr char checksum_method[] = "SHA1";
|
|
|
|
constexpr char checksum_value[] =
|
|
|
|
"\xbd\xb7\xf3\x4a\x59\xdf\xa1\x59\x2c\xe7\xf5\x2e\x99\xf9\x8c\x57\x0c\x52"
|
|
|
|
"\x5c\xbd";
|
|
|
|
|
|
|
|
addition.AddBlobFile(blob_file_number, total_blob_count, total_blob_bytes,
|
|
|
|
checksum_method, checksum_value);
|
|
|
|
|
|
|
|
// Add dummy table file to ensure the blob file is referenced.
|
|
|
|
constexpr uint64_t table_file_number = 1;
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
AddDummyFileToEdit(&addition, table_file_number, blob_file_number,
|
|
|
|
1 /*epoch_number*/);
|
|
|
|
|
|
|
|
ASSERT_OK(builder.Apply(&addition));
|
|
|
|
|
|
|
|
{
|
|
|
|
// Garbage blob count overflow
|
|
|
|
constexpr uint64_t garbage_blob_count = 5679;
|
|
|
|
constexpr uint64_t garbage_blob_bytes = 999999;
|
|
|
|
|
|
|
|
VersionEdit garbage;
|
|
|
|
|
|
|
|
garbage.AddBlobFileGarbage(blob_file_number, garbage_blob_count,
|
|
|
|
garbage_blob_bytes);
|
|
|
|
|
|
|
|
const Status s = builder.Apply(&garbage);
|
|
|
|
ASSERT_TRUE(s.IsCorruption());
|
|
|
|
ASSERT_TRUE(
|
|
|
|
std::strstr(s.getState(), "Garbage overflow for blob file #1234"));
|
|
|
|
}
|
|
|
|
|
|
|
|
{
|
|
|
|
// Garbage blob bytes overflow
|
|
|
|
constexpr uint64_t garbage_blob_count = 5678;
|
|
|
|
constexpr uint64_t garbage_blob_bytes = 1000000;
|
|
|
|
|
|
|
|
VersionEdit garbage;
|
|
|
|
|
|
|
|
garbage.AddBlobFileGarbage(blob_file_number, garbage_blob_count,
|
|
|
|
garbage_blob_bytes);
|
|
|
|
|
|
|
|
const Status s = builder.Apply(&garbage);
|
|
|
|
ASSERT_TRUE(s.IsCorruption());
|
|
|
|
ASSERT_TRUE(
|
|
|
|
std::strstr(s.getState(), "Garbage overflow for blob file #1234"));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
TEST_F(VersionBuilderTest, SaveBlobFilesTo) {
|
|
|
|
// Add three blob files to base version.
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
for (uint64_t i = 1; i <= 3; ++i) {
|
|
|
|
const uint64_t table_file_number = 2 * i;
|
|
|
|
const uint64_t blob_file_number = 2 * i + 1;
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
const uint64_t total_blob_count = i * 1000;
|
|
|
|
const uint64_t total_blob_bytes = i * 1000000;
|
|
|
|
const uint64_t garbage_blob_count = i * 100;
|
|
|
|
const uint64_t garbage_blob_bytes = i * 20000;
|
|
|
|
|
|
|
|
AddBlob(blob_file_number, total_blob_count, total_blob_bytes,
|
|
|
|
/* checksum_method */ std::string(),
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
/* checksum_value */ std::string(),
|
|
|
|
BlobFileMetaData::LinkedSsts{table_file_number}, garbage_blob_count,
|
|
|
|
garbage_blob_bytes);
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
}
|
|
|
|
|
|
|
|
// Add dummy table files to ensure the blob files are referenced.
|
|
|
|
// Note: files are added to L0, so they have to be added in reverse order
|
|
|
|
// (newest first).
|
|
|
|
for (uint64_t i = 3; i >= 1; --i) {
|
|
|
|
const uint64_t table_file_number = 2 * i;
|
|
|
|
const uint64_t blob_file_number = 2 * i + 1;
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
AddDummyFile(table_file_number, blob_file_number, i /*epoch_number*/);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
}
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
|
|
|
|
|
|
|
VersionBuilder builder(env_options, &ioptions_, table_cache, &vstorage_,
|
|
|
|
version_set);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
|
|
|
VersionEdit edit;
|
|
|
|
|
|
|
|
// Add some garbage to the second and third blob files. The second blob file
|
|
|
|
// remains valid since it does not consist entirely of garbage yet. The third
|
|
|
|
// blob file is all garbage after the edit and will not be part of the new
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
// version. The corresponding dummy table file is also removed for
|
|
|
|
// consistency.
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
edit.AddBlobFileGarbage(/* blob_file_number */ 5,
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
/* garbage_blob_count */ 200,
|
|
|
|
/* garbage_blob_bytes */ 100000);
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
edit.AddBlobFileGarbage(/* blob_file_number */ 7,
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
/* garbage_blob_count */ 2700,
|
|
|
|
/* garbage_blob_bytes */ 2940000);
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
edit.DeleteFile(/* level */ 0, /* file_number */ 6);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
|
|
|
// Add a fourth blob file.
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
edit.AddBlobFile(/* blob_file_number */ 9, /* total_blob_count */ 4000,
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
/* total_blob_bytes */ 4000000,
|
|
|
|
/* checksum_method */ std::string(),
|
|
|
|
/* checksum_value */ std::string());
|
|
|
|
|
|
|
|
ASSERT_OK(builder.Apply(&edit));
|
|
|
|
|
|
|
|
constexpr bool force_consistency_checks = false;
|
|
|
|
VersionStorageInfo new_vstorage(&icmp_, ucmp_, options_.num_levels,
|
|
|
|
kCompactionStyleLevel, &vstorage_,
|
|
|
|
force_consistency_checks);
|
|
|
|
|
|
|
|
ASSERT_OK(builder.SaveTo(&new_vstorage));
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo(&new_vstorage);
|
|
|
|
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
const auto& new_blob_files = new_vstorage.GetBlobFiles();
|
|
|
|
ASSERT_EQ(new_blob_files.size(), 3);
|
|
|
|
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
const auto meta3 = new_vstorage.GetBlobFileMetaData(/* blob_file_number */ 3);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
ASSERT_NE(meta3, nullptr);
|
|
|
|
ASSERT_EQ(meta3->GetBlobFileNumber(), 3);
|
|
|
|
ASSERT_EQ(meta3->GetTotalBlobCount(), 1000);
|
|
|
|
ASSERT_EQ(meta3->GetTotalBlobBytes(), 1000000);
|
|
|
|
ASSERT_EQ(meta3->GetGarbageBlobCount(), 100);
|
|
|
|
ASSERT_EQ(meta3->GetGarbageBlobBytes(), 20000);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
const auto meta5 = new_vstorage.GetBlobFileMetaData(/* blob_file_number */ 5);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
ASSERT_NE(meta5, nullptr);
|
|
|
|
ASSERT_EQ(meta5->GetBlobFileNumber(), 5);
|
|
|
|
ASSERT_EQ(meta5->GetTotalBlobCount(), 2000);
|
|
|
|
ASSERT_EQ(meta5->GetTotalBlobBytes(), 2000000);
|
|
|
|
ASSERT_EQ(meta5->GetGarbageBlobCount(), 400);
|
|
|
|
ASSERT_EQ(meta5->GetGarbageBlobBytes(), 140000);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
const auto meta9 = new_vstorage.GetBlobFileMetaData(/* blob_file_number */ 9);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
ASSERT_NE(meta9, nullptr);
|
|
|
|
ASSERT_EQ(meta9->GetBlobFileNumber(), 9);
|
|
|
|
ASSERT_EQ(meta9->GetTotalBlobCount(), 4000);
|
|
|
|
ASSERT_EQ(meta9->GetTotalBlobBytes(), 4000000);
|
|
|
|
ASSERT_EQ(meta9->GetGarbageBlobCount(), 0);
|
|
|
|
ASSERT_EQ(meta9->GetGarbageBlobBytes(), 0);
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
|
|
|
|
// Delete the first table file, which makes the first blob file obsolete
|
|
|
|
// since it's at the head and unreferenced.
|
|
|
|
VersionBuilder second_builder(env_options, &ioptions_, table_cache,
|
|
|
|
&new_vstorage, version_set);
|
|
|
|
|
|
|
|
VersionEdit second_edit;
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
second_edit.DeleteFile(/* level */ 0, /* file_number */ 2);
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
|
|
|
|
ASSERT_OK(second_builder.Apply(&second_edit));
|
|
|
|
|
|
|
|
VersionStorageInfo newer_vstorage(&icmp_, ucmp_, options_.num_levels,
|
|
|
|
kCompactionStyleLevel, &new_vstorage,
|
|
|
|
force_consistency_checks);
|
|
|
|
|
|
|
|
ASSERT_OK(second_builder.SaveTo(&newer_vstorage));
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo(&newer_vstorage);
|
|
|
|
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
const auto& newer_blob_files = newer_vstorage.GetBlobFiles();
|
|
|
|
ASSERT_EQ(newer_blob_files.size(), 2);
|
|
|
|
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
const auto newer_meta3 =
|
|
|
|
newer_vstorage.GetBlobFileMetaData(/* blob_file_number */ 3);
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
ASSERT_EQ(newer_meta3, nullptr);
|
Clean up blob files based on the linked SST set (#7001)
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
5 years ago
|
|
|
|
|
|
|
UnrefFilesInVersion(&newer_vstorage);
|
|
|
|
UnrefFilesInVersion(&new_vstorage);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(VersionBuilderTest, SaveBlobFilesToConcurrentJobs) {
|
|
|
|
// When multiple background jobs (flushes/compactions) are executing in
|
|
|
|
// parallel, it is possible for the VersionEdit adding blob file K to be
|
|
|
|
// applied *after* the VersionEdit adding blob file N (for N > K). This test
|
|
|
|
// case makes sure this is handled correctly.
|
|
|
|
|
|
|
|
// Add blob file #4 (referenced by table file #3) to base version.
|
|
|
|
constexpr uint64_t base_table_file_number = 3;
|
|
|
|
constexpr uint64_t base_blob_file_number = 4;
|
|
|
|
constexpr uint64_t base_total_blob_count = 100;
|
|
|
|
constexpr uint64_t base_total_blob_bytes = 1 << 20;
|
|
|
|
|
|
|
|
constexpr char checksum_method[] = "SHA1";
|
|
|
|
constexpr char checksum_value[] = "\xfa\xce\xb0\x0c";
|
|
|
|
constexpr uint64_t garbage_blob_count = 0;
|
|
|
|
constexpr uint64_t garbage_blob_bytes = 0;
|
|
|
|
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
AddDummyFile(base_table_file_number, base_blob_file_number,
|
|
|
|
1 /*epoch_number*/);
|
|
|
|
AddBlob(base_blob_file_number, base_total_blob_count, base_total_blob_bytes,
|
|
|
|
checksum_method, checksum_value,
|
|
|
|
BlobFileMetaData::LinkedSsts{base_table_file_number},
|
|
|
|
garbage_blob_count, garbage_blob_bytes);
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
|
|
|
|
|
|
|
VersionBuilder builder(env_options, &ioptions_, table_cache, &vstorage_,
|
|
|
|
version_set);
|
|
|
|
|
|
|
|
VersionEdit edit;
|
|
|
|
|
|
|
|
// Add blob file #2 (referenced by table file #1).
|
|
|
|
constexpr int level = 0;
|
|
|
|
constexpr uint64_t table_file_number = 1;
|
|
|
|
constexpr uint32_t path_id = 0;
|
|
|
|
constexpr uint64_t file_size = 1 << 12;
|
|
|
|
constexpr char smallest[] = "key1";
|
|
|
|
constexpr char largest[] = "key987";
|
|
|
|
constexpr SequenceNumber smallest_seqno = 0;
|
|
|
|
constexpr SequenceNumber largest_seqno = 0;
|
|
|
|
constexpr bool marked_for_compaction = false;
|
|
|
|
|
|
|
|
constexpr uint64_t blob_file_number = 2;
|
|
|
|
static_assert(blob_file_number < base_blob_file_number,
|
|
|
|
"Added blob file should have a smaller file number");
|
|
|
|
|
|
|
|
constexpr uint64_t total_blob_count = 234;
|
|
|
|
constexpr uint64_t total_blob_bytes = 1 << 22;
|
|
|
|
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
edit.AddFile(
|
|
|
|
level, table_file_number, path_id, file_size, GetInternalKey(smallest),
|
|
|
|
GetInternalKey(largest), smallest_seqno, largest_seqno,
|
|
|
|
marked_for_compaction, Temperature::kUnknown, blob_file_number,
|
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime, 2 /*epoch_number*/,
|
|
|
|
checksum_value, checksum_method, kNullUniqueId64x2, 0);
|
|
|
|
edit.AddBlobFile(blob_file_number, total_blob_count, total_blob_bytes,
|
|
|
|
checksum_method, checksum_value);
|
|
|
|
|
|
|
|
ASSERT_OK(builder.Apply(&edit));
|
|
|
|
|
|
|
|
constexpr bool force_consistency_checks = true;
|
|
|
|
VersionStorageInfo new_vstorage(&icmp_, ucmp_, options_.num_levels,
|
|
|
|
kCompactionStyleLevel, &vstorage_,
|
|
|
|
force_consistency_checks);
|
|
|
|
|
|
|
|
ASSERT_OK(builder.SaveTo(&new_vstorage));
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo(&new_vstorage);
|
|
|
|
|
|
|
|
const auto& new_blob_files = new_vstorage.GetBlobFiles();
|
|
|
|
ASSERT_EQ(new_blob_files.size(), 2);
|
|
|
|
|
|
|
|
const auto base_meta =
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
new_vstorage.GetBlobFileMetaData(base_blob_file_number);
|
|
|
|
|
|
|
|
ASSERT_NE(base_meta, nullptr);
|
|
|
|
ASSERT_EQ(base_meta->GetBlobFileNumber(), base_blob_file_number);
|
|
|
|
ASSERT_EQ(base_meta->GetTotalBlobCount(), base_total_blob_count);
|
|
|
|
ASSERT_EQ(base_meta->GetTotalBlobBytes(), base_total_blob_bytes);
|
|
|
|
ASSERT_EQ(base_meta->GetGarbageBlobCount(), garbage_blob_count);
|
|
|
|
ASSERT_EQ(base_meta->GetGarbageBlobBytes(), garbage_blob_bytes);
|
|
|
|
ASSERT_EQ(base_meta->GetChecksumMethod(), checksum_method);
|
|
|
|
ASSERT_EQ(base_meta->GetChecksumValue(), checksum_value);
|
|
|
|
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
const auto added_meta = new_vstorage.GetBlobFileMetaData(blob_file_number);
|
|
|
|
|
|
|
|
ASSERT_NE(added_meta, nullptr);
|
|
|
|
ASSERT_EQ(added_meta->GetBlobFileNumber(), blob_file_number);
|
|
|
|
ASSERT_EQ(added_meta->GetTotalBlobCount(), total_blob_count);
|
|
|
|
ASSERT_EQ(added_meta->GetTotalBlobBytes(), total_blob_bytes);
|
|
|
|
ASSERT_EQ(added_meta->GetGarbageBlobCount(), garbage_blob_count);
|
|
|
|
ASSERT_EQ(added_meta->GetGarbageBlobBytes(), garbage_blob_bytes);
|
|
|
|
ASSERT_EQ(added_meta->GetChecksumMethod(), checksum_method);
|
|
|
|
ASSERT_EQ(added_meta->GetChecksumValue(), checksum_value);
|
|
|
|
|
|
|
|
UnrefFilesInVersion(&new_vstorage);
|
|
|
|
}
|
|
|
|
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
TEST_F(VersionBuilderTest, CheckConsistencyForBlobFiles) {
|
|
|
|
// Initialize base version. The first table file points to a valid blob file
|
|
|
|
// in this version; the second one does not refer to any blob files.
|
|
|
|
|
|
|
|
Add(/* level */ 1, /* file_number */ 1, /* smallest */ "150",
|
|
|
|
/* largest */ "200", /* file_size */ 100,
|
|
|
|
/* path_id */ 0, /* smallest_seq */ 100, /* largest_seq */ 100,
|
|
|
|
/* num_entries */ 0, /* num_deletions */ 0,
|
|
|
|
/* sampled */ false, /* smallest_seqno */ 100, /* largest_seqno */ 100,
|
|
|
|
/* oldest_blob_file_number */ 16);
|
|
|
|
Add(/* level */ 1, /* file_number */ 23, /* smallest */ "201",
|
|
|
|
/* largest */ "300", /* file_size */ 100,
|
|
|
|
/* path_id */ 0, /* smallest_seq */ 200, /* largest_seq */ 200,
|
|
|
|
/* num_entries */ 0, /* num_deletions */ 0,
|
|
|
|
/* sampled */ false, /* smallest_seqno */ 200, /* largest_seqno */ 200,
|
|
|
|
kInvalidBlobFileNumber);
|
|
|
|
|
|
|
|
AddBlob(/* blob_file_number */ 16, /* total_blob_count */ 1000,
|
|
|
|
/* total_blob_bytes */ 1000000,
|
|
|
|
/* checksum_method */ std::string(),
|
|
|
|
/* checksum_value */ std::string(), BlobFileMetaData::LinkedSsts{1},
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
/* garbage_blob_count */ 500, /* garbage_blob_bytes */ 300000);
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
|
|
|
// Add a new table file that points to the existing blob file, and add a
|
|
|
|
// new table file--blob file pair.
|
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
|
|
|
|
|
|
|
VersionBuilder builder(env_options, &ioptions_, table_cache, &vstorage_,
|
|
|
|
version_set);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
|
|
|
VersionEdit edit;
|
|
|
|
|
|
|
|
edit.AddFile(/* level */ 1, /* file_number */ 606, /* path_id */ 0,
|
|
|
|
/* file_size */ 100, /* smallest */ GetInternalKey("701"),
|
|
|
|
/* largest */ GetInternalKey("750"), /* smallest_seqno */ 200,
|
|
|
|
/* largest_seqno */ 200, /* marked_for_compaction */ false,
|
|
|
|
Temperature::kUnknown,
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
/* oldest_blob_file_number */ 16, kUnknownOldestAncesterTime,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownFileCreationTime, kUnknownEpochNumber,
|
|
|
|
kUnknownFileChecksum, kUnknownFileChecksumFuncName,
|
|
|
|
kNullUniqueId64x2, 0);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
|
|
|
edit.AddFile(/* level */ 1, /* file_number */ 700, /* path_id */ 0,
|
|
|
|
/* file_size */ 100, /* smallest */ GetInternalKey("801"),
|
|
|
|
/* largest */ GetInternalKey("850"), /* smallest_seqno */ 200,
|
|
|
|
/* largest_seqno */ 200, /* marked_for_compaction */ false,
|
|
|
|
Temperature::kUnknown,
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
/* oldest_blob_file_number */ 1000, kUnknownOldestAncesterTime,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownFileCreationTime, kUnknownEpochNumber,
|
|
|
|
kUnknownFileChecksum, kUnknownFileChecksumFuncName,
|
|
|
|
kNullUniqueId64x2, 0);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
edit.AddBlobFile(/* blob_file_number */ 1000, /* total_blob_count */ 2000,
|
|
|
|
/* total_blob_bytes */ 200000,
|
|
|
|
/* checksum_method */ std::string(),
|
|
|
|
/* checksum_value */ std::string());
|
|
|
|
|
|
|
|
ASSERT_OK(builder.Apply(&edit));
|
|
|
|
|
|
|
|
// Save to a new version in order to trigger consistency checks.
|
|
|
|
constexpr bool force_consistency_checks = true;
|
|
|
|
VersionStorageInfo new_vstorage(&icmp_, ucmp_, options_.num_levels,
|
|
|
|
kCompactionStyleLevel, &vstorage_,
|
|
|
|
force_consistency_checks);
|
|
|
|
|
|
|
|
ASSERT_OK(builder.SaveTo(&new_vstorage));
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo(&new_vstorage);
|
|
|
|
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
UnrefFilesInVersion(&new_vstorage);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(VersionBuilderTest, CheckConsistencyForBlobFilesInconsistentLinks) {
|
|
|
|
// Initialize base version. Links between the table file and the blob file
|
|
|
|
// are inconsistent.
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
|
|
|
Add(/* level */ 1, /* file_number */ 1, /* smallest */ "150",
|
|
|
|
/* largest */ "200", /* file_size */ 100,
|
|
|
|
/* path_id */ 0, /* smallest_seq */ 100, /* largest_seq */ 100,
|
|
|
|
/* num_entries */ 0, /* num_deletions */ 0,
|
|
|
|
/* sampled */ false, /* smallest_seqno */ 100, /* largest_seqno */ 100,
|
|
|
|
/* oldest_blob_file_number */ 256);
|
|
|
|
|
|
|
|
AddBlob(/* blob_file_number */ 16, /* total_blob_count */ 1000,
|
|
|
|
/* total_blob_bytes */ 1000000,
|
|
|
|
/* checksum_method */ std::string(),
|
|
|
|
/* checksum_value */ std::string(), BlobFileMetaData::LinkedSsts{1},
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
/* garbage_blob_count */ 500, /* garbage_blob_bytes */ 300000);
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
|
|
|
|
|
|
|
VersionBuilder builder(env_options, &ioptions_, table_cache, &vstorage_,
|
|
|
|
version_set);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
|
|
|
// Save to a new version in order to trigger consistency checks.
|
|
|
|
constexpr bool force_consistency_checks = true;
|
|
|
|
VersionStorageInfo new_vstorage(&icmp_, ucmp_, options_.num_levels,
|
|
|
|
kCompactionStyleLevel, &vstorage_,
|
|
|
|
force_consistency_checks);
|
|
|
|
|
|
|
|
const Status s = builder.SaveTo(&new_vstorage);
|
|
|
|
ASSERT_TRUE(s.IsCorruption());
|
|
|
|
ASSERT_TRUE(std::strstr(
|
|
|
|
s.getState(),
|
|
|
|
"Links are inconsistent between table files and blob file #16"));
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
|
|
|
UnrefFilesInVersion(&new_vstorage);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(VersionBuilderTest, CheckConsistencyForBlobFilesAllGarbage) {
|
|
|
|
// Initialize base version. The table file points to a blob file that is
|
|
|
|
// all garbage.
|
|
|
|
|
|
|
|
Add(/* level */ 1, /* file_number */ 1, /* smallest */ "150",
|
|
|
|
/* largest */ "200", /* file_size */ 100,
|
|
|
|
/* path_id */ 0, /* smallest_seq */ 100, /* largest_seq */ 100,
|
|
|
|
/* num_entries */ 0, /* num_deletions */ 0,
|
|
|
|
/* sampled */ false, /* smallest_seqno */ 100, /* largest_seqno */ 100,
|
|
|
|
/* oldest_blob_file_number */ 16);
|
|
|
|
|
|
|
|
AddBlob(/* blob_file_number */ 16, /* total_blob_count */ 1000,
|
|
|
|
/* total_blob_bytes */ 1000000,
|
|
|
|
/* checksum_method */ std::string(),
|
|
|
|
/* checksum_value */ std::string(), BlobFileMetaData::LinkedSsts{1},
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
/* garbage_blob_count */ 1000, /* garbage_blob_bytes */ 1000000);
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
|
|
|
|
|
|
|
VersionBuilder builder(env_options, &ioptions_, table_cache, &vstorage_,
|
|
|
|
version_set);
|
Add blob files to VersionStorageInfo/VersionBuilder (#6597)
Summary:
The patch adds a couple of classes to represent metadata about
blob files: `SharedBlobFileMetaData` contains the information elements
that are immutable (once the blob file is closed), e.g. blob file number,
total number and size of blob files, checksum method/value, while
`BlobFileMetaData` contains attributes that can vary across versions like
the amount of garbage in the file. There is a single `SharedBlobFileMetaData`
for each blob file, which is jointly owned by the `BlobFileMetaData` objects
that point to it; `BlobFileMetaData` objects, in turn, are owned by `Version`s
and can also be shared if the (immutable _and_ mutable) state of the blob file
is the same in two versions.
In addition, the patch adds the blob file metadata to `VersionStorageInfo`, and extends
`VersionBuilder` so that it can apply blob file related `VersionEdit`s (i.e. those
containing `BlobFileAddition`s and/or `BlobFileGarbage`), and save blob file metadata
to a new `VersionStorageInfo`. Consistency checks are also extended to ensure
that table files point to blob files that are part of the `Version`, and that all blob files
that are part of any given `Version` have at least some _non_-garbage data in them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6597
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D20656803
Pulled By: ltamasi
fbshipit-source-id: f1f74d135045b3b42d0146f03ee576ef0a4bfd80
5 years ago
|
|
|
|
|
|
|
// Save to a new version in order to trigger consistency checks.
|
|
|
|
constexpr bool force_consistency_checks = true;
|
|
|
|
VersionStorageInfo new_vstorage(&icmp_, ucmp_, options_.num_levels,
|
|
|
|
kCompactionStyleLevel, &vstorage_,
|
|
|
|
force_consistency_checks);
|
|
|
|
|
|
|
|
const Status s = builder.SaveTo(&new_vstorage);
|
|
|
|
ASSERT_TRUE(s.IsCorruption());
|
|
|
|
ASSERT_TRUE(
|
|
|
|
std::strstr(s.getState(), "Blob file #16 consists entirely of garbage"));
|
|
|
|
|
|
|
|
UnrefFilesInVersion(&new_vstorage);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(VersionBuilderTest, CheckConsistencyForBlobFilesAllGarbageLinkedSsts) {
|
|
|
|
// Initialize base version, with a table file pointing to a blob file
|
|
|
|
// that has no garbage at this point.
|
|
|
|
|
|
|
|
Add(/* level */ 1, /* file_number */ 1, /* smallest */ "150",
|
|
|
|
/* largest */ "200", /* file_size */ 100,
|
|
|
|
/* path_id */ 0, /* smallest_seq */ 100, /* largest_seq */ 100,
|
|
|
|
/* num_entries */ 0, /* num_deletions */ 0,
|
|
|
|
/* sampled */ false, /* smallest_seqno */ 100, /* largest_seqno */ 100,
|
|
|
|
/* oldest_blob_file_number */ 16);
|
|
|
|
|
|
|
|
AddBlob(/* blob_file_number */ 16, /* total_blob_count */ 1000,
|
|
|
|
/* total_blob_bytes */ 1000000,
|
|
|
|
/* checksum_method */ std::string(),
|
|
|
|
/* checksum_value */ std::string(), BlobFileMetaData::LinkedSsts{1},
|
|
|
|
/* garbage_blob_count */ 0, /* garbage_blob_bytes */ 0);
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
|
|
|
// Mark the entire blob file garbage but do not remove the linked SST.
|
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
|
|
|
|
|
|
|
VersionBuilder builder(env_options, &ioptions_, table_cache, &vstorage_,
|
|
|
|
version_set);
|
|
|
|
|
|
|
|
VersionEdit edit;
|
|
|
|
|
|
|
|
edit.AddBlobFileGarbage(/* blob_file_number */ 16,
|
|
|
|
/* garbage_blob_count */ 1000,
|
|
|
|
/* garbage_blob_bytes */ 1000000);
|
|
|
|
|
|
|
|
ASSERT_OK(builder.Apply(&edit));
|
|
|
|
|
|
|
|
// Save to a new version in order to trigger consistency checks.
|
|
|
|
constexpr bool force_consistency_checks = true;
|
|
|
|
VersionStorageInfo new_vstorage(&icmp_, ucmp_, options_.num_levels,
|
|
|
|
kCompactionStyleLevel, &vstorage_,
|
|
|
|
force_consistency_checks);
|
|
|
|
|
|
|
|
const Status s = builder.SaveTo(&new_vstorage);
|
|
|
|
ASSERT_TRUE(s.IsCorruption());
|
|
|
|
ASSERT_TRUE(
|
|
|
|
std::strstr(s.getState(), "Blob file #16 consists entirely of garbage"));
|
|
|
|
|
|
|
|
UnrefFilesInVersion(&new_vstorage);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(VersionBuilderTest, MaintainLinkedSstsForBlobFiles) {
|
|
|
|
// Initialize base version. Table files 1..10 are linked to blob files 1..5,
|
|
|
|
// while table files 11..20 are not linked to any blob files.
|
|
|
|
|
|
|
|
for (uint64_t i = 1; i <= 10; ++i) {
|
|
|
|
std::ostringstream oss;
|
|
|
|
oss << std::setw(2) << std::setfill('0') << i;
|
|
|
|
|
|
|
|
const std::string key = oss.str();
|
|
|
|
|
|
|
|
Add(/* level */ 1, /* file_number */ i, /* smallest */ key.c_str(),
|
|
|
|
/* largest */ key.c_str(), /* file_size */ 100,
|
|
|
|
/* path_id */ 0, /* smallest_seq */ i * 100, /* largest_seq */ i * 100,
|
|
|
|
/* num_entries */ 0, /* num_deletions */ 0,
|
|
|
|
/* sampled */ false, /* smallest_seqno */ i * 100,
|
|
|
|
/* largest_seqno */ i * 100,
|
|
|
|
/* oldest_blob_file_number */ ((i - 1) % 5) + 1);
|
|
|
|
}
|
|
|
|
|
|
|
|
for (uint64_t i = 1; i <= 5; ++i) {
|
|
|
|
AddBlob(/* blob_file_number */ i, /* total_blob_count */ 2000,
|
|
|
|
/* total_blob_bytes */ 2000000,
|
|
|
|
/* checksum_method */ std::string(),
|
|
|
|
/* checksum_value */ std::string(),
|
|
|
|
BlobFileMetaData::LinkedSsts{i, i + 5},
|
|
|
|
/* garbage_blob_count */ 1000, /* garbage_blob_bytes */ 1000000);
|
|
|
|
}
|
|
|
|
|
|
|
|
for (uint64_t i = 11; i <= 20; ++i) {
|
|
|
|
std::ostringstream oss;
|
|
|
|
oss << std::setw(2) << std::setfill('0') << i;
|
|
|
|
|
|
|
|
const std::string key = oss.str();
|
|
|
|
|
|
|
|
Add(/* level */ 1, /* file_number */ i, /* smallest */ key.c_str(),
|
|
|
|
/* largest */ key.c_str(), /* file_size */ 100,
|
|
|
|
/* path_id */ 0, /* smallest_seq */ i * 100, /* largest_seq */ i * 100,
|
|
|
|
/* num_entries */ 0, /* num_deletions */ 0,
|
|
|
|
/* sampled */ false, /* smallest_seqno */ i * 100,
|
|
|
|
/* largest_seqno */ i * 100, kInvalidBlobFileNumber);
|
|
|
|
}
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
|
|
|
{
|
|
|
|
const auto& blob_files = vstorage_.GetBlobFiles();
|
|
|
|
ASSERT_EQ(blob_files.size(), 5);
|
|
|
|
|
|
|
|
const std::vector<BlobFileMetaData::LinkedSsts> expected_linked_ssts{
|
|
|
|
{1, 6}, {2, 7}, {3, 8}, {4, 9}, {5, 10}};
|
|
|
|
|
|
|
|
for (size_t i = 0; i < 5; ++i) {
|
|
|
|
const auto meta =
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
vstorage_.GetBlobFileMetaData(/* blob_file_number */ i + 1);
|
|
|
|
ASSERT_NE(meta, nullptr);
|
|
|
|
ASSERT_EQ(meta->GetLinkedSsts(), expected_linked_ssts[i]);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
VersionEdit edit;
|
|
|
|
|
|
|
|
// Add an SST that references a blob file.
|
|
|
|
edit.AddFile(
|
|
|
|
/* level */ 1, /* file_number */ 21, /* path_id */ 0,
|
|
|
|
/* file_size */ 100, /* smallest */ GetInternalKey("21", 2100),
|
|
|
|
/* largest */ GetInternalKey("21", 2100), /* smallest_seqno */ 2100,
|
|
|
|
/* largest_seqno */ 2100, /* marked_for_compaction */ false,
|
|
|
|
Temperature::kUnknown,
|
|
|
|
/* oldest_blob_file_number */ 1, kUnknownOldestAncesterTime,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownFileCreationTime, kUnknownEpochNumber, kUnknownFileChecksum,
|
|
|
|
kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
|
|
|
|
|
|
|
// Add an SST that does not reference any blob files.
|
|
|
|
edit.AddFile(
|
|
|
|
/* level */ 1, /* file_number */ 22, /* path_id */ 0,
|
|
|
|
/* file_size */ 100, /* smallest */ GetInternalKey("22", 2200),
|
|
|
|
/* largest */ GetInternalKey("22", 2200), /* smallest_seqno */ 2200,
|
|
|
|
/* largest_seqno */ 2200, /* marked_for_compaction */ false,
|
|
|
|
Temperature::kUnknown, kInvalidBlobFileNumber, kUnknownOldestAncesterTime,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownFileCreationTime, kUnknownEpochNumber, kUnknownFileChecksum,
|
|
|
|
kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
|
|
|
|
|
|
|
// Delete a file that references a blob file.
|
|
|
|
edit.DeleteFile(/* level */ 1, /* file_number */ 6);
|
|
|
|
|
|
|
|
// Delete a file that does not reference any blob files.
|
|
|
|
edit.DeleteFile(/* level */ 1, /* file_number */ 16);
|
|
|
|
|
|
|
|
// Trivially move a file that references a blob file. Note that we save
|
|
|
|
// the original BlobFileMetaData object so we can check that no new object
|
|
|
|
// gets created.
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
auto meta3 = vstorage_.GetBlobFileMetaData(/* blob_file_number */ 3);
|
|
|
|
|
|
|
|
edit.DeleteFile(/* level */ 1, /* file_number */ 3);
|
|
|
|
edit.AddFile(/* level */ 2, /* file_number */ 3, /* path_id */ 0,
|
|
|
|
/* file_size */ 100, /* smallest */ GetInternalKey("03", 300),
|
|
|
|
/* largest */ GetInternalKey("03", 300),
|
|
|
|
/* smallest_seqno */ 300,
|
|
|
|
/* largest_seqno */ 300, /* marked_for_compaction */ false,
|
|
|
|
Temperature::kUnknown,
|
|
|
|
/* oldest_blob_file_number */ 3, kUnknownOldestAncesterTime,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownFileCreationTime, kUnknownEpochNumber,
|
|
|
|
kUnknownFileChecksum, kUnknownFileChecksumFuncName,
|
|
|
|
kNullUniqueId64x2, 0);
|
|
|
|
|
|
|
|
// Trivially move a file that does not reference any blob files.
|
|
|
|
edit.DeleteFile(/* level */ 1, /* file_number */ 13);
|
|
|
|
edit.AddFile(/* level */ 2, /* file_number */ 13, /* path_id */ 0,
|
|
|
|
/* file_size */ 100, /* smallest */ GetInternalKey("13", 1300),
|
|
|
|
/* largest */ GetInternalKey("13", 1300),
|
|
|
|
/* smallest_seqno */ 1300,
|
|
|
|
/* largest_seqno */ 1300, /* marked_for_compaction */ false,
|
|
|
|
Temperature::kUnknown, kInvalidBlobFileNumber,
|
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownEpochNumber, kUnknownFileChecksum,
|
|
|
|
kUnknownFileChecksumFuncName, kNullUniqueId64x2, 0);
|
|
|
|
|
|
|
|
// Add one more SST file that references a blob file, then promptly
|
|
|
|
// delete it in a second version edit before the new version gets saved.
|
|
|
|
// This file should not show up as linked to the blob file in the new version.
|
|
|
|
edit.AddFile(/* level */ 1, /* file_number */ 23, /* path_id */ 0,
|
|
|
|
/* file_size */ 100, /* smallest */ GetInternalKey("23", 2300),
|
|
|
|
/* largest */ GetInternalKey("23", 2300),
|
|
|
|
/* smallest_seqno */ 2300,
|
|
|
|
/* largest_seqno */ 2300, /* marked_for_compaction */ false,
|
|
|
|
Temperature::kUnknown,
|
|
|
|
/* oldest_blob_file_number */ 5, kUnknownOldestAncesterTime,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
kUnknownFileCreationTime, kUnknownEpochNumber,
|
|
|
|
kUnknownFileChecksum, kUnknownFileChecksumFuncName,
|
|
|
|
kNullUniqueId64x2, 0);
|
|
|
|
|
|
|
|
VersionEdit edit2;
|
|
|
|
|
|
|
|
edit2.DeleteFile(/* level */ 1, /* file_number */ 23);
|
|
|
|
|
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
|
|
|
|
|
|
|
VersionBuilder builder(env_options, &ioptions_, table_cache, &vstorage_,
|
|
|
|
version_set);
|
|
|
|
|
|
|
|
ASSERT_OK(builder.Apply(&edit));
|
|
|
|
ASSERT_OK(builder.Apply(&edit2));
|
|
|
|
|
|
|
|
constexpr bool force_consistency_checks = true;
|
|
|
|
VersionStorageInfo new_vstorage(&icmp_, ucmp_, options_.num_levels,
|
|
|
|
kCompactionStyleLevel, &vstorage_,
|
|
|
|
force_consistency_checks);
|
|
|
|
|
|
|
|
ASSERT_OK(builder.SaveTo(&new_vstorage));
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo(&new_vstorage);
|
|
|
|
|
|
|
|
{
|
|
|
|
const auto& blob_files = new_vstorage.GetBlobFiles();
|
|
|
|
ASSERT_EQ(blob_files.size(), 5);
|
|
|
|
|
|
|
|
const std::vector<BlobFileMetaData::LinkedSsts> expected_linked_ssts{
|
|
|
|
{1, 21}, {2, 7}, {3, 8}, {4, 9}, {5, 10}};
|
|
|
|
|
|
|
|
for (size_t i = 0; i < 5; ++i) {
|
|
|
|
const auto meta =
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
new_vstorage.GetBlobFileMetaData(/* blob_file_number */ i + 1);
|
|
|
|
ASSERT_NE(meta, nullptr);
|
|
|
|
ASSERT_EQ(meta->GetLinkedSsts(), expected_linked_ssts[i]);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Make sure that no new BlobFileMetaData got created for the blob file
|
|
|
|
// affected by the trivial move.
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
ASSERT_EQ(new_vstorage.GetBlobFileMetaData(/* blob_file_number */ 3),
|
|
|
|
meta3);
|
|
|
|
}
|
|
|
|
|
|
|
|
UnrefFilesInVersion(&new_vstorage);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(VersionBuilderTest, CheckConsistencyForFileDeletedTwice) {
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
Add(0, 1U, "150", "200", 100, /*path_id*/ 0,
|
|
|
|
/*smallest_seq*/ 100, /*largest_seq*/ 100,
|
|
|
|
/*num_entries*/ 0, /*num_deletions*/ 0,
|
|
|
|
/*sampled*/ false, /*smallest_seqno*/ 0,
|
|
|
|
/*largest_seqno*/ 0,
|
|
|
|
/*oldest_blob_file_number*/ kInvalidBlobFileNumber,
|
|
|
|
/*epoch_number*/ 1);
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo();
|
|
|
|
|
|
|
|
VersionEdit version_edit;
|
|
|
|
version_edit.DeleteFile(0, 1U);
|
|
|
|
|
|
|
|
EnvOptions env_options;
|
|
|
|
constexpr TableCache* table_cache = nullptr;
|
|
|
|
constexpr VersionSet* version_set = nullptr;
|
|
|
|
|
|
|
|
VersionBuilder version_builder(env_options, &ioptions_, table_cache,
|
|
|
|
&vstorage_, version_set);
|
|
|
|
VersionStorageInfo new_vstorage(&icmp_, ucmp_, options_.num_levels,
|
|
|
|
kCompactionStyleLevel, nullptr,
|
|
|
|
true /* force_consistency_checks */);
|
|
|
|
ASSERT_OK(version_builder.Apply(&version_edit));
|
|
|
|
ASSERT_OK(version_builder.SaveTo(&new_vstorage));
|
|
|
|
|
|
|
|
UpdateVersionStorageInfo(&new_vstorage);
|
|
|
|
|
|
|
|
VersionBuilder version_builder2(env_options, &ioptions_, table_cache,
|
|
|
|
&new_vstorage, version_set);
|
|
|
|
VersionStorageInfo new_vstorage2(&icmp_, ucmp_, options_.num_levels,
|
|
|
|
kCompactionStyleLevel, nullptr,
|
|
|
|
true /* force_consistency_checks */);
|
|
|
|
ASSERT_NOK(version_builder2.Apply(&version_edit));
|
|
|
|
|
|
|
|
UnrefFilesInVersion(&new_vstorage);
|
|
|
|
UnrefFilesInVersion(&new_vstorage2);
|
|
|
|
}
|
|
|
|
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
TEST_F(VersionBuilderTest, CheckConsistencyForL0FilesSortedByEpochNumber) {
|
|
|
|
Status s;
|
|
|
|
// To verify files of same epoch number of overlapping ranges are caught as
|
|
|
|
// corrupted
|
|
|
|
VersionEdit version_edit_1;
|
|
|
|
version_edit_1.AddFile(
|
|
|
|
/* level */ 0, /* file_number */ 1U, /* path_id */ 0,
|
|
|
|
/* file_size */ 100, /* smallest */ GetInternalKey("a", 1),
|
|
|
|
/* largest */ GetInternalKey("c", 3), /* smallest_seqno */ 1,
|
|
|
|
/* largest_seqno */ 3, /* marked_for_compaction */ false,
|
|
|
|
Temperature::kUnknown,
|
|
|
|
/* oldest_blob_file_number */ kInvalidBlobFileNumber,
|
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime,
|
|
|
|
1 /* epoch_number */, kUnknownFileChecksum, kUnknownFileChecksumFuncName,
|
|
|
|
kNullUniqueId64x2, 0);
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
version_edit_1.AddFile(
|
|
|
|
/* level */ 0, /* file_number */ 2U, /* path_id */ 0,
|
|
|
|
/* file_size */ 100, /* smallest */ GetInternalKey("b", 2),
|
|
|
|
/* largest */ GetInternalKey("d", 4), /* smallest_seqno */ 2,
|
|
|
|
/* largest_seqno */ 4, /* marked_for_compaction */ false,
|
|
|
|
Temperature::kUnknown,
|
|
|
|
/* oldest_blob_file_number */ kInvalidBlobFileNumber,
|
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime,
|
|
|
|
1 /* epoch_number */, kUnknownFileChecksum, kUnknownFileChecksumFuncName,
|
|
|
|
kNullUniqueId64x2, 0);
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
|
|
|
|
VersionBuilder version_builder_1(EnvOptions(), &ioptions_,
|
|
|
|
nullptr /* table_cache */, &vstorage_,
|
|
|
|
nullptr /* file_metadata_cache_res_mgr */);
|
|
|
|
VersionStorageInfo new_vstorage_1(
|
|
|
|
&icmp_, ucmp_, options_.num_levels, kCompactionStyleLevel,
|
|
|
|
nullptr /* src_vstorage */, true /* force_consistency_checks */);
|
|
|
|
|
|
|
|
ASSERT_OK(version_builder_1.Apply(&version_edit_1));
|
|
|
|
s = version_builder_1.SaveTo(&new_vstorage_1);
|
|
|
|
EXPECT_TRUE(s.IsCorruption());
|
|
|
|
EXPECT_TRUE(std::strstr(
|
|
|
|
s.getState(), "L0 files of same epoch number but overlapping range"));
|
|
|
|
UnrefFilesInVersion(&new_vstorage_1);
|
|
|
|
|
|
|
|
// To verify L0 files not sorted by epoch_number are caught as corrupted
|
|
|
|
VersionEdit version_edit_2;
|
|
|
|
version_edit_2.AddFile(
|
|
|
|
/* level */ 0, /* file_number */ 1U, /* path_id */ 0,
|
|
|
|
/* file_size */ 100, /* smallest */ GetInternalKey("a", 1),
|
|
|
|
/* largest */ GetInternalKey("a", 1), /* smallest_seqno */ 1,
|
|
|
|
/* largest_seqno */ 1, /* marked_for_compaction */ false,
|
|
|
|
Temperature::kUnknown,
|
|
|
|
/* oldest_blob_file_number */ kInvalidBlobFileNumber,
|
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime,
|
|
|
|
1 /* epoch_number */, kUnknownFileChecksum, kUnknownFileChecksumFuncName,
|
|
|
|
kNullUniqueId64x2, 0);
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
version_edit_2.AddFile(
|
|
|
|
/* level */ 0, /* file_number */ 2U, /* path_id */ 0,
|
|
|
|
/* file_size */ 100, /* smallest */ GetInternalKey("b", 2),
|
|
|
|
/* largest */ GetInternalKey("b", 2), /* smallest_seqno */ 2,
|
|
|
|
/* largest_seqno */ 2, /* marked_for_compaction */ false,
|
|
|
|
Temperature::kUnknown,
|
|
|
|
/* oldest_blob_file_number */ kInvalidBlobFileNumber,
|
|
|
|
kUnknownOldestAncesterTime, kUnknownFileCreationTime,
|
|
|
|
2 /* epoch_number */, kUnknownFileChecksum, kUnknownFileChecksumFuncName,
|
|
|
|
kNullUniqueId64x2, 0);
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
|
|
|
|
|
|
|
VersionBuilder version_builder_2(EnvOptions(), &ioptions_,
|
|
|
|
nullptr /* table_cache */, &vstorage_,
|
|
|
|
nullptr /* file_metadata_cache_res_mgr */);
|
|
|
|
VersionStorageInfo new_vstorage_2(
|
|
|
|
&icmp_, ucmp_, options_.num_levels, kCompactionStyleLevel,
|
|
|
|
nullptr /* src_vstorage */, true /* force_consistency_checks */);
|
|
|
|
|
|
|
|
ASSERT_OK(version_builder_2.Apply(&version_edit_2));
|
|
|
|
s = version_builder_2.SaveTo(&new_vstorage_2);
|
|
|
|
ASSERT_TRUE(s.ok());
|
|
|
|
|
|
|
|
const std::vector<FileMetaData*>& l0_files = new_vstorage_2.LevelFiles(0);
|
|
|
|
ASSERT_EQ(l0_files.size(), 2);
|
|
|
|
// Manually corrupt L0 files's epoch_number
|
|
|
|
l0_files[0]->epoch_number = 1;
|
|
|
|
l0_files[1]->epoch_number = 2;
|
|
|
|
|
|
|
|
// To surface corruption error by applying dummy version edit
|
|
|
|
VersionEdit dummy_version_edit;
|
|
|
|
VersionBuilder dummy_version_builder(
|
|
|
|
EnvOptions(), &ioptions_, nullptr /* table_cache */, &vstorage_,
|
|
|
|
nullptr /* file_metadata_cache_res_mgr */);
|
|
|
|
ASSERT_OK(dummy_version_builder.Apply(&dummy_version_edit));
|
|
|
|
s = dummy_version_builder.SaveTo(&new_vstorage_2);
|
|
|
|
EXPECT_TRUE(s.IsCorruption());
|
|
|
|
EXPECT_TRUE(std::strstr(s.getState(), "L0 files are not sorted properly"));
|
|
|
|
|
|
|
|
UnrefFilesInVersion(&new_vstorage_2);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(VersionBuilderTest, EstimatedActiveKeys) {
|
|
|
|
const uint32_t kTotalSamples = 20;
|
|
|
|
const uint32_t kNumLevels = 5;
|
|
|
|
const uint32_t kFilesPerLevel = 8;
|
|
|
|
const uint32_t kNumFiles = kNumLevels * kFilesPerLevel;
|
|
|
|
const uint32_t kEntriesPerFile = 1000;
|
|
|
|
const uint32_t kDeletionsPerFile = 100;
|
|
|
|
for (uint32_t i = 0; i < kNumFiles; ++i) {
|
|
|
|
Add(static_cast<int>(i / kFilesPerLevel), i + 1,
|
|
|
|
std::to_string((i + 100) * 1000).c_str(),
|
|
|
|
std::to_string((i + 100) * 1000 + 999).c_str(), 100U, 0, 100, 100,
|
|
|
|
kEntriesPerFile, kDeletionsPerFile, (i < kTotalSamples));
|
|
|
|
}
|
|
|
|
// minus 2X for the number of deletion entries because:
|
|
|
|
// 1x for deletion entry does not count as a data entry.
|
|
|
|
// 1x for each deletion entry will actually remove one data entry.
|
|
|
|
ASSERT_EQ(vstorage_.GetEstimatedActiveKeys(),
|
|
|
|
(kEntriesPerFile - 2 * kDeletionsPerFile) * kNumFiles);
|
|
|
|
}
|
|
|
|
|
|
|
|
} // namespace ROCKSDB_NAMESPACE
|
|
|
|
|
|
|
|
int main(int argc, char** argv) {
|
|
|
|
ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
|
|
|
|
::testing::InitGoogleTest(&argc, argv);
|
|
|
|
return RUN_ALL_TESTS();
|
|
|
|
}
|