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

2494 lines
93 KiB

// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
// This source code is licensed under both the GPLv2 (found in the
// COPYING file in the root directory) and Apache 2.0 License
// (found in the LICENSE.Apache file in the root directory).
//
// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file. See the AUTHORS file for names of contributors.
#include <cinttypes>
#include "db/db_impl/db_impl.h"
#include "db/error_handler.h"
#include "db/event_helpers.h"
#include "logging/logging.h"
#include "monitoring/perf_context_imp.h"
#include "options/options_helper.h"
#include "test_util/sync_point.h"
#include "util/cast_util.h"
namespace ROCKSDB_NAMESPACE {
// Convenience methods
Status DBImpl::Put(const WriteOptions& o, ColumnFamilyHandle* column_family,
const Slice& key, const Slice& val) {
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
const Status s = FailIfCfHasTs(column_family);
if (!s.ok()) {
return s;
}
return DB::Put(o, column_family, key, val);
}
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
Status DBImpl::Put(const WriteOptions& o, ColumnFamilyHandle* column_family,
const Slice& key, const Slice& ts, const Slice& val) {
const Status s = FailIfTsMismatchCf(column_family, ts, /*ts_for_read=*/false);
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
if (!s.ok()) {
return s;
}
return DB::Put(o, column_family, key, ts, val);
}
Status DBImpl::PutEntity(const WriteOptions& options,
ColumnFamilyHandle* column_family, const Slice& key,
const WideColumns& columns) {
const Status s = FailIfCfHasTs(column_family);
if (!s.ok()) {
return s;
}
return DB::PutEntity(options, column_family, key, columns);
}
Status DBImpl::Merge(const WriteOptions& o, ColumnFamilyHandle* column_family,
const Slice& key, const Slice& val) {
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
const Status s = FailIfCfHasTs(column_family);
if (!s.ok()) {
return s;
}
auto cfh = static_cast_with_check<ColumnFamilyHandleImpl>(column_family);
if (!cfh->cfd()->ioptions()->merge_operator) {
return Status::NotSupported("Provide a merge_operator when opening DB");
} else {
return DB::Merge(o, column_family, key, val);
}
}
Status DBImpl::Merge(const WriteOptions& o, ColumnFamilyHandle* column_family,
const Slice& key, const Slice& ts, const Slice& val) {
const Status s = FailIfTsMismatchCf(column_family, ts, /*ts_for_read=*/false);
if (!s.ok()) {
return s;
}
return DB::Merge(o, column_family, key, ts, val);
}
Status DBImpl::Delete(const WriteOptions& write_options,
ColumnFamilyHandle* column_family, const Slice& key) {
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
const Status s = FailIfCfHasTs(column_family);
if (!s.ok()) {
return s;
}
return DB::Delete(write_options, column_family, key);
}
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
Status DBImpl::Delete(const WriteOptions& write_options,
ColumnFamilyHandle* column_family, const Slice& key,
const Slice& ts) {
const Status s = FailIfTsMismatchCf(column_family, ts, /*ts_for_read=*/false);
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
if (!s.ok()) {
return s;
}
return DB::Delete(write_options, column_family, key, ts);
}
Status DBImpl::SingleDelete(const WriteOptions& write_options,
ColumnFamilyHandle* column_family,
const Slice& key) {
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
const Status s = FailIfCfHasTs(column_family);
if (!s.ok()) {
return s;
}
return DB::SingleDelete(write_options, column_family, key);
}
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
Status DBImpl::SingleDelete(const WriteOptions& write_options,
ColumnFamilyHandle* column_family, const Slice& key,
const Slice& ts) {
const Status s = FailIfTsMismatchCf(column_family, ts, /*ts_for_read=*/false);
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
if (!s.ok()) {
return s;
}
return DB::SingleDelete(write_options, column_family, key, ts);
}
Status DBImpl::DeleteRange(const WriteOptions& write_options,
ColumnFamilyHandle* column_family,
const Slice& begin_key, const Slice& end_key) {
const Status s = FailIfCfHasTs(column_family);
if (!s.ok()) {
return s;
}
return DB::DeleteRange(write_options, column_family, begin_key, end_key);
}
User-defined timestamp support for `DeleteRange()` (#10661) Summary: Add user-defined timestamp support for range deletion. The new API is `DeleteRange(opt, cf, begin_key, end_key, ts)`. Most of the change is to update the comparator to compare without timestamp. Other than that, major changes are - internal range tombstone data structures (`FragmentedRangeTombstoneList`, `RangeTombstone`, etc.) to store timestamps. - Garbage collection of range tombstones and range tombstone covered keys during compaction. - Get()/MultiGet() to return the timestamp of a range tombstone when needed. - Get/Iterator with range tombstones bounded by readoptions.timestamp. - timestamp crash test now issues DeleteRange by default. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10661 Test Plan: - Added unit test: `make check` - Stress test: `python3 tools/db_crashtest.py --enable_ts whitebox --readpercent=57 --prefixpercent=4 --writepercent=25 -delpercent=5 --iterpercent=5 --delrangepercent=4` - Ran `db_bench` to measure regression when timestamp is not enabled. The tests are for write (with some range deletion) and iterate with DB fitting in memory: `./db_bench--benchmarks=fillrandom,seekrandom --writes_per_range_tombstone=200 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=500000 --reads=500000 --seek_nexts=10 --disable_auto_compactions -disable_wal=true --max_num_range_tombstones=1000`. Did not see consistent regression in no timestamp case. | micros/op | fillrandom | seekrandom | | --- | --- | --- | |main| 2.58 |10.96| |PR 10661| 2.68 |10.63| Reviewed By: riversand963 Differential Revision: D39441192 Pulled By: cbi42 fbshipit-source-id: f05aca3c41605caf110daf0ff405919f300ddec2
2 years ago
Status DBImpl::DeleteRange(const WriteOptions& write_options,
ColumnFamilyHandle* column_family,
const Slice& begin_key, const Slice& end_key,
const Slice& ts) {
const Status s = FailIfTsMismatchCf(column_family, ts, /*ts_for_read=*/false);
if (!s.ok()) {
return s;
}
return DB::DeleteRange(write_options, column_family, begin_key, end_key, ts);
}
void DBImpl::SetRecoverableStatePreReleaseCallback(
PreReleaseCallback* callback) {
recoverable_state_pre_release_callback_.reset(callback);
}
Status DBImpl::Write(const WriteOptions& write_options, WriteBatch* my_batch) {
Status s;
if (write_options.protection_bytes_per_key > 0) {
s = WriteBatchInternal::UpdateProtectionInfo(
my_batch, write_options.protection_bytes_per_key);
}
if (s.ok()) {
s = WriteImpl(write_options, my_batch, /*callback=*/nullptr,
/*log_used=*/nullptr);
}
return s;
}
Status DBImpl::WriteWithCallback(const WriteOptions& write_options,
WriteBatch* my_batch,
WriteCallback* callback) {
Status s;
if (write_options.protection_bytes_per_key > 0) {
s = WriteBatchInternal::UpdateProtectionInfo(
my_batch, write_options.protection_bytes_per_key);
}
if (s.ok()) {
s = WriteImpl(write_options, my_batch, callback, nullptr);
}
return s;
}
// The main write queue. This is the only write queue that updates LastSequence.
// When using one write queue, the same sequence also indicates the last
// published sequence.
Status DBImpl::WriteImpl(const WriteOptions& write_options,
WriteBatch* my_batch, WriteCallback* callback,
uint64_t* log_used, uint64_t log_ref,
bool disable_memtable, uint64_t* seq_used,
size_t batch_cnt,
Snapshots with user-specified timestamps (#9879) Summary: In RocksDB, keys are associated with (internal) sequence numbers which denote when the keys are written to the database. Sequence numbers in different RocksDB instances are unrelated, thus not comparable. It is nice if we can associate sequence numbers with their corresponding actual timestamps. One thing we can do is to support user-defined timestamp, which allows the applications to specify the format of custom timestamps and encode a timestamp with each key. More details can be found at https://github.com/facebook/rocksdb/wiki/User-defined-Timestamp-%28Experimental%29. This PR provides a different but complementary approach. We can associate rocksdb snapshots (defined in https://github.com/facebook/rocksdb/blob/7.2.fb/include/rocksdb/snapshot.h#L20) with **user-specified** timestamps. Since a snapshot is essentially an object representing a sequence number, this PR establishes a bi-directional mapping between sequence numbers and timestamps. In the past, snapshots are usually taken by readers. The current super-version is grabbed, and a `rocksdb::Snapshot` object is created with the last published sequence number of the super-version. You can see that the reader actually has no good idea of what timestamp to assign to this snapshot, because by the time the `GetSnapshot()` is called, an arbitrarily long period of time may have already elapsed since the last write, which is when the last published sequence number is written. This observation motivates the creation of "timestamped" snapshots on the write path. Currently, this functionality is exposed only to the layer of `TransactionDB`. Application can tell RocksDB to create a snapshot when a transaction commits, effectively associating the last sequence number with a timestamp. It is also assumed that application will ensure any two snapshots with timestamps should satisfy the following: ``` snapshot1.seq < snapshot2.seq iff. snapshot1.ts < snapshot2.ts ``` If the application can guarantee that when a reader takes a timestamped snapshot, there is no active writes going on in the database, then we also allow the user to use a new API `TransactionDB::CreateTimestampedSnapshot()` to create a snapshot with associated timestamp. Code example ```cpp // Create a timestamped snapshot when committing transaction. txn->SetCommitTimestamp(100); txn->SetSnapshotOnNextOperation(); txn->Commit(); // A wrapper API for convenience Status Transaction::CommitAndTryCreateSnapshot( std::shared_ptr<TransactionNotifier> notifier, TxnTimestamp ts, std::shared_ptr<const Snapshot>* ret); // Create a timestamped snapshot if caller guarantees no concurrent writes std::pair<Status, std::shared_ptr<const Snapshot>> snapshot = txn_db->CreateTimestampedSnapshot(100); ``` The snapshots created in this way will be managed by RocksDB with ref-counting and potentially shared with other readers. We provide the following APIs for readers to retrieve a snapshot given a timestamp. ```cpp // Return the timestamped snapshot correponding to given timestamp. If ts is // kMaxTxnTimestamp, then we return the latest timestamped snapshot if present. // Othersise, we return the snapshot whose timestamp is equal to `ts`. If no // such snapshot exists, then we return null. std::shared_ptr<const Snapshot> TransactionDB::GetTimestampedSnapshot(TxnTimestamp ts) const; // Return the latest timestamped snapshot if present. std::shared_ptr<const Snapshot> TransactionDB::GetLatestTimestampedSnapshot() const; ``` We also provide two additional APIs for stats collection and reporting purposes. ```cpp Status TransactionDB::GetAllTimestampedSnapshots( std::vector<std::shared_ptr<const Snapshot>>& snapshots) const; // Return timestamped snapshots whose timestamps fall in [ts_lb, ts_ub) and store them in `snapshots`. Status TransactionDB::GetTimestampedSnapshots( TxnTimestamp ts_lb, TxnTimestamp ts_ub, std::vector<std::shared_ptr<const Snapshot>>& snapshots) const; ``` To prevent the number of timestamped snapshots from growing infinitely, we provide the following API to release timestamped snapshots whose timestamps are older than or equal to a given threshold. ```cpp void TransactionDB::ReleaseTimestampedSnapshotsOlderThan(TxnTimestamp ts); ``` Before shutdown, RocksDB will release all timestamped snapshots. Comparison with user-defined timestamp and how they can be combined: User-defined timestamp persists every key with a timestamp, while timestamped snapshots maintain a volatile mapping between snapshots (sequence numbers) and timestamps. Different internal keys with the same user key but different timestamps will be treated as different by compaction, thus a newer version will not hide older versions (with smaller timestamps) unless they are eligible for garbage collection. In contrast, taking a timestamped snapshot at a certain sequence number and timestamp prevents all the keys visible in this snapshot from been dropped by compaction. Here, visible means (seq < snapshot and most recent). The timestamped snapshot supports the semantics of reading at an exact point in time. Timestamped snapshots can also be used with user-defined timestamp. Pull Request resolved: https://github.com/facebook/rocksdb/pull/9879 Test Plan: ``` make check TEST_TMPDIR=/dev/shm make crash_test_with_txn ``` Reviewed By: siying Differential Revision: D35783919 Pulled By: riversand963 fbshipit-source-id: 586ad905e169189e19d3bfc0cb0177a7239d1bd4
3 years ago
PreReleaseCallback* pre_release_callback,
PostMemTableCallback* post_memtable_callback) {
assert(!seq_per_batch_ || batch_cnt != 0);
assert(my_batch == nullptr || my_batch->Count() == 0 ||
write_options.protection_bytes_per_key == 0 ||
write_options.protection_bytes_per_key ==
my_batch->GetProtectionBytesPerKey());
if (my_batch == nullptr) {
return Status::InvalidArgument("Batch is nullptr!");
Support user-defined timestamps in write-committed txns (#9629) Summary: Pull Request resolved: https://github.com/facebook/rocksdb/pull/9629 Pessimistic transactions use pessimistic concurrency control, i.e. locking. Keys are locked upon first operation that writes the key or has the intention of writing. For example, `PessimisticTransaction::Put()`, `PessimisticTransaction::Delete()`, `PessimisticTransaction::SingleDelete()` will write to or delete a key, while `PessimisticTransaction::GetForUpdate()` is used by application to indicate to RocksDB that the transaction has the intention of performing write operation later in the same transaction. Pessimistic transactions support two-phase commit (2PC). A transaction can be `Prepared()`'ed and then `Commit()`. The prepare phase is similar to a promise: once `Prepare()` succeeds, the transaction has acquired the necessary resources to commit. The resources include locks, persistence of WAL, etc. Write-committed transaction is the default pessimistic transaction implementation. In RocksDB write-committed transaction, `Prepare()` will write data to the WAL as a prepare section. `Commit()` will write a commit marker to the WAL and then write data to the memtables. While writing to the memtables, different keys in the transaction's write batch will be assigned different sequence numbers in ascending order. Until commit/rollback, the transaction holds locks on the keys so that no other transaction can write to the same keys. Furthermore, the keys' sequence numbers represent the order in which they are committed and should be made visible. This is convenient for us to implement support for user-defined timestamps. Since column families with and without timestamps can co-exist in the same database, a transaction may or may not involve timestamps. Based on this observation, we add two optional members to each `PessimisticTransaction`, `read_timestamp_` and `commit_timestamp_`. If no key in the transaction's write batch has timestamp, then setting these two variables do not have any effect. For the rest of this commit, we discuss only the cases when these two variables are meaningful. read_timestamp_ is used mainly for validation, and should be set before first call to `GetForUpdate()`. Otherwise, the latter will return non-ok status. `GetForUpdate()` calls `TryLock()` that can verify if another transaction has written the same key since `read_timestamp_` till this call to `GetForUpdate()`. If another transaction has indeed written the same key, then validation fails, and RocksDB allows this transaction to refine `read_timestamp_` by increasing it. Note that a transaction can still use `Get()` with a different timestamp to read, but the result of the read should not be used to determine data that will be written later. commit_timestamp_ must be set after finishing writing and before transaction commit. This applies to both 2PC and non-2PC cases. In the case of 2PC, it's usually set after prepare phase succeeds. We currently require that the commit timestamp be chosen after all keys are locked. This means we disallow the `TransactionDB`-level APIs if user-defined timestamp is used by the transaction. Specifically, calling `PessimisticTransactionDB::Put()`, `PessimisticTransactionDB::Delete()`, `PessimisticTransactionDB::SingleDelete()`, etc. will return non-ok status because they specify timestamps before locking the keys. Users are also prompted to use the `Transaction` APIs when they receive the non-ok status. Reviewed By: ltamasi Differential Revision: D31822445 fbshipit-source-id: b82abf8e230216dc89cc519564a588224a88fd43
3 years ago
} else if (!disable_memtable &&
WriteBatchInternal::TimestampsUpdateNeeded(*my_batch)) {
// If writing to memtable, then we require the caller to set/update the
// timestamps for the keys in the write batch.
// Otherwise, it means we are just writing to the WAL, and we allow
// timestamps unset for the keys in the write batch. This can happen if we
// use TransactionDB with write-committed policy, and we currently do not
// support user-defined timestamp with other policies.
// In the prepare phase, a transaction can write the batch to the WAL
// without inserting to memtable. The keys in the batch do not have to be
// assigned timestamps because they will be used only during recovery if
// there is a commit marker which includes their commit timestamp.
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
return Status::InvalidArgument("write batch must have timestamp(s) set");
Rate-limit automatic WAL flush after each user write (#9607) Summary: **Context:** WAL flush is currently not rate-limited by `Options::rate_limiter`. This PR is to provide rate-limiting to auto WAL flush, the one that automatically happen after each user write operation (i.e, `Options::manual_wal_flush == false`), by adding `WriteOptions::rate_limiter_options`. Note that we are NOT rate-limiting WAL flush that do NOT automatically happen after each user write, such as `Options::manual_wal_flush == true + manual FlushWAL()` (rate-limiting multiple WAL flushes), for the benefits of: - being consistent with [ReadOptions::rate_limiter_priority](https://github.com/facebook/rocksdb/blob/7.0.fb/include/rocksdb/options.h#L515) - being able to turn off some WAL flush's rate-limiting but not all (e.g, turn off specific the WAL flush of a critical user write like a service's heartbeat) `WriteOptions::rate_limiter_options` only accept `Env::IO_USER` and `Env::IO_TOTAL` currently due to an implementation constraint. - The constraint is that we currently queue parallel writes (including WAL writes) based on FIFO policy which does not factor rate limiter priority into this layer's scheduling. If we allow lower priorities such as `Env::IO_HIGH/MID/LOW` and such writes specified with lower priorities occurs before ones specified with higher priorities (even just by a tiny bit in arrival time), the former would have blocked the latter, leading to a "priority inversion" issue and contradictory to what we promise for rate-limiting priority. Therefore we only allow `Env::IO_USER` and `Env::IO_TOTAL` right now before improving that scheduling. A pre-requisite to this feature is to support operation-level rate limiting in `WritableFileWriter`, which is also included in this PR. **Summary:** - Renamed test suite `DBRateLimiterTest to DBRateLimiterOnReadTest` for adding a new test suite - Accept `rate_limiter_priority` in `WritableFileWriter`'s private and public write functions - Passed `WriteOptions::rate_limiter_options` to `WritableFileWriter` in the path of automatic WAL flush. Pull Request resolved: https://github.com/facebook/rocksdb/pull/9607 Test Plan: - Added new unit test to verify existing flush/compaction rate-limiting does not break, since `DBTest, RateLimitingTest` is disabled and current db-level rate-limiting tests focus on read only (e.g, `db_rate_limiter_test`, `DBTest2, RateLimitedCompactionReads`). - Added new unit test `DBRateLimiterOnWriteWALTest, AutoWalFlush` - `strace -ftt -e trace=write ./db_bench -benchmarks=fillseq -db=/dev/shm/testdb -rate_limit_auto_wal_flush=1 -rate_limiter_bytes_per_sec=15 -rate_limiter_refill_period_us=1000000 -write_buffer_size=100000000 -disable_auto_compactions=1 -num=100` - verified that WAL flush(i.e, system-call _write_) were chunked into 15 bytes and each _write_ was roughly 1 second apart - verified the chunking disappeared when `-rate_limit_auto_wal_flush=0` - crash test: `python3 tools/db_crashtest.py blackbox --disable_wal=0 --rate_limit_auto_wal_flush=1 --rate_limiter_bytes_per_sec=10485760 --interval=10` killed as normal **Benchmarked on flush/compaction to ensure no performance regression:** - compaction with rate-limiting (see table 1, avg over 1280-run): pre-change: **915635 micros/op**; post-change: **907350 micros/op (improved by 0.106%)** ``` #!/bin/bash TEST_TMPDIR=/dev/shm/testdb START=1 NUM_DATA_ENTRY=8 N=10 rm -f compact_bmk_output.txt compact_bmk_output_2.txt dont_care_output.txt for i in $(eval echo "{$START..$NUM_DATA_ENTRY}") do NUM_RUN=$(($N*(2**($i-1)))) for j in $(eval echo "{$START..$NUM_RUN}") do ./db_bench --benchmarks=fillrandom -db=$TEST_TMPDIR -disable_auto_compactions=1 -write_buffer_size=6710886 > dont_care_output.txt && ./db_bench --benchmarks=compact -use_existing_db=1 -db=$TEST_TMPDIR -level0_file_num_compaction_trigger=1 -rate_limiter_bytes_per_sec=100000000 | egrep 'compact' done > compact_bmk_output.txt && awk -v NUM_RUN=$NUM_RUN '{sum+=$3;sum_sqrt+=$3^2}END{print sum/NUM_RUN, sqrt(sum_sqrt/NUM_RUN-(sum/NUM_RUN)^2)}' compact_bmk_output.txt >> compact_bmk_output_2.txt done ``` - compaction w/o rate-limiting (see table 2, avg over 640-run): pre-change: **822197 micros/op**; post-change: **823148 micros/op (regressed by 0.12%)** ``` Same as above script, except that -rate_limiter_bytes_per_sec=0 ``` - flush with rate-limiting (see table 3, avg over 320-run, run on the [patch](https://github.com/hx235/rocksdb/commit/ee5c6023a9f6533fab9afdc681568daa21da4953) to augment current db_bench ): pre-change: **745752 micros/op**; post-change: **745331 micros/op (regressed by 0.06 %)** ``` #!/bin/bash TEST_TMPDIR=/dev/shm/testdb START=1 NUM_DATA_ENTRY=8 N=10 rm -f flush_bmk_output.txt flush_bmk_output_2.txt for i in $(eval echo "{$START..$NUM_DATA_ENTRY}") do NUM_RUN=$(($N*(2**($i-1)))) for j in $(eval echo "{$START..$NUM_RUN}") do ./db_bench -db=$TEST_TMPDIR -write_buffer_size=1048576000 -num=1000000 -rate_limiter_bytes_per_sec=100000000 -benchmarks=fillseq,flush | egrep 'flush' done > flush_bmk_output.txt && awk -v NUM_RUN=$NUM_RUN '{sum+=$3;sum_sqrt+=$3^2}END{print sum/NUM_RUN, sqrt(sum_sqrt/NUM_RUN-(sum/NUM_RUN)^2)}' flush_bmk_output.txt >> flush_bmk_output_2.txt done ``` - flush w/o rate-limiting (see table 4, avg over 320-run, run on the [patch](https://github.com/hx235/rocksdb/commit/ee5c6023a9f6533fab9afdc681568daa21da4953) to augment current db_bench): pre-change: **487512 micros/op**, post-change: **485856 micors/ops (improved by 0.34%)** ``` Same as above script, except that -rate_limiter_bytes_per_sec=0 ``` | table 1 - compact with rate-limiting| #-run | (pre-change) avg micros/op | std micros/op | (post-change) avg micros/op | std micros/op | change in avg micros/op (%) -- | -- | -- | -- | -- | -- 10 | 896978 | 16046.9 | 901242 | 15670.9 | 0.475373978 20 | 893718 | 15813 | 886505 | 17544.7 | -0.8070778478 40 | 900426 | 23882.2 | 894958 | 15104.5 | -0.6072681153 80 | 906635 | 21761.5 | 903332 | 23948.3 | -0.3643141948 160 | 898632 | 21098.9 | 907583 | 21145 | 0.9960695813 3.20E+02 | 905252 | 22785.5 | 908106 | 25325.5 | 0.3152713278 6.40E+02 | 905213 | 23598.6 | 906741 | 21370.5 | 0.1688000504 **1.28E+03** | **908316** | **23533.1** | **907350** | **24626.8** | **-0.1063506533** average over #-run | 901896.25 | 21064.9625 | 901977.125 | 20592.025 | 0.008967217682 | table 2 - compact w/o rate-limiting| #-run | (pre-change) avg micros/op | std micros/op | (post-change) avg micros/op | std micros/op | change in avg micros/op (%) -- | -- | -- | -- | -- | -- 10 | 811211 | 26996.7 | 807586 | 28456.4 | -0.4468627768 20 | 815465 | 14803.7 | 814608 | 28719.7 | -0.105093413 40 | 809203 | 26187.1 | 797835 | 25492.1 | -1.404839082 80 | 822088 | 28765.3 | 822192 | 32840.4 | 0.01265071379 160 | 821719 | 36344.7 | 821664 | 29544.9 | -0.006693285661 3.20E+02 | 820921 | 27756.4 | 821403 | 28347.7 | 0.05871454135 **6.40E+02** | **822197** | **28960.6** | **823148** | **30055.1** | **0.1156657103** average over #-run | 8.18E+05 | 2.71E+04 | 8.15E+05 | 2.91E+04 | -0.25 | table 3 - flush with rate-limiting| #-run | (pre-change) avg micros/op | std micros/op | (post-change) avg micros/op | std micros/op | change in avg micros/op (%) -- | -- | -- | -- | -- | -- 10 | 741721 | 11770.8 | 740345 | 5949.76 | -0.1855144994 20 | 735169 | 3561.83 | 743199 | 9755.77 | 1.09226586 40 | 743368 | 8891.03 | 742102 | 8683.22 | -0.1703059588 80 | 742129 | 8148.51 | 743417 | 9631.58| 0.1735547324 160 | 749045 | 9757.21 | 746256 | 9191.86 | -0.3723407806 **3.20E+02** | **745752** | **9819.65** | **745331** | **9840.62** | **-0.0564530836** 6.40E+02 | 749006 | 11080.5 | 748173 | 10578.7 | -0.1112140624 average over #-run | 743741.4286 | 9004.218571 | 744117.5714 | 9090.215714 | 0.05057441238 | table 4 - flush w/o rate-limiting| #-run | (pre-change) avg micros/op | std micros/op | (post-change) avg micros/op | std micros/op | change in avg micros/op (%) -- | -- | -- | -- | -- | -- 10 | 477283 | 24719.6 | 473864 | 12379 | -0.7163464863 20 | 486743 | 20175.2 | 502296 | 23931.3 | 3.195320734 40 | 482846 | 15309.2 | 489820 | 22259.5 | 1.444352858 80 | 491490 | 21883.1 | 490071 | 23085.7 | -0.2887139108 160 | 493347 | 28074.3 | 483609 | 21211.7 | -1.973864238 **3.20E+02** | **487512** | **21401.5** | **485856** | **22195.2** | **-0.3396839462** 6.40E+02 | 490307 | 25418.6 | 485435 | 22405.2 | -0.9936631539 average over #-run | 4.87E+05 | 2.24E+04 | 4.87E+05 | 2.11E+04 | 0.00E+00 Reviewed By: ajkr Differential Revision: D34442441 Pulled By: hx235 fbshipit-source-id: 4790f13e1e5c0a95ae1d1cc93ffcf69dc6e78bdd
3 years ago
} else if (write_options.rate_limiter_priority != Env::IO_TOTAL &&
write_options.rate_limiter_priority != Env::IO_USER) {
return Status::InvalidArgument(
"WriteOptions::rate_limiter_priority only allows "
"Env::IO_TOTAL and Env::IO_USER due to implementation constraints");
} else if (write_options.rate_limiter_priority != Env::IO_TOTAL &&
(write_options.disableWAL || manual_wal_flush_)) {
return Status::InvalidArgument(
"WriteOptions::rate_limiter_priority currently only supports "
"rate-limiting automatic WAL flush, which requires "
"`WriteOptions::disableWAL` and "
"`DBOptions::manual_wal_flush` both set to false");
} else if (write_options.protection_bytes_per_key != 0 &&
write_options.protection_bytes_per_key != 8) {
return Status::InvalidArgument(
"`WriteOptions::protection_bytes_per_key` must be zero or eight");
}
// TODO: this use of operator bool on `tracer_` can avoid unnecessary lock
// grabs but does not seem thread-safe.
if (tracer_) {
InstrumentedMutexLock lock(&trace_mutex_);
if (tracer_ && !tracer_->IsWriteOrderPreserved()) {
// We don't have to preserve write order so can trace anywhere. It's more
// efficient to trace here than to add latency to a phase of the log/apply
// pipeline.
// TODO: maybe handle the tracing status?
tracer_->Write(my_batch).PermitUncheckedError();
}
}
if (write_options.sync && write_options.disableWAL) {
return Status::InvalidArgument("Sync writes has to enable WAL.");
}
if (two_write_queues_ && immutable_db_options_.enable_pipelined_write) {
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
return Status::NotSupported(
"pipelined_writes is not compatible with concurrent prepares");
}
if (seq_per_batch_ && immutable_db_options_.enable_pipelined_write) {
// TODO(yiwu): update pipeline write with seq_per_batch and batch_cnt
return Status::NotSupported(
"pipelined_writes is not compatible with seq_per_batch");
}
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
if (immutable_db_options_.unordered_write &&
immutable_db_options_.enable_pipelined_write) {
return Status::NotSupported(
"pipelined_writes is not compatible with unordered_write");
}
Snapshots with user-specified timestamps (#9879) Summary: In RocksDB, keys are associated with (internal) sequence numbers which denote when the keys are written to the database. Sequence numbers in different RocksDB instances are unrelated, thus not comparable. It is nice if we can associate sequence numbers with their corresponding actual timestamps. One thing we can do is to support user-defined timestamp, which allows the applications to specify the format of custom timestamps and encode a timestamp with each key. More details can be found at https://github.com/facebook/rocksdb/wiki/User-defined-Timestamp-%28Experimental%29. This PR provides a different but complementary approach. We can associate rocksdb snapshots (defined in https://github.com/facebook/rocksdb/blob/7.2.fb/include/rocksdb/snapshot.h#L20) with **user-specified** timestamps. Since a snapshot is essentially an object representing a sequence number, this PR establishes a bi-directional mapping between sequence numbers and timestamps. In the past, snapshots are usually taken by readers. The current super-version is grabbed, and a `rocksdb::Snapshot` object is created with the last published sequence number of the super-version. You can see that the reader actually has no good idea of what timestamp to assign to this snapshot, because by the time the `GetSnapshot()` is called, an arbitrarily long period of time may have already elapsed since the last write, which is when the last published sequence number is written. This observation motivates the creation of "timestamped" snapshots on the write path. Currently, this functionality is exposed only to the layer of `TransactionDB`. Application can tell RocksDB to create a snapshot when a transaction commits, effectively associating the last sequence number with a timestamp. It is also assumed that application will ensure any two snapshots with timestamps should satisfy the following: ``` snapshot1.seq < snapshot2.seq iff. snapshot1.ts < snapshot2.ts ``` If the application can guarantee that when a reader takes a timestamped snapshot, there is no active writes going on in the database, then we also allow the user to use a new API `TransactionDB::CreateTimestampedSnapshot()` to create a snapshot with associated timestamp. Code example ```cpp // Create a timestamped snapshot when committing transaction. txn->SetCommitTimestamp(100); txn->SetSnapshotOnNextOperation(); txn->Commit(); // A wrapper API for convenience Status Transaction::CommitAndTryCreateSnapshot( std::shared_ptr<TransactionNotifier> notifier, TxnTimestamp ts, std::shared_ptr<const Snapshot>* ret); // Create a timestamped snapshot if caller guarantees no concurrent writes std::pair<Status, std::shared_ptr<const Snapshot>> snapshot = txn_db->CreateTimestampedSnapshot(100); ``` The snapshots created in this way will be managed by RocksDB with ref-counting and potentially shared with other readers. We provide the following APIs for readers to retrieve a snapshot given a timestamp. ```cpp // Return the timestamped snapshot correponding to given timestamp. If ts is // kMaxTxnTimestamp, then we return the latest timestamped snapshot if present. // Othersise, we return the snapshot whose timestamp is equal to `ts`. If no // such snapshot exists, then we return null. std::shared_ptr<const Snapshot> TransactionDB::GetTimestampedSnapshot(TxnTimestamp ts) const; // Return the latest timestamped snapshot if present. std::shared_ptr<const Snapshot> TransactionDB::GetLatestTimestampedSnapshot() const; ``` We also provide two additional APIs for stats collection and reporting purposes. ```cpp Status TransactionDB::GetAllTimestampedSnapshots( std::vector<std::shared_ptr<const Snapshot>>& snapshots) const; // Return timestamped snapshots whose timestamps fall in [ts_lb, ts_ub) and store them in `snapshots`. Status TransactionDB::GetTimestampedSnapshots( TxnTimestamp ts_lb, TxnTimestamp ts_ub, std::vector<std::shared_ptr<const Snapshot>>& snapshots) const; ``` To prevent the number of timestamped snapshots from growing infinitely, we provide the following API to release timestamped snapshots whose timestamps are older than or equal to a given threshold. ```cpp void TransactionDB::ReleaseTimestampedSnapshotsOlderThan(TxnTimestamp ts); ``` Before shutdown, RocksDB will release all timestamped snapshots. Comparison with user-defined timestamp and how they can be combined: User-defined timestamp persists every key with a timestamp, while timestamped snapshots maintain a volatile mapping between snapshots (sequence numbers) and timestamps. Different internal keys with the same user key but different timestamps will be treated as different by compaction, thus a newer version will not hide older versions (with smaller timestamps) unless they are eligible for garbage collection. In contrast, taking a timestamped snapshot at a certain sequence number and timestamp prevents all the keys visible in this snapshot from been dropped by compaction. Here, visible means (seq < snapshot and most recent). The timestamped snapshot supports the semantics of reading at an exact point in time. Timestamped snapshots can also be used with user-defined timestamp. Pull Request resolved: https://github.com/facebook/rocksdb/pull/9879 Test Plan: ``` make check TEST_TMPDIR=/dev/shm make crash_test_with_txn ``` Reviewed By: siying Differential Revision: D35783919 Pulled By: riversand963 fbshipit-source-id: 586ad905e169189e19d3bfc0cb0177a7239d1bd4
3 years ago
if (immutable_db_options_.enable_pipelined_write &&
post_memtable_callback != nullptr) {
return Status::NotSupported(
"pipelined write currently does not honor post_memtable_callback");
}
if (seq_per_batch_ && post_memtable_callback != nullptr) {
return Status::NotSupported(
"seq_per_batch currently does not honor post_memtable_callback");
}
// Otherwise IsLatestPersistentState optimization does not make sense
assert(!WriteBatchInternal::IsLatestPersistentState(my_batch) ||
disable_memtable);
if (write_options.low_pri) {
Status s = ThrottleLowPriWritesIfNeeded(write_options, my_batch);
if (!s.ok()) {
return s;
}
}
if (two_write_queues_ && disable_memtable) {
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
AssignOrder assign_order =
seq_per_batch_ ? kDoAssignOrder : kDontAssignOrder;
// Otherwise it is WAL-only Prepare batches in WriteCommitted policy and
// they don't consume sequence.
return WriteImplWALOnly(&nonmem_write_thread_, write_options, my_batch,
callback, log_used, log_ref, seq_used, batch_cnt,
pre_release_callback, assign_order,
kDontPublishLastSeq, disable_memtable);
}
if (immutable_db_options_.unordered_write) {
const size_t sub_batch_cnt = batch_cnt != 0
? batch_cnt
// every key is a sub-batch consuming a seq
: WriteBatchInternal::Count(my_batch);
uint64_t seq = 0;
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
// Use a write thread to i) optimize for WAL write, ii) publish last
// sequence in in increasing order, iii) call pre_release_callback serially
Status status = WriteImplWALOnly(
&write_thread_, write_options, my_batch, callback, log_used, log_ref,
&seq, sub_batch_cnt, pre_release_callback, kDoAssignOrder,
kDoPublishLastSeq, disable_memtable);
TEST_SYNC_POINT("DBImpl::WriteImpl:UnorderedWriteAfterWriteWAL");
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
if (!status.ok()) {
return status;
}
if (seq_used) {
*seq_used = seq;
}
if (!disable_memtable) {
TEST_SYNC_POINT("DBImpl::WriteImpl:BeforeUnorderedWriteMemtable");
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
status = UnorderedWriteMemtable(write_options, my_batch, callback,
log_ref, seq, sub_batch_cnt);
}
return status;
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
}
if (immutable_db_options_.enable_pipelined_write) {
return PipelinedWriteImpl(write_options, my_batch, callback, log_used,
log_ref, disable_memtable, seq_used);
}
PERF_TIMER_GUARD(write_pre_and_post_process_time);
WriteThread::Writer w(write_options, my_batch, callback, log_ref,
Snapshots with user-specified timestamps (#9879) Summary: In RocksDB, keys are associated with (internal) sequence numbers which denote when the keys are written to the database. Sequence numbers in different RocksDB instances are unrelated, thus not comparable. It is nice if we can associate sequence numbers with their corresponding actual timestamps. One thing we can do is to support user-defined timestamp, which allows the applications to specify the format of custom timestamps and encode a timestamp with each key. More details can be found at https://github.com/facebook/rocksdb/wiki/User-defined-Timestamp-%28Experimental%29. This PR provides a different but complementary approach. We can associate rocksdb snapshots (defined in https://github.com/facebook/rocksdb/blob/7.2.fb/include/rocksdb/snapshot.h#L20) with **user-specified** timestamps. Since a snapshot is essentially an object representing a sequence number, this PR establishes a bi-directional mapping between sequence numbers and timestamps. In the past, snapshots are usually taken by readers. The current super-version is grabbed, and a `rocksdb::Snapshot` object is created with the last published sequence number of the super-version. You can see that the reader actually has no good idea of what timestamp to assign to this snapshot, because by the time the `GetSnapshot()` is called, an arbitrarily long period of time may have already elapsed since the last write, which is when the last published sequence number is written. This observation motivates the creation of "timestamped" snapshots on the write path. Currently, this functionality is exposed only to the layer of `TransactionDB`. Application can tell RocksDB to create a snapshot when a transaction commits, effectively associating the last sequence number with a timestamp. It is also assumed that application will ensure any two snapshots with timestamps should satisfy the following: ``` snapshot1.seq < snapshot2.seq iff. snapshot1.ts < snapshot2.ts ``` If the application can guarantee that when a reader takes a timestamped snapshot, there is no active writes going on in the database, then we also allow the user to use a new API `TransactionDB::CreateTimestampedSnapshot()` to create a snapshot with associated timestamp. Code example ```cpp // Create a timestamped snapshot when committing transaction. txn->SetCommitTimestamp(100); txn->SetSnapshotOnNextOperation(); txn->Commit(); // A wrapper API for convenience Status Transaction::CommitAndTryCreateSnapshot( std::shared_ptr<TransactionNotifier> notifier, TxnTimestamp ts, std::shared_ptr<const Snapshot>* ret); // Create a timestamped snapshot if caller guarantees no concurrent writes std::pair<Status, std::shared_ptr<const Snapshot>> snapshot = txn_db->CreateTimestampedSnapshot(100); ``` The snapshots created in this way will be managed by RocksDB with ref-counting and potentially shared with other readers. We provide the following APIs for readers to retrieve a snapshot given a timestamp. ```cpp // Return the timestamped snapshot correponding to given timestamp. If ts is // kMaxTxnTimestamp, then we return the latest timestamped snapshot if present. // Othersise, we return the snapshot whose timestamp is equal to `ts`. If no // such snapshot exists, then we return null. std::shared_ptr<const Snapshot> TransactionDB::GetTimestampedSnapshot(TxnTimestamp ts) const; // Return the latest timestamped snapshot if present. std::shared_ptr<const Snapshot> TransactionDB::GetLatestTimestampedSnapshot() const; ``` We also provide two additional APIs for stats collection and reporting purposes. ```cpp Status TransactionDB::GetAllTimestampedSnapshots( std::vector<std::shared_ptr<const Snapshot>>& snapshots) const; // Return timestamped snapshots whose timestamps fall in [ts_lb, ts_ub) and store them in `snapshots`. Status TransactionDB::GetTimestampedSnapshots( TxnTimestamp ts_lb, TxnTimestamp ts_ub, std::vector<std::shared_ptr<const Snapshot>>& snapshots) const; ``` To prevent the number of timestamped snapshots from growing infinitely, we provide the following API to release timestamped snapshots whose timestamps are older than or equal to a given threshold. ```cpp void TransactionDB::ReleaseTimestampedSnapshotsOlderThan(TxnTimestamp ts); ``` Before shutdown, RocksDB will release all timestamped snapshots. Comparison with user-defined timestamp and how they can be combined: User-defined timestamp persists every key with a timestamp, while timestamped snapshots maintain a volatile mapping between snapshots (sequence numbers) and timestamps. Different internal keys with the same user key but different timestamps will be treated as different by compaction, thus a newer version will not hide older versions (with smaller timestamps) unless they are eligible for garbage collection. In contrast, taking a timestamped snapshot at a certain sequence number and timestamp prevents all the keys visible in this snapshot from been dropped by compaction. Here, visible means (seq < snapshot and most recent). The timestamped snapshot supports the semantics of reading at an exact point in time. Timestamped snapshots can also be used with user-defined timestamp. Pull Request resolved: https://github.com/facebook/rocksdb/pull/9879 Test Plan: ``` make check TEST_TMPDIR=/dev/shm make crash_test_with_txn ``` Reviewed By: siying Differential Revision: D35783919 Pulled By: riversand963 fbshipit-source-id: 586ad905e169189e19d3bfc0cb0177a7239d1bd4
3 years ago
disable_memtable, batch_cnt, pre_release_callback,
post_memtable_callback);
StopWatch write_sw(immutable_db_options_.clock, stats_, DB_WRITE);
write_thread_.JoinBatchGroup(&w);
if (w.state == WriteThread::STATE_PARALLEL_MEMTABLE_WRITER) {
// we are a non-leader in a parallel group
if (w.ShouldWriteToMemtable()) {
PERF_TIMER_STOP(write_pre_and_post_process_time);
PERF_TIMER_GUARD(write_memtable_time);
ColumnFamilyMemTablesImpl column_family_memtables(
versions_->GetColumnFamilySet());
w.status = WriteBatchInternal::InsertInto(
&w, w.sequence, &column_family_memtables, &flush_scheduler_,
Refactor trimming logic for immutable memtables (#5022) Summary: MyRocks currently sets `max_write_buffer_number_to_maintain` in order to maintain enough history for transaction conflict checking. The effectiveness of this approach depends on the size of memtables. When memtables are small, it may not keep enough history; when memtables are large, this may consume too much memory. We are proposing a new way to configure memtable list history: by limiting the memory usage of immutable memtables. The new option is `max_write_buffer_size_to_maintain` and it will take precedence over the old `max_write_buffer_number_to_maintain` if they are both set to non-zero values. The new option accounts for the total memory usage of flushed immutable memtables and mutable memtable. When the total usage exceeds the limit, RocksDB may start dropping immutable memtables (which is also called trimming history), starting from the oldest one. The semantics of the old option actually works both as an upper bound and lower bound. History trimming will start if number of immutable memtables exceeds the limit, but it will never go below (limit-1) due to history trimming. In order the mimic the behavior with the new option, history trimming will stop if dropping the next immutable memtable causes the total memory usage go below the size limit. For example, assuming the size limit is set to 64MB, and there are 3 immutable memtables with sizes of 20, 30, 30. Although the total memory usage is 80MB > 64MB, dropping the oldest memtable will reduce the memory usage to 60MB < 64MB, so in this case no memtable will be dropped. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5022 Differential Revision: D14394062 Pulled By: miasantreble fbshipit-source-id: 60457a509c6af89d0993f988c9b5c2aa9e45f5c5
5 years ago
&trim_history_scheduler_,
write_options.ignore_missing_column_families, 0 /*log_number*/, this,
true /*concurrent_memtable_writes*/, seq_per_batch_, w.batch_cnt,
batch_per_txn_, write_options.memtable_insert_hint_per_batch);
PERF_TIMER_START(write_pre_and_post_process_time);
}
if (write_thread_.CompleteParallelMemTableWriter(&w)) {
// we're responsible for exit batch group
// TODO(myabandeh): propagate status to write_group
auto last_sequence = w.write_group->last_sequence;
Snapshots with user-specified timestamps (#9879) Summary: In RocksDB, keys are associated with (internal) sequence numbers which denote when the keys are written to the database. Sequence numbers in different RocksDB instances are unrelated, thus not comparable. It is nice if we can associate sequence numbers with their corresponding actual timestamps. One thing we can do is to support user-defined timestamp, which allows the applications to specify the format of custom timestamps and encode a timestamp with each key. More details can be found at https://github.com/facebook/rocksdb/wiki/User-defined-Timestamp-%28Experimental%29. This PR provides a different but complementary approach. We can associate rocksdb snapshots (defined in https://github.com/facebook/rocksdb/blob/7.2.fb/include/rocksdb/snapshot.h#L20) with **user-specified** timestamps. Since a snapshot is essentially an object representing a sequence number, this PR establishes a bi-directional mapping between sequence numbers and timestamps. In the past, snapshots are usually taken by readers. The current super-version is grabbed, and a `rocksdb::Snapshot` object is created with the last published sequence number of the super-version. You can see that the reader actually has no good idea of what timestamp to assign to this snapshot, because by the time the `GetSnapshot()` is called, an arbitrarily long period of time may have already elapsed since the last write, which is when the last published sequence number is written. This observation motivates the creation of "timestamped" snapshots on the write path. Currently, this functionality is exposed only to the layer of `TransactionDB`. Application can tell RocksDB to create a snapshot when a transaction commits, effectively associating the last sequence number with a timestamp. It is also assumed that application will ensure any two snapshots with timestamps should satisfy the following: ``` snapshot1.seq < snapshot2.seq iff. snapshot1.ts < snapshot2.ts ``` If the application can guarantee that when a reader takes a timestamped snapshot, there is no active writes going on in the database, then we also allow the user to use a new API `TransactionDB::CreateTimestampedSnapshot()` to create a snapshot with associated timestamp. Code example ```cpp // Create a timestamped snapshot when committing transaction. txn->SetCommitTimestamp(100); txn->SetSnapshotOnNextOperation(); txn->Commit(); // A wrapper API for convenience Status Transaction::CommitAndTryCreateSnapshot( std::shared_ptr<TransactionNotifier> notifier, TxnTimestamp ts, std::shared_ptr<const Snapshot>* ret); // Create a timestamped snapshot if caller guarantees no concurrent writes std::pair<Status, std::shared_ptr<const Snapshot>> snapshot = txn_db->CreateTimestampedSnapshot(100); ``` The snapshots created in this way will be managed by RocksDB with ref-counting and potentially shared with other readers. We provide the following APIs for readers to retrieve a snapshot given a timestamp. ```cpp // Return the timestamped snapshot correponding to given timestamp. If ts is // kMaxTxnTimestamp, then we return the latest timestamped snapshot if present. // Othersise, we return the snapshot whose timestamp is equal to `ts`. If no // such snapshot exists, then we return null. std::shared_ptr<const Snapshot> TransactionDB::GetTimestampedSnapshot(TxnTimestamp ts) const; // Return the latest timestamped snapshot if present. std::shared_ptr<const Snapshot> TransactionDB::GetLatestTimestampedSnapshot() const; ``` We also provide two additional APIs for stats collection and reporting purposes. ```cpp Status TransactionDB::GetAllTimestampedSnapshots( std::vector<std::shared_ptr<const Snapshot>>& snapshots) const; // Return timestamped snapshots whose timestamps fall in [ts_lb, ts_ub) and store them in `snapshots`. Status TransactionDB::GetTimestampedSnapshots( TxnTimestamp ts_lb, TxnTimestamp ts_ub, std::vector<std::shared_ptr<const Snapshot>>& snapshots) const; ``` To prevent the number of timestamped snapshots from growing infinitely, we provide the following API to release timestamped snapshots whose timestamps are older than or equal to a given threshold. ```cpp void TransactionDB::ReleaseTimestampedSnapshotsOlderThan(TxnTimestamp ts); ``` Before shutdown, RocksDB will release all timestamped snapshots. Comparison with user-defined timestamp and how they can be combined: User-defined timestamp persists every key with a timestamp, while timestamped snapshots maintain a volatile mapping between snapshots (sequence numbers) and timestamps. Different internal keys with the same user key but different timestamps will be treated as different by compaction, thus a newer version will not hide older versions (with smaller timestamps) unless they are eligible for garbage collection. In contrast, taking a timestamped snapshot at a certain sequence number and timestamp prevents all the keys visible in this snapshot from been dropped by compaction. Here, visible means (seq < snapshot and most recent). The timestamped snapshot supports the semantics of reading at an exact point in time. Timestamped snapshots can also be used with user-defined timestamp. Pull Request resolved: https://github.com/facebook/rocksdb/pull/9879 Test Plan: ``` make check TEST_TMPDIR=/dev/shm make crash_test_with_txn ``` Reviewed By: siying Differential Revision: D35783919 Pulled By: riversand963 fbshipit-source-id: 586ad905e169189e19d3bfc0cb0177a7239d1bd4
3 years ago
for (auto* tmp_w : *(w.write_group)) {
assert(tmp_w);
if (tmp_w->post_memtable_callback) {
Status tmp_s =
(*tmp_w->post_memtable_callback)(last_sequence, disable_memtable);
// TODO: propagate the execution status of post_memtable_callback to
// caller.
assert(tmp_s.ok());
}
}
versions_->SetLastSequence(last_sequence);
MemTableInsertStatusCheck(w.status);
write_thread_.ExitAsBatchGroupFollower(&w);
}
assert(w.state == WriteThread::STATE_COMPLETED);
// STATE_COMPLETED conditional below handles exit
}
if (w.state == WriteThread::STATE_COMPLETED) {
if (log_used != nullptr) {
*log_used = w.log_used;
}
if (seq_used != nullptr) {
*seq_used = w.sequence;
}
// write is complete and leader has updated sequence
return w.FinalStatus();
}
// else we are the leader of the write batch group
assert(w.state == WriteThread::STATE_GROUP_LEADER);
Status status;
// Once reaches this point, the current writer "w" will try to do its write
// job. It may also pick up some of the remaining writers in the "writers_"
// when it finds suitable, and finish them in the same write batch.
// This is how a write job could be done by the other writer.
WriteContext write_context;
LogContext log_context(write_options.sync);
WriteThread::WriteGroup write_group;
bool in_parallel_group = false;
uint64_t last_sequence = kMaxSequenceNumber;
assert(!two_write_queues_ || !disable_memtable);
{
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
// With concurrent writes we do preprocess only in the write thread that
// also does write to memtable to avoid sync issue on shared data structure
// with the other thread
// PreprocessWrite does its own perf timing.
PERF_TIMER_STOP(write_pre_and_post_process_time);
status = PreprocessWrite(write_options, &log_context, &write_context);
if (!two_write_queues_) {
// Assign it after ::PreprocessWrite since the sequence might advance
// inside it by WriteRecoverableState
last_sequence = versions_->LastSequence();
}
PERF_TIMER_START(write_pre_and_post_process_time);
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
}
// Add to log and apply to memtable. We can release the lock
// during this phase since &w is currently responsible for logging
// and protects against concurrent loggers and concurrent writes
// into memtables
TEST_SYNC_POINT("DBImpl::WriteImpl:BeforeLeaderEnters");
last_batch_group_size_ =
write_thread_.EnterAsBatchGroupLeader(&w, &write_group);
IOStatus io_s;
Status pre_release_cb_status;
if (status.ok()) {
// TODO: this use of operator bool on `tracer_` can avoid unnecessary lock
// grabs but does not seem thread-safe.
if (tracer_) {
InstrumentedMutexLock lock(&trace_mutex_);
if (tracer_ && tracer_->IsWriteOrderPreserved()) {
for (auto* writer : write_group) {
// TODO: maybe handle the tracing status?
tracer_->Write(writer->batch).PermitUncheckedError();
}
}
}
// Rules for when we can update the memtable concurrently
// 1. supported by memtable
// 2. Puts are not okay if inplace_update_support
// 3. Merges are not okay
//
// Rules 1..2 are enforced by checking the options
// during startup (CheckConcurrentWritesSupported), so if
// options.allow_concurrent_memtable_write is true then they can be
// assumed to be true. Rule 3 is checked for each batch. We could
// relax rules 2 if we could prevent write batches from referring
// more than once to a particular key.
bool parallel = immutable_db_options_.allow_concurrent_memtable_write &&
write_group.size > 1;
size_t total_count = 0;
size_t valid_batches = 0;
size_t total_byte_size = 0;
size_t pre_release_callback_cnt = 0;
for (auto* writer : write_group) {
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
assert(writer);
if (writer->CheckCallback(this)) {
valid_batches += writer->batch_cnt;
if (writer->ShouldWriteToMemtable()) {
total_count += WriteBatchInternal::Count(writer->batch);
parallel = parallel && !writer->batch->HasMerge();
}
total_byte_size = WriteBatchInternal::AppendedByteSize(
total_byte_size, WriteBatchInternal::ByteSize(writer->batch));
if (writer->pre_release_callback) {
pre_release_callback_cnt++;
}
}
}
// Note about seq_per_batch_: either disableWAL is set for the entire write
// group or not. In either case we inc seq for each write batch with no
// failed callback. This means that there could be a batch with
// disalbe_memtable in between; although we do not write this batch to
// memtable it still consumes a seq. Otherwise, if !seq_per_batch_, we inc
// the seq per valid written key to mem.
size_t seq_inc = seq_per_batch_ ? valid_batches : total_count;
const bool concurrent_update = two_write_queues_;
// Update stats while we are an exclusive group leader, so we know
// that nobody else can be writing to these particular stats.
// We're optimistic, updating the stats before we successfully
// commit. That lets us release our leader status early.
auto stats = default_cf_internal_stats_;
stats->AddDBStats(InternalStats::kIntStatsNumKeysWritten, total_count,
concurrent_update);
RecordTick(stats_, NUMBER_KEYS_WRITTEN, total_count);
stats->AddDBStats(InternalStats::kIntStatsBytesWritten, total_byte_size,
concurrent_update);
RecordTick(stats_, BYTES_WRITTEN, total_byte_size);
stats->AddDBStats(InternalStats::kIntStatsWriteDoneBySelf, 1,
concurrent_update);
RecordTick(stats_, WRITE_DONE_BY_SELF);
auto write_done_by_other = write_group.size - 1;
if (write_done_by_other > 0) {
stats->AddDBStats(InternalStats::kIntStatsWriteDoneByOther,
write_done_by_other, concurrent_update);
RecordTick(stats_, WRITE_DONE_BY_OTHER, write_done_by_other);
}
RecordInHistogram(stats_, BYTES_PER_WRITE, total_byte_size);
if (write_options.disableWAL) {
has_unpersisted_data_.store(true, std::memory_order_relaxed);
}
PERF_TIMER_STOP(write_pre_and_post_process_time);
if (!two_write_queues_) {
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
if (status.ok() && !write_options.disableWAL) {
assert(log_context.log_file_number_size);
LogFileNumberSize& log_file_number_size =
*(log_context.log_file_number_size);
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
PERF_TIMER_GUARD(write_wal_time);
io_s =
WriteToWAL(write_group, log_context.writer, log_used,
log_context.need_log_sync, log_context.need_log_dir_sync,
last_sequence + 1, log_file_number_size);
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
}
} else {
if (status.ok() && !write_options.disableWAL) {
PERF_TIMER_GUARD(write_wal_time);
// LastAllocatedSequence is increased inside WriteToWAL under
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
// wal_write_mutex_ to ensure ordered events in WAL
io_s = ConcurrentWriteToWAL(write_group, log_used, &last_sequence,
seq_inc);
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
} else {
// Otherwise we inc seq number for memtable writes
last_sequence = versions_->FetchAddLastAllocatedSequence(seq_inc);
}
}
status = io_s;
assert(last_sequence != kMaxSequenceNumber);
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
const SequenceNumber current_sequence = last_sequence + 1;
last_sequence += seq_inc;
// PreReleaseCallback is called after WAL write and before memtable write
if (status.ok()) {
SequenceNumber next_sequence = current_sequence;
size_t index = 0;
// Note: the logic for advancing seq here must be consistent with the
// logic in WriteBatchInternal::InsertInto(write_group...) as well as
// with WriteBatchInternal::InsertInto(write_batch...) that is called on
// the merged batch during recovery from the WAL.
for (auto* writer : write_group) {
if (writer->CallbackFailed()) {
continue;
}
writer->sequence = next_sequence;
if (writer->pre_release_callback) {
Status ws = writer->pre_release_callback->Callback(
writer->sequence, disable_memtable, writer->log_used, index++,
pre_release_callback_cnt);
if (!ws.ok()) {
status = pre_release_cb_status = ws;
break;
}
}
if (seq_per_batch_) {
assert(writer->batch_cnt);
next_sequence += writer->batch_cnt;
} else if (writer->ShouldWriteToMemtable()) {
next_sequence += WriteBatchInternal::Count(writer->batch);
}
}
}
if (status.ok()) {
PERF_TIMER_GUARD(write_memtable_time);
if (!parallel) {
// w.sequence will be set inside InsertInto
w.status = WriteBatchInternal::InsertInto(
write_group, current_sequence, column_family_memtables_.get(),
Refactor trimming logic for immutable memtables (#5022) Summary: MyRocks currently sets `max_write_buffer_number_to_maintain` in order to maintain enough history for transaction conflict checking. The effectiveness of this approach depends on the size of memtables. When memtables are small, it may not keep enough history; when memtables are large, this may consume too much memory. We are proposing a new way to configure memtable list history: by limiting the memory usage of immutable memtables. The new option is `max_write_buffer_size_to_maintain` and it will take precedence over the old `max_write_buffer_number_to_maintain` if they are both set to non-zero values. The new option accounts for the total memory usage of flushed immutable memtables and mutable memtable. When the total usage exceeds the limit, RocksDB may start dropping immutable memtables (which is also called trimming history), starting from the oldest one. The semantics of the old option actually works both as an upper bound and lower bound. History trimming will start if number of immutable memtables exceeds the limit, but it will never go below (limit-1) due to history trimming. In order the mimic the behavior with the new option, history trimming will stop if dropping the next immutable memtable causes the total memory usage go below the size limit. For example, assuming the size limit is set to 64MB, and there are 3 immutable memtables with sizes of 20, 30, 30. Although the total memory usage is 80MB > 64MB, dropping the oldest memtable will reduce the memory usage to 60MB < 64MB, so in this case no memtable will be dropped. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5022 Differential Revision: D14394062 Pulled By: miasantreble fbshipit-source-id: 60457a509c6af89d0993f988c9b5c2aa9e45f5c5
5 years ago
&flush_scheduler_, &trim_history_scheduler_,
write_options.ignore_missing_column_families,
0 /*recovery_log_number*/, this, parallel, seq_per_batch_,
batch_per_txn_);
} else {
write_group.last_sequence = last_sequence;
write_thread_.LaunchParallelMemTableWriters(&write_group);
in_parallel_group = true;
// Each parallel follower is doing each own writes. The leader should
// also do its own.
if (w.ShouldWriteToMemtable()) {
ColumnFamilyMemTablesImpl column_family_memtables(
versions_->GetColumnFamilySet());
assert(w.sequence == current_sequence);
w.status = WriteBatchInternal::InsertInto(
&w, w.sequence, &column_family_memtables, &flush_scheduler_,
Refactor trimming logic for immutable memtables (#5022) Summary: MyRocks currently sets `max_write_buffer_number_to_maintain` in order to maintain enough history for transaction conflict checking. The effectiveness of this approach depends on the size of memtables. When memtables are small, it may not keep enough history; when memtables are large, this may consume too much memory. We are proposing a new way to configure memtable list history: by limiting the memory usage of immutable memtables. The new option is `max_write_buffer_size_to_maintain` and it will take precedence over the old `max_write_buffer_number_to_maintain` if they are both set to non-zero values. The new option accounts for the total memory usage of flushed immutable memtables and mutable memtable. When the total usage exceeds the limit, RocksDB may start dropping immutable memtables (which is also called trimming history), starting from the oldest one. The semantics of the old option actually works both as an upper bound and lower bound. History trimming will start if number of immutable memtables exceeds the limit, but it will never go below (limit-1) due to history trimming. In order the mimic the behavior with the new option, history trimming will stop if dropping the next immutable memtable causes the total memory usage go below the size limit. For example, assuming the size limit is set to 64MB, and there are 3 immutable memtables with sizes of 20, 30, 30. Although the total memory usage is 80MB > 64MB, dropping the oldest memtable will reduce the memory usage to 60MB < 64MB, so in this case no memtable will be dropped. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5022 Differential Revision: D14394062 Pulled By: miasantreble fbshipit-source-id: 60457a509c6af89d0993f988c9b5c2aa9e45f5c5
5 years ago
&trim_history_scheduler_,
write_options.ignore_missing_column_families, 0 /*log_number*/,
this, true /*concurrent_memtable_writes*/, seq_per_batch_,
w.batch_cnt, batch_per_txn_,
write_options.memtable_insert_hint_per_batch);
}
}
if (seq_used != nullptr) {
*seq_used = w.sequence;
}
}
}
PERF_TIMER_START(write_pre_and_post_process_time);
if (!io_s.ok()) {
// Check WriteToWAL status
IOStatusCheck(io_s);
}
if (!w.CallbackFailed()) {
if (!io_s.ok()) {
assert(pre_release_cb_status.ok());
} else {
WriteStatusCheck(pre_release_cb_status);
}
} else {
assert(pre_release_cb_status.ok());
}
if (log_context.need_log_sync) {
VersionEdit synced_wals;
log_write_mutex_.Lock();
if (status.ok()) {
MarkLogsSynced(logfile_number_, log_context.need_log_dir_sync,
&synced_wals);
} else {
MarkLogsNotSynced(logfile_number_);
}
log_write_mutex_.Unlock();
if (status.ok() && synced_wals.IsWalAddition()) {
InstrumentedMutexLock l(&mutex_);
Group rocksdb.sst.read.micros stat by IOActivity flush and compaction (#11288) Summary: **Context:** The existing stat rocksdb.sst.read.micros does not reflect each of compaction and flush cases but aggregate them, which is not so helpful for us to understand IO read behavior of each of them. **Summary** - Update `StopWatch` and `RandomAccessFileReader` to record `rocksdb.sst.read.micros` and `rocksdb.file.{flush/compaction}.read.micros` - Fixed the default histogram in `RandomAccessFileReader` - New field `ReadOptions/IOOptions::io_activity`; Pass `ReadOptions` through paths under db open, flush and compaction to where we can prepare `IOOptions` and pass it to `RandomAccessFileReader` - Use `thread_status_util` for assertion in `DbStressFSWrapper` for continuous testing on we are passing correct `io_activity` under db open, flush and compaction Pull Request resolved: https://github.com/facebook/rocksdb/pull/11288 Test Plan: - **Stress test** - **Db bench 1: rocksdb.sst.read.micros COUNT ≈ sum of rocksdb.file.read.flush.micros's and rocksdb.file.read.compaction.micros's.** (without blob) - May not be exactly the same due to `HistogramStat::Add` only guarantees atomic not accuracy across threads. ``` ./db_bench -db=/dev/shm/testdb/ -statistics=true -benchmarks="fillseq" -key_size=32 -value_size=512 -num=50000 -write_buffer_size=655 -target_file_size_base=655 -disable_auto_compactions=false -compression_type=none -bloom_bits=3 (-use_plain_table=1 -prefix_size=10) ``` ``` // BlockBasedTable rocksdb.sst.read.micros P50 : 2.009374 P95 : 4.968548 P99 : 8.110362 P100 : 43.000000 COUNT : 40456 SUM : 114805 rocksdb.file.read.flush.micros P50 : 1.871841 P95 : 3.872407 P99 : 5.540541 P100 : 43.000000 COUNT : 2250 SUM : 6116 rocksdb.file.read.compaction.micros P50 : 2.023109 P95 : 5.029149 P99 : 8.196910 P100 : 26.000000 COUNT : 38206 SUM : 108689 // PlainTable Does not apply ``` - **Db bench 2: performance** **Read** SETUP: db with 900 files ``` ./db_bench -db=/dev/shm/testdb/ -benchmarks="fillseq" -key_size=32 -value_size=512 -num=50000 -write_buffer_size=655 -disable_auto_compactions=true -target_file_size_base=655 -compression_type=none ```run till convergence ``` ./db_bench -seed=1678564177044286 -use_existing_db=true -db=/dev/shm/testdb -benchmarks=readrandom[-X60] -statistics=true -num=1000000 -disable_auto_compactions=true -compression_type=none -bloom_bits=3 ``` Pre-change `readrandom [AVG 60 runs] : 21568 (± 248) ops/sec` Post-change (no regression, -0.3%) `readrandom [AVG 60 runs] : 21486 (± 236) ops/sec` **Compaction/Flush**run till convergence ``` ./db_bench -db=/dev/shm/testdb2/ -seed=1678564177044286 -benchmarks="fillseq[-X60]" -key_size=32 -value_size=512 -num=50000 -write_buffer_size=655 -disable_auto_compactions=false -target_file_size_base=655 -compression_type=none rocksdb.sst.read.micros COUNT : 33820 rocksdb.sst.read.flush.micros COUNT : 1800 rocksdb.sst.read.compaction.micros COUNT : 32020 ``` Pre-change `fillseq [AVG 46 runs] : 1391 (± 214) ops/sec; 0.7 (± 0.1) MB/sec` Post-change (no regression, ~-0.4%) `fillseq [AVG 46 runs] : 1385 (± 216) ops/sec; 0.7 (± 0.1) MB/sec` Reviewed By: ajkr Differential Revision: D44007011 Pulled By: hx235 fbshipit-source-id: a54c89e4846dfc9a135389edf3f3eedfea257132
2 years ago
// TODO: plumb Env::IOActivity
const ReadOptions read_options;
status = ApplyWALToManifest(read_options, &synced_wals);
}
// Requesting sync with two_write_queues_ is expected to be very rare. We
// hence provide a simple implementation that is not necessarily efficient.
if (two_write_queues_) {
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
if (manual_wal_flush_) {
status = FlushWAL(true);
} else {
status = SyncWAL();
}
}
}
bool should_exit_batch_group = true;
if (in_parallel_group) {
// CompleteParallelWorker returns true if this thread should
// handle exit, false means somebody else did
should_exit_batch_group = write_thread_.CompleteParallelMemTableWriter(&w);
}
if (should_exit_batch_group) {
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
if (status.ok()) {
Snapshots with user-specified timestamps (#9879) Summary: In RocksDB, keys are associated with (internal) sequence numbers which denote when the keys are written to the database. Sequence numbers in different RocksDB instances are unrelated, thus not comparable. It is nice if we can associate sequence numbers with their corresponding actual timestamps. One thing we can do is to support user-defined timestamp, which allows the applications to specify the format of custom timestamps and encode a timestamp with each key. More details can be found at https://github.com/facebook/rocksdb/wiki/User-defined-Timestamp-%28Experimental%29. This PR provides a different but complementary approach. We can associate rocksdb snapshots (defined in https://github.com/facebook/rocksdb/blob/7.2.fb/include/rocksdb/snapshot.h#L20) with **user-specified** timestamps. Since a snapshot is essentially an object representing a sequence number, this PR establishes a bi-directional mapping between sequence numbers and timestamps. In the past, snapshots are usually taken by readers. The current super-version is grabbed, and a `rocksdb::Snapshot` object is created with the last published sequence number of the super-version. You can see that the reader actually has no good idea of what timestamp to assign to this snapshot, because by the time the `GetSnapshot()` is called, an arbitrarily long period of time may have already elapsed since the last write, which is when the last published sequence number is written. This observation motivates the creation of "timestamped" snapshots on the write path. Currently, this functionality is exposed only to the layer of `TransactionDB`. Application can tell RocksDB to create a snapshot when a transaction commits, effectively associating the last sequence number with a timestamp. It is also assumed that application will ensure any two snapshots with timestamps should satisfy the following: ``` snapshot1.seq < snapshot2.seq iff. snapshot1.ts < snapshot2.ts ``` If the application can guarantee that when a reader takes a timestamped snapshot, there is no active writes going on in the database, then we also allow the user to use a new API `TransactionDB::CreateTimestampedSnapshot()` to create a snapshot with associated timestamp. Code example ```cpp // Create a timestamped snapshot when committing transaction. txn->SetCommitTimestamp(100); txn->SetSnapshotOnNextOperation(); txn->Commit(); // A wrapper API for convenience Status Transaction::CommitAndTryCreateSnapshot( std::shared_ptr<TransactionNotifier> notifier, TxnTimestamp ts, std::shared_ptr<const Snapshot>* ret); // Create a timestamped snapshot if caller guarantees no concurrent writes std::pair<Status, std::shared_ptr<const Snapshot>> snapshot = txn_db->CreateTimestampedSnapshot(100); ``` The snapshots created in this way will be managed by RocksDB with ref-counting and potentially shared with other readers. We provide the following APIs for readers to retrieve a snapshot given a timestamp. ```cpp // Return the timestamped snapshot correponding to given timestamp. If ts is // kMaxTxnTimestamp, then we return the latest timestamped snapshot if present. // Othersise, we return the snapshot whose timestamp is equal to `ts`. If no // such snapshot exists, then we return null. std::shared_ptr<const Snapshot> TransactionDB::GetTimestampedSnapshot(TxnTimestamp ts) const; // Return the latest timestamped snapshot if present. std::shared_ptr<const Snapshot> TransactionDB::GetLatestTimestampedSnapshot() const; ``` We also provide two additional APIs for stats collection and reporting purposes. ```cpp Status TransactionDB::GetAllTimestampedSnapshots( std::vector<std::shared_ptr<const Snapshot>>& snapshots) const; // Return timestamped snapshots whose timestamps fall in [ts_lb, ts_ub) and store them in `snapshots`. Status TransactionDB::GetTimestampedSnapshots( TxnTimestamp ts_lb, TxnTimestamp ts_ub, std::vector<std::shared_ptr<const Snapshot>>& snapshots) const; ``` To prevent the number of timestamped snapshots from growing infinitely, we provide the following API to release timestamped snapshots whose timestamps are older than or equal to a given threshold. ```cpp void TransactionDB::ReleaseTimestampedSnapshotsOlderThan(TxnTimestamp ts); ``` Before shutdown, RocksDB will release all timestamped snapshots. Comparison with user-defined timestamp and how they can be combined: User-defined timestamp persists every key with a timestamp, while timestamped snapshots maintain a volatile mapping between snapshots (sequence numbers) and timestamps. Different internal keys with the same user key but different timestamps will be treated as different by compaction, thus a newer version will not hide older versions (with smaller timestamps) unless they are eligible for garbage collection. In contrast, taking a timestamped snapshot at a certain sequence number and timestamp prevents all the keys visible in this snapshot from been dropped by compaction. Here, visible means (seq < snapshot and most recent). The timestamped snapshot supports the semantics of reading at an exact point in time. Timestamped snapshots can also be used with user-defined timestamp. Pull Request resolved: https://github.com/facebook/rocksdb/pull/9879 Test Plan: ``` make check TEST_TMPDIR=/dev/shm make crash_test_with_txn ``` Reviewed By: siying Differential Revision: D35783919 Pulled By: riversand963 fbshipit-source-id: 586ad905e169189e19d3bfc0cb0177a7239d1bd4
3 years ago
for (auto* tmp_w : write_group) {
assert(tmp_w);
if (tmp_w->post_memtable_callback) {
Status tmp_s =
(*tmp_w->post_memtable_callback)(last_sequence, disable_memtable);
// TODO: propagate the execution status of post_memtable_callback to
// caller.
assert(tmp_s.ok());
}
}
// Note: if we are to resume after non-OK statuses we need to revisit how
// we reacts to non-OK statuses here.
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
versions_->SetLastSequence(last_sequence);
}
MemTableInsertStatusCheck(w.status);
write_thread_.ExitAsBatchGroupLeader(write_group, status);
}
if (status.ok()) {
status = w.FinalStatus();
}
return status;
}
Status DBImpl::PipelinedWriteImpl(const WriteOptions& write_options,
WriteBatch* my_batch, WriteCallback* callback,
uint64_t* log_used, uint64_t log_ref,
bool disable_memtable, uint64_t* seq_used) {
PERF_TIMER_GUARD(write_pre_and_post_process_time);
StopWatch write_sw(immutable_db_options_.clock, stats_, DB_WRITE);
WriteContext write_context;
WriteThread::Writer w(write_options, my_batch, callback, log_ref,
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
disable_memtable, /*_batch_cnt=*/0,
/*_pre_release_callback=*/nullptr);
write_thread_.JoinBatchGroup(&w);
TEST_SYNC_POINT("DBImplWrite::PipelinedWriteImpl:AfterJoinBatchGroup");
if (w.state == WriteThread::STATE_GROUP_LEADER) {
WriteThread::WriteGroup wal_write_group;
if (w.callback && !w.callback->AllowWriteBatching()) {
write_thread_.WaitForMemTableWriters();
}
LogContext log_context(!write_options.disableWAL && write_options.sync);
// PreprocessWrite does its own perf timing.
PERF_TIMER_STOP(write_pre_and_post_process_time);
w.status = PreprocessWrite(write_options, &log_context, &write_context);
PERF_TIMER_START(write_pre_and_post_process_time);
// This can set non-OK status if callback fail.
last_batch_group_size_ =
write_thread_.EnterAsBatchGroupLeader(&w, &wal_write_group);
const SequenceNumber current_sequence =
write_thread_.UpdateLastSequence(versions_->LastSequence()) + 1;
size_t total_count = 0;
size_t total_byte_size = 0;
if (w.status.ok()) {
// TODO: this use of operator bool on `tracer_` can avoid unnecessary lock
// grabs but does not seem thread-safe.
if (tracer_) {
InstrumentedMutexLock lock(&trace_mutex_);
if (tracer_ != nullptr && tracer_->IsWriteOrderPreserved()) {
for (auto* writer : wal_write_group) {
// TODO: maybe handle the tracing status?
tracer_->Write(writer->batch).PermitUncheckedError();
}
}
}
SequenceNumber next_sequence = current_sequence;
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
for (auto* writer : wal_write_group) {
assert(writer);
if (writer->CheckCallback(this)) {
if (writer->ShouldWriteToMemtable()) {
writer->sequence = next_sequence;
size_t count = WriteBatchInternal::Count(writer->batch);
next_sequence += count;
total_count += count;
}
total_byte_size = WriteBatchInternal::AppendedByteSize(
total_byte_size, WriteBatchInternal::ByteSize(writer->batch));
}
}
if (w.disable_wal) {
has_unpersisted_data_.store(true, std::memory_order_relaxed);
}
write_thread_.UpdateLastSequence(current_sequence + total_count - 1);
}
auto stats = default_cf_internal_stats_;
stats->AddDBStats(InternalStats::kIntStatsNumKeysWritten, total_count);
RecordTick(stats_, NUMBER_KEYS_WRITTEN, total_count);
stats->AddDBStats(InternalStats::kIntStatsBytesWritten, total_byte_size);
RecordTick(stats_, BYTES_WRITTEN, total_byte_size);
RecordInHistogram(stats_, BYTES_PER_WRITE, total_byte_size);
PERF_TIMER_STOP(write_pre_and_post_process_time);
IOStatus io_s;
io_s.PermitUncheckedError(); // Allow io_s to be uninitialized
if (w.status.ok() && !write_options.disableWAL) {
PERF_TIMER_GUARD(write_wal_time);
stats->AddDBStats(InternalStats::kIntStatsWriteDoneBySelf, 1);
RecordTick(stats_, WRITE_DONE_BY_SELF, 1);
if (wal_write_group.size > 1) {
stats->AddDBStats(InternalStats::kIntStatsWriteDoneByOther,
wal_write_group.size - 1);
RecordTick(stats_, WRITE_DONE_BY_OTHER, wal_write_group.size - 1);
}
assert(log_context.log_file_number_size);
LogFileNumberSize& log_file_number_size =
*(log_context.log_file_number_size);
io_s =
WriteToWAL(wal_write_group, log_context.writer, log_used,
log_context.need_log_sync, log_context.need_log_dir_sync,
current_sequence, log_file_number_size);
w.status = io_s;
}
if (!io_s.ok()) {
// Check WriteToWAL status
IOStatusCheck(io_s);
} else if (!w.CallbackFailed()) {
WriteStatusCheck(w.status);
}
VersionEdit synced_wals;
if (log_context.need_log_sync) {
InstrumentedMutexLock l(&log_write_mutex_);
if (w.status.ok()) {
MarkLogsSynced(logfile_number_, log_context.need_log_dir_sync,
&synced_wals);
} else {
MarkLogsNotSynced(logfile_number_);
}
}
if (w.status.ok() && synced_wals.IsWalAddition()) {
InstrumentedMutexLock l(&mutex_);
Group rocksdb.sst.read.micros stat by IOActivity flush and compaction (#11288) Summary: **Context:** The existing stat rocksdb.sst.read.micros does not reflect each of compaction and flush cases but aggregate them, which is not so helpful for us to understand IO read behavior of each of them. **Summary** - Update `StopWatch` and `RandomAccessFileReader` to record `rocksdb.sst.read.micros` and `rocksdb.file.{flush/compaction}.read.micros` - Fixed the default histogram in `RandomAccessFileReader` - New field `ReadOptions/IOOptions::io_activity`; Pass `ReadOptions` through paths under db open, flush and compaction to where we can prepare `IOOptions` and pass it to `RandomAccessFileReader` - Use `thread_status_util` for assertion in `DbStressFSWrapper` for continuous testing on we are passing correct `io_activity` under db open, flush and compaction Pull Request resolved: https://github.com/facebook/rocksdb/pull/11288 Test Plan: - **Stress test** - **Db bench 1: rocksdb.sst.read.micros COUNT ≈ sum of rocksdb.file.read.flush.micros's and rocksdb.file.read.compaction.micros's.** (without blob) - May not be exactly the same due to `HistogramStat::Add` only guarantees atomic not accuracy across threads. ``` ./db_bench -db=/dev/shm/testdb/ -statistics=true -benchmarks="fillseq" -key_size=32 -value_size=512 -num=50000 -write_buffer_size=655 -target_file_size_base=655 -disable_auto_compactions=false -compression_type=none -bloom_bits=3 (-use_plain_table=1 -prefix_size=10) ``` ``` // BlockBasedTable rocksdb.sst.read.micros P50 : 2.009374 P95 : 4.968548 P99 : 8.110362 P100 : 43.000000 COUNT : 40456 SUM : 114805 rocksdb.file.read.flush.micros P50 : 1.871841 P95 : 3.872407 P99 : 5.540541 P100 : 43.000000 COUNT : 2250 SUM : 6116 rocksdb.file.read.compaction.micros P50 : 2.023109 P95 : 5.029149 P99 : 8.196910 P100 : 26.000000 COUNT : 38206 SUM : 108689 // PlainTable Does not apply ``` - **Db bench 2: performance** **Read** SETUP: db with 900 files ``` ./db_bench -db=/dev/shm/testdb/ -benchmarks="fillseq" -key_size=32 -value_size=512 -num=50000 -write_buffer_size=655 -disable_auto_compactions=true -target_file_size_base=655 -compression_type=none ```run till convergence ``` ./db_bench -seed=1678564177044286 -use_existing_db=true -db=/dev/shm/testdb -benchmarks=readrandom[-X60] -statistics=true -num=1000000 -disable_auto_compactions=true -compression_type=none -bloom_bits=3 ``` Pre-change `readrandom [AVG 60 runs] : 21568 (± 248) ops/sec` Post-change (no regression, -0.3%) `readrandom [AVG 60 runs] : 21486 (± 236) ops/sec` **Compaction/Flush**run till convergence ``` ./db_bench -db=/dev/shm/testdb2/ -seed=1678564177044286 -benchmarks="fillseq[-X60]" -key_size=32 -value_size=512 -num=50000 -write_buffer_size=655 -disable_auto_compactions=false -target_file_size_base=655 -compression_type=none rocksdb.sst.read.micros COUNT : 33820 rocksdb.sst.read.flush.micros COUNT : 1800 rocksdb.sst.read.compaction.micros COUNT : 32020 ``` Pre-change `fillseq [AVG 46 runs] : 1391 (± 214) ops/sec; 0.7 (± 0.1) MB/sec` Post-change (no regression, ~-0.4%) `fillseq [AVG 46 runs] : 1385 (± 216) ops/sec; 0.7 (± 0.1) MB/sec` Reviewed By: ajkr Differential Revision: D44007011 Pulled By: hx235 fbshipit-source-id: a54c89e4846dfc9a135389edf3f3eedfea257132
2 years ago
// TODO: plumb Env::IOActivity
const ReadOptions read_options;
w.status = ApplyWALToManifest(read_options, &synced_wals);
}
write_thread_.ExitAsBatchGroupLeader(wal_write_group, w.status);
}
// NOTE: the memtable_write_group is declared before the following
// `if` statement because its lifetime needs to be longer
// that the inner context of the `if` as a reference to it
// may be used further below within the outer _write_thread
WriteThread::WriteGroup memtable_write_group;
if (w.state == WriteThread::STATE_MEMTABLE_WRITER_LEADER) {
PERF_TIMER_GUARD(write_memtable_time);
assert(w.ShouldWriteToMemtable());
write_thread_.EnterAsMemTableWriter(&w, &memtable_write_group);
if (memtable_write_group.size > 1 &&
immutable_db_options_.allow_concurrent_memtable_write) {
write_thread_.LaunchParallelMemTableWriters(&memtable_write_group);
} else {
memtable_write_group.status = WriteBatchInternal::InsertInto(
memtable_write_group, w.sequence, column_family_memtables_.get(),
Refactor trimming logic for immutable memtables (#5022) Summary: MyRocks currently sets `max_write_buffer_number_to_maintain` in order to maintain enough history for transaction conflict checking. The effectiveness of this approach depends on the size of memtables. When memtables are small, it may not keep enough history; when memtables are large, this may consume too much memory. We are proposing a new way to configure memtable list history: by limiting the memory usage of immutable memtables. The new option is `max_write_buffer_size_to_maintain` and it will take precedence over the old `max_write_buffer_number_to_maintain` if they are both set to non-zero values. The new option accounts for the total memory usage of flushed immutable memtables and mutable memtable. When the total usage exceeds the limit, RocksDB may start dropping immutable memtables (which is also called trimming history), starting from the oldest one. The semantics of the old option actually works both as an upper bound and lower bound. History trimming will start if number of immutable memtables exceeds the limit, but it will never go below (limit-1) due to history trimming. In order the mimic the behavior with the new option, history trimming will stop if dropping the next immutable memtable causes the total memory usage go below the size limit. For example, assuming the size limit is set to 64MB, and there are 3 immutable memtables with sizes of 20, 30, 30. Although the total memory usage is 80MB > 64MB, dropping the oldest memtable will reduce the memory usage to 60MB < 64MB, so in this case no memtable will be dropped. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5022 Differential Revision: D14394062 Pulled By: miasantreble fbshipit-source-id: 60457a509c6af89d0993f988c9b5c2aa9e45f5c5
5 years ago
&flush_scheduler_, &trim_history_scheduler_,
write_options.ignore_missing_column_families, 0 /*log_number*/, this,
false /*concurrent_memtable_writes*/, seq_per_batch_, batch_per_txn_);
versions_->SetLastSequence(memtable_write_group.last_sequence);
write_thread_.ExitAsMemTableWriter(&w, memtable_write_group);
}
} else {
// NOTE: the memtable_write_group is never really used,
// so we need to set its status to pass ASSERT_STATUS_CHECKED
memtable_write_group.status.PermitUncheckedError();
}
if (w.state == WriteThread::STATE_PARALLEL_MEMTABLE_WRITER) {
assert(w.ShouldWriteToMemtable());
ColumnFamilyMemTablesImpl column_family_memtables(
versions_->GetColumnFamilySet());
w.status = WriteBatchInternal::InsertInto(
&w, w.sequence, &column_family_memtables, &flush_scheduler_,
Refactor trimming logic for immutable memtables (#5022) Summary: MyRocks currently sets `max_write_buffer_number_to_maintain` in order to maintain enough history for transaction conflict checking. The effectiveness of this approach depends on the size of memtables. When memtables are small, it may not keep enough history; when memtables are large, this may consume too much memory. We are proposing a new way to configure memtable list history: by limiting the memory usage of immutable memtables. The new option is `max_write_buffer_size_to_maintain` and it will take precedence over the old `max_write_buffer_number_to_maintain` if they are both set to non-zero values. The new option accounts for the total memory usage of flushed immutable memtables and mutable memtable. When the total usage exceeds the limit, RocksDB may start dropping immutable memtables (which is also called trimming history), starting from the oldest one. The semantics of the old option actually works both as an upper bound and lower bound. History trimming will start if number of immutable memtables exceeds the limit, but it will never go below (limit-1) due to history trimming. In order the mimic the behavior with the new option, history trimming will stop if dropping the next immutable memtable causes the total memory usage go below the size limit. For example, assuming the size limit is set to 64MB, and there are 3 immutable memtables with sizes of 20, 30, 30. Although the total memory usage is 80MB > 64MB, dropping the oldest memtable will reduce the memory usage to 60MB < 64MB, so in this case no memtable will be dropped. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5022 Differential Revision: D14394062 Pulled By: miasantreble fbshipit-source-id: 60457a509c6af89d0993f988c9b5c2aa9e45f5c5
5 years ago
&trim_history_scheduler_, write_options.ignore_missing_column_families,
0 /*log_number*/, this, true /*concurrent_memtable_writes*/,
false /*seq_per_batch*/, 0 /*batch_cnt*/, true /*batch_per_txn*/,
write_options.memtable_insert_hint_per_batch);
if (write_thread_.CompleteParallelMemTableWriter(&w)) {
MemTableInsertStatusCheck(w.status);
versions_->SetLastSequence(w.write_group->last_sequence);
write_thread_.ExitAsMemTableWriter(&w, *w.write_group);
}
}
if (seq_used != nullptr) {
*seq_used = w.sequence;
}
assert(w.state == WriteThread::STATE_COMPLETED);
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
return w.FinalStatus();
}
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
Status DBImpl::UnorderedWriteMemtable(const WriteOptions& write_options,
WriteBatch* my_batch,
WriteCallback* callback, uint64_t log_ref,
SequenceNumber seq,
const size_t sub_batch_cnt) {
PERF_TIMER_GUARD(write_pre_and_post_process_time);
StopWatch write_sw(immutable_db_options_.clock, stats_, DB_WRITE);
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
WriteThread::Writer w(write_options, my_batch, callback, log_ref,
false /*disable_memtable*/);
if (w.CheckCallback(this) && w.ShouldWriteToMemtable()) {
w.sequence = seq;
size_t total_count = WriteBatchInternal::Count(my_batch);
InternalStats* stats = default_cf_internal_stats_;
stats->AddDBStats(InternalStats::kIntStatsNumKeysWritten, total_count);
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
RecordTick(stats_, NUMBER_KEYS_WRITTEN, total_count);
ColumnFamilyMemTablesImpl column_family_memtables(
versions_->GetColumnFamilySet());
w.status = WriteBatchInternal::InsertInto(
&w, w.sequence, &column_family_memtables, &flush_scheduler_,
Refactor trimming logic for immutable memtables (#5022) Summary: MyRocks currently sets `max_write_buffer_number_to_maintain` in order to maintain enough history for transaction conflict checking. The effectiveness of this approach depends on the size of memtables. When memtables are small, it may not keep enough history; when memtables are large, this may consume too much memory. We are proposing a new way to configure memtable list history: by limiting the memory usage of immutable memtables. The new option is `max_write_buffer_size_to_maintain` and it will take precedence over the old `max_write_buffer_number_to_maintain` if they are both set to non-zero values. The new option accounts for the total memory usage of flushed immutable memtables and mutable memtable. When the total usage exceeds the limit, RocksDB may start dropping immutable memtables (which is also called trimming history), starting from the oldest one. The semantics of the old option actually works both as an upper bound and lower bound. History trimming will start if number of immutable memtables exceeds the limit, but it will never go below (limit-1) due to history trimming. In order the mimic the behavior with the new option, history trimming will stop if dropping the next immutable memtable causes the total memory usage go below the size limit. For example, assuming the size limit is set to 64MB, and there are 3 immutable memtables with sizes of 20, 30, 30. Although the total memory usage is 80MB > 64MB, dropping the oldest memtable will reduce the memory usage to 60MB < 64MB, so in this case no memtable will be dropped. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5022 Differential Revision: D14394062 Pulled By: miasantreble fbshipit-source-id: 60457a509c6af89d0993f988c9b5c2aa9e45f5c5
5 years ago
&trim_history_scheduler_, write_options.ignore_missing_column_families,
0 /*log_number*/, this, true /*concurrent_memtable_writes*/,
seq_per_batch_, sub_batch_cnt, true /*batch_per_txn*/,
write_options.memtable_insert_hint_per_batch);
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
if (write_options.disableWAL) {
has_unpersisted_data_.store(true, std::memory_order_relaxed);
}
}
size_t pending_cnt = pending_memtable_writes_.fetch_sub(1) - 1;
if (pending_cnt == 0) {
// switch_cv_ waits until pending_memtable_writes_ = 0. Locking its mutex
// before notify ensures that cv is in waiting state when it is notified
// thus not missing the update to pending_memtable_writes_ even though it is
// not modified under the mutex.
std::lock_guard<std::mutex> lck(switch_mutex_);
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
switch_cv_.notify_all();
}
WriteStatusCheck(w.status);
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
if (!w.FinalStatus().ok()) {
return w.FinalStatus();
}
return Status::OK();
}
// The 2nd write queue. If enabled it will be used only for WAL-only writes.
// This is the only queue that updates LastPublishedSequence which is only
// applicable in a two-queue setting.
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
Status DBImpl::WriteImplWALOnly(
WriteThread* write_thread, const WriteOptions& write_options,
WriteBatch* my_batch, WriteCallback* callback, uint64_t* log_used,
const uint64_t log_ref, uint64_t* seq_used, const size_t sub_batch_cnt,
PreReleaseCallback* pre_release_callback, const AssignOrder assign_order,
const PublishLastSeq publish_last_seq, const bool disable_memtable) {
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
PERF_TIMER_GUARD(write_pre_and_post_process_time);
WriteThread::Writer w(write_options, my_batch, callback, log_ref,
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
disable_memtable, sub_batch_cnt, pre_release_callback);
StopWatch write_sw(immutable_db_options_.clock, stats_, DB_WRITE);
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
write_thread->JoinBatchGroup(&w);
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
assert(w.state != WriteThread::STATE_PARALLEL_MEMTABLE_WRITER);
if (w.state == WriteThread::STATE_COMPLETED) {
if (log_used != nullptr) {
*log_used = w.log_used;
}
if (seq_used != nullptr) {
*seq_used = w.sequence;
}
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
return w.FinalStatus();
}
// else we are the leader of the write batch group
assert(w.state == WriteThread::STATE_GROUP_LEADER);
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
if (publish_last_seq == kDoPublishLastSeq) {
Status status;
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
// Currently we only use kDoPublishLastSeq in unordered_write
assert(immutable_db_options_.unordered_write);
WriteContext write_context;
if (error_handler_.IsDBStopped()) {
status = error_handler_.GetBGError();
}
// TODO(myabandeh): Make preliminary checks thread-safe so we could do them
// without paying the cost of obtaining the mutex.
if (status.ok()) {
LogContext log_context;
status = PreprocessWrite(write_options, &log_context, &write_context);
WriteStatusCheckOnLocked(status);
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
}
if (!status.ok()) {
WriteThread::WriteGroup write_group;
write_thread->EnterAsBatchGroupLeader(&w, &write_group);
write_thread->ExitAsBatchGroupLeader(write_group, status);
return status;
}
Revise LockWAL/UnlockWAL implementation (#11020) Summary: RocksDB has two public APIs: `DB::LockWAL()`/`DB::UnlockWAL()`. The current implementation acquires and releases the internal `DBImpl::log_write_mutex_`. According to the comment on `DBImpl::log_write_mutex_`: https://github.com/facebook/rocksdb/blob/7.8.fb/db/db_impl/db_impl.h#L2287:L2288 > Note: to avoid dealock, if needed to acquire both log_write_mutex_ and mutex_, the order should be first mutex_ and then log_write_mutex_. This puts limitations on how applications can use the `LockWAL()` API. After `LockWAL()` returns ok, then application should not perform any operation that acquires `mutex_`. Currently, the use case of `LockWAL()` is MyRocks implementing the MySQL storage engine handlerton `lock_hton_log` interface. The operation that MyRocks performs after `LockWAL()` is `GetSortedWalFiless()` which not only acquires mutex_, but also `log_write_mutex_`. There are two issues: 1. Applications using these two APIs may hang if one thread calls `GetSortedWalFiles()` after calling `LockWAL()` because log_write_mutex is not recursive. 2. Two threads may dead lock due to lock order inversion. To fix these issues, we can modify the implementation of LockWAL so that it does not keep `log_write_mutex_` held until UnlockWAL. To achieve the goal of locking the WAL, we can instead manually inject a write stall so that all future writes will be stopped. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11020 Test Plan: make check Reviewed By: ajkr Differential Revision: D41785203 Pulled By: riversand963 fbshipit-source-id: 5ccb7a9c6eb9a2c3fa80fd2c399cc2568b8f89ce
2 years ago
} else {
InstrumentedMutexLock lock(&mutex_);
Fix DelayWrite() calls for two_write_queues (#11130) Summary: PR https://github.com/facebook/rocksdb/issues/11020 fixed a case where it was easy to deadlock the DB with LockWAL() but introduced a bug showing up as a rare assertion failure in the stress test. Specifically, `assert(w->state == STATE_INIT)` in `WriteThread::LinkOne()` called from `BeginWriteStall()`, `DelayWrite()`, `WriteImplWALOnly()`. I haven't been about to generate a unit test that reproduces this failure but I believe the root cause is that DelayWrite() was never meant to be re-entrant, only called from the DB's write_thread_ leader. https://github.com/facebook/rocksdb/issues/11020 introduced a call to DelayWrite() from the nonmem_write_thread_ group leader. This fix is to make DelayWrite() apply to the specific write queue that it is being called from (inject a dummy write stall entry to the head of the appropriate write queue). WriteController is re-entrant, based on polling and state changes signalled with bg_cv_, so can manage stalling two queues. The only anticipated complication (called out by Andrew in previous PR) is that we don't want timed write delays being injected in parallel for the two queues, because that dimishes the intended throttling effect. Thus, we only allow timed delays for the primary write queue. HISTORY not updated because this is intended for the same release where the bug was introduced. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11130 Test Plan: Although I was not able to reproduce the assertion failure, I was able to reproduce a distinct flaw with what I believe is the same root cause: a kind of deadlock if both write queues need to wake up from stopped writes. Only one will be waiting on bg_cv_ (the other waiting in `LinkOne()` for the write queue to open up), so a single SignalAll() will only unblock one of the queues, with the other re-instating the stop until another signal on bg_cv_. A simple unit test is added for this case. Will also run crash_test_with_multiops_wc_txn for a while looking for issues. Reviewed By: ajkr Differential Revision: D42749330 Pulled By: pdillinger fbshipit-source-id: 4317dd899a93d57c26fd5af7143038f82d4d4d1b
2 years ago
Status status =
DelayWrite(/*num_bytes=*/0ull, *write_thread, write_options);
Revise LockWAL/UnlockWAL implementation (#11020) Summary: RocksDB has two public APIs: `DB::LockWAL()`/`DB::UnlockWAL()`. The current implementation acquires and releases the internal `DBImpl::log_write_mutex_`. According to the comment on `DBImpl::log_write_mutex_`: https://github.com/facebook/rocksdb/blob/7.8.fb/db/db_impl/db_impl.h#L2287:L2288 > Note: to avoid dealock, if needed to acquire both log_write_mutex_ and mutex_, the order should be first mutex_ and then log_write_mutex_. This puts limitations on how applications can use the `LockWAL()` API. After `LockWAL()` returns ok, then application should not perform any operation that acquires `mutex_`. Currently, the use case of `LockWAL()` is MyRocks implementing the MySQL storage engine handlerton `lock_hton_log` interface. The operation that MyRocks performs after `LockWAL()` is `GetSortedWalFiless()` which not only acquires mutex_, but also `log_write_mutex_`. There are two issues: 1. Applications using these two APIs may hang if one thread calls `GetSortedWalFiles()` after calling `LockWAL()` because log_write_mutex is not recursive. 2. Two threads may dead lock due to lock order inversion. To fix these issues, we can modify the implementation of LockWAL so that it does not keep `log_write_mutex_` held until UnlockWAL. To achieve the goal of locking the WAL, we can instead manually inject a write stall so that all future writes will be stopped. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11020 Test Plan: make check Reviewed By: ajkr Differential Revision: D41785203 Pulled By: riversand963 fbshipit-source-id: 5ccb7a9c6eb9a2c3fa80fd2c399cc2568b8f89ce
2 years ago
if (!status.ok()) {
WriteThread::WriteGroup write_group;
write_thread->EnterAsBatchGroupLeader(&w, &write_group);
write_thread->ExitAsBatchGroupLeader(write_group, status);
return status;
}
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
}
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
WriteThread::WriteGroup write_group;
uint64_t last_sequence;
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
write_thread->EnterAsBatchGroupLeader(&w, &write_group);
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
// Note: no need to update last_batch_group_size_ here since the batch writes
// to WAL only
// TODO: this use of operator bool on `tracer_` can avoid unnecessary lock
// grabs but does not seem thread-safe.
if (tracer_) {
InstrumentedMutexLock lock(&trace_mutex_);
if (tracer_ != nullptr && tracer_->IsWriteOrderPreserved()) {
for (auto* writer : write_group) {
// TODO: maybe handle the tracing status?
tracer_->Write(writer->batch).PermitUncheckedError();
}
}
}
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
size_t pre_release_callback_cnt = 0;
size_t total_byte_size = 0;
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
for (auto* writer : write_group) {
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
assert(writer);
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
if (writer->CheckCallback(this)) {
total_byte_size = WriteBatchInternal::AppendedByteSize(
total_byte_size, WriteBatchInternal::ByteSize(writer->batch));
if (writer->pre_release_callback) {
pre_release_callback_cnt++;
}
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
}
}
const bool concurrent_update = true;
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
// Update stats while we are an exclusive group leader, so we know
// that nobody else can be writing to these particular stats.
// We're optimistic, updating the stats before we successfully
// commit. That lets us release our leader status early.
auto stats = default_cf_internal_stats_;
stats->AddDBStats(InternalStats::kIntStatsBytesWritten, total_byte_size,
concurrent_update);
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
RecordTick(stats_, BYTES_WRITTEN, total_byte_size);
stats->AddDBStats(InternalStats::kIntStatsWriteDoneBySelf, 1,
concurrent_update);
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
RecordTick(stats_, WRITE_DONE_BY_SELF);
auto write_done_by_other = write_group.size - 1;
if (write_done_by_other > 0) {
stats->AddDBStats(InternalStats::kIntStatsWriteDoneByOther,
write_done_by_other, concurrent_update);
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
RecordTick(stats_, WRITE_DONE_BY_OTHER, write_done_by_other);
}
RecordInHistogram(stats_, BYTES_PER_WRITE, total_byte_size);
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
PERF_TIMER_STOP(write_pre_and_post_process_time);
PERF_TIMER_GUARD(write_wal_time);
// LastAllocatedSequence is increased inside WriteToWAL under
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
// wal_write_mutex_ to ensure ordered events in WAL
size_t seq_inc = 0 /* total_count */;
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
if (assign_order == kDoAssignOrder) {
size_t total_batch_cnt = 0;
for (auto* writer : write_group) {
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
assert(writer->batch_cnt || !seq_per_batch_);
if (!writer->CallbackFailed()) {
total_batch_cnt += writer->batch_cnt;
}
}
seq_inc = total_batch_cnt;
}
Status status;
if (!write_options.disableWAL) {
IOStatus io_s =
ConcurrentWriteToWAL(write_group, log_used, &last_sequence, seq_inc);
status = io_s;
// last_sequence may not be set if there is an error
// This error checking and return is moved up to avoid using uninitialized
// last_sequence.
if (!io_s.ok()) {
IOStatusCheck(io_s);
write_thread->ExitAsBatchGroupLeader(write_group, status);
return status;
}
} else {
// Otherwise we inc seq number to do solely the seq allocation
last_sequence = versions_->FetchAddLastAllocatedSequence(seq_inc);
}
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
size_t memtable_write_cnt = 0;
auto curr_seq = last_sequence + 1;
for (auto* writer : write_group) {
if (writer->CallbackFailed()) {
continue;
}
writer->sequence = curr_seq;
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
if (assign_order == kDoAssignOrder) {
assert(writer->batch_cnt || !seq_per_batch_);
curr_seq += writer->batch_cnt;
}
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
if (!writer->disable_memtable) {
memtable_write_cnt++;
}
// else seq advances only by memtable writes
}
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
if (status.ok() && write_options.sync) {
assert(!write_options.disableWAL);
// Requesting sync with two_write_queues_ is expected to be very rare. We
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
// hance provide a simple implementation that is not necessarily efficient.
if (manual_wal_flush_) {
status = FlushWAL(true);
} else {
status = SyncWAL();
}
}
PERF_TIMER_START(write_pre_and_post_process_time);
if (!w.CallbackFailed()) {
WriteStatusCheck(status);
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
}
if (status.ok()) {
size_t index = 0;
for (auto* writer : write_group) {
if (!writer->CallbackFailed() && writer->pre_release_callback) {
assert(writer->sequence != kMaxSequenceNumber);
Status ws = writer->pre_release_callback->Callback(
writer->sequence, disable_memtable, writer->log_used, index++,
pre_release_callback_cnt);
if (!ws.ok()) {
status = ws;
break;
}
}
}
}
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
if (publish_last_seq == kDoPublishLastSeq) {
versions_->SetLastSequence(last_sequence + seq_inc);
// Currently we only use kDoPublishLastSeq in unordered_write
assert(immutable_db_options_.unordered_write);
}
if (immutable_db_options_.unordered_write && status.ok()) {
pending_memtable_writes_ += memtable_write_cnt;
}
write_thread->ExitAsBatchGroupLeader(write_group, status);
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
if (status.ok()) {
status = w.FinalStatus();
}
if (seq_used != nullptr) {
*seq_used = w.sequence;
}
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
return status;
}
void DBImpl::WriteStatusCheckOnLocked(const Status& status) {
// Is setting bg_error_ enough here? This will at least stop
// compaction and fail any further writes.
InstrumentedMutexLock l(&mutex_);
assert(!status.IsIOFenced() || !error_handler_.GetBGError().ok());
if (immutable_db_options_.paranoid_checks && !status.ok() &&
!status.IsBusy() && !status.IsIncomplete()) {
// Maybe change the return status to void?
error_handler_.SetBGError(status, BackgroundErrorReason::kWriteCallback);
}
}
void DBImpl::WriteStatusCheck(const Status& status) {
// Is setting bg_error_ enough here? This will at least stop
// compaction and fail any further writes.
assert(!status.IsIOFenced() || !error_handler_.GetBGError().ok());
if (immutable_db_options_.paranoid_checks && !status.ok() &&
!status.IsBusy() && !status.IsIncomplete()) {
mutex_.Lock();
// Maybe change the return status to void?
error_handler_.SetBGError(status, BackgroundErrorReason::kWriteCallback);
mutex_.Unlock();
}
}
void DBImpl::IOStatusCheck(const IOStatus& io_status) {
// Is setting bg_error_ enough here? This will at least stop
// compaction and fail any further writes.
if ((immutable_db_options_.paranoid_checks && !io_status.ok() &&
!io_status.IsBusy() && !io_status.IsIncomplete()) ||
io_status.IsIOFenced()) {
mutex_.Lock();
// Maybe change the return status to void?
error_handler_.SetBGError(io_status, BackgroundErrorReason::kWriteCallback);
mutex_.Unlock();
} else {
// Force writable file to be continue writable.
logs_.back().writer->file()->reset_seen_error();
}
}
void DBImpl::MemTableInsertStatusCheck(const Status& status) {
// A non-OK status here indicates that the state implied by the
// WAL has diverged from the in-memory state. This could be
// because of a corrupt write_batch (very bad), or because the
// client specified an invalid column family and didn't specify
// ignore_missing_column_families.
if (!status.ok()) {
mutex_.Lock();
assert(!error_handler_.IsBGWorkStopped());
// Maybe change the return status to void?
error_handler_.SetBGError(status, BackgroundErrorReason::kMemTable)
.PermitUncheckedError();
mutex_.Unlock();
}
}
Status DBImpl::PreprocessWrite(const WriteOptions& write_options,
LogContext* log_context,
WriteContext* write_context) {
assert(write_context != nullptr && log_context != nullptr);
Status status;
Auto recovery from out of space errors (#4164) Summary: This commit implements automatic recovery from a Status::NoSpace() error during background operations such as write callback, flush and compaction. The broad design is as follows - 1. Compaction errors are treated as soft errors and don't put the database in read-only mode. A compaction is delayed until enough free disk space is available to accomodate the compaction outputs, which is estimated based on the input size. This means that users can continue to write, and we rely on the WriteController to delay or stop writes if the compaction debt becomes too high due to persistent low disk space condition 2. Errors during write callback and flush are treated as hard errors, i.e the database is put in read-only mode and goes back to read-write only fater certain recovery actions are taken. 3. Both types of recovery rely on the SstFileManagerImpl to poll for sufficient disk space. We assume that there is a 1-1 mapping between an SFM and the underlying OS storage container. For cases where multiple DBs are hosted on a single storage container, the user is expected to allocate a single SFM instance and use the same one for all the DBs. If no SFM is specified by the user, DBImpl::Open() will allocate one, but this will be one per DB and each DB will recover independently. The recovery implemented by SFM is as follows - a) On the first occurance of an out of space error during compaction, subsequent compactions will be delayed until the disk free space check indicates enough available space. The required space is computed as the sum of input sizes. b) The free space check requirement will be removed once the amount of free space is greater than the size reserved by in progress compactions when the first error occured c) If the out of space error is a hard error, a background thread in SFM will poll for sufficient headroom before triggering the recovery of the database and putting it in write-only mode. The headroom is calculated as the sum of the write_buffer_size of all the DB instances associated with the SFM 4. EventListener callbacks will be called at the start and completion of automatic recovery. Users can disable the auto recov ery in the start callback, and later initiate it manually by calling DB::Resume() Todo: 1. More extensive testing 2. Add disk full condition to db_stress (follow-on PR) Pull Request resolved: https://github.com/facebook/rocksdb/pull/4164 Differential Revision: D9846378 Pulled By: anand1976 fbshipit-source-id: 80ea875dbd7f00205e19c82215ff6e37da10da4a
6 years ago
if (error_handler_.IsDBStopped()) {
InstrumentedMutexLock l(&mutex_);
Auto recovery from out of space errors (#4164) Summary: This commit implements automatic recovery from a Status::NoSpace() error during background operations such as write callback, flush and compaction. The broad design is as follows - 1. Compaction errors are treated as soft errors and don't put the database in read-only mode. A compaction is delayed until enough free disk space is available to accomodate the compaction outputs, which is estimated based on the input size. This means that users can continue to write, and we rely on the WriteController to delay or stop writes if the compaction debt becomes too high due to persistent low disk space condition 2. Errors during write callback and flush are treated as hard errors, i.e the database is put in read-only mode and goes back to read-write only fater certain recovery actions are taken. 3. Both types of recovery rely on the SstFileManagerImpl to poll for sufficient disk space. We assume that there is a 1-1 mapping between an SFM and the underlying OS storage container. For cases where multiple DBs are hosted on a single storage container, the user is expected to allocate a single SFM instance and use the same one for all the DBs. If no SFM is specified by the user, DBImpl::Open() will allocate one, but this will be one per DB and each DB will recover independently. The recovery implemented by SFM is as follows - a) On the first occurance of an out of space error during compaction, subsequent compactions will be delayed until the disk free space check indicates enough available space. The required space is computed as the sum of input sizes. b) The free space check requirement will be removed once the amount of free space is greater than the size reserved by in progress compactions when the first error occured c) If the out of space error is a hard error, a background thread in SFM will poll for sufficient headroom before triggering the recovery of the database and putting it in write-only mode. The headroom is calculated as the sum of the write_buffer_size of all the DB instances associated with the SFM 4. EventListener callbacks will be called at the start and completion of automatic recovery. Users can disable the auto recov ery in the start callback, and later initiate it manually by calling DB::Resume() Todo: 1. More extensive testing 2. Add disk full condition to db_stress (follow-on PR) Pull Request resolved: https://github.com/facebook/rocksdb/pull/4164 Differential Revision: D9846378 Pulled By: anand1976 fbshipit-source-id: 80ea875dbd7f00205e19c82215ff6e37da10da4a
6 years ago
status = error_handler_.GetBGError();
}
PERF_TIMER_GUARD(write_scheduling_flushes_compactions_time);
if (UNLIKELY(status.ok() && total_log_size_ > GetMaxTotalWalSize())) {
assert(versions_);
InstrumentedMutexLock l(&mutex_);
const ColumnFamilySet* const column_families =
versions_->GetColumnFamilySet();
assert(column_families);
size_t num_cfs = column_families->NumberOfColumnFamilies();
assert(num_cfs >= 1);
if (num_cfs > 1) {
WaitForPendingWrites();
status = SwitchWAL(write_context);
}
}
if (UNLIKELY(status.ok() && write_buffer_manager_->ShouldFlush())) {
// Before a new memtable is added in SwitchMemtable(),
// write_buffer_manager_->ShouldFlush() will keep returning true. If another
// thread is writing to another DB with the same write buffer, they may also
// be flushed. We may end up with flushing much more DBs than needed. It's
// suboptimal but still correct.
InstrumentedMutexLock l(&mutex_);
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
WaitForPendingWrites();
status = HandleWriteBufferManagerFlush(write_context);
}
Refactor trimming logic for immutable memtables (#5022) Summary: MyRocks currently sets `max_write_buffer_number_to_maintain` in order to maintain enough history for transaction conflict checking. The effectiveness of this approach depends on the size of memtables. When memtables are small, it may not keep enough history; when memtables are large, this may consume too much memory. We are proposing a new way to configure memtable list history: by limiting the memory usage of immutable memtables. The new option is `max_write_buffer_size_to_maintain` and it will take precedence over the old `max_write_buffer_number_to_maintain` if they are both set to non-zero values. The new option accounts for the total memory usage of flushed immutable memtables and mutable memtable. When the total usage exceeds the limit, RocksDB may start dropping immutable memtables (which is also called trimming history), starting from the oldest one. The semantics of the old option actually works both as an upper bound and lower bound. History trimming will start if number of immutable memtables exceeds the limit, but it will never go below (limit-1) due to history trimming. In order the mimic the behavior with the new option, history trimming will stop if dropping the next immutable memtable causes the total memory usage go below the size limit. For example, assuming the size limit is set to 64MB, and there are 3 immutable memtables with sizes of 20, 30, 30. Although the total memory usage is 80MB > 64MB, dropping the oldest memtable will reduce the memory usage to 60MB < 64MB, so in this case no memtable will be dropped. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5022 Differential Revision: D14394062 Pulled By: miasantreble fbshipit-source-id: 60457a509c6af89d0993f988c9b5c2aa9e45f5c5
5 years ago
if (UNLIKELY(status.ok() && !trim_history_scheduler_.Empty())) {
InstrumentedMutexLock l(&mutex_);
Refactor trimming logic for immutable memtables (#5022) Summary: MyRocks currently sets `max_write_buffer_number_to_maintain` in order to maintain enough history for transaction conflict checking. The effectiveness of this approach depends on the size of memtables. When memtables are small, it may not keep enough history; when memtables are large, this may consume too much memory. We are proposing a new way to configure memtable list history: by limiting the memory usage of immutable memtables. The new option is `max_write_buffer_size_to_maintain` and it will take precedence over the old `max_write_buffer_number_to_maintain` if they are both set to non-zero values. The new option accounts for the total memory usage of flushed immutable memtables and mutable memtable. When the total usage exceeds the limit, RocksDB may start dropping immutable memtables (which is also called trimming history), starting from the oldest one. The semantics of the old option actually works both as an upper bound and lower bound. History trimming will start if number of immutable memtables exceeds the limit, but it will never go below (limit-1) due to history trimming. In order the mimic the behavior with the new option, history trimming will stop if dropping the next immutable memtable causes the total memory usage go below the size limit. For example, assuming the size limit is set to 64MB, and there are 3 immutable memtables with sizes of 20, 30, 30. Although the total memory usage is 80MB > 64MB, dropping the oldest memtable will reduce the memory usage to 60MB < 64MB, so in this case no memtable will be dropped. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5022 Differential Revision: D14394062 Pulled By: miasantreble fbshipit-source-id: 60457a509c6af89d0993f988c9b5c2aa9e45f5c5
5 years ago
status = TrimMemtableHistory(write_context);
}
if (UNLIKELY(status.ok() && !flush_scheduler_.Empty())) {
InstrumentedMutexLock l(&mutex_);
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
6 years ago
WaitForPendingWrites();
status = ScheduleFlushes(write_context);
}
PERF_TIMER_STOP(write_scheduling_flushes_compactions_time);
PERF_TIMER_GUARD(write_pre_and_post_process_time);
if (UNLIKELY(status.ok() && (write_controller_.IsStopped() ||
write_controller_.NeedsDelay()))) {
PERF_TIMER_STOP(write_pre_and_post_process_time);
PERF_TIMER_GUARD(write_delay_time);
// We don't know size of curent batch so that we always use the size
// for previous one. It might create a fairness issue that expiration
// might happen for smaller writes but larger writes can go through.
// Can optimize it if it is an issue.
InstrumentedMutexLock l(&mutex_);
Fix DelayWrite() calls for two_write_queues (#11130) Summary: PR https://github.com/facebook/rocksdb/issues/11020 fixed a case where it was easy to deadlock the DB with LockWAL() but introduced a bug showing up as a rare assertion failure in the stress test. Specifically, `assert(w->state == STATE_INIT)` in `WriteThread::LinkOne()` called from `BeginWriteStall()`, `DelayWrite()`, `WriteImplWALOnly()`. I haven't been about to generate a unit test that reproduces this failure but I believe the root cause is that DelayWrite() was never meant to be re-entrant, only called from the DB's write_thread_ leader. https://github.com/facebook/rocksdb/issues/11020 introduced a call to DelayWrite() from the nonmem_write_thread_ group leader. This fix is to make DelayWrite() apply to the specific write queue that it is being called from (inject a dummy write stall entry to the head of the appropriate write queue). WriteController is re-entrant, based on polling and state changes signalled with bg_cv_, so can manage stalling two queues. The only anticipated complication (called out by Andrew in previous PR) is that we don't want timed write delays being injected in parallel for the two queues, because that dimishes the intended throttling effect. Thus, we only allow timed delays for the primary write queue. HISTORY not updated because this is intended for the same release where the bug was introduced. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11130 Test Plan: Although I was not able to reproduce the assertion failure, I was able to reproduce a distinct flaw with what I believe is the same root cause: a kind of deadlock if both write queues need to wake up from stopped writes. Only one will be waiting on bg_cv_ (the other waiting in `LinkOne()` for the write queue to open up), so a single SignalAll() will only unblock one of the queues, with the other re-instating the stop until another signal on bg_cv_. A simple unit test is added for this case. Will also run crash_test_with_multiops_wc_txn for a while looking for issues. Reviewed By: ajkr Differential Revision: D42749330 Pulled By: pdillinger fbshipit-source-id: 4317dd899a93d57c26fd5af7143038f82d4d4d1b
2 years ago
status = DelayWrite(last_batch_group_size_, write_thread_, write_options);
PERF_TIMER_START(write_pre_and_post_process_time);
}
Stall writes in WriteBufferManager when memory_usage exceeds buffer_size (#7898) Summary: When WriteBufferManager is shared across DBs and column families to maintain memory usage under a limit, OOMs have been observed when flush cannot finish but writes continuously insert to memtables. In order to avoid OOMs, when memory usage goes beyond buffer_limit_ and DBs tries to write, this change will stall incoming writers until flush is completed and memory_usage drops. Design: Stall condition: When total memory usage exceeds WriteBufferManager::buffer_size_ (memory_usage() >= buffer_size_) WriterBufferManager::ShouldStall() returns true. DBImpl first block incoming/future writers by calling write_thread_.BeginWriteStall() (which adds dummy stall object to the writer's queue). Then DB is blocked on a state State::Blocked (current write doesn't go through). WBStallInterface object maintained by every DB instance is added to the queue of WriteBufferManager. If multiple DBs tries to write during this stall, they will also be blocked when check WriteBufferManager::ShouldStall() returns true. End Stall condition: When flush is finished and memory usage goes down, stall will end only if memory waiting to be flushed is less than buffer_size/2. This lower limit will give time for flush to complete and avoid continous stalling if memory usage remains close to buffer_size. WriterBufferManager::EndWriteStall() is called, which removes all instances from its queue and signal them to continue. Their state is changed to State::Running and they are unblocked. DBImpl then signal all incoming writers of that DB to continue by calling write_thread_.EndWriteStall() (which removes dummy stall object from the queue). DB instance creates WBMStallInterface which is an interface to block and signal DBs during stall. When DB needs to be blocked or signalled by WriteBufferManager, state_for_wbm_ state is changed accordingly (RUNNING or BLOCKED). Pull Request resolved: https://github.com/facebook/rocksdb/pull/7898 Test Plan: Added a new test db/db_write_buffer_manager_test.cc Reviewed By: anand1976 Differential Revision: D26093227 Pulled By: akankshamahajan15 fbshipit-source-id: 2bbd982a3fb7033f6de6153aa92a221249861aae
4 years ago
// If memory usage exceeded beyond a certain threshold,
// write_buffer_manager_->ShouldStall() returns true to all threads writing to
// all DBs and writers will be stalled.
// It does soft checking because WriteBufferManager::buffer_limit_ has already
// exceeded at this point so no new write (including current one) will go
// through until memory usage is decreased.
if (UNLIKELY(status.ok() && write_buffer_manager_->ShouldStall())) {
New stat rocksdb.{cf|db}-write-stall-stats exposed in a structural way (#11300) Summary: **Context/Summary:** Users are interested in figuring out what has caused write stall. - Refactor write stall related stats from property `kCFStats` into its own db property `rocksdb.cf-write-stall-stats` as a map or string. For now, this only contains count of different combination of (CF-scope `WriteStallCause`) + (`WriteStallCondition`) - Add new `WriteStallCause::kWriteBufferManagerLimit` to reflect write stall caused by write buffer manager - Add new `rocksdb.db-write-stall-stats`. For now, this only contains `WriteStallCause::kWriteBufferManagerLimit` + `WriteStallCondition::kStopped` - Expose functions in new class `WriteStallStatsMapKeys` for examining the above two properties returned as map - Misc: rename/comment some write stall InternalStats for clarity Pull Request resolved: https://github.com/facebook/rocksdb/pull/11300 Test Plan: - New UT - Stress test `python3 tools/db_crashtest.py blackbox --simple --get_property_one_in=1` - Perf test: Both converge very slowly at similar rates but post-change has higher average ops/sec than pre-change even though they are run at the same time. ``` ./db_bench -seed=1679014417652004 -db=/dev/shm/testdb/ -statistics=false -benchmarks="fillseq[-X60]" -key_size=32 -value_size=512 -num=100000 -db_write_buffer_size=655 -target_file_size_base=655 -disable_auto_compactions=false -compression_type=none -bloom_bits=3 ``` pre-change: ``` fillseq [AVG 15 runs] : 1176 (± 732) ops/sec; 0.6 (± 0.4) MB/sec fillseq : 1052.671 micros/op 949 ops/sec 105.267 seconds 100000 operations; 0.5 MB/s fillseq [AVG 16 runs] : 1162 (± 685) ops/sec; 0.6 (± 0.4) MB/sec fillseq : 1387.330 micros/op 720 ops/sec 138.733 seconds 100000 operations; 0.4 MB/s fillseq [AVG 17 runs] : 1136 (± 646) ops/sec; 0.6 (± 0.3) MB/sec fillseq : 1232.011 micros/op 811 ops/sec 123.201 seconds 100000 operations; 0.4 MB/s fillseq [AVG 18 runs] : 1118 (± 610) ops/sec; 0.6 (± 0.3) MB/sec fillseq : 1282.567 micros/op 779 ops/sec 128.257 seconds 100000 operations; 0.4 MB/s fillseq [AVG 19 runs] : 1100 (± 578) ops/sec; 0.6 (± 0.3) MB/sec fillseq : 1914.336 micros/op 522 ops/sec 191.434 seconds 100000 operations; 0.3 MB/s fillseq [AVG 20 runs] : 1071 (± 551) ops/sec; 0.6 (± 0.3) MB/sec fillseq : 1227.510 micros/op 814 ops/sec 122.751 seconds 100000 operations; 0.4 MB/s fillseq [AVG 21 runs] : 1059 (± 525) ops/sec; 0.5 (± 0.3) MB/sec ``` post-change: ``` fillseq [AVG 15 runs] : 1226 (± 732) ops/sec; 0.6 (± 0.4) MB/sec fillseq : 1323.825 micros/op 755 ops/sec 132.383 seconds 100000 operations; 0.4 MB/s fillseq [AVG 16 runs] : 1196 (± 687) ops/sec; 0.6 (± 0.4) MB/sec fillseq : 1223.905 micros/op 817 ops/sec 122.391 seconds 100000 operations; 0.4 MB/s fillseq [AVG 17 runs] : 1174 (± 647) ops/sec; 0.6 (± 0.3) MB/sec fillseq : 1168.996 micros/op 855 ops/sec 116.900 seconds 100000 operations; 0.4 MB/s fillseq [AVG 18 runs] : 1156 (± 611) ops/sec; 0.6 (± 0.3) MB/sec fillseq : 1348.729 micros/op 741 ops/sec 134.873 seconds 100000 operations; 0.4 MB/s fillseq [AVG 19 runs] : 1134 (± 579) ops/sec; 0.6 (± 0.3) MB/sec fillseq : 1196.887 micros/op 835 ops/sec 119.689 seconds 100000 operations; 0.4 MB/s fillseq [AVG 20 runs] : 1119 (± 550) ops/sec; 0.6 (± 0.3) MB/sec fillseq : 1193.697 micros/op 837 ops/sec 119.370 seconds 100000 operations; 0.4 MB/s fillseq [AVG 21 runs] : 1106 (± 524) ops/sec; 0.6 (± 0.3) MB/sec ``` Reviewed By: ajkr Differential Revision: D44159541 Pulled By: hx235 fbshipit-source-id: 8d29efb70001fdc52d34535eeb3364fc3e71e40b
2 years ago
default_cf_internal_stats_->AddDBStats(
InternalStats::kIntStatsWriteBufferManagerLimitStopsCounts, 1,
true /* concurrent */);
Stall writes in WriteBufferManager when memory_usage exceeds buffer_size (#7898) Summary: When WriteBufferManager is shared across DBs and column families to maintain memory usage under a limit, OOMs have been observed when flush cannot finish but writes continuously insert to memtables. In order to avoid OOMs, when memory usage goes beyond buffer_limit_ and DBs tries to write, this change will stall incoming writers until flush is completed and memory_usage drops. Design: Stall condition: When total memory usage exceeds WriteBufferManager::buffer_size_ (memory_usage() >= buffer_size_) WriterBufferManager::ShouldStall() returns true. DBImpl first block incoming/future writers by calling write_thread_.BeginWriteStall() (which adds dummy stall object to the writer's queue). Then DB is blocked on a state State::Blocked (current write doesn't go through). WBStallInterface object maintained by every DB instance is added to the queue of WriteBufferManager. If multiple DBs tries to write during this stall, they will also be blocked when check WriteBufferManager::ShouldStall() returns true. End Stall condition: When flush is finished and memory usage goes down, stall will end only if memory waiting to be flushed is less than buffer_size/2. This lower limit will give time for flush to complete and avoid continous stalling if memory usage remains close to buffer_size. WriterBufferManager::EndWriteStall() is called, which removes all instances from its queue and signal them to continue. Their state is changed to State::Running and they are unblocked. DBImpl then signal all incoming writers of that DB to continue by calling write_thread_.EndWriteStall() (which removes dummy stall object from the queue). DB instance creates WBMStallInterface which is an interface to block and signal DBs during stall. When DB needs to be blocked or signalled by WriteBufferManager, state_for_wbm_ state is changed accordingly (RUNNING or BLOCKED). Pull Request resolved: https://github.com/facebook/rocksdb/pull/7898 Test Plan: Added a new test db/db_write_buffer_manager_test.cc Reviewed By: anand1976 Differential Revision: D26093227 Pulled By: akankshamahajan15 fbshipit-source-id: 2bbd982a3fb7033f6de6153aa92a221249861aae
4 years ago
if (write_options.no_slowdown) {
status = Status::Incomplete("Write stall");
} else {
InstrumentedMutexLock l(&mutex_);
Stall writes in WriteBufferManager when memory_usage exceeds buffer_size (#7898) Summary: When WriteBufferManager is shared across DBs and column families to maintain memory usage under a limit, OOMs have been observed when flush cannot finish but writes continuously insert to memtables. In order to avoid OOMs, when memory usage goes beyond buffer_limit_ and DBs tries to write, this change will stall incoming writers until flush is completed and memory_usage drops. Design: Stall condition: When total memory usage exceeds WriteBufferManager::buffer_size_ (memory_usage() >= buffer_size_) WriterBufferManager::ShouldStall() returns true. DBImpl first block incoming/future writers by calling write_thread_.BeginWriteStall() (which adds dummy stall object to the writer's queue). Then DB is blocked on a state State::Blocked (current write doesn't go through). WBStallInterface object maintained by every DB instance is added to the queue of WriteBufferManager. If multiple DBs tries to write during this stall, they will also be blocked when check WriteBufferManager::ShouldStall() returns true. End Stall condition: When flush is finished and memory usage goes down, stall will end only if memory waiting to be flushed is less than buffer_size/2. This lower limit will give time for flush to complete and avoid continous stalling if memory usage remains close to buffer_size. WriterBufferManager::EndWriteStall() is called, which removes all instances from its queue and signal them to continue. Their state is changed to State::Running and they are unblocked. DBImpl then signal all incoming writers of that DB to continue by calling write_thread_.EndWriteStall() (which removes dummy stall object from the queue). DB instance creates WBMStallInterface which is an interface to block and signal DBs during stall. When DB needs to be blocked or signalled by WriteBufferManager, state_for_wbm_ state is changed accordingly (RUNNING or BLOCKED). Pull Request resolved: https://github.com/facebook/rocksdb/pull/7898 Test Plan: Added a new test db/db_write_buffer_manager_test.cc Reviewed By: anand1976 Differential Revision: D26093227 Pulled By: akankshamahajan15 fbshipit-source-id: 2bbd982a3fb7033f6de6153aa92a221249861aae
4 years ago
WriteBufferManagerStallWrites();
}
}
InstrumentedMutexLock l(&log_write_mutex_);
if (status.ok() && log_context->need_log_sync) {
// Wait until the parallel syncs are finished. Any sync process has to sync
// the front log too so it is enough to check the status of front()
// We do a while loop since log_sync_cv_ is signalled when any sync is
// finished
// Note: there does not seem to be a reason to wait for parallel sync at
// this early step but it is not important since parallel sync (SyncWAL) and
// need_log_sync are usually not used together.
while (logs_.front().IsSyncing()) {
log_sync_cv_.Wait();
}
for (auto& log : logs_) {
// This is just to prevent the logs to be synced by a parallel SyncWAL
// call. We will do the actual syncing later after we will write to the
// WAL.
// Note: there does not seem to be a reason to set this early before we
// actually write to the WAL
log.PrepareForSync();
}
} else {
log_context->need_log_sync = false;
}
log_context->writer = logs_.back().writer;
log_context->need_log_dir_sync =
log_context->need_log_dir_sync && !log_dir_synced_;
log_context->log_file_number_size = std::addressof(alive_log_files_.back());
return status;
}
Status DBImpl::MergeBatch(const WriteThread::WriteGroup& write_group,
WriteBatch* tmp_batch, WriteBatch** merged_batch,
size_t* write_with_wal,
WriteBatch** to_be_cached_state) {
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
assert(write_with_wal != nullptr);
assert(tmp_batch != nullptr);
assert(*to_be_cached_state == nullptr);
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
*write_with_wal = 0;
auto* leader = write_group.leader;
assert(!leader->disable_wal); // Same holds for all in the batch group
if (write_group.size == 1 && !leader->CallbackFailed() &&
leader->batch->GetWalTerminationPoint().is_cleared()) {
// we simply write the first WriteBatch to WAL if the group only
// contains one batch, that batch should be written to the WAL,
// and the batch is not wanting to be truncated
*merged_batch = leader->batch;
if (WriteBatchInternal::IsLatestPersistentState(*merged_batch)) {
*to_be_cached_state = *merged_batch;
}
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
*write_with_wal = 1;
} else {
// WAL needs all of the batches flattened into a single batch.
// We could avoid copying here with an iov-like AddRecord
// interface
*merged_batch = tmp_batch;
for (auto writer : write_group) {
if (!writer->CallbackFailed()) {
Status s = WriteBatchInternal::Append(*merged_batch, writer->batch,
/*WAL_only*/ true);
if (!s.ok()) {
tmp_batch->Clear();
return s;
}
if (WriteBatchInternal::IsLatestPersistentState(writer->batch)) {
// We only need to cache the last of such write batch
*to_be_cached_state = writer->batch;
}
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
(*write_with_wal)++;
}
}
}
// return merged_batch;
return Status::OK();
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
}
// When two_write_queues_ is disabled, this function is called from the only
// write thread. Otherwise this must be called holding log_write_mutex_.
IOStatus DBImpl::WriteToWAL(const WriteBatch& merged_batch,
log::Writer* log_writer, uint64_t* log_used,
Rate-limit automatic WAL flush after each user write (#9607) Summary: **Context:** WAL flush is currently not rate-limited by `Options::rate_limiter`. This PR is to provide rate-limiting to auto WAL flush, the one that automatically happen after each user write operation (i.e, `Options::manual_wal_flush == false`), by adding `WriteOptions::rate_limiter_options`. Note that we are NOT rate-limiting WAL flush that do NOT automatically happen after each user write, such as `Options::manual_wal_flush == true + manual FlushWAL()` (rate-limiting multiple WAL flushes), for the benefits of: - being consistent with [ReadOptions::rate_limiter_priority](https://github.com/facebook/rocksdb/blob/7.0.fb/include/rocksdb/options.h#L515) - being able to turn off some WAL flush's rate-limiting but not all (e.g, turn off specific the WAL flush of a critical user write like a service's heartbeat) `WriteOptions::rate_limiter_options` only accept `Env::IO_USER` and `Env::IO_TOTAL` currently due to an implementation constraint. - The constraint is that we currently queue parallel writes (including WAL writes) based on FIFO policy which does not factor rate limiter priority into this layer's scheduling. If we allow lower priorities such as `Env::IO_HIGH/MID/LOW` and such writes specified with lower priorities occurs before ones specified with higher priorities (even just by a tiny bit in arrival time), the former would have blocked the latter, leading to a "priority inversion" issue and contradictory to what we promise for rate-limiting priority. Therefore we only allow `Env::IO_USER` and `Env::IO_TOTAL` right now before improving that scheduling. A pre-requisite to this feature is to support operation-level rate limiting in `WritableFileWriter`, which is also included in this PR. **Summary:** - Renamed test suite `DBRateLimiterTest to DBRateLimiterOnReadTest` for adding a new test suite - Accept `rate_limiter_priority` in `WritableFileWriter`'s private and public write functions - Passed `WriteOptions::rate_limiter_options` to `WritableFileWriter` in the path of automatic WAL flush. Pull Request resolved: https://github.com/facebook/rocksdb/pull/9607 Test Plan: - Added new unit test to verify existing flush/compaction rate-limiting does not break, since `DBTest, RateLimitingTest` is disabled and current db-level rate-limiting tests focus on read only (e.g, `db_rate_limiter_test`, `DBTest2, RateLimitedCompactionReads`). - Added new unit test `DBRateLimiterOnWriteWALTest, AutoWalFlush` - `strace -ftt -e trace=write ./db_bench -benchmarks=fillseq -db=/dev/shm/testdb -rate_limit_auto_wal_flush=1 -rate_limiter_bytes_per_sec=15 -rate_limiter_refill_period_us=1000000 -write_buffer_size=100000000 -disable_auto_compactions=1 -num=100` - verified that WAL flush(i.e, system-call _write_) were chunked into 15 bytes and each _write_ was roughly 1 second apart - verified the chunking disappeared when `-rate_limit_auto_wal_flush=0` - crash test: `python3 tools/db_crashtest.py blackbox --disable_wal=0 --rate_limit_auto_wal_flush=1 --rate_limiter_bytes_per_sec=10485760 --interval=10` killed as normal **Benchmarked on flush/compaction to ensure no performance regression:** - compaction with rate-limiting (see table 1, avg over 1280-run): pre-change: **915635 micros/op**; post-change: **907350 micros/op (improved by 0.106%)** ``` #!/bin/bash TEST_TMPDIR=/dev/shm/testdb START=1 NUM_DATA_ENTRY=8 N=10 rm -f compact_bmk_output.txt compact_bmk_output_2.txt dont_care_output.txt for i in $(eval echo "{$START..$NUM_DATA_ENTRY}") do NUM_RUN=$(($N*(2**($i-1)))) for j in $(eval echo "{$START..$NUM_RUN}") do ./db_bench --benchmarks=fillrandom -db=$TEST_TMPDIR -disable_auto_compactions=1 -write_buffer_size=6710886 > dont_care_output.txt && ./db_bench --benchmarks=compact -use_existing_db=1 -db=$TEST_TMPDIR -level0_file_num_compaction_trigger=1 -rate_limiter_bytes_per_sec=100000000 | egrep 'compact' done > compact_bmk_output.txt && awk -v NUM_RUN=$NUM_RUN '{sum+=$3;sum_sqrt+=$3^2}END{print sum/NUM_RUN, sqrt(sum_sqrt/NUM_RUN-(sum/NUM_RUN)^2)}' compact_bmk_output.txt >> compact_bmk_output_2.txt done ``` - compaction w/o rate-limiting (see table 2, avg over 640-run): pre-change: **822197 micros/op**; post-change: **823148 micros/op (regressed by 0.12%)** ``` Same as above script, except that -rate_limiter_bytes_per_sec=0 ``` - flush with rate-limiting (see table 3, avg over 320-run, run on the [patch](https://github.com/hx235/rocksdb/commit/ee5c6023a9f6533fab9afdc681568daa21da4953) to augment current db_bench ): pre-change: **745752 micros/op**; post-change: **745331 micros/op (regressed by 0.06 %)** ``` #!/bin/bash TEST_TMPDIR=/dev/shm/testdb START=1 NUM_DATA_ENTRY=8 N=10 rm -f flush_bmk_output.txt flush_bmk_output_2.txt for i in $(eval echo "{$START..$NUM_DATA_ENTRY}") do NUM_RUN=$(($N*(2**($i-1)))) for j in $(eval echo "{$START..$NUM_RUN}") do ./db_bench -db=$TEST_TMPDIR -write_buffer_size=1048576000 -num=1000000 -rate_limiter_bytes_per_sec=100000000 -benchmarks=fillseq,flush | egrep 'flush' done > flush_bmk_output.txt && awk -v NUM_RUN=$NUM_RUN '{sum+=$3;sum_sqrt+=$3^2}END{print sum/NUM_RUN, sqrt(sum_sqrt/NUM_RUN-(sum/NUM_RUN)^2)}' flush_bmk_output.txt >> flush_bmk_output_2.txt done ``` - flush w/o rate-limiting (see table 4, avg over 320-run, run on the [patch](https://github.com/hx235/rocksdb/commit/ee5c6023a9f6533fab9afdc681568daa21da4953) to augment current db_bench): pre-change: **487512 micros/op**, post-change: **485856 micors/ops (improved by 0.34%)** ``` Same as above script, except that -rate_limiter_bytes_per_sec=0 ``` | table 1 - compact with rate-limiting| #-run | (pre-change) avg micros/op | std micros/op | (post-change) avg micros/op | std micros/op | change in avg micros/op (%) -- | -- | -- | -- | -- | -- 10 | 896978 | 16046.9 | 901242 | 15670.9 | 0.475373978 20 | 893718 | 15813 | 886505 | 17544.7 | -0.8070778478 40 | 900426 | 23882.2 | 894958 | 15104.5 | -0.6072681153 80 | 906635 | 21761.5 | 903332 | 23948.3 | -0.3643141948 160 | 898632 | 21098.9 | 907583 | 21145 | 0.9960695813 3.20E+02 | 905252 | 22785.5 | 908106 | 25325.5 | 0.3152713278 6.40E+02 | 905213 | 23598.6 | 906741 | 21370.5 | 0.1688000504 **1.28E+03** | **908316** | **23533.1** | **907350** | **24626.8** | **-0.1063506533** average over #-run | 901896.25 | 21064.9625 | 901977.125 | 20592.025 | 0.008967217682 | table 2 - compact w/o rate-limiting| #-run | (pre-change) avg micros/op | std micros/op | (post-change) avg micros/op | std micros/op | change in avg micros/op (%) -- | -- | -- | -- | -- | -- 10 | 811211 | 26996.7 | 807586 | 28456.4 | -0.4468627768 20 | 815465 | 14803.7 | 814608 | 28719.7 | -0.105093413 40 | 809203 | 26187.1 | 797835 | 25492.1 | -1.404839082 80 | 822088 | 28765.3 | 822192 | 32840.4 | 0.01265071379 160 | 821719 | 36344.7 | 821664 | 29544.9 | -0.006693285661 3.20E+02 | 820921 | 27756.4 | 821403 | 28347.7 | 0.05871454135 **6.40E+02** | **822197** | **28960.6** | **823148** | **30055.1** | **0.1156657103** average over #-run | 8.18E+05 | 2.71E+04 | 8.15E+05 | 2.91E+04 | -0.25 | table 3 - flush with rate-limiting| #-run | (pre-change) avg micros/op | std micros/op | (post-change) avg micros/op | std micros/op | change in avg micros/op (%) -- | -- | -- | -- | -- | -- 10 | 741721 | 11770.8 | 740345 | 5949.76 | -0.1855144994 20 | 735169 | 3561.83 | 743199 | 9755.77 | 1.09226586 40 | 743368 | 8891.03 | 742102 | 8683.22 | -0.1703059588 80 | 742129 | 8148.51 | 743417 | 9631.58| 0.1735547324 160 | 749045 | 9757.21 | 746256 | 9191.86 | -0.3723407806 **3.20E+02** | **745752** | **9819.65** | **745331** | **9840.62** | **-0.0564530836** 6.40E+02 | 749006 | 11080.5 | 748173 | 10578.7 | -0.1112140624 average over #-run | 743741.4286 | 9004.218571 | 744117.5714 | 9090.215714 | 0.05057441238 | table 4 - flush w/o rate-limiting| #-run | (pre-change) avg micros/op | std micros/op | (post-change) avg micros/op | std micros/op | change in avg micros/op (%) -- | -- | -- | -- | -- | -- 10 | 477283 | 24719.6 | 473864 | 12379 | -0.7163464863 20 | 486743 | 20175.2 | 502296 | 23931.3 | 3.195320734 40 | 482846 | 15309.2 | 489820 | 22259.5 | 1.444352858 80 | 491490 | 21883.1 | 490071 | 23085.7 | -0.2887139108 160 | 493347 | 28074.3 | 483609 | 21211.7 | -1.973864238 **3.20E+02** | **487512** | **21401.5** | **485856** | **22195.2** | **-0.3396839462** 6.40E+02 | 490307 | 25418.6 | 485435 | 22405.2 | -0.9936631539 average over #-run | 4.87E+05 | 2.24E+04 | 4.87E+05 | 2.11E+04 | 0.00E+00 Reviewed By: ajkr Differential Revision: D34442441 Pulled By: hx235 fbshipit-source-id: 4790f13e1e5c0a95ae1d1cc93ffcf69dc6e78bdd
3 years ago
uint64_t* log_size,
Fix a TSAN-reported bug caused by concurrent accesss to std::deque (#9686) Summary: Pull Request resolved: https://github.com/facebook/rocksdb/pull/9686 According to https://www.cplusplus.com/reference/deque/deque/back/, " The container is accessed (neither the const nor the non-const versions modify the container). The last element is potentially accessed or modified by the caller. Concurrently accessing or modifying other elements is safe. " Also according to https://www.cplusplus.com/reference/deque/deque/pop_front/, " The container is modified. The first element is modified. Concurrently accessing or modifying other elements is safe (although see iterator validity above). " In RocksDB, we never pop the last element of `DBImpl::alive_log_files_`. We have been exploiting this fact and the above two properties when ensuring correctness when `DBImpl::alive_log_files_` may be accessed concurrently. Specifically, it can be accessed in the write path when db mutex is released. Sometimes, the log_mute_ is held. It can also be accessed in `FindObsoleteFiles()` when db mutex is always held. It can also be accessed during recovery when db mutex is also held. Given the fact that we never pop the last element of alive_log_files_, we currently do not acquire additional locks when accessing it in `WriteToWAL()` as follows ``` alive_log_files_.back().AddSize(log_entry.size()); ``` This is problematic. Check source code of deque.h ``` back() _GLIBCXX_NOEXCEPT { __glibcxx_requires_nonempty(); ... } pop_front() _GLIBCXX_NOEXCEPT { ... if (this->_M_impl._M_start._M_cur != this->_M_impl._M_start._M_last - 1) { ... ++this->_M_impl._M_start._M_cur; } ... } ``` `back()` will actually call `__glibcxx_requires_nonempty()` first. If `__glibcxx_requires_nonempty()` is enabled and not an empty macro, it will call `empty()` ``` bool empty() { return this->_M_impl._M_finish == this->_M_impl._M_start; } ``` You can see that it will access `this->_M_impl._M_start`, racing with `pop_front()`. Therefore, TSAN will actually catch the bug in this case. To be able to use TSAN on our library and unit tests, we should always coordinate concurrent accesses to STL containers properly. We need to pass information about db mutex and log mutex into `WriteToWAL()`, otherwise it's impossible to know which mutex to acquire inside the function. To fix this, we can catch the tail of `alive_log_files_` by reference, so that we do not have to call `back()` in `WriteToWAL()`. Reviewed By: pdillinger Differential Revision: D34780309 fbshipit-source-id: 1def9821f0c437f2736c6a26445d75890377889b
3 years ago
Env::IOPriority rate_limiter_priority,
LogFileNumberSize& log_file_number_size) {
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
assert(log_size != nullptr);
Fix a TSAN-reported bug caused by concurrent accesss to std::deque (#9686) Summary: Pull Request resolved: https://github.com/facebook/rocksdb/pull/9686 According to https://www.cplusplus.com/reference/deque/deque/back/, " The container is accessed (neither the const nor the non-const versions modify the container). The last element is potentially accessed or modified by the caller. Concurrently accessing or modifying other elements is safe. " Also according to https://www.cplusplus.com/reference/deque/deque/pop_front/, " The container is modified. The first element is modified. Concurrently accessing or modifying other elements is safe (although see iterator validity above). " In RocksDB, we never pop the last element of `DBImpl::alive_log_files_`. We have been exploiting this fact and the above two properties when ensuring correctness when `DBImpl::alive_log_files_` may be accessed concurrently. Specifically, it can be accessed in the write path when db mutex is released. Sometimes, the log_mute_ is held. It can also be accessed in `FindObsoleteFiles()` when db mutex is always held. It can also be accessed during recovery when db mutex is also held. Given the fact that we never pop the last element of alive_log_files_, we currently do not acquire additional locks when accessing it in `WriteToWAL()` as follows ``` alive_log_files_.back().AddSize(log_entry.size()); ``` This is problematic. Check source code of deque.h ``` back() _GLIBCXX_NOEXCEPT { __glibcxx_requires_nonempty(); ... } pop_front() _GLIBCXX_NOEXCEPT { ... if (this->_M_impl._M_start._M_cur != this->_M_impl._M_start._M_last - 1) { ... ++this->_M_impl._M_start._M_cur; } ... } ``` `back()` will actually call `__glibcxx_requires_nonempty()` first. If `__glibcxx_requires_nonempty()` is enabled and not an empty macro, it will call `empty()` ``` bool empty() { return this->_M_impl._M_finish == this->_M_impl._M_start; } ``` You can see that it will access `this->_M_impl._M_start`, racing with `pop_front()`. Therefore, TSAN will actually catch the bug in this case. To be able to use TSAN on our library and unit tests, we should always coordinate concurrent accesses to STL containers properly. We need to pass information about db mutex and log mutex into `WriteToWAL()`, otherwise it's impossible to know which mutex to acquire inside the function. To fix this, we can catch the tail of `alive_log_files_` by reference, so that we do not have to call `back()` in `WriteToWAL()`. Reviewed By: pdillinger Differential Revision: D34780309 fbshipit-source-id: 1def9821f0c437f2736c6a26445d75890377889b
3 years ago
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
Slice log_entry = WriteBatchInternal::Contents(&merged_batch);
TEST_SYNC_POINT_CALLBACK("DBImpl::WriteToWAL:log_entry", &log_entry);
auto s = merged_batch.VerifyChecksum();
if (!s.ok()) {
return status_to_io_status(std::move(s));
}
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
*log_size = log_entry.size();
// When two_write_queues_ WriteToWAL has to be protected from concurretn calls
// from the two queues anyway and log_write_mutex_ is already held. Otherwise
// if manual_wal_flush_ is enabled we need to protect log_writer->AddRecord
// from possible concurrent calls via the FlushWAL by the application.
const bool needs_locking = manual_wal_flush_ && !two_write_queues_;
// Due to performance cocerns of missed branch prediction penalize the new
// manual_wal_flush_ feature (by UNLIKELY) instead of the more common case
// when we do not need any locking.
if (UNLIKELY(needs_locking)) {
log_write_mutex_.Lock();
}
IOStatus io_s = log_writer->MaybeAddUserDefinedTimestampSizeRecord(
versions_->GetColumnFamiliesTimestampSizeForRecord(),
rate_limiter_priority);
if (!io_s.ok()) {
return io_s;
}
io_s = log_writer->AddRecord(log_entry, rate_limiter_priority);
if (UNLIKELY(needs_locking)) {
log_write_mutex_.Unlock();
}
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
if (log_used != nullptr) {
*log_used = logfile_number_;
}
total_log_size_ += log_entry.size();
log_file_number_size.AddSize(*log_size);
log_empty_ = false;
return io_s;
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
}
IOStatus DBImpl::WriteToWAL(const WriteThread::WriteGroup& write_group,
log::Writer* log_writer, uint64_t* log_used,
bool need_log_sync, bool need_log_dir_sync,
SequenceNumber sequence,
LogFileNumberSize& log_file_number_size) {
IOStatus io_s;
Fix a TSAN-reported bug caused by concurrent accesss to std::deque (#9686) Summary: Pull Request resolved: https://github.com/facebook/rocksdb/pull/9686 According to https://www.cplusplus.com/reference/deque/deque/back/, " The container is accessed (neither the const nor the non-const versions modify the container). The last element is potentially accessed or modified by the caller. Concurrently accessing or modifying other elements is safe. " Also according to https://www.cplusplus.com/reference/deque/deque/pop_front/, " The container is modified. The first element is modified. Concurrently accessing or modifying other elements is safe (although see iterator validity above). " In RocksDB, we never pop the last element of `DBImpl::alive_log_files_`. We have been exploiting this fact and the above two properties when ensuring correctness when `DBImpl::alive_log_files_` may be accessed concurrently. Specifically, it can be accessed in the write path when db mutex is released. Sometimes, the log_mute_ is held. It can also be accessed in `FindObsoleteFiles()` when db mutex is always held. It can also be accessed during recovery when db mutex is also held. Given the fact that we never pop the last element of alive_log_files_, we currently do not acquire additional locks when accessing it in `WriteToWAL()` as follows ``` alive_log_files_.back().AddSize(log_entry.size()); ``` This is problematic. Check source code of deque.h ``` back() _GLIBCXX_NOEXCEPT { __glibcxx_requires_nonempty(); ... } pop_front() _GLIBCXX_NOEXCEPT { ... if (this->_M_impl._M_start._M_cur != this->_M_impl._M_start._M_last - 1) { ... ++this->_M_impl._M_start._M_cur; } ... } ``` `back()` will actually call `__glibcxx_requires_nonempty()` first. If `__glibcxx_requires_nonempty()` is enabled and not an empty macro, it will call `empty()` ``` bool empty() { return this->_M_impl._M_finish == this->_M_impl._M_start; } ``` You can see that it will access `this->_M_impl._M_start`, racing with `pop_front()`. Therefore, TSAN will actually catch the bug in this case. To be able to use TSAN on our library and unit tests, we should always coordinate concurrent accesses to STL containers properly. We need to pass information about db mutex and log mutex into `WriteToWAL()`, otherwise it's impossible to know which mutex to acquire inside the function. To fix this, we can catch the tail of `alive_log_files_` by reference, so that we do not have to call `back()` in `WriteToWAL()`. Reviewed By: pdillinger Differential Revision: D34780309 fbshipit-source-id: 1def9821f0c437f2736c6a26445d75890377889b
3 years ago
assert(!two_write_queues_);
assert(!write_group.leader->disable_wal);
// Same holds for all in the batch group
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
size_t write_with_wal = 0;
WriteBatch* to_be_cached_state = nullptr;
WriteBatch* merged_batch;
io_s = status_to_io_status(MergeBatch(write_group, &tmp_batch_, &merged_batch,
&write_with_wal, &to_be_cached_state));
if (UNLIKELY(!io_s.ok())) {
return io_s;
}
if (merged_batch == write_group.leader->batch) {
write_group.leader->log_used = logfile_number_;
} else if (write_with_wal > 1) {
for (auto writer : write_group) {
writer->log_used = logfile_number_;
}
}
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
WriteBatchInternal::SetSequence(merged_batch, sequence);
uint64_t log_size;
Rate-limit automatic WAL flush after each user write (#9607) Summary: **Context:** WAL flush is currently not rate-limited by `Options::rate_limiter`. This PR is to provide rate-limiting to auto WAL flush, the one that automatically happen after each user write operation (i.e, `Options::manual_wal_flush == false`), by adding `WriteOptions::rate_limiter_options`. Note that we are NOT rate-limiting WAL flush that do NOT automatically happen after each user write, such as `Options::manual_wal_flush == true + manual FlushWAL()` (rate-limiting multiple WAL flushes), for the benefits of: - being consistent with [ReadOptions::rate_limiter_priority](https://github.com/facebook/rocksdb/blob/7.0.fb/include/rocksdb/options.h#L515) - being able to turn off some WAL flush's rate-limiting but not all (e.g, turn off specific the WAL flush of a critical user write like a service's heartbeat) `WriteOptions::rate_limiter_options` only accept `Env::IO_USER` and `Env::IO_TOTAL` currently due to an implementation constraint. - The constraint is that we currently queue parallel writes (including WAL writes) based on FIFO policy which does not factor rate limiter priority into this layer's scheduling. If we allow lower priorities such as `Env::IO_HIGH/MID/LOW` and such writes specified with lower priorities occurs before ones specified with higher priorities (even just by a tiny bit in arrival time), the former would have blocked the latter, leading to a "priority inversion" issue and contradictory to what we promise for rate-limiting priority. Therefore we only allow `Env::IO_USER` and `Env::IO_TOTAL` right now before improving that scheduling. A pre-requisite to this feature is to support operation-level rate limiting in `WritableFileWriter`, which is also included in this PR. **Summary:** - Renamed test suite `DBRateLimiterTest to DBRateLimiterOnReadTest` for adding a new test suite - Accept `rate_limiter_priority` in `WritableFileWriter`'s private and public write functions - Passed `WriteOptions::rate_limiter_options` to `WritableFileWriter` in the path of automatic WAL flush. Pull Request resolved: https://github.com/facebook/rocksdb/pull/9607 Test Plan: - Added new unit test to verify existing flush/compaction rate-limiting does not break, since `DBTest, RateLimitingTest` is disabled and current db-level rate-limiting tests focus on read only (e.g, `db_rate_limiter_test`, `DBTest2, RateLimitedCompactionReads`). - Added new unit test `DBRateLimiterOnWriteWALTest, AutoWalFlush` - `strace -ftt -e trace=write ./db_bench -benchmarks=fillseq -db=/dev/shm/testdb -rate_limit_auto_wal_flush=1 -rate_limiter_bytes_per_sec=15 -rate_limiter_refill_period_us=1000000 -write_buffer_size=100000000 -disable_auto_compactions=1 -num=100` - verified that WAL flush(i.e, system-call _write_) were chunked into 15 bytes and each _write_ was roughly 1 second apart - verified the chunking disappeared when `-rate_limit_auto_wal_flush=0` - crash test: `python3 tools/db_crashtest.py blackbox --disable_wal=0 --rate_limit_auto_wal_flush=1 --rate_limiter_bytes_per_sec=10485760 --interval=10` killed as normal **Benchmarked on flush/compaction to ensure no performance regression:** - compaction with rate-limiting (see table 1, avg over 1280-run): pre-change: **915635 micros/op**; post-change: **907350 micros/op (improved by 0.106%)** ``` #!/bin/bash TEST_TMPDIR=/dev/shm/testdb START=1 NUM_DATA_ENTRY=8 N=10 rm -f compact_bmk_output.txt compact_bmk_output_2.txt dont_care_output.txt for i in $(eval echo "{$START..$NUM_DATA_ENTRY}") do NUM_RUN=$(($N*(2**($i-1)))) for j in $(eval echo "{$START..$NUM_RUN}") do ./db_bench --benchmarks=fillrandom -db=$TEST_TMPDIR -disable_auto_compactions=1 -write_buffer_size=6710886 > dont_care_output.txt && ./db_bench --benchmarks=compact -use_existing_db=1 -db=$TEST_TMPDIR -level0_file_num_compaction_trigger=1 -rate_limiter_bytes_per_sec=100000000 | egrep 'compact' done > compact_bmk_output.txt && awk -v NUM_RUN=$NUM_RUN '{sum+=$3;sum_sqrt+=$3^2}END{print sum/NUM_RUN, sqrt(sum_sqrt/NUM_RUN-(sum/NUM_RUN)^2)}' compact_bmk_output.txt >> compact_bmk_output_2.txt done ``` - compaction w/o rate-limiting (see table 2, avg over 640-run): pre-change: **822197 micros/op**; post-change: **823148 micros/op (regressed by 0.12%)** ``` Same as above script, except that -rate_limiter_bytes_per_sec=0 ``` - flush with rate-limiting (see table 3, avg over 320-run, run on the [patch](https://github.com/hx235/rocksdb/commit/ee5c6023a9f6533fab9afdc681568daa21da4953) to augment current db_bench ): pre-change: **745752 micros/op**; post-change: **745331 micros/op (regressed by 0.06 %)** ``` #!/bin/bash TEST_TMPDIR=/dev/shm/testdb START=1 NUM_DATA_ENTRY=8 N=10 rm -f flush_bmk_output.txt flush_bmk_output_2.txt for i in $(eval echo "{$START..$NUM_DATA_ENTRY}") do NUM_RUN=$(($N*(2**($i-1)))) for j in $(eval echo "{$START..$NUM_RUN}") do ./db_bench -db=$TEST_TMPDIR -write_buffer_size=1048576000 -num=1000000 -rate_limiter_bytes_per_sec=100000000 -benchmarks=fillseq,flush | egrep 'flush' done > flush_bmk_output.txt && awk -v NUM_RUN=$NUM_RUN '{sum+=$3;sum_sqrt+=$3^2}END{print sum/NUM_RUN, sqrt(sum_sqrt/NUM_RUN-(sum/NUM_RUN)^2)}' flush_bmk_output.txt >> flush_bmk_output_2.txt done ``` - flush w/o rate-limiting (see table 4, avg over 320-run, run on the [patch](https://github.com/hx235/rocksdb/commit/ee5c6023a9f6533fab9afdc681568daa21da4953) to augment current db_bench): pre-change: **487512 micros/op**, post-change: **485856 micors/ops (improved by 0.34%)** ``` Same as above script, except that -rate_limiter_bytes_per_sec=0 ``` | table 1 - compact with rate-limiting| #-run | (pre-change) avg micros/op | std micros/op | (post-change) avg micros/op | std micros/op | change in avg micros/op (%) -- | -- | -- | -- | -- | -- 10 | 896978 | 16046.9 | 901242 | 15670.9 | 0.475373978 20 | 893718 | 15813 | 886505 | 17544.7 | -0.8070778478 40 | 900426 | 23882.2 | 894958 | 15104.5 | -0.6072681153 80 | 906635 | 21761.5 | 903332 | 23948.3 | -0.3643141948 160 | 898632 | 21098.9 | 907583 | 21145 | 0.9960695813 3.20E+02 | 905252 | 22785.5 | 908106 | 25325.5 | 0.3152713278 6.40E+02 | 905213 | 23598.6 | 906741 | 21370.5 | 0.1688000504 **1.28E+03** | **908316** | **23533.1** | **907350** | **24626.8** | **-0.1063506533** average over #-run | 901896.25 | 21064.9625 | 901977.125 | 20592.025 | 0.008967217682 | table 2 - compact w/o rate-limiting| #-run | (pre-change) avg micros/op | std micros/op | (post-change) avg micros/op | std micros/op | change in avg micros/op (%) -- | -- | -- | -- | -- | -- 10 | 811211 | 26996.7 | 807586 | 28456.4 | -0.4468627768 20 | 815465 | 14803.7 | 814608 | 28719.7 | -0.105093413 40 | 809203 | 26187.1 | 797835 | 25492.1 | -1.404839082 80 | 822088 | 28765.3 | 822192 | 32840.4 | 0.01265071379 160 | 821719 | 36344.7 | 821664 | 29544.9 | -0.006693285661 3.20E+02 | 820921 | 27756.4 | 821403 | 28347.7 | 0.05871454135 **6.40E+02** | **822197** | **28960.6** | **823148** | **30055.1** | **0.1156657103** average over #-run | 8.18E+05 | 2.71E+04 | 8.15E+05 | 2.91E+04 | -0.25 | table 3 - flush with rate-limiting| #-run | (pre-change) avg micros/op | std micros/op | (post-change) avg micros/op | std micros/op | change in avg micros/op (%) -- | -- | -- | -- | -- | -- 10 | 741721 | 11770.8 | 740345 | 5949.76 | -0.1855144994 20 | 735169 | 3561.83 | 743199 | 9755.77 | 1.09226586 40 | 743368 | 8891.03 | 742102 | 8683.22 | -0.1703059588 80 | 742129 | 8148.51 | 743417 | 9631.58| 0.1735547324 160 | 749045 | 9757.21 | 746256 | 9191.86 | -0.3723407806 **3.20E+02** | **745752** | **9819.65** | **745331** | **9840.62** | **-0.0564530836** 6.40E+02 | 749006 | 11080.5 | 748173 | 10578.7 | -0.1112140624 average over #-run | 743741.4286 | 9004.218571 | 744117.5714 | 9090.215714 | 0.05057441238 | table 4 - flush w/o rate-limiting| #-run | (pre-change) avg micros/op | std micros/op | (post-change) avg micros/op | std micros/op | change in avg micros/op (%) -- | -- | -- | -- | -- | -- 10 | 477283 | 24719.6 | 473864 | 12379 | -0.7163464863 20 | 486743 | 20175.2 | 502296 | 23931.3 | 3.195320734 40 | 482846 | 15309.2 | 489820 | 22259.5 | 1.444352858 80 | 491490 | 21883.1 | 490071 | 23085.7 | -0.2887139108 160 | 493347 | 28074.3 | 483609 | 21211.7 | -1.973864238 **3.20E+02** | **487512** | **21401.5** | **485856** | **22195.2** | **-0.3396839462** 6.40E+02 | 490307 | 25418.6 | 485435 | 22405.2 | -0.9936631539 average over #-run | 4.87E+05 | 2.24E+04 | 4.87E+05 | 2.11E+04 | 0.00E+00 Reviewed By: ajkr Differential Revision: D34442441 Pulled By: hx235 fbshipit-source-id: 4790f13e1e5c0a95ae1d1cc93ffcf69dc6e78bdd
3 years ago
io_s = WriteToWAL(*merged_batch, log_writer, log_used, &log_size,
write_group.leader->rate_limiter_priority,
log_file_number_size);
if (to_be_cached_state) {
cached_recoverable_state_ = *to_be_cached_state;
cached_recoverable_state_empty_ = false;
}
if (io_s.ok() && need_log_sync) {
StopWatch sw(immutable_db_options_.clock, stats_, WAL_FILE_SYNC_MICROS);
// It's safe to access logs_ with unlocked mutex_ here because:
// - we've set getting_synced=true for all logs,
// so other threads won't pop from logs_ while we're here,
// - only writer thread can push to logs_, and we're in
// writer thread, so no one will push to logs_,
// - as long as other threads don't modify it, it's safe to read
// from std::deque from multiple threads concurrently.
//
// Sync operation should work with locked log_write_mutex_, because:
// when DBOptions.manual_wal_flush_ is set,
// FlushWAL function will be invoked by another thread.
// if without locked log_write_mutex_, the log file may get data
// corruption
const bool needs_locking = manual_wal_flush_ && !two_write_queues_;
if (UNLIKELY(needs_locking)) {
log_write_mutex_.Lock();
}
for (auto& log : logs_) {
io_s = log.writer->file()->Sync(immutable_db_options_.use_fsync);
if (!io_s.ok()) {
break;
}
}
if (UNLIKELY(needs_locking)) {
log_write_mutex_.Unlock();
}
if (io_s.ok() && need_log_dir_sync) {
// We only sync WAL directory the first time WAL syncing is
// requested, so that in case users never turn on WAL sync,
// we can avoid the disk I/O in the write code path.
io_s = directories_.GetWalDir()->FsyncWithDirOptions(
IOOptions(), nullptr,
DirFsyncOptions(DirFsyncOptions::FsyncReason::kNewFileSynced));
}
}
if (merged_batch == &tmp_batch_) {
tmp_batch_.Clear();
}
if (io_s.ok()) {
auto stats = default_cf_internal_stats_;
if (need_log_sync) {
stats->AddDBStats(InternalStats::kIntStatsWalFileSynced, 1);
RecordTick(stats_, WAL_FILE_SYNCED);
}
stats->AddDBStats(InternalStats::kIntStatsWalFileBytes, log_size);
RecordTick(stats_, WAL_FILE_BYTES, log_size);
stats->AddDBStats(InternalStats::kIntStatsWriteWithWal, write_with_wal);
RecordTick(stats_, WRITE_WITH_WAL, write_with_wal);
}
return io_s;
}
IOStatus DBImpl::ConcurrentWriteToWAL(
const WriteThread::WriteGroup& write_group, uint64_t* log_used,
SequenceNumber* last_sequence, size_t seq_inc) {
IOStatus io_s;
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
Fix a TSAN-reported bug caused by concurrent accesss to std::deque (#9686) Summary: Pull Request resolved: https://github.com/facebook/rocksdb/pull/9686 According to https://www.cplusplus.com/reference/deque/deque/back/, " The container is accessed (neither the const nor the non-const versions modify the container). The last element is potentially accessed or modified by the caller. Concurrently accessing or modifying other elements is safe. " Also according to https://www.cplusplus.com/reference/deque/deque/pop_front/, " The container is modified. The first element is modified. Concurrently accessing or modifying other elements is safe (although see iterator validity above). " In RocksDB, we never pop the last element of `DBImpl::alive_log_files_`. We have been exploiting this fact and the above two properties when ensuring correctness when `DBImpl::alive_log_files_` may be accessed concurrently. Specifically, it can be accessed in the write path when db mutex is released. Sometimes, the log_mute_ is held. It can also be accessed in `FindObsoleteFiles()` when db mutex is always held. It can also be accessed during recovery when db mutex is also held. Given the fact that we never pop the last element of alive_log_files_, we currently do not acquire additional locks when accessing it in `WriteToWAL()` as follows ``` alive_log_files_.back().AddSize(log_entry.size()); ``` This is problematic. Check source code of deque.h ``` back() _GLIBCXX_NOEXCEPT { __glibcxx_requires_nonempty(); ... } pop_front() _GLIBCXX_NOEXCEPT { ... if (this->_M_impl._M_start._M_cur != this->_M_impl._M_start._M_last - 1) { ... ++this->_M_impl._M_start._M_cur; } ... } ``` `back()` will actually call `__glibcxx_requires_nonempty()` first. If `__glibcxx_requires_nonempty()` is enabled and not an empty macro, it will call `empty()` ``` bool empty() { return this->_M_impl._M_finish == this->_M_impl._M_start; } ``` You can see that it will access `this->_M_impl._M_start`, racing with `pop_front()`. Therefore, TSAN will actually catch the bug in this case. To be able to use TSAN on our library and unit tests, we should always coordinate concurrent accesses to STL containers properly. We need to pass information about db mutex and log mutex into `WriteToWAL()`, otherwise it's impossible to know which mutex to acquire inside the function. To fix this, we can catch the tail of `alive_log_files_` by reference, so that we do not have to call `back()` in `WriteToWAL()`. Reviewed By: pdillinger Differential Revision: D34780309 fbshipit-source-id: 1def9821f0c437f2736c6a26445d75890377889b
3 years ago
assert(two_write_queues_ || immutable_db_options_.unordered_write);
assert(!write_group.leader->disable_wal);
// Same holds for all in the batch group
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
WriteBatch tmp_batch;
size_t write_with_wal = 0;
WriteBatch* to_be_cached_state = nullptr;
WriteBatch* merged_batch;
io_s = status_to_io_status(MergeBatch(write_group, &tmp_batch, &merged_batch,
&write_with_wal, &to_be_cached_state));
if (UNLIKELY(!io_s.ok())) {
return io_s;
}
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
// We need to lock log_write_mutex_ since logs_ and alive_log_files might be
// pushed back concurrently
log_write_mutex_.Lock();
if (merged_batch == write_group.leader->batch) {
write_group.leader->log_used = logfile_number_;
} else if (write_with_wal > 1) {
for (auto writer : write_group) {
writer->log_used = logfile_number_;
}
}
*last_sequence = versions_->FetchAddLastAllocatedSequence(seq_inc);
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
auto sequence = *last_sequence + 1;
WriteBatchInternal::SetSequence(merged_batch, sequence);
log::Writer* log_writer = logs_.back().writer;
LogFileNumberSize& log_file_number_size = alive_log_files_.back();
assert(log_writer->get_log_number() == log_file_number_size.number);
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
uint64_t log_size;
Rate-limit automatic WAL flush after each user write (#9607) Summary: **Context:** WAL flush is currently not rate-limited by `Options::rate_limiter`. This PR is to provide rate-limiting to auto WAL flush, the one that automatically happen after each user write operation (i.e, `Options::manual_wal_flush == false`), by adding `WriteOptions::rate_limiter_options`. Note that we are NOT rate-limiting WAL flush that do NOT automatically happen after each user write, such as `Options::manual_wal_flush == true + manual FlushWAL()` (rate-limiting multiple WAL flushes), for the benefits of: - being consistent with [ReadOptions::rate_limiter_priority](https://github.com/facebook/rocksdb/blob/7.0.fb/include/rocksdb/options.h#L515) - being able to turn off some WAL flush's rate-limiting but not all (e.g, turn off specific the WAL flush of a critical user write like a service's heartbeat) `WriteOptions::rate_limiter_options` only accept `Env::IO_USER` and `Env::IO_TOTAL` currently due to an implementation constraint. - The constraint is that we currently queue parallel writes (including WAL writes) based on FIFO policy which does not factor rate limiter priority into this layer's scheduling. If we allow lower priorities such as `Env::IO_HIGH/MID/LOW` and such writes specified with lower priorities occurs before ones specified with higher priorities (even just by a tiny bit in arrival time), the former would have blocked the latter, leading to a "priority inversion" issue and contradictory to what we promise for rate-limiting priority. Therefore we only allow `Env::IO_USER` and `Env::IO_TOTAL` right now before improving that scheduling. A pre-requisite to this feature is to support operation-level rate limiting in `WritableFileWriter`, which is also included in this PR. **Summary:** - Renamed test suite `DBRateLimiterTest to DBRateLimiterOnReadTest` for adding a new test suite - Accept `rate_limiter_priority` in `WritableFileWriter`'s private and public write functions - Passed `WriteOptions::rate_limiter_options` to `WritableFileWriter` in the path of automatic WAL flush. Pull Request resolved: https://github.com/facebook/rocksdb/pull/9607 Test Plan: - Added new unit test to verify existing flush/compaction rate-limiting does not break, since `DBTest, RateLimitingTest` is disabled and current db-level rate-limiting tests focus on read only (e.g, `db_rate_limiter_test`, `DBTest2, RateLimitedCompactionReads`). - Added new unit test `DBRateLimiterOnWriteWALTest, AutoWalFlush` - `strace -ftt -e trace=write ./db_bench -benchmarks=fillseq -db=/dev/shm/testdb -rate_limit_auto_wal_flush=1 -rate_limiter_bytes_per_sec=15 -rate_limiter_refill_period_us=1000000 -write_buffer_size=100000000 -disable_auto_compactions=1 -num=100` - verified that WAL flush(i.e, system-call _write_) were chunked into 15 bytes and each _write_ was roughly 1 second apart - verified the chunking disappeared when `-rate_limit_auto_wal_flush=0` - crash test: `python3 tools/db_crashtest.py blackbox --disable_wal=0 --rate_limit_auto_wal_flush=1 --rate_limiter_bytes_per_sec=10485760 --interval=10` killed as normal **Benchmarked on flush/compaction to ensure no performance regression:** - compaction with rate-limiting (see table 1, avg over 1280-run): pre-change: **915635 micros/op**; post-change: **907350 micros/op (improved by 0.106%)** ``` #!/bin/bash TEST_TMPDIR=/dev/shm/testdb START=1 NUM_DATA_ENTRY=8 N=10 rm -f compact_bmk_output.txt compact_bmk_output_2.txt dont_care_output.txt for i in $(eval echo "{$START..$NUM_DATA_ENTRY}") do NUM_RUN=$(($N*(2**($i-1)))) for j in $(eval echo "{$START..$NUM_RUN}") do ./db_bench --benchmarks=fillrandom -db=$TEST_TMPDIR -disable_auto_compactions=1 -write_buffer_size=6710886 > dont_care_output.txt && ./db_bench --benchmarks=compact -use_existing_db=1 -db=$TEST_TMPDIR -level0_file_num_compaction_trigger=1 -rate_limiter_bytes_per_sec=100000000 | egrep 'compact' done > compact_bmk_output.txt && awk -v NUM_RUN=$NUM_RUN '{sum+=$3;sum_sqrt+=$3^2}END{print sum/NUM_RUN, sqrt(sum_sqrt/NUM_RUN-(sum/NUM_RUN)^2)}' compact_bmk_output.txt >> compact_bmk_output_2.txt done ``` - compaction w/o rate-limiting (see table 2, avg over 640-run): pre-change: **822197 micros/op**; post-change: **823148 micros/op (regressed by 0.12%)** ``` Same as above script, except that -rate_limiter_bytes_per_sec=0 ``` - flush with rate-limiting (see table 3, avg over 320-run, run on the [patch](https://github.com/hx235/rocksdb/commit/ee5c6023a9f6533fab9afdc681568daa21da4953) to augment current db_bench ): pre-change: **745752 micros/op**; post-change: **745331 micros/op (regressed by 0.06 %)** ``` #!/bin/bash TEST_TMPDIR=/dev/shm/testdb START=1 NUM_DATA_ENTRY=8 N=10 rm -f flush_bmk_output.txt flush_bmk_output_2.txt for i in $(eval echo "{$START..$NUM_DATA_ENTRY}") do NUM_RUN=$(($N*(2**($i-1)))) for j in $(eval echo "{$START..$NUM_RUN}") do ./db_bench -db=$TEST_TMPDIR -write_buffer_size=1048576000 -num=1000000 -rate_limiter_bytes_per_sec=100000000 -benchmarks=fillseq,flush | egrep 'flush' done > flush_bmk_output.txt && awk -v NUM_RUN=$NUM_RUN '{sum+=$3;sum_sqrt+=$3^2}END{print sum/NUM_RUN, sqrt(sum_sqrt/NUM_RUN-(sum/NUM_RUN)^2)}' flush_bmk_output.txt >> flush_bmk_output_2.txt done ``` - flush w/o rate-limiting (see table 4, avg over 320-run, run on the [patch](https://github.com/hx235/rocksdb/commit/ee5c6023a9f6533fab9afdc681568daa21da4953) to augment current db_bench): pre-change: **487512 micros/op**, post-change: **485856 micors/ops (improved by 0.34%)** ``` Same as above script, except that -rate_limiter_bytes_per_sec=0 ``` | table 1 - compact with rate-limiting| #-run | (pre-change) avg micros/op | std micros/op | (post-change) avg micros/op | std micros/op | change in avg micros/op (%) -- | -- | -- | -- | -- | -- 10 | 896978 | 16046.9 | 901242 | 15670.9 | 0.475373978 20 | 893718 | 15813 | 886505 | 17544.7 | -0.8070778478 40 | 900426 | 23882.2 | 894958 | 15104.5 | -0.6072681153 80 | 906635 | 21761.5 | 903332 | 23948.3 | -0.3643141948 160 | 898632 | 21098.9 | 907583 | 21145 | 0.9960695813 3.20E+02 | 905252 | 22785.5 | 908106 | 25325.5 | 0.3152713278 6.40E+02 | 905213 | 23598.6 | 906741 | 21370.5 | 0.1688000504 **1.28E+03** | **908316** | **23533.1** | **907350** | **24626.8** | **-0.1063506533** average over #-run | 901896.25 | 21064.9625 | 901977.125 | 20592.025 | 0.008967217682 | table 2 - compact w/o rate-limiting| #-run | (pre-change) avg micros/op | std micros/op | (post-change) avg micros/op | std micros/op | change in avg micros/op (%) -- | -- | -- | -- | -- | -- 10 | 811211 | 26996.7 | 807586 | 28456.4 | -0.4468627768 20 | 815465 | 14803.7 | 814608 | 28719.7 | -0.105093413 40 | 809203 | 26187.1 | 797835 | 25492.1 | -1.404839082 80 | 822088 | 28765.3 | 822192 | 32840.4 | 0.01265071379 160 | 821719 | 36344.7 | 821664 | 29544.9 | -0.006693285661 3.20E+02 | 820921 | 27756.4 | 821403 | 28347.7 | 0.05871454135 **6.40E+02** | **822197** | **28960.6** | **823148** | **30055.1** | **0.1156657103** average over #-run | 8.18E+05 | 2.71E+04 | 8.15E+05 | 2.91E+04 | -0.25 | table 3 - flush with rate-limiting| #-run | (pre-change) avg micros/op | std micros/op | (post-change) avg micros/op | std micros/op | change in avg micros/op (%) -- | -- | -- | -- | -- | -- 10 | 741721 | 11770.8 | 740345 | 5949.76 | -0.1855144994 20 | 735169 | 3561.83 | 743199 | 9755.77 | 1.09226586 40 | 743368 | 8891.03 | 742102 | 8683.22 | -0.1703059588 80 | 742129 | 8148.51 | 743417 | 9631.58| 0.1735547324 160 | 749045 | 9757.21 | 746256 | 9191.86 | -0.3723407806 **3.20E+02** | **745752** | **9819.65** | **745331** | **9840.62** | **-0.0564530836** 6.40E+02 | 749006 | 11080.5 | 748173 | 10578.7 | -0.1112140624 average over #-run | 743741.4286 | 9004.218571 | 744117.5714 | 9090.215714 | 0.05057441238 | table 4 - flush w/o rate-limiting| #-run | (pre-change) avg micros/op | std micros/op | (post-change) avg micros/op | std micros/op | change in avg micros/op (%) -- | -- | -- | -- | -- | -- 10 | 477283 | 24719.6 | 473864 | 12379 | -0.7163464863 20 | 486743 | 20175.2 | 502296 | 23931.3 | 3.195320734 40 | 482846 | 15309.2 | 489820 | 22259.5 | 1.444352858 80 | 491490 | 21883.1 | 490071 | 23085.7 | -0.2887139108 160 | 493347 | 28074.3 | 483609 | 21211.7 | -1.973864238 **3.20E+02** | **487512** | **21401.5** | **485856** | **22195.2** | **-0.3396839462** 6.40E+02 | 490307 | 25418.6 | 485435 | 22405.2 | -0.9936631539 average over #-run | 4.87E+05 | 2.24E+04 | 4.87E+05 | 2.11E+04 | 0.00E+00 Reviewed By: ajkr Differential Revision: D34442441 Pulled By: hx235 fbshipit-source-id: 4790f13e1e5c0a95ae1d1cc93ffcf69dc6e78bdd
3 years ago
io_s = WriteToWAL(*merged_batch, log_writer, log_used, &log_size,
Fix a TSAN-reported bug caused by concurrent accesss to std::deque (#9686) Summary: Pull Request resolved: https://github.com/facebook/rocksdb/pull/9686 According to https://www.cplusplus.com/reference/deque/deque/back/, " The container is accessed (neither the const nor the non-const versions modify the container). The last element is potentially accessed or modified by the caller. Concurrently accessing or modifying other elements is safe. " Also according to https://www.cplusplus.com/reference/deque/deque/pop_front/, " The container is modified. The first element is modified. Concurrently accessing or modifying other elements is safe (although see iterator validity above). " In RocksDB, we never pop the last element of `DBImpl::alive_log_files_`. We have been exploiting this fact and the above two properties when ensuring correctness when `DBImpl::alive_log_files_` may be accessed concurrently. Specifically, it can be accessed in the write path when db mutex is released. Sometimes, the log_mute_ is held. It can also be accessed in `FindObsoleteFiles()` when db mutex is always held. It can also be accessed during recovery when db mutex is also held. Given the fact that we never pop the last element of alive_log_files_, we currently do not acquire additional locks when accessing it in `WriteToWAL()` as follows ``` alive_log_files_.back().AddSize(log_entry.size()); ``` This is problematic. Check source code of deque.h ``` back() _GLIBCXX_NOEXCEPT { __glibcxx_requires_nonempty(); ... } pop_front() _GLIBCXX_NOEXCEPT { ... if (this->_M_impl._M_start._M_cur != this->_M_impl._M_start._M_last - 1) { ... ++this->_M_impl._M_start._M_cur; } ... } ``` `back()` will actually call `__glibcxx_requires_nonempty()` first. If `__glibcxx_requires_nonempty()` is enabled and not an empty macro, it will call `empty()` ``` bool empty() { return this->_M_impl._M_finish == this->_M_impl._M_start; } ``` You can see that it will access `this->_M_impl._M_start`, racing with `pop_front()`. Therefore, TSAN will actually catch the bug in this case. To be able to use TSAN on our library and unit tests, we should always coordinate concurrent accesses to STL containers properly. We need to pass information about db mutex and log mutex into `WriteToWAL()`, otherwise it's impossible to know which mutex to acquire inside the function. To fix this, we can catch the tail of `alive_log_files_` by reference, so that we do not have to call `back()` in `WriteToWAL()`. Reviewed By: pdillinger Differential Revision: D34780309 fbshipit-source-id: 1def9821f0c437f2736c6a26445d75890377889b
3 years ago
write_group.leader->rate_limiter_priority,
log_file_number_size);
if (to_be_cached_state) {
cached_recoverable_state_ = *to_be_cached_state;
cached_recoverable_state_empty_ = false;
}
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
log_write_mutex_.Unlock();
if (io_s.ok()) {
const bool concurrent = true;
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
auto stats = default_cf_internal_stats_;
stats->AddDBStats(InternalStats::kIntStatsWalFileBytes, log_size,
concurrent);
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
RecordTick(stats_, WAL_FILE_BYTES, log_size);
stats->AddDBStats(InternalStats::kIntStatsWriteWithWal, write_with_wal,
concurrent);
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
RecordTick(stats_, WRITE_WITH_WAL, write_with_wal);
}
return io_s;
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
}
Status DBImpl::WriteRecoverableState() {
mutex_.AssertHeld();
if (!cached_recoverable_state_empty_) {
bool dont_care_bool;
SequenceNumber next_seq;
if (two_write_queues_) {
log_write_mutex_.Lock();
}
SequenceNumber seq;
if (two_write_queues_) {
seq = versions_->FetchAddLastAllocatedSequence(0);
} else {
seq = versions_->LastSequence();
}
WriteBatchInternal::SetSequence(&cached_recoverable_state_, seq + 1);
auto status = WriteBatchInternal::InsertInto(
&cached_recoverable_state_, column_family_memtables_.get(),
Refactor trimming logic for immutable memtables (#5022) Summary: MyRocks currently sets `max_write_buffer_number_to_maintain` in order to maintain enough history for transaction conflict checking. The effectiveness of this approach depends on the size of memtables. When memtables are small, it may not keep enough history; when memtables are large, this may consume too much memory. We are proposing a new way to configure memtable list history: by limiting the memory usage of immutable memtables. The new option is `max_write_buffer_size_to_maintain` and it will take precedence over the old `max_write_buffer_number_to_maintain` if they are both set to non-zero values. The new option accounts for the total memory usage of flushed immutable memtables and mutable memtable. When the total usage exceeds the limit, RocksDB may start dropping immutable memtables (which is also called trimming history), starting from the oldest one. The semantics of the old option actually works both as an upper bound and lower bound. History trimming will start if number of immutable memtables exceeds the limit, but it will never go below (limit-1) due to history trimming. In order the mimic the behavior with the new option, history trimming will stop if dropping the next immutable memtable causes the total memory usage go below the size limit. For example, assuming the size limit is set to 64MB, and there are 3 immutable memtables with sizes of 20, 30, 30. Although the total memory usage is 80MB > 64MB, dropping the oldest memtable will reduce the memory usage to 60MB < 64MB, so in this case no memtable will be dropped. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5022 Differential Revision: D14394062 Pulled By: miasantreble fbshipit-source-id: 60457a509c6af89d0993f988c9b5c2aa9e45f5c5
5 years ago
&flush_scheduler_, &trim_history_scheduler_, true,
0 /*recovery_log_number*/, this, false /* concurrent_memtable_writes */,
&next_seq, &dont_care_bool, seq_per_batch_);
auto last_seq = next_seq - 1;
if (two_write_queues_) {
versions_->FetchAddLastAllocatedSequence(last_seq - seq);
versions_->SetLastPublishedSequence(last_seq);
}
versions_->SetLastSequence(last_seq);
if (two_write_queues_) {
log_write_mutex_.Unlock();
}
if (status.ok() && recoverable_state_pre_release_callback_) {
const bool DISABLE_MEMTABLE = true;
for (uint64_t sub_batch_seq = seq + 1;
sub_batch_seq < next_seq && status.ok(); sub_batch_seq++) {
uint64_t const no_log_num = 0;
// Unlock it since the callback might end up locking mutex. e.g.,
// AddCommitted -> AdvanceMaxEvictedSeq -> GetSnapshotListFromDB
mutex_.Unlock();
status = recoverable_state_pre_release_callback_->Callback(
sub_batch_seq, !DISABLE_MEMTABLE, no_log_num, 0, 1);
mutex_.Lock();
}
}
if (status.ok()) {
cached_recoverable_state_.Clear();
cached_recoverable_state_empty_ = true;
}
return status;
}
return Status::OK();
}
void DBImpl::SelectColumnFamiliesForAtomicFlush(
Fix bug of prematurely excluded CF in atomic flush contains unflushed data that should've been included in the atomic flush (#11148) Summary: **Context:** Atomic flush should guarantee recoverability of all data of seqno up to the max seqno of the flush. It achieves this by ensuring all such data are flushed by the time this atomic flush finishes through `SelectColumnFamiliesForAtomicFlush()`. However, our crash test exposed the following case where an excluded CF from an atomic flush contains unflushed data of seqno less than the max seqno of that atomic flush and loses its data with `WriteOptions::DisableWAL=true` in face of a crash right after the atomic flush finishes . ``` ./db_stress --preserve_unverified_changes=1 --reopen=0 --acquire_snapshot_one_in=0 --adaptive_readahead=1 --allow_data_in_errors=True --async_io=1 --atomic_flush=1 --avoid_flush_during_recovery=0 --avoid_unnecessary_blocking_io=0 --backup_max_size=104857600 --backup_one_in=0 --batch_protection_bytes_per_key=0 --block_size=16384 --bloom_bits=15 --bottommost_compression_type=none --bytes_per_sync=262144 --cache_index_and_filter_blocks=0 --cache_size=8388608 --cache_type=lru_cache --charge_compression_dictionary_building_buffer=0 --charge_file_metadata=1 --charge_filter_construction=0 --charge_table_reader=0 --checkpoint_one_in=0 --checksum_type=kXXH3 --clear_column_family_one_in=0 --compact_files_one_in=0 --compact_range_one_in=0 --compaction_pri=1 --compaction_ttl=100 --compression_max_dict_buffer_bytes=134217727 --compression_max_dict_bytes=16384 --compression_parallel_threads=1 --compression_type=lz4hc --compression_use_zstd_dict_trainer=0 --compression_zstd_max_train_bytes=0 --continuous_verification_interval=0 --data_block_index_type=0 --db=$db --db_write_buffer_size=1048576 --delpercent=4 --delrangepercent=1 --destroy_db_initially=0 --detect_filter_construct_corruption=0 --disable_wal=1 --enable_compaction_filter=0 --enable_pipelined_write=0 --expected_values_dir=$exp --fail_if_options_file_error=0 --fifo_allow_compaction=0 --file_checksum_impl=none --flush_one_in=0 --format_version=5 --get_current_wal_file_one_in=0 --get_live_files_one_in=100 --get_property_one_in=0 --get_sorted_wal_files_one_in=0 --index_block_restart_interval=2 --index_type=0 --ingest_external_file_one_in=0 --initial_auto_readahead_size=524288 --iterpercent=10 --key_len_percent_dist=1,30,69 --level_compaction_dynamic_level_bytes=True --long_running_snapshots=1 --manual_wal_flush_one_in=100 --mark_for_compaction_one_file_in=0 --max_auto_readahead_size=0 --max_background_compactions=20 --max_bytes_for_level_base=10485760 --max_key=10000 --max_key_len=3 --max_manifest_file_size=1073741824 --max_write_batch_group_size_bytes=64 --max_write_buffer_number=3 --max_write_buffer_size_to_maintain=0 --memtable_prefix_bloom_size_ratio=0.01 --memtable_protection_bytes_per_key=4 --memtable_whole_key_filtering=0 --memtablerep=skip_list --min_write_buffer_number_to_merge=2 --mmap_read=1 --mock_direct_io=False --nooverwritepercent=1 --num_file_reads_for_auto_readahead=0 --open_files=-1 --open_metadata_write_fault_one_in=0 --open_read_fault_one_in=0 --open_write_fault_one_in=0 --ops_per_thread=100000000 --optimize_filters_for_memory=1 --paranoid_file_checks=1 --partition_filters=0 --partition_pinning=3 --pause_background_one_in=0 --periodic_compaction_seconds=100 --prefix_size=8 --prefixpercent=5 --prepopulate_block_cache=0 --preserve_internal_time_seconds=3600 --progress_reports=0 --read_fault_one_in=32 --readahead_size=16384 --readpercent=50 --recycle_log_file_num=0 --ribbon_starting_level=6 --secondary_cache_fault_one_in=0 --set_options_one_in=10000 --snapshot_hold_ops=100000 --sst_file_manager_bytes_per_sec=104857600 --sst_file_manager_bytes_per_truncate=1048576 --stats_dump_period_sec=10 --subcompactions=1 --sync=0 --sync_fault_injection=0 --target_file_size_base=524288 --target_file_size_multiplier=2 --test_batches_snapshots=0 --top_level_index_pinning=0 --unpartitioned_pinning=1 --use_direct_io_for_flush_and_compaction=0 --use_direct_reads=0 --use_full_merge_v1=0 --use_merge=0 --use_multiget=1 --use_put_entity_one_in=0 --user_timestamp_size=0 --value_size_mult=32 --verify_checksum=1 --verify_checksum_one_in=0 --verify_db_one_in=1000 --verify_sst_unique_id_in_manifest=1 --wal_bytes_per_sync=524288 --wal_compression=none --write_buffer_size=524288 --write_dbid_to_manifest=1 --write_fault_one_in=0 --writepercent=30 & pid=$! sleep 0.2 sleep 10 kill $pid sleep 0.2 ./db_stress --ops_per_thread=1 --preserve_unverified_changes=1 --reopen=0 --acquire_snapshot_one_in=0 --adaptive_readahead=1 --allow_data_in_errors=True --async_io=1 --atomic_flush=1 --avoid_flush_during_recovery=0 --avoid_unnecessary_blocking_io=0 --backup_max_size=104857600 --backup_one_in=0 --batch_protection_bytes_per_key=0 --block_size=16384 --bloom_bits=15 --bottommost_compression_type=none --bytes_per_sync=262144 --cache_index_and_filter_blocks=0 --cache_size=8388608 --cache_type=lru_cache --charge_compression_dictionary_building_buffer=0 --charge_file_metadata=1 --charge_filter_construction=0 --charge_table_reader=0 --checkpoint_one_in=0 --checksum_type=kXXH3 --clear_column_family_one_in=0 --compact_files_one_in=0 --compact_range_one_in=0 --compaction_pri=1 --compaction_ttl=100 --compression_max_dict_buffer_bytes=134217727 --compression_max_dict_bytes=16384 --compression_parallel_threads=1 --compression_type=lz4hc --compression_use_zstd_dict_trainer=0 --compression_zstd_max_train_bytes=0 --continuous_verification_interval=0 --data_block_index_type=0 --db=$db --db_write_buffer_size=1048576 --delpercent=4 --delrangepercent=1 --destroy_db_initially=0 --detect_filter_construct_corruption=0 --disable_wal=1 --enable_compaction_filter=0 --enable_pipelined_write=0 --expected_values_dir=$exp --fail_if_options_file_error=0 --fifo_allow_compaction=0 --file_checksum_impl=none --flush_one_in=0 --format_version=5 --get_current_wal_file_one_in=0 --get_live_files_one_in=100 --get_property_one_in=0 --get_sorted_wal_files_one_in=0 --index_block_restart_interval=2 --index_type=0 --ingest_external_file_one_in=0 --initial_auto_readahead_size=524288 --iterpercent=10 --key_len_percent_dist=1,30,69 --level_compaction_dynamic_level_bytes=True --long_running_snapshots=1 --manual_wal_flush_one_in=100 --mark_for_compaction_one_file_in=0 --max_auto_readahead_size=0 --max_background_compactions=20 --max_bytes_for_level_base=10485760 --max_key=10000 --max_key_len=3 --max_manifest_file_size=1073741824 --max_write_batch_group_size_bytes=64 --max_write_buffer_number=3 --max_write_buffer_size_to_maintain=0 --memtable_prefix_bloom_size_ratio=0.01 --memtable_protection_bytes_per_key=4 --memtable_whole_key_filtering=0 --memtablerep=skip_list --min_write_buffer_number_to_merge=2 --mmap_read=1 --mock_direct_io=False --nooverwritepercent=1 --num_file_reads_for_auto_readahead=0 --open_files=-1 --open_metadata_write_fault_one_in=0 --open_read_fault_one_in=0 --open_write_fault_one_in=0 --ops_per_thread=100000000 --optimize_filters_for_memory=1 --paranoid_file_checks=1 --partition_filters=0 --partition_pinning=3 --pause_background_one_in=0 --periodic_compaction_seconds=100 --prefix_size=8 --prefixpercent=5 --prepopulate_block_cache=0 --preserve_internal_time_seconds=3600 --progress_reports=0 --read_fault_one_in=32 --readahead_size=16384 --readpercent=50 --recycle_log_file_num=0 --ribbon_starting_level=6 --secondary_cache_fault_one_in=0 --set_options_one_in=10000 --snapshot_hold_ops=100000 --sst_file_manager_bytes_per_sec=104857600 --sst_file_manager_bytes_per_truncate=1048576 --stats_dump_period_sec=10 --subcompactions=1 --sync=0 --sync_fault_injection=0 --target_file_size_base=524288 --target_file_size_multiplier=2 --test_batches_snapshots=0 --top_level_index_pinning=0 --unpartitioned_pinning=1 --use_direct_io_for_flush_and_compaction=0 --use_direct_reads=0 --use_full_merge_v1=0 --use_merge=0 --use_multiget=1 --use_put_entity_one_in=0 --user_timestamp_size=0 --value_size_mult=32 --verify_checksum=1 --verify_checksum_one_in=0 --verify_db_one_in=1000 --verify_sst_unique_id_in_manifest=1 --wal_bytes_per_sync=524288 --wal_compression=none --write_buffer_size=524288 --write_dbid_to_manifest=1 --write_fault_one_in=0 --writepercent=30 & pid=$! sleep 0.2 sleep 40 kill $pid sleep 0.2 Verification failed for column family 6 key 0000000000000239000000000000012B0000000000000138 (56622): value_from_db: , value_from_expected: 4A6331754E4F4C4D42434041464744455A5B58595E5F5C5D5253505156575455, msg: Value not found: NotFound: Crash-recovery verification failed :( No writes or ops? Verification failed :( ``` The bug is due to the following: - When atomic flush is used, an empty CF is legally [excluded](https://github.com/facebook/rocksdb/blob/7.10.fb/db/db_filesnapshot.cc#L39) in `SelectColumnFamiliesForAtomicFlush` as the first step of `DBImpl::FlushForGetLiveFiles` before [passing](https://github.com/facebook/rocksdb/blob/7.10.fb/db/db_filesnapshot.cc#L42) the included CFDs to `AtomicFlushMemTables`. - But [later](https://github.com/facebook/rocksdb/blob/7.10.fb/db/db_impl/db_impl_compaction_flush.cc#L2133) in `AtomicFlushMemTables`, `WaitUntilFlushWouldNotStallWrites` will [release the db mutex](https://github.com/facebook/rocksdb/blob/7.10.fb/db/db_impl/db_impl_compaction_flush.cc#L2403), during which data@seqno N can be inserted into the excluded CF and data@seqno M can be inserted into one of the included CFs, where M > N. - However, data@seqno N in an already-excluded CF is thus excluded from this atomic flush while we seqno N is less than seqno M. **Summary:** - Replace `SelectColumnFamiliesForAtomicFlush()`-before-`AtomicFlushMemTables()` with `SelectColumnFamiliesForAtomicFlush()`-after-wait-within-`AtomicFlushMemTables()` so we ensure no write affecting the recoverability of this atomic job (i.e, change to max seqno of this atomic flush or insertion of data with less seqno than the max seqno of the atomic flush to excluded CF) can happen after calling `SelectColumnFamiliesForAtomicFlush()`. - For above, refactored and clarified comments on `SelectColumnFamiliesForAtomicFlush()` and `AtomicFlushMemTables()` for clearer semantics of passed-in CFDs to atomic-flush Pull Request resolved: https://github.com/facebook/rocksdb/pull/11148 Test Plan: - New unit test failed before the fix and passes after - Make check - Rehearsal stress test Reviewed By: ajkr Differential Revision: D42799871 Pulled By: hx235 fbshipit-source-id: 13636b63e9c25c5895857afc36ea580d57f6d644
2 years ago
autovector<ColumnFamilyData*>* selected_cfds,
const autovector<ColumnFamilyData*>& provided_candidate_cfds) {
mutex_.AssertHeld();
assert(selected_cfds);
autovector<ColumnFamilyData*> candidate_cfds;
// Generate candidate cfds if not provided
if (provided_candidate_cfds.empty()) {
for (ColumnFamilyData* cfd : *versions_->GetColumnFamilySet()) {
if (!cfd->IsDropped() && cfd->initialized()) {
cfd->Ref();
candidate_cfds.push_back(cfd);
}
}
} else {
candidate_cfds = provided_candidate_cfds;
}
for (ColumnFamilyData* cfd : candidate_cfds) {
if (cfd->IsDropped()) {
continue;
}
if (cfd->imm()->NumNotFlushed() != 0 || !cfd->mem()->IsEmpty() ||
!cached_recoverable_state_empty_.load()) {
Fix bug of prematurely excluded CF in atomic flush contains unflushed data that should've been included in the atomic flush (#11148) Summary: **Context:** Atomic flush should guarantee recoverability of all data of seqno up to the max seqno of the flush. It achieves this by ensuring all such data are flushed by the time this atomic flush finishes through `SelectColumnFamiliesForAtomicFlush()`. However, our crash test exposed the following case where an excluded CF from an atomic flush contains unflushed data of seqno less than the max seqno of that atomic flush and loses its data with `WriteOptions::DisableWAL=true` in face of a crash right after the atomic flush finishes . ``` ./db_stress --preserve_unverified_changes=1 --reopen=0 --acquire_snapshot_one_in=0 --adaptive_readahead=1 --allow_data_in_errors=True --async_io=1 --atomic_flush=1 --avoid_flush_during_recovery=0 --avoid_unnecessary_blocking_io=0 --backup_max_size=104857600 --backup_one_in=0 --batch_protection_bytes_per_key=0 --block_size=16384 --bloom_bits=15 --bottommost_compression_type=none --bytes_per_sync=262144 --cache_index_and_filter_blocks=0 --cache_size=8388608 --cache_type=lru_cache --charge_compression_dictionary_building_buffer=0 --charge_file_metadata=1 --charge_filter_construction=0 --charge_table_reader=0 --checkpoint_one_in=0 --checksum_type=kXXH3 --clear_column_family_one_in=0 --compact_files_one_in=0 --compact_range_one_in=0 --compaction_pri=1 --compaction_ttl=100 --compression_max_dict_buffer_bytes=134217727 --compression_max_dict_bytes=16384 --compression_parallel_threads=1 --compression_type=lz4hc --compression_use_zstd_dict_trainer=0 --compression_zstd_max_train_bytes=0 --continuous_verification_interval=0 --data_block_index_type=0 --db=$db --db_write_buffer_size=1048576 --delpercent=4 --delrangepercent=1 --destroy_db_initially=0 --detect_filter_construct_corruption=0 --disable_wal=1 --enable_compaction_filter=0 --enable_pipelined_write=0 --expected_values_dir=$exp --fail_if_options_file_error=0 --fifo_allow_compaction=0 --file_checksum_impl=none --flush_one_in=0 --format_version=5 --get_current_wal_file_one_in=0 --get_live_files_one_in=100 --get_property_one_in=0 --get_sorted_wal_files_one_in=0 --index_block_restart_interval=2 --index_type=0 --ingest_external_file_one_in=0 --initial_auto_readahead_size=524288 --iterpercent=10 --key_len_percent_dist=1,30,69 --level_compaction_dynamic_level_bytes=True --long_running_snapshots=1 --manual_wal_flush_one_in=100 --mark_for_compaction_one_file_in=0 --max_auto_readahead_size=0 --max_background_compactions=20 --max_bytes_for_level_base=10485760 --max_key=10000 --max_key_len=3 --max_manifest_file_size=1073741824 --max_write_batch_group_size_bytes=64 --max_write_buffer_number=3 --max_write_buffer_size_to_maintain=0 --memtable_prefix_bloom_size_ratio=0.01 --memtable_protection_bytes_per_key=4 --memtable_whole_key_filtering=0 --memtablerep=skip_list --min_write_buffer_number_to_merge=2 --mmap_read=1 --mock_direct_io=False --nooverwritepercent=1 --num_file_reads_for_auto_readahead=0 --open_files=-1 --open_metadata_write_fault_one_in=0 --open_read_fault_one_in=0 --open_write_fault_one_in=0 --ops_per_thread=100000000 --optimize_filters_for_memory=1 --paranoid_file_checks=1 --partition_filters=0 --partition_pinning=3 --pause_background_one_in=0 --periodic_compaction_seconds=100 --prefix_size=8 --prefixpercent=5 --prepopulate_block_cache=0 --preserve_internal_time_seconds=3600 --progress_reports=0 --read_fault_one_in=32 --readahead_size=16384 --readpercent=50 --recycle_log_file_num=0 --ribbon_starting_level=6 --secondary_cache_fault_one_in=0 --set_options_one_in=10000 --snapshot_hold_ops=100000 --sst_file_manager_bytes_per_sec=104857600 --sst_file_manager_bytes_per_truncate=1048576 --stats_dump_period_sec=10 --subcompactions=1 --sync=0 --sync_fault_injection=0 --target_file_size_base=524288 --target_file_size_multiplier=2 --test_batches_snapshots=0 --top_level_index_pinning=0 --unpartitioned_pinning=1 --use_direct_io_for_flush_and_compaction=0 --use_direct_reads=0 --use_full_merge_v1=0 --use_merge=0 --use_multiget=1 --use_put_entity_one_in=0 --user_timestamp_size=0 --value_size_mult=32 --verify_checksum=1 --verify_checksum_one_in=0 --verify_db_one_in=1000 --verify_sst_unique_id_in_manifest=1 --wal_bytes_per_sync=524288 --wal_compression=none --write_buffer_size=524288 --write_dbid_to_manifest=1 --write_fault_one_in=0 --writepercent=30 & pid=$! sleep 0.2 sleep 10 kill $pid sleep 0.2 ./db_stress --ops_per_thread=1 --preserve_unverified_changes=1 --reopen=0 --acquire_snapshot_one_in=0 --adaptive_readahead=1 --allow_data_in_errors=True --async_io=1 --atomic_flush=1 --avoid_flush_during_recovery=0 --avoid_unnecessary_blocking_io=0 --backup_max_size=104857600 --backup_one_in=0 --batch_protection_bytes_per_key=0 --block_size=16384 --bloom_bits=15 --bottommost_compression_type=none --bytes_per_sync=262144 --cache_index_and_filter_blocks=0 --cache_size=8388608 --cache_type=lru_cache --charge_compression_dictionary_building_buffer=0 --charge_file_metadata=1 --charge_filter_construction=0 --charge_table_reader=0 --checkpoint_one_in=0 --checksum_type=kXXH3 --clear_column_family_one_in=0 --compact_files_one_in=0 --compact_range_one_in=0 --compaction_pri=1 --compaction_ttl=100 --compression_max_dict_buffer_bytes=134217727 --compression_max_dict_bytes=16384 --compression_parallel_threads=1 --compression_type=lz4hc --compression_use_zstd_dict_trainer=0 --compression_zstd_max_train_bytes=0 --continuous_verification_interval=0 --data_block_index_type=0 --db=$db --db_write_buffer_size=1048576 --delpercent=4 --delrangepercent=1 --destroy_db_initially=0 --detect_filter_construct_corruption=0 --disable_wal=1 --enable_compaction_filter=0 --enable_pipelined_write=0 --expected_values_dir=$exp --fail_if_options_file_error=0 --fifo_allow_compaction=0 --file_checksum_impl=none --flush_one_in=0 --format_version=5 --get_current_wal_file_one_in=0 --get_live_files_one_in=100 --get_property_one_in=0 --get_sorted_wal_files_one_in=0 --index_block_restart_interval=2 --index_type=0 --ingest_external_file_one_in=0 --initial_auto_readahead_size=524288 --iterpercent=10 --key_len_percent_dist=1,30,69 --level_compaction_dynamic_level_bytes=True --long_running_snapshots=1 --manual_wal_flush_one_in=100 --mark_for_compaction_one_file_in=0 --max_auto_readahead_size=0 --max_background_compactions=20 --max_bytes_for_level_base=10485760 --max_key=10000 --max_key_len=3 --max_manifest_file_size=1073741824 --max_write_batch_group_size_bytes=64 --max_write_buffer_number=3 --max_write_buffer_size_to_maintain=0 --memtable_prefix_bloom_size_ratio=0.01 --memtable_protection_bytes_per_key=4 --memtable_whole_key_filtering=0 --memtablerep=skip_list --min_write_buffer_number_to_merge=2 --mmap_read=1 --mock_direct_io=False --nooverwritepercent=1 --num_file_reads_for_auto_readahead=0 --open_files=-1 --open_metadata_write_fault_one_in=0 --open_read_fault_one_in=0 --open_write_fault_one_in=0 --ops_per_thread=100000000 --optimize_filters_for_memory=1 --paranoid_file_checks=1 --partition_filters=0 --partition_pinning=3 --pause_background_one_in=0 --periodic_compaction_seconds=100 --prefix_size=8 --prefixpercent=5 --prepopulate_block_cache=0 --preserve_internal_time_seconds=3600 --progress_reports=0 --read_fault_one_in=32 --readahead_size=16384 --readpercent=50 --recycle_log_file_num=0 --ribbon_starting_level=6 --secondary_cache_fault_one_in=0 --set_options_one_in=10000 --snapshot_hold_ops=100000 --sst_file_manager_bytes_per_sec=104857600 --sst_file_manager_bytes_per_truncate=1048576 --stats_dump_period_sec=10 --subcompactions=1 --sync=0 --sync_fault_injection=0 --target_file_size_base=524288 --target_file_size_multiplier=2 --test_batches_snapshots=0 --top_level_index_pinning=0 --unpartitioned_pinning=1 --use_direct_io_for_flush_and_compaction=0 --use_direct_reads=0 --use_full_merge_v1=0 --use_merge=0 --use_multiget=1 --use_put_entity_one_in=0 --user_timestamp_size=0 --value_size_mult=32 --verify_checksum=1 --verify_checksum_one_in=0 --verify_db_one_in=1000 --verify_sst_unique_id_in_manifest=1 --wal_bytes_per_sync=524288 --wal_compression=none --write_buffer_size=524288 --write_dbid_to_manifest=1 --write_fault_one_in=0 --writepercent=30 & pid=$! sleep 0.2 sleep 40 kill $pid sleep 0.2 Verification failed for column family 6 key 0000000000000239000000000000012B0000000000000138 (56622): value_from_db: , value_from_expected: 4A6331754E4F4C4D42434041464744455A5B58595E5F5C5D5253505156575455, msg: Value not found: NotFound: Crash-recovery verification failed :( No writes or ops? Verification failed :( ``` The bug is due to the following: - When atomic flush is used, an empty CF is legally [excluded](https://github.com/facebook/rocksdb/blob/7.10.fb/db/db_filesnapshot.cc#L39) in `SelectColumnFamiliesForAtomicFlush` as the first step of `DBImpl::FlushForGetLiveFiles` before [passing](https://github.com/facebook/rocksdb/blob/7.10.fb/db/db_filesnapshot.cc#L42) the included CFDs to `AtomicFlushMemTables`. - But [later](https://github.com/facebook/rocksdb/blob/7.10.fb/db/db_impl/db_impl_compaction_flush.cc#L2133) in `AtomicFlushMemTables`, `WaitUntilFlushWouldNotStallWrites` will [release the db mutex](https://github.com/facebook/rocksdb/blob/7.10.fb/db/db_impl/db_impl_compaction_flush.cc#L2403), during which data@seqno N can be inserted into the excluded CF and data@seqno M can be inserted into one of the included CFs, where M > N. - However, data@seqno N in an already-excluded CF is thus excluded from this atomic flush while we seqno N is less than seqno M. **Summary:** - Replace `SelectColumnFamiliesForAtomicFlush()`-before-`AtomicFlushMemTables()` with `SelectColumnFamiliesForAtomicFlush()`-after-wait-within-`AtomicFlushMemTables()` so we ensure no write affecting the recoverability of this atomic job (i.e, change to max seqno of this atomic flush or insertion of data with less seqno than the max seqno of the atomic flush to excluded CF) can happen after calling `SelectColumnFamiliesForAtomicFlush()`. - For above, refactored and clarified comments on `SelectColumnFamiliesForAtomicFlush()` and `AtomicFlushMemTables()` for clearer semantics of passed-in CFDs to atomic-flush Pull Request resolved: https://github.com/facebook/rocksdb/pull/11148 Test Plan: - New unit test failed before the fix and passes after - Make check - Rehearsal stress test Reviewed By: ajkr Differential Revision: D42799871 Pulled By: hx235 fbshipit-source-id: 13636b63e9c25c5895857afc36ea580d57f6d644
2 years ago
selected_cfds->push_back(cfd);
}
}
// Unref the newly generated candidate cfds (when not provided) in
// `candidate_cfds`
if (provided_candidate_cfds.empty()) {
for (auto candidate_cfd : candidate_cfds) {
candidate_cfd->UnrefAndTryDelete();
}
}
}
// Assign sequence number for atomic flush.
void DBImpl::AssignAtomicFlushSeq(const autovector<ColumnFamilyData*>& cfds) {
assert(immutable_db_options_.atomic_flush);
auto seq = versions_->LastSequence();
for (auto cfd : cfds) {
cfd->imm()->AssignAtomicFlushSeq(seq);
}
}
Status DBImpl::SwitchWAL(WriteContext* write_context) {
mutex_.AssertHeld();
assert(write_context != nullptr);
Status status;
if (alive_log_files_.begin()->getting_flushed) {
return status;
}
auto oldest_alive_log = alive_log_files_.begin()->number;
Skip deleted WALs during recovery Summary: This patch record min log number to keep to the manifest while flushing SST files to ignore them and any WAL older than them during recovery. This is to avoid scenarios when we have a gap between the WAL files are fed to the recovery procedure. The gap could happen by for example out-of-order WAL deletion. Such gap could cause problems in 2PC recovery where the prepared and commit entry are placed into two separate WAL and gap in the WALs could result into not processing the WAL with the commit entry and hence breaking the 2PC recovery logic. Before the commit, for 2PC case, we determined which log number to keep in FindObsoleteFiles(). We looked at the earliest logs with outstanding prepare entries, or prepare entries whose respective commit or abort are in memtable. With the commit, the same calculation is done while we apply the SST flush. Just before installing the flush file, we precompute the earliest log file to keep after the flush finishes using the same logic (but skipping the memtables just flushed), record this information to the manifest entry for this new flushed SST file. This pre-computed value is also remembered in memory, and will later be used to determine whether a log file can be deleted. This value is unlikely to change until next flush because the commit entry will stay in memtable. (In WritePrepared, we could have removed the older log files as soon as all prepared entries are committed. It's not yet done anyway. Even if we do it, the only thing we loss with this new approach is earlier log deletion between two flushes, which does not guarantee to happen anyway because the obsolete file clean-up function is only executed after flush or compaction) This min log number to keep is stored in the manifest using the safely-ignore customized field of AddFile entry, in order to guarantee that the DB generated using newer release can be opened by previous releases no older than 4.2. Closes https://github.com/facebook/rocksdb/pull/3765 Differential Revision: D7747618 Pulled By: siying fbshipit-source-id: d00c92105b4f83852e9754a1b70d6b64cb590729
7 years ago
bool flush_wont_release_oldest_log = false;
if (allow_2pc()) {
auto oldest_log_with_uncommitted_prep =
Skip deleted WALs during recovery Summary: This patch record min log number to keep to the manifest while flushing SST files to ignore them and any WAL older than them during recovery. This is to avoid scenarios when we have a gap between the WAL files are fed to the recovery procedure. The gap could happen by for example out-of-order WAL deletion. Such gap could cause problems in 2PC recovery where the prepared and commit entry are placed into two separate WAL and gap in the WALs could result into not processing the WAL with the commit entry and hence breaking the 2PC recovery logic. Before the commit, for 2PC case, we determined which log number to keep in FindObsoleteFiles(). We looked at the earliest logs with outstanding prepare entries, or prepare entries whose respective commit or abort are in memtable. With the commit, the same calculation is done while we apply the SST flush. Just before installing the flush file, we precompute the earliest log file to keep after the flush finishes using the same logic (but skipping the memtables just flushed), record this information to the manifest entry for this new flushed SST file. This pre-computed value is also remembered in memory, and will later be used to determine whether a log file can be deleted. This value is unlikely to change until next flush because the commit entry will stay in memtable. (In WritePrepared, we could have removed the older log files as soon as all prepared entries are committed. It's not yet done anyway. Even if we do it, the only thing we loss with this new approach is earlier log deletion between two flushes, which does not guarantee to happen anyway because the obsolete file clean-up function is only executed after flush or compaction) This min log number to keep is stored in the manifest using the safely-ignore customized field of AddFile entry, in order to guarantee that the DB generated using newer release can be opened by previous releases no older than 4.2. Closes https://github.com/facebook/rocksdb/pull/3765 Differential Revision: D7747618 Pulled By: siying fbshipit-source-id: d00c92105b4f83852e9754a1b70d6b64cb590729
7 years ago
logs_with_prep_tracker_.FindMinLogContainingOutstandingPrep();
assert(oldest_log_with_uncommitted_prep == 0 ||
oldest_log_with_uncommitted_prep >= oldest_alive_log);
if (oldest_log_with_uncommitted_prep > 0 &&
oldest_log_with_uncommitted_prep == oldest_alive_log) {
Skip deleted WALs during recovery Summary: This patch record min log number to keep to the manifest while flushing SST files to ignore them and any WAL older than them during recovery. This is to avoid scenarios when we have a gap between the WAL files are fed to the recovery procedure. The gap could happen by for example out-of-order WAL deletion. Such gap could cause problems in 2PC recovery where the prepared and commit entry are placed into two separate WAL and gap in the WALs could result into not processing the WAL with the commit entry and hence breaking the 2PC recovery logic. Before the commit, for 2PC case, we determined which log number to keep in FindObsoleteFiles(). We looked at the earliest logs with outstanding prepare entries, or prepare entries whose respective commit or abort are in memtable. With the commit, the same calculation is done while we apply the SST flush. Just before installing the flush file, we precompute the earliest log file to keep after the flush finishes using the same logic (but skipping the memtables just flushed), record this information to the manifest entry for this new flushed SST file. This pre-computed value is also remembered in memory, and will later be used to determine whether a log file can be deleted. This value is unlikely to change until next flush because the commit entry will stay in memtable. (In WritePrepared, we could have removed the older log files as soon as all prepared entries are committed. It's not yet done anyway. Even if we do it, the only thing we loss with this new approach is earlier log deletion between two flushes, which does not guarantee to happen anyway because the obsolete file clean-up function is only executed after flush or compaction) This min log number to keep is stored in the manifest using the safely-ignore customized field of AddFile entry, in order to guarantee that the DB generated using newer release can be opened by previous releases no older than 4.2. Closes https://github.com/facebook/rocksdb/pull/3765 Differential Revision: D7747618 Pulled By: siying fbshipit-source-id: d00c92105b4f83852e9754a1b70d6b64cb590729
7 years ago
if (unable_to_release_oldest_log_) {
// we already attempted to flush all column families dependent on
// the oldest alive log but the log still contained uncommitted
Skip deleted WALs during recovery Summary: This patch record min log number to keep to the manifest while flushing SST files to ignore them and any WAL older than them during recovery. This is to avoid scenarios when we have a gap between the WAL files are fed to the recovery procedure. The gap could happen by for example out-of-order WAL deletion. Such gap could cause problems in 2PC recovery where the prepared and commit entry are placed into two separate WAL and gap in the WALs could result into not processing the WAL with the commit entry and hence breaking the 2PC recovery logic. Before the commit, for 2PC case, we determined which log number to keep in FindObsoleteFiles(). We looked at the earliest logs with outstanding prepare entries, or prepare entries whose respective commit or abort are in memtable. With the commit, the same calculation is done while we apply the SST flush. Just before installing the flush file, we precompute the earliest log file to keep after the flush finishes using the same logic (but skipping the memtables just flushed), record this information to the manifest entry for this new flushed SST file. This pre-computed value is also remembered in memory, and will later be used to determine whether a log file can be deleted. This value is unlikely to change until next flush because the commit entry will stay in memtable. (In WritePrepared, we could have removed the older log files as soon as all prepared entries are committed. It's not yet done anyway. Even if we do it, the only thing we loss with this new approach is earlier log deletion between two flushes, which does not guarantee to happen anyway because the obsolete file clean-up function is only executed after flush or compaction) This min log number to keep is stored in the manifest using the safely-ignore customized field of AddFile entry, in order to guarantee that the DB generated using newer release can be opened by previous releases no older than 4.2. Closes https://github.com/facebook/rocksdb/pull/3765 Differential Revision: D7747618 Pulled By: siying fbshipit-source-id: d00c92105b4f83852e9754a1b70d6b64cb590729
7 years ago
// transactions so there is still nothing that we can do.
return status;
Skip deleted WALs during recovery Summary: This patch record min log number to keep to the manifest while flushing SST files to ignore them and any WAL older than them during recovery. This is to avoid scenarios when we have a gap between the WAL files are fed to the recovery procedure. The gap could happen by for example out-of-order WAL deletion. Such gap could cause problems in 2PC recovery where the prepared and commit entry are placed into two separate WAL and gap in the WALs could result into not processing the WAL with the commit entry and hence breaking the 2PC recovery logic. Before the commit, for 2PC case, we determined which log number to keep in FindObsoleteFiles(). We looked at the earliest logs with outstanding prepare entries, or prepare entries whose respective commit or abort are in memtable. With the commit, the same calculation is done while we apply the SST flush. Just before installing the flush file, we precompute the earliest log file to keep after the flush finishes using the same logic (but skipping the memtables just flushed), record this information to the manifest entry for this new flushed SST file. This pre-computed value is also remembered in memory, and will later be used to determine whether a log file can be deleted. This value is unlikely to change until next flush because the commit entry will stay in memtable. (In WritePrepared, we could have removed the older log files as soon as all prepared entries are committed. It's not yet done anyway. Even if we do it, the only thing we loss with this new approach is earlier log deletion between two flushes, which does not guarantee to happen anyway because the obsolete file clean-up function is only executed after flush or compaction) This min log number to keep is stored in the manifest using the safely-ignore customized field of AddFile entry, in order to guarantee that the DB generated using newer release can be opened by previous releases no older than 4.2. Closes https://github.com/facebook/rocksdb/pull/3765 Differential Revision: D7747618 Pulled By: siying fbshipit-source-id: d00c92105b4f83852e9754a1b70d6b64cb590729
7 years ago
} else {
ROCKS_LOG_WARN(
immutable_db_options_.info_log,
"Unable to release oldest log due to uncommitted transaction");
Skip deleted WALs during recovery Summary: This patch record min log number to keep to the manifest while flushing SST files to ignore them and any WAL older than them during recovery. This is to avoid scenarios when we have a gap between the WAL files are fed to the recovery procedure. The gap could happen by for example out-of-order WAL deletion. Such gap could cause problems in 2PC recovery where the prepared and commit entry are placed into two separate WAL and gap in the WALs could result into not processing the WAL with the commit entry and hence breaking the 2PC recovery logic. Before the commit, for 2PC case, we determined which log number to keep in FindObsoleteFiles(). We looked at the earliest logs with outstanding prepare entries, or prepare entries whose respective commit or abort are in memtable. With the commit, the same calculation is done while we apply the SST flush. Just before installing the flush file, we precompute the earliest log file to keep after the flush finishes using the same logic (but skipping the memtables just flushed), record this information to the manifest entry for this new flushed SST file. This pre-computed value is also remembered in memory, and will later be used to determine whether a log file can be deleted. This value is unlikely to change until next flush because the commit entry will stay in memtable. (In WritePrepared, we could have removed the older log files as soon as all prepared entries are committed. It's not yet done anyway. Even if we do it, the only thing we loss with this new approach is earlier log deletion between two flushes, which does not guarantee to happen anyway because the obsolete file clean-up function is only executed after flush or compaction) This min log number to keep is stored in the manifest using the safely-ignore customized field of AddFile entry, in order to guarantee that the DB generated using newer release can be opened by previous releases no older than 4.2. Closes https://github.com/facebook/rocksdb/pull/3765 Differential Revision: D7747618 Pulled By: siying fbshipit-source-id: d00c92105b4f83852e9754a1b70d6b64cb590729
7 years ago
unable_to_release_oldest_log_ = true;
flush_wont_release_oldest_log = true;
}
}
Skip deleted WALs during recovery Summary: This patch record min log number to keep to the manifest while flushing SST files to ignore them and any WAL older than them during recovery. This is to avoid scenarios when we have a gap between the WAL files are fed to the recovery procedure. The gap could happen by for example out-of-order WAL deletion. Such gap could cause problems in 2PC recovery where the prepared and commit entry are placed into two separate WAL and gap in the WALs could result into not processing the WAL with the commit entry and hence breaking the 2PC recovery logic. Before the commit, for 2PC case, we determined which log number to keep in FindObsoleteFiles(). We looked at the earliest logs with outstanding prepare entries, or prepare entries whose respective commit or abort are in memtable. With the commit, the same calculation is done while we apply the SST flush. Just before installing the flush file, we precompute the earliest log file to keep after the flush finishes using the same logic (but skipping the memtables just flushed), record this information to the manifest entry for this new flushed SST file. This pre-computed value is also remembered in memory, and will later be used to determine whether a log file can be deleted. This value is unlikely to change until next flush because the commit entry will stay in memtable. (In WritePrepared, we could have removed the older log files as soon as all prepared entries are committed. It's not yet done anyway. Even if we do it, the only thing we loss with this new approach is earlier log deletion between two flushes, which does not guarantee to happen anyway because the obsolete file clean-up function is only executed after flush or compaction) This min log number to keep is stored in the manifest using the safely-ignore customized field of AddFile entry, in order to guarantee that the DB generated using newer release can be opened by previous releases no older than 4.2. Closes https://github.com/facebook/rocksdb/pull/3765 Differential Revision: D7747618 Pulled By: siying fbshipit-source-id: d00c92105b4f83852e9754a1b70d6b64cb590729
7 years ago
}
if (!flush_wont_release_oldest_log) {
// we only mark this log as getting flushed if we have successfully
// flushed all data in this log. If this log contains outstanding prepared
// transactions then we cannot flush this log until those transactions are
// commited.
Skip deleted WALs during recovery Summary: This patch record min log number to keep to the manifest while flushing SST files to ignore them and any WAL older than them during recovery. This is to avoid scenarios when we have a gap between the WAL files are fed to the recovery procedure. The gap could happen by for example out-of-order WAL deletion. Such gap could cause problems in 2PC recovery where the prepared and commit entry are placed into two separate WAL and gap in the WALs could result into not processing the WAL with the commit entry and hence breaking the 2PC recovery logic. Before the commit, for 2PC case, we determined which log number to keep in FindObsoleteFiles(). We looked at the earliest logs with outstanding prepare entries, or prepare entries whose respective commit or abort are in memtable. With the commit, the same calculation is done while we apply the SST flush. Just before installing the flush file, we precompute the earliest log file to keep after the flush finishes using the same logic (but skipping the memtables just flushed), record this information to the manifest entry for this new flushed SST file. This pre-computed value is also remembered in memory, and will later be used to determine whether a log file can be deleted. This value is unlikely to change until next flush because the commit entry will stay in memtable. (In WritePrepared, we could have removed the older log files as soon as all prepared entries are committed. It's not yet done anyway. Even if we do it, the only thing we loss with this new approach is earlier log deletion between two flushes, which does not guarantee to happen anyway because the obsolete file clean-up function is only executed after flush or compaction) This min log number to keep is stored in the manifest using the safely-ignore customized field of AddFile entry, in order to guarantee that the DB generated using newer release can be opened by previous releases no older than 4.2. Closes https://github.com/facebook/rocksdb/pull/3765 Differential Revision: D7747618 Pulled By: siying fbshipit-source-id: d00c92105b4f83852e9754a1b70d6b64cb590729
7 years ago
unable_to_release_oldest_log_ = false;
alive_log_files_.begin()->getting_flushed = true;
}
ROCKS_LOG_INFO(
immutable_db_options_.info_log,
"Flushing all column families with data in WAL number %" PRIu64
". Total log size is %" PRIu64 " while max_total_wal_size is %" PRIu64,
oldest_alive_log, total_log_size_.load(), GetMaxTotalWalSize());
// no need to refcount because drop is happening in write thread, so can't
// happen while we're in the write thread
autovector<ColumnFamilyData*> cfds;
if (immutable_db_options_.atomic_flush) {
SelectColumnFamiliesForAtomicFlush(&cfds);
} else {
for (auto cfd : *versions_->GetColumnFamilySet()) {
if (cfd->IsDropped()) {
continue;
}
if (cfd->OldestLogToKeep() <= oldest_alive_log) {
cfds.push_back(cfd);
}
}
MaybeFlushStatsCF(&cfds);
}
WriteThread::Writer nonmem_w;
if (two_write_queues_) {
nonmem_write_thread_.EnterUnbatched(&nonmem_w, &mutex_);
}
for (const auto cfd : cfds) {
cfd->Ref();
status = SwitchMemtable(cfd, write_context);
cfd->UnrefAndTryDelete();
if (!status.ok()) {
break;
}
}
if (two_write_queues_) {
nonmem_write_thread_.ExitUnbatched(&nonmem_w);
}
if (status.ok()) {
if (immutable_db_options_.atomic_flush) {
AssignAtomicFlushSeq(cfds);
}
for (auto cfd : cfds) {
cfd->imm()->FlushRequested();
if (!immutable_db_options_.atomic_flush) {
FlushRequest flush_req;
GenerateFlushRequest({cfd}, FlushReason::kWalFull, &flush_req);
SchedulePendingFlush(flush_req);
}
}
if (immutable_db_options_.atomic_flush) {
FlushRequest flush_req;
GenerateFlushRequest(cfds, FlushReason::kWalFull, &flush_req);
SchedulePendingFlush(flush_req);
}
MaybeScheduleFlushOrCompaction();
}
return status;
}
Status DBImpl::HandleWriteBufferManagerFlush(WriteContext* write_context) {
mutex_.AssertHeld();
assert(write_context != nullptr);
Status status;
// Before a new memtable is added in SwitchMemtable(),
// write_buffer_manager_->ShouldFlush() will keep returning true. If another
// thread is writing to another DB with the same write buffer, they may also
// be flushed. We may end up with flushing much more DBs than needed. It's
// suboptimal but still correct.
// no need to refcount because drop is happening in write thread, so can't
// happen while we're in the write thread
autovector<ColumnFamilyData*> cfds;
if (immutable_db_options_.atomic_flush) {
SelectColumnFamiliesForAtomicFlush(&cfds);
} else {
ColumnFamilyData* cfd_picked = nullptr;
SequenceNumber seq_num_for_cf_picked = kMaxSequenceNumber;
for (auto cfd : *versions_->GetColumnFamilySet()) {
if (cfd->IsDropped()) {
continue;
}
Prevent a case of WriteBufferManager flush thrashing (#6364) Summary: Previously, the flushes triggered by `WriteBufferManager` could affect the same CF repeatedly if it happens to get consecutive writes. Such flushes are not particularly useful for reducing memory usage since they switch nearly-empty memtables to immutable while they've just begun filling their first arena block. In fact they may not even reduce the mutable memory count if they involve replacing one mutable memtable containing one arena block with a new mutable memtable containing one arena block. Further, if such switches happen even a few times before a flush finishes, the immutable memtable limit will be reached and writes will stall. This PR adds a heuristic to not switch memtables to immutable for CFs that already have one or more immutable memtables awaiting flush. There is a memory usage regression if the user continues writing to the same CF, that DB does not have any CFs eligible for switching, flushes are not finishing, and the `WriteBufferManager` was constructed with `allow_stall=false`. Before, it would grow by switching nearly empty memtables until writes stall. Now, it would grow by filling memtables until writes stall. This feels like an acceptable behavior change because users who prefer to stall over violate the memory limit should be using `allow_stall=true`, which is unaffected by this PR. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6364 Test Plan: - Command: `rm -rf /dev/shm/dbbench/ && TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num_multi_db=8 -num_column_families=2 -write_buffer_size=4194304 -db_write_buffer_size=16777216 -compression_type=none -statistics=true -target_file_size_base=4194304 -max_bytes_for_level_base=16777216` - `rocksdb.db.write.stall` count before this PR: 175 - `rocksdb.db.write.stall` count after this PR: 0 Reviewed By: jay-zhuang Differential Revision: D20167197 Pulled By: ajkr fbshipit-source-id: 4a64064e9bc33d57c0a35f15547542d0191d0cb7
2 years ago
if (!cfd->mem()->IsEmpty() && !cfd->imm()->IsFlushPendingOrRunning()) {
// We only consider flush on CFs with bytes in the mutable memtable,
// and no immutable memtables for which flush has yet to finish. If
// we triggered flush on CFs already trying to flush, we would risk
// creating too many immutable memtables leading to write stalls.
uint64_t seq = cfd->mem()->GetCreationSeq();
if (cfd_picked == nullptr || seq < seq_num_for_cf_picked) {
cfd_picked = cfd;
seq_num_for_cf_picked = seq;
}
}
}
if (cfd_picked != nullptr) {
cfds.push_back(cfd_picked);
}
MaybeFlushStatsCF(&cfds);
}
Prevent a case of WriteBufferManager flush thrashing (#6364) Summary: Previously, the flushes triggered by `WriteBufferManager` could affect the same CF repeatedly if it happens to get consecutive writes. Such flushes are not particularly useful for reducing memory usage since they switch nearly-empty memtables to immutable while they've just begun filling their first arena block. In fact they may not even reduce the mutable memory count if they involve replacing one mutable memtable containing one arena block with a new mutable memtable containing one arena block. Further, if such switches happen even a few times before a flush finishes, the immutable memtable limit will be reached and writes will stall. This PR adds a heuristic to not switch memtables to immutable for CFs that already have one or more immutable memtables awaiting flush. There is a memory usage regression if the user continues writing to the same CF, that DB does not have any CFs eligible for switching, flushes are not finishing, and the `WriteBufferManager` was constructed with `allow_stall=false`. Before, it would grow by switching nearly empty memtables until writes stall. Now, it would grow by filling memtables until writes stall. This feels like an acceptable behavior change because users who prefer to stall over violate the memory limit should be using `allow_stall=true`, which is unaffected by this PR. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6364 Test Plan: - Command: `rm -rf /dev/shm/dbbench/ && TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num_multi_db=8 -num_column_families=2 -write_buffer_size=4194304 -db_write_buffer_size=16777216 -compression_type=none -statistics=true -target_file_size_base=4194304 -max_bytes_for_level_base=16777216` - `rocksdb.db.write.stall` count before this PR: 175 - `rocksdb.db.write.stall` count after this PR: 0 Reviewed By: jay-zhuang Differential Revision: D20167197 Pulled By: ajkr fbshipit-source-id: 4a64064e9bc33d57c0a35f15547542d0191d0cb7
2 years ago
if (!cfds.empty()) {
ROCKS_LOG_INFO(
immutable_db_options_.info_log,
"Flushing triggered to alleviate write buffer memory usage. Write "
"buffer is using %" ROCKSDB_PRIszt
" bytes out of a total of %" ROCKSDB_PRIszt ".",
write_buffer_manager_->memory_usage(),
write_buffer_manager_->buffer_size());
}
WriteThread::Writer nonmem_w;
if (two_write_queues_) {
nonmem_write_thread_.EnterUnbatched(&nonmem_w, &mutex_);
}
for (const auto cfd : cfds) {
if (cfd->mem()->IsEmpty()) {
continue;
}
cfd->Ref();
status = SwitchMemtable(cfd, write_context);
cfd->UnrefAndTryDelete();
if (!status.ok()) {
break;
}
}
if (two_write_queues_) {
nonmem_write_thread_.ExitUnbatched(&nonmem_w);
}
if (status.ok()) {
if (immutable_db_options_.atomic_flush) {
AssignAtomicFlushSeq(cfds);
}
for (const auto cfd : cfds) {
cfd->imm()->FlushRequested();
if (!immutable_db_options_.atomic_flush) {
FlushRequest flush_req;
GenerateFlushRequest({cfd}, FlushReason::kWriteBufferManager,
&flush_req);
SchedulePendingFlush(flush_req);
}
}
if (immutable_db_options_.atomic_flush) {
FlushRequest flush_req;
GenerateFlushRequest(cfds, FlushReason::kWriteBufferManager, &flush_req);
SchedulePendingFlush(flush_req);
}
MaybeScheduleFlushOrCompaction();
}
return status;
}
uint64_t DBImpl::GetMaxTotalWalSize() const {
uint64_t max_total_wal_size =
max_total_wal_size_.load(std::memory_order_acquire);
if (max_total_wal_size > 0) {
return max_total_wal_size;
}
return 4 * max_total_in_memory_state_.load(std::memory_order_acquire);
}
// REQUIRES: mutex_ is held
Fix DelayWrite() calls for two_write_queues (#11130) Summary: PR https://github.com/facebook/rocksdb/issues/11020 fixed a case where it was easy to deadlock the DB with LockWAL() but introduced a bug showing up as a rare assertion failure in the stress test. Specifically, `assert(w->state == STATE_INIT)` in `WriteThread::LinkOne()` called from `BeginWriteStall()`, `DelayWrite()`, `WriteImplWALOnly()`. I haven't been about to generate a unit test that reproduces this failure but I believe the root cause is that DelayWrite() was never meant to be re-entrant, only called from the DB's write_thread_ leader. https://github.com/facebook/rocksdb/issues/11020 introduced a call to DelayWrite() from the nonmem_write_thread_ group leader. This fix is to make DelayWrite() apply to the specific write queue that it is being called from (inject a dummy write stall entry to the head of the appropriate write queue). WriteController is re-entrant, based on polling and state changes signalled with bg_cv_, so can manage stalling two queues. The only anticipated complication (called out by Andrew in previous PR) is that we don't want timed write delays being injected in parallel for the two queues, because that dimishes the intended throttling effect. Thus, we only allow timed delays for the primary write queue. HISTORY not updated because this is intended for the same release where the bug was introduced. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11130 Test Plan: Although I was not able to reproduce the assertion failure, I was able to reproduce a distinct flaw with what I believe is the same root cause: a kind of deadlock if both write queues need to wake up from stopped writes. Only one will be waiting on bg_cv_ (the other waiting in `LinkOne()` for the write queue to open up), so a single SignalAll() will only unblock one of the queues, with the other re-instating the stop until another signal on bg_cv_. A simple unit test is added for this case. Will also run crash_test_with_multiops_wc_txn for a while looking for issues. Reviewed By: ajkr Differential Revision: D42749330 Pulled By: pdillinger fbshipit-source-id: 4317dd899a93d57c26fd5af7143038f82d4d4d1b
2 years ago
// REQUIRES: this thread is currently at the leader for write_thread
Status DBImpl::DelayWrite(uint64_t num_bytes, WriteThread& write_thread,
const WriteOptions& write_options) {
Revise LockWAL/UnlockWAL implementation (#11020) Summary: RocksDB has two public APIs: `DB::LockWAL()`/`DB::UnlockWAL()`. The current implementation acquires and releases the internal `DBImpl::log_write_mutex_`. According to the comment on `DBImpl::log_write_mutex_`: https://github.com/facebook/rocksdb/blob/7.8.fb/db/db_impl/db_impl.h#L2287:L2288 > Note: to avoid dealock, if needed to acquire both log_write_mutex_ and mutex_, the order should be first mutex_ and then log_write_mutex_. This puts limitations on how applications can use the `LockWAL()` API. After `LockWAL()` returns ok, then application should not perform any operation that acquires `mutex_`. Currently, the use case of `LockWAL()` is MyRocks implementing the MySQL storage engine handlerton `lock_hton_log` interface. The operation that MyRocks performs after `LockWAL()` is `GetSortedWalFiless()` which not only acquires mutex_, but also `log_write_mutex_`. There are two issues: 1. Applications using these two APIs may hang if one thread calls `GetSortedWalFiles()` after calling `LockWAL()` because log_write_mutex is not recursive. 2. Two threads may dead lock due to lock order inversion. To fix these issues, we can modify the implementation of LockWAL so that it does not keep `log_write_mutex_` held until UnlockWAL. To achieve the goal of locking the WAL, we can instead manually inject a write stall so that all future writes will be stopped. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11020 Test Plan: make check Reviewed By: ajkr Differential Revision: D41785203 Pulled By: riversand963 fbshipit-source-id: 5ccb7a9c6eb9a2c3fa80fd2c399cc2568b8f89ce
2 years ago
mutex_.AssertHeld();
uint64_t time_delayed = 0;
bool delayed = false;
{
StopWatch sw(immutable_db_options_.clock, stats_, WRITE_STALL,
Group rocksdb.sst.read.micros stat by IOActivity flush and compaction (#11288) Summary: **Context:** The existing stat rocksdb.sst.read.micros does not reflect each of compaction and flush cases but aggregate them, which is not so helpful for us to understand IO read behavior of each of them. **Summary** - Update `StopWatch` and `RandomAccessFileReader` to record `rocksdb.sst.read.micros` and `rocksdb.file.{flush/compaction}.read.micros` - Fixed the default histogram in `RandomAccessFileReader` - New field `ReadOptions/IOOptions::io_activity`; Pass `ReadOptions` through paths under db open, flush and compaction to where we can prepare `IOOptions` and pass it to `RandomAccessFileReader` - Use `thread_status_util` for assertion in `DbStressFSWrapper` for continuous testing on we are passing correct `io_activity` under db open, flush and compaction Pull Request resolved: https://github.com/facebook/rocksdb/pull/11288 Test Plan: - **Stress test** - **Db bench 1: rocksdb.sst.read.micros COUNT ≈ sum of rocksdb.file.read.flush.micros's and rocksdb.file.read.compaction.micros's.** (without blob) - May not be exactly the same due to `HistogramStat::Add` only guarantees atomic not accuracy across threads. ``` ./db_bench -db=/dev/shm/testdb/ -statistics=true -benchmarks="fillseq" -key_size=32 -value_size=512 -num=50000 -write_buffer_size=655 -target_file_size_base=655 -disable_auto_compactions=false -compression_type=none -bloom_bits=3 (-use_plain_table=1 -prefix_size=10) ``` ``` // BlockBasedTable rocksdb.sst.read.micros P50 : 2.009374 P95 : 4.968548 P99 : 8.110362 P100 : 43.000000 COUNT : 40456 SUM : 114805 rocksdb.file.read.flush.micros P50 : 1.871841 P95 : 3.872407 P99 : 5.540541 P100 : 43.000000 COUNT : 2250 SUM : 6116 rocksdb.file.read.compaction.micros P50 : 2.023109 P95 : 5.029149 P99 : 8.196910 P100 : 26.000000 COUNT : 38206 SUM : 108689 // PlainTable Does not apply ``` - **Db bench 2: performance** **Read** SETUP: db with 900 files ``` ./db_bench -db=/dev/shm/testdb/ -benchmarks="fillseq" -key_size=32 -value_size=512 -num=50000 -write_buffer_size=655 -disable_auto_compactions=true -target_file_size_base=655 -compression_type=none ```run till convergence ``` ./db_bench -seed=1678564177044286 -use_existing_db=true -db=/dev/shm/testdb -benchmarks=readrandom[-X60] -statistics=true -num=1000000 -disable_auto_compactions=true -compression_type=none -bloom_bits=3 ``` Pre-change `readrandom [AVG 60 runs] : 21568 (± 248) ops/sec` Post-change (no regression, -0.3%) `readrandom [AVG 60 runs] : 21486 (± 236) ops/sec` **Compaction/Flush**run till convergence ``` ./db_bench -db=/dev/shm/testdb2/ -seed=1678564177044286 -benchmarks="fillseq[-X60]" -key_size=32 -value_size=512 -num=50000 -write_buffer_size=655 -disable_auto_compactions=false -target_file_size_base=655 -compression_type=none rocksdb.sst.read.micros COUNT : 33820 rocksdb.sst.read.flush.micros COUNT : 1800 rocksdb.sst.read.compaction.micros COUNT : 32020 ``` Pre-change `fillseq [AVG 46 runs] : 1391 (± 214) ops/sec; 0.7 (± 0.1) MB/sec` Post-change (no regression, ~-0.4%) `fillseq [AVG 46 runs] : 1385 (± 216) ops/sec; 0.7 (± 0.1) MB/sec` Reviewed By: ajkr Differential Revision: D44007011 Pulled By: hx235 fbshipit-source-id: a54c89e4846dfc9a135389edf3f3eedfea257132
2 years ago
Histograms::HISTOGRAM_ENUM_MAX, &time_delayed);
Fix DelayWrite() calls for two_write_queues (#11130) Summary: PR https://github.com/facebook/rocksdb/issues/11020 fixed a case where it was easy to deadlock the DB with LockWAL() but introduced a bug showing up as a rare assertion failure in the stress test. Specifically, `assert(w->state == STATE_INIT)` in `WriteThread::LinkOne()` called from `BeginWriteStall()`, `DelayWrite()`, `WriteImplWALOnly()`. I haven't been about to generate a unit test that reproduces this failure but I believe the root cause is that DelayWrite() was never meant to be re-entrant, only called from the DB's write_thread_ leader. https://github.com/facebook/rocksdb/issues/11020 introduced a call to DelayWrite() from the nonmem_write_thread_ group leader. This fix is to make DelayWrite() apply to the specific write queue that it is being called from (inject a dummy write stall entry to the head of the appropriate write queue). WriteController is re-entrant, based on polling and state changes signalled with bg_cv_, so can manage stalling two queues. The only anticipated complication (called out by Andrew in previous PR) is that we don't want timed write delays being injected in parallel for the two queues, because that dimishes the intended throttling effect. Thus, we only allow timed delays for the primary write queue. HISTORY not updated because this is intended for the same release where the bug was introduced. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11130 Test Plan: Although I was not able to reproduce the assertion failure, I was able to reproduce a distinct flaw with what I believe is the same root cause: a kind of deadlock if both write queues need to wake up from stopped writes. Only one will be waiting on bg_cv_ (the other waiting in `LinkOne()` for the write queue to open up), so a single SignalAll() will only unblock one of the queues, with the other re-instating the stop until another signal on bg_cv_. A simple unit test is added for this case. Will also run crash_test_with_multiops_wc_txn for a while looking for issues. Reviewed By: ajkr Differential Revision: D42749330 Pulled By: pdillinger fbshipit-source-id: 4317dd899a93d57c26fd5af7143038f82d4d4d1b
2 years ago
// To avoid parallel timed delays (bad throttling), only support them
// on the primary write queue.
uint64_t delay;
if (&write_thread == &write_thread_) {
delay =
write_controller_.GetDelay(immutable_db_options_.clock, num_bytes);
} else {
assert(num_bytes == 0);
delay = 0;
}
TEST_SYNC_POINT("DBImpl::DelayWrite:Start");
if (delay > 0) {
if (write_options.no_slowdown) {
return Status::Incomplete("Write stall");
}
TEST_SYNC_POINT("DBImpl::DelayWrite:Sleep");
Fix DelayWrite() calls for two_write_queues (#11130) Summary: PR https://github.com/facebook/rocksdb/issues/11020 fixed a case where it was easy to deadlock the DB with LockWAL() but introduced a bug showing up as a rare assertion failure in the stress test. Specifically, `assert(w->state == STATE_INIT)` in `WriteThread::LinkOne()` called from `BeginWriteStall()`, `DelayWrite()`, `WriteImplWALOnly()`. I haven't been about to generate a unit test that reproduces this failure but I believe the root cause is that DelayWrite() was never meant to be re-entrant, only called from the DB's write_thread_ leader. https://github.com/facebook/rocksdb/issues/11020 introduced a call to DelayWrite() from the nonmem_write_thread_ group leader. This fix is to make DelayWrite() apply to the specific write queue that it is being called from (inject a dummy write stall entry to the head of the appropriate write queue). WriteController is re-entrant, based on polling and state changes signalled with bg_cv_, so can manage stalling two queues. The only anticipated complication (called out by Andrew in previous PR) is that we don't want timed write delays being injected in parallel for the two queues, because that dimishes the intended throttling effect. Thus, we only allow timed delays for the primary write queue. HISTORY not updated because this is intended for the same release where the bug was introduced. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11130 Test Plan: Although I was not able to reproduce the assertion failure, I was able to reproduce a distinct flaw with what I believe is the same root cause: a kind of deadlock if both write queues need to wake up from stopped writes. Only one will be waiting on bg_cv_ (the other waiting in `LinkOne()` for the write queue to open up), so a single SignalAll() will only unblock one of the queues, with the other re-instating the stop until another signal on bg_cv_. A simple unit test is added for this case. Will also run crash_test_with_multiops_wc_txn for a while looking for issues. Reviewed By: ajkr Differential Revision: D42749330 Pulled By: pdillinger fbshipit-source-id: 4317dd899a93d57c26fd5af7143038f82d4d4d1b
2 years ago
// Notify write_thread about the stall so it can setup a barrier and
// fail any pending writers with no_slowdown
Fix DelayWrite() calls for two_write_queues (#11130) Summary: PR https://github.com/facebook/rocksdb/issues/11020 fixed a case where it was easy to deadlock the DB with LockWAL() but introduced a bug showing up as a rare assertion failure in the stress test. Specifically, `assert(w->state == STATE_INIT)` in `WriteThread::LinkOne()` called from `BeginWriteStall()`, `DelayWrite()`, `WriteImplWALOnly()`. I haven't been about to generate a unit test that reproduces this failure but I believe the root cause is that DelayWrite() was never meant to be re-entrant, only called from the DB's write_thread_ leader. https://github.com/facebook/rocksdb/issues/11020 introduced a call to DelayWrite() from the nonmem_write_thread_ group leader. This fix is to make DelayWrite() apply to the specific write queue that it is being called from (inject a dummy write stall entry to the head of the appropriate write queue). WriteController is re-entrant, based on polling and state changes signalled with bg_cv_, so can manage stalling two queues. The only anticipated complication (called out by Andrew in previous PR) is that we don't want timed write delays being injected in parallel for the two queues, because that dimishes the intended throttling effect. Thus, we only allow timed delays for the primary write queue. HISTORY not updated because this is intended for the same release where the bug was introduced. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11130 Test Plan: Although I was not able to reproduce the assertion failure, I was able to reproduce a distinct flaw with what I believe is the same root cause: a kind of deadlock if both write queues need to wake up from stopped writes. Only one will be waiting on bg_cv_ (the other waiting in `LinkOne()` for the write queue to open up), so a single SignalAll() will only unblock one of the queues, with the other re-instating the stop until another signal on bg_cv_. A simple unit test is added for this case. Will also run crash_test_with_multiops_wc_txn for a while looking for issues. Reviewed By: ajkr Differential Revision: D42749330 Pulled By: pdillinger fbshipit-source-id: 4317dd899a93d57c26fd5af7143038f82d4d4d1b
2 years ago
write_thread.BeginWriteStall();
mutex_.Unlock();
TEST_SYNC_POINT("DBImpl::DelayWrite:BeginWriteStallDone");
// We will delay the write until we have slept for `delay` microseconds
// or we don't need a delay anymore. We check for cancellation every 1ms
// (slightly longer because WriteController minimum delay is 1ms, in
// case of sleep imprecision, rounding, etc.)
const uint64_t kDelayInterval = 1001;
uint64_t stall_end = sw.start_time() + delay;
while (write_controller_.NeedsDelay()) {
if (immutable_db_options_.clock->NowMicros() >= stall_end) {
// We already delayed this write `delay` microseconds
break;
}
delayed = true;
// Sleep for 0.001 seconds
immutable_db_options_.clock->SleepForMicroseconds(kDelayInterval);
}
mutex_.Lock();
Fix DelayWrite() calls for two_write_queues (#11130) Summary: PR https://github.com/facebook/rocksdb/issues/11020 fixed a case where it was easy to deadlock the DB with LockWAL() but introduced a bug showing up as a rare assertion failure in the stress test. Specifically, `assert(w->state == STATE_INIT)` in `WriteThread::LinkOne()` called from `BeginWriteStall()`, `DelayWrite()`, `WriteImplWALOnly()`. I haven't been about to generate a unit test that reproduces this failure but I believe the root cause is that DelayWrite() was never meant to be re-entrant, only called from the DB's write_thread_ leader. https://github.com/facebook/rocksdb/issues/11020 introduced a call to DelayWrite() from the nonmem_write_thread_ group leader. This fix is to make DelayWrite() apply to the specific write queue that it is being called from (inject a dummy write stall entry to the head of the appropriate write queue). WriteController is re-entrant, based on polling and state changes signalled with bg_cv_, so can manage stalling two queues. The only anticipated complication (called out by Andrew in previous PR) is that we don't want timed write delays being injected in parallel for the two queues, because that dimishes the intended throttling effect. Thus, we only allow timed delays for the primary write queue. HISTORY not updated because this is intended for the same release where the bug was introduced. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11130 Test Plan: Although I was not able to reproduce the assertion failure, I was able to reproduce a distinct flaw with what I believe is the same root cause: a kind of deadlock if both write queues need to wake up from stopped writes. Only one will be waiting on bg_cv_ (the other waiting in `LinkOne()` for the write queue to open up), so a single SignalAll() will only unblock one of the queues, with the other re-instating the stop until another signal on bg_cv_. A simple unit test is added for this case. Will also run crash_test_with_multiops_wc_txn for a while looking for issues. Reviewed By: ajkr Differential Revision: D42749330 Pulled By: pdillinger fbshipit-source-id: 4317dd899a93d57c26fd5af7143038f82d4d4d1b
2 years ago
write_thread.EndWriteStall();
}
Auto recovery from out of space errors (#4164) Summary: This commit implements automatic recovery from a Status::NoSpace() error during background operations such as write callback, flush and compaction. The broad design is as follows - 1. Compaction errors are treated as soft errors and don't put the database in read-only mode. A compaction is delayed until enough free disk space is available to accomodate the compaction outputs, which is estimated based on the input size. This means that users can continue to write, and we rely on the WriteController to delay or stop writes if the compaction debt becomes too high due to persistent low disk space condition 2. Errors during write callback and flush are treated as hard errors, i.e the database is put in read-only mode and goes back to read-write only fater certain recovery actions are taken. 3. Both types of recovery rely on the SstFileManagerImpl to poll for sufficient disk space. We assume that there is a 1-1 mapping between an SFM and the underlying OS storage container. For cases where multiple DBs are hosted on a single storage container, the user is expected to allocate a single SFM instance and use the same one for all the DBs. If no SFM is specified by the user, DBImpl::Open() will allocate one, but this will be one per DB and each DB will recover independently. The recovery implemented by SFM is as follows - a) On the first occurance of an out of space error during compaction, subsequent compactions will be delayed until the disk free space check indicates enough available space. The required space is computed as the sum of input sizes. b) The free space check requirement will be removed once the amount of free space is greater than the size reserved by in progress compactions when the first error occured c) If the out of space error is a hard error, a background thread in SFM will poll for sufficient headroom before triggering the recovery of the database and putting it in write-only mode. The headroom is calculated as the sum of the write_buffer_size of all the DB instances associated with the SFM 4. EventListener callbacks will be called at the start and completion of automatic recovery. Users can disable the auto recov ery in the start callback, and later initiate it manually by calling DB::Resume() Todo: 1. More extensive testing 2. Add disk full condition to db_stress (follow-on PR) Pull Request resolved: https://github.com/facebook/rocksdb/pull/4164 Differential Revision: D9846378 Pulled By: anand1976 fbshipit-source-id: 80ea875dbd7f00205e19c82215ff6e37da10da4a
6 years ago
// Don't wait if there's a background error, even if its a soft error. We
// might wait here indefinitely as the background compaction may never
// finish successfully, resulting in the stall condition lasting
// indefinitely
while (error_handler_.GetBGError().ok() && write_controller_.IsStopped() &&
!shutting_down_.load(std::memory_order_relaxed)) {
if (write_options.no_slowdown) {
return Status::Incomplete("Write stall");
}
delayed = true;
Fix DelayWrite() calls for two_write_queues (#11130) Summary: PR https://github.com/facebook/rocksdb/issues/11020 fixed a case where it was easy to deadlock the DB with LockWAL() but introduced a bug showing up as a rare assertion failure in the stress test. Specifically, `assert(w->state == STATE_INIT)` in `WriteThread::LinkOne()` called from `BeginWriteStall()`, `DelayWrite()`, `WriteImplWALOnly()`. I haven't been about to generate a unit test that reproduces this failure but I believe the root cause is that DelayWrite() was never meant to be re-entrant, only called from the DB's write_thread_ leader. https://github.com/facebook/rocksdb/issues/11020 introduced a call to DelayWrite() from the nonmem_write_thread_ group leader. This fix is to make DelayWrite() apply to the specific write queue that it is being called from (inject a dummy write stall entry to the head of the appropriate write queue). WriteController is re-entrant, based on polling and state changes signalled with bg_cv_, so can manage stalling two queues. The only anticipated complication (called out by Andrew in previous PR) is that we don't want timed write delays being injected in parallel for the two queues, because that dimishes the intended throttling effect. Thus, we only allow timed delays for the primary write queue. HISTORY not updated because this is intended for the same release where the bug was introduced. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11130 Test Plan: Although I was not able to reproduce the assertion failure, I was able to reproduce a distinct flaw with what I believe is the same root cause: a kind of deadlock if both write queues need to wake up from stopped writes. Only one will be waiting on bg_cv_ (the other waiting in `LinkOne()` for the write queue to open up), so a single SignalAll() will only unblock one of the queues, with the other re-instating the stop until another signal on bg_cv_. A simple unit test is added for this case. Will also run crash_test_with_multiops_wc_txn for a while looking for issues. Reviewed By: ajkr Differential Revision: D42749330 Pulled By: pdillinger fbshipit-source-id: 4317dd899a93d57c26fd5af7143038f82d4d4d1b
2 years ago
// Notify write_thread about the stall so it can setup a barrier and
// fail any pending writers with no_slowdown
Fix DelayWrite() calls for two_write_queues (#11130) Summary: PR https://github.com/facebook/rocksdb/issues/11020 fixed a case where it was easy to deadlock the DB with LockWAL() but introduced a bug showing up as a rare assertion failure in the stress test. Specifically, `assert(w->state == STATE_INIT)` in `WriteThread::LinkOne()` called from `BeginWriteStall()`, `DelayWrite()`, `WriteImplWALOnly()`. I haven't been about to generate a unit test that reproduces this failure but I believe the root cause is that DelayWrite() was never meant to be re-entrant, only called from the DB's write_thread_ leader. https://github.com/facebook/rocksdb/issues/11020 introduced a call to DelayWrite() from the nonmem_write_thread_ group leader. This fix is to make DelayWrite() apply to the specific write queue that it is being called from (inject a dummy write stall entry to the head of the appropriate write queue). WriteController is re-entrant, based on polling and state changes signalled with bg_cv_, so can manage stalling two queues. The only anticipated complication (called out by Andrew in previous PR) is that we don't want timed write delays being injected in parallel for the two queues, because that dimishes the intended throttling effect. Thus, we only allow timed delays for the primary write queue. HISTORY not updated because this is intended for the same release where the bug was introduced. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11130 Test Plan: Although I was not able to reproduce the assertion failure, I was able to reproduce a distinct flaw with what I believe is the same root cause: a kind of deadlock if both write queues need to wake up from stopped writes. Only one will be waiting on bg_cv_ (the other waiting in `LinkOne()` for the write queue to open up), so a single SignalAll() will only unblock one of the queues, with the other re-instating the stop until another signal on bg_cv_. A simple unit test is added for this case. Will also run crash_test_with_multiops_wc_txn for a while looking for issues. Reviewed By: ajkr Differential Revision: D42749330 Pulled By: pdillinger fbshipit-source-id: 4317dd899a93d57c26fd5af7143038f82d4d4d1b
2 years ago
write_thread.BeginWriteStall();
if (&write_thread == &write_thread_) {
TEST_SYNC_POINT("DBImpl::DelayWrite:Wait");
} else {
TEST_SYNC_POINT("DBImpl::DelayWrite:NonmemWait");
}
bg_cv_.Wait();
TEST_SYNC_POINT_CALLBACK("DBImpl::DelayWrite:AfterWait", &mutex_);
Fix DelayWrite() calls for two_write_queues (#11130) Summary: PR https://github.com/facebook/rocksdb/issues/11020 fixed a case where it was easy to deadlock the DB with LockWAL() but introduced a bug showing up as a rare assertion failure in the stress test. Specifically, `assert(w->state == STATE_INIT)` in `WriteThread::LinkOne()` called from `BeginWriteStall()`, `DelayWrite()`, `WriteImplWALOnly()`. I haven't been about to generate a unit test that reproduces this failure but I believe the root cause is that DelayWrite() was never meant to be re-entrant, only called from the DB's write_thread_ leader. https://github.com/facebook/rocksdb/issues/11020 introduced a call to DelayWrite() from the nonmem_write_thread_ group leader. This fix is to make DelayWrite() apply to the specific write queue that it is being called from (inject a dummy write stall entry to the head of the appropriate write queue). WriteController is re-entrant, based on polling and state changes signalled with bg_cv_, so can manage stalling two queues. The only anticipated complication (called out by Andrew in previous PR) is that we don't want timed write delays being injected in parallel for the two queues, because that dimishes the intended throttling effect. Thus, we only allow timed delays for the primary write queue. HISTORY not updated because this is intended for the same release where the bug was introduced. Pull Request resolved: https://github.com/facebook/rocksdb/pull/11130 Test Plan: Although I was not able to reproduce the assertion failure, I was able to reproduce a distinct flaw with what I believe is the same root cause: a kind of deadlock if both write queues need to wake up from stopped writes. Only one will be waiting on bg_cv_ (the other waiting in `LinkOne()` for the write queue to open up), so a single SignalAll() will only unblock one of the queues, with the other re-instating the stop until another signal on bg_cv_. A simple unit test is added for this case. Will also run crash_test_with_multiops_wc_txn for a while looking for issues. Reviewed By: ajkr Differential Revision: D42749330 Pulled By: pdillinger fbshipit-source-id: 4317dd899a93d57c26fd5af7143038f82d4d4d1b
2 years ago
write_thread.EndWriteStall();
}
}
assert(!delayed || !write_options.no_slowdown);
if (delayed) {
default_cf_internal_stats_->AddDBStats(
InternalStats::kIntStatsWriteStallMicros, time_delayed);
RecordTick(stats_, STALL_MICROS, time_delayed);
}
Auto recovery from out of space errors (#4164) Summary: This commit implements automatic recovery from a Status::NoSpace() error during background operations such as write callback, flush and compaction. The broad design is as follows - 1. Compaction errors are treated as soft errors and don't put the database in read-only mode. A compaction is delayed until enough free disk space is available to accomodate the compaction outputs, which is estimated based on the input size. This means that users can continue to write, and we rely on the WriteController to delay or stop writes if the compaction debt becomes too high due to persistent low disk space condition 2. Errors during write callback and flush are treated as hard errors, i.e the database is put in read-only mode and goes back to read-write only fater certain recovery actions are taken. 3. Both types of recovery rely on the SstFileManagerImpl to poll for sufficient disk space. We assume that there is a 1-1 mapping between an SFM and the underlying OS storage container. For cases where multiple DBs are hosted on a single storage container, the user is expected to allocate a single SFM instance and use the same one for all the DBs. If no SFM is specified by the user, DBImpl::Open() will allocate one, but this will be one per DB and each DB will recover independently. The recovery implemented by SFM is as follows - a) On the first occurance of an out of space error during compaction, subsequent compactions will be delayed until the disk free space check indicates enough available space. The required space is computed as the sum of input sizes. b) The free space check requirement will be removed once the amount of free space is greater than the size reserved by in progress compactions when the first error occured c) If the out of space error is a hard error, a background thread in SFM will poll for sufficient headroom before triggering the recovery of the database and putting it in write-only mode. The headroom is calculated as the sum of the write_buffer_size of all the DB instances associated with the SFM 4. EventListener callbacks will be called at the start and completion of automatic recovery. Users can disable the auto recov ery in the start callback, and later initiate it manually by calling DB::Resume() Todo: 1. More extensive testing 2. Add disk full condition to db_stress (follow-on PR) Pull Request resolved: https://github.com/facebook/rocksdb/pull/4164 Differential Revision: D9846378 Pulled By: anand1976 fbshipit-source-id: 80ea875dbd7f00205e19c82215ff6e37da10da4a
6 years ago
// If DB is not in read-only mode and write_controller is not stopping
// writes, we can ignore any background errors and allow the write to
// proceed
Status s;
if (write_controller_.IsStopped()) {
if (!shutting_down_.load(std::memory_order_relaxed)) {
// If writes are still stopped and db not shutdown, it means we bailed
// due to a background error
s = Status::Incomplete(error_handler_.GetBGError().ToString());
} else {
s = Status::ShutdownInProgress("stalled writes");
}
Auto recovery from out of space errors (#4164) Summary: This commit implements automatic recovery from a Status::NoSpace() error during background operations such as write callback, flush and compaction. The broad design is as follows - 1. Compaction errors are treated as soft errors and don't put the database in read-only mode. A compaction is delayed until enough free disk space is available to accomodate the compaction outputs, which is estimated based on the input size. This means that users can continue to write, and we rely on the WriteController to delay or stop writes if the compaction debt becomes too high due to persistent low disk space condition 2. Errors during write callback and flush are treated as hard errors, i.e the database is put in read-only mode and goes back to read-write only fater certain recovery actions are taken. 3. Both types of recovery rely on the SstFileManagerImpl to poll for sufficient disk space. We assume that there is a 1-1 mapping between an SFM and the underlying OS storage container. For cases where multiple DBs are hosted on a single storage container, the user is expected to allocate a single SFM instance and use the same one for all the DBs. If no SFM is specified by the user, DBImpl::Open() will allocate one, but this will be one per DB and each DB will recover independently. The recovery implemented by SFM is as follows - a) On the first occurance of an out of space error during compaction, subsequent compactions will be delayed until the disk free space check indicates enough available space. The required space is computed as the sum of input sizes. b) The free space check requirement will be removed once the amount of free space is greater than the size reserved by in progress compactions when the first error occured c) If the out of space error is a hard error, a background thread in SFM will poll for sufficient headroom before triggering the recovery of the database and putting it in write-only mode. The headroom is calculated as the sum of the write_buffer_size of all the DB instances associated with the SFM 4. EventListener callbacks will be called at the start and completion of automatic recovery. Users can disable the auto recov ery in the start callback, and later initiate it manually by calling DB::Resume() Todo: 1. More extensive testing 2. Add disk full condition to db_stress (follow-on PR) Pull Request resolved: https://github.com/facebook/rocksdb/pull/4164 Differential Revision: D9846378 Pulled By: anand1976 fbshipit-source-id: 80ea875dbd7f00205e19c82215ff6e37da10da4a
6 years ago
}
if (error_handler_.IsDBStopped()) {
s = error_handler_.GetBGError();
}
return s;
}
Stall writes in WriteBufferManager when memory_usage exceeds buffer_size (#7898) Summary: When WriteBufferManager is shared across DBs and column families to maintain memory usage under a limit, OOMs have been observed when flush cannot finish but writes continuously insert to memtables. In order to avoid OOMs, when memory usage goes beyond buffer_limit_ and DBs tries to write, this change will stall incoming writers until flush is completed and memory_usage drops. Design: Stall condition: When total memory usage exceeds WriteBufferManager::buffer_size_ (memory_usage() >= buffer_size_) WriterBufferManager::ShouldStall() returns true. DBImpl first block incoming/future writers by calling write_thread_.BeginWriteStall() (which adds dummy stall object to the writer's queue). Then DB is blocked on a state State::Blocked (current write doesn't go through). WBStallInterface object maintained by every DB instance is added to the queue of WriteBufferManager. If multiple DBs tries to write during this stall, they will also be blocked when check WriteBufferManager::ShouldStall() returns true. End Stall condition: When flush is finished and memory usage goes down, stall will end only if memory waiting to be flushed is less than buffer_size/2. This lower limit will give time for flush to complete and avoid continous stalling if memory usage remains close to buffer_size. WriterBufferManager::EndWriteStall() is called, which removes all instances from its queue and signal them to continue. Their state is changed to State::Running and they are unblocked. DBImpl then signal all incoming writers of that DB to continue by calling write_thread_.EndWriteStall() (which removes dummy stall object from the queue). DB instance creates WBMStallInterface which is an interface to block and signal DBs during stall. When DB needs to be blocked or signalled by WriteBufferManager, state_for_wbm_ state is changed accordingly (RUNNING or BLOCKED). Pull Request resolved: https://github.com/facebook/rocksdb/pull/7898 Test Plan: Added a new test db/db_write_buffer_manager_test.cc Reviewed By: anand1976 Differential Revision: D26093227 Pulled By: akankshamahajan15 fbshipit-source-id: 2bbd982a3fb7033f6de6153aa92a221249861aae
4 years ago
// REQUIRES: mutex_ is held
// REQUIRES: this thread is currently at the front of the writer queue
void DBImpl::WriteBufferManagerStallWrites() {
mutex_.AssertHeld();
// First block future writer threads who want to add themselves to the queue
// of WriteThread.
write_thread_.BeginWriteStall();
mutex_.Unlock();
// Change the state to State::Blocked.
static_cast<WBMStallInterface*>(wbm_stall_.get())
->SetState(WBMStallInterface::State::BLOCKED);
// Then WriteBufferManager will add DB instance to its queue
// and block this thread by calling WBMStallInterface::Block().
write_buffer_manager_->BeginWriteStall(wbm_stall_.get());
wbm_stall_->Block();
mutex_.Lock();
// Stall has ended. Signal writer threads so that they can add
// themselves to the WriteThread queue for writes.
write_thread_.EndWriteStall();
}
Status DBImpl::ThrottleLowPriWritesIfNeeded(const WriteOptions& write_options,
WriteBatch* my_batch) {
assert(write_options.low_pri);
// This is called outside the DB mutex. Although it is safe to make the call,
// the consistency condition is not guaranteed to hold. It's OK to live with
// it in this case.
// If we need to speed compaction, it means the compaction is left behind
// and we start to limit low pri writes to a limit.
if (write_controller_.NeedSpeedupCompaction()) {
if (allow_2pc() && (my_batch->HasCommit() || my_batch->HasRollback())) {
// For 2PC, we only rate limit prepare, not commit.
return Status::OK();
}
if (write_options.no_slowdown) {
return Status::Incomplete("Low priority write stall");
} else {
assert(my_batch != nullptr);
// Rate limit those writes. The reason that we don't completely wait
// is that in case the write is heavy, low pri writes may never have
// a chance to run. Now we guarantee we are still slowly making
// progress.
PERF_TIMER_GUARD(write_delay_time);
write_controller_.low_pri_rate_limiter()->Request(
my_batch->GetDataSize(), Env::IO_HIGH, nullptr /* stats */,
RateLimiter::OpType::kWrite);
}
}
return Status::OK();
}
void DBImpl::MaybeFlushStatsCF(autovector<ColumnFamilyData*>* cfds) {
assert(cfds != nullptr);
if (!cfds->empty() && immutable_db_options_.persist_stats_to_disk) {
ColumnFamilyData* cfd_stats =
versions_->GetColumnFamilySet()->GetColumnFamily(
kPersistentStatsColumnFamilyName);
if (cfd_stats != nullptr && !cfd_stats->mem()->IsEmpty()) {
for (ColumnFamilyData* cfd : *cfds) {
if (cfd == cfd_stats) {
// stats CF already included in cfds
return;
}
}
// force flush stats CF when its log number is less than all other CF's
// log numbers
bool force_flush_stats_cf = true;
for (auto* loop_cfd : *versions_->GetColumnFamilySet()) {
if (loop_cfd == cfd_stats) {
continue;
}
if (loop_cfd->GetLogNumber() <= cfd_stats->GetLogNumber()) {
force_flush_stats_cf = false;
}
}
if (force_flush_stats_cf) {
cfds->push_back(cfd_stats);
ROCKS_LOG_INFO(immutable_db_options_.info_log,
"Force flushing stats CF with automated flush "
"to avoid holding old logs");
}
}
}
}
Refactor trimming logic for immutable memtables (#5022) Summary: MyRocks currently sets `max_write_buffer_number_to_maintain` in order to maintain enough history for transaction conflict checking. The effectiveness of this approach depends on the size of memtables. When memtables are small, it may not keep enough history; when memtables are large, this may consume too much memory. We are proposing a new way to configure memtable list history: by limiting the memory usage of immutable memtables. The new option is `max_write_buffer_size_to_maintain` and it will take precedence over the old `max_write_buffer_number_to_maintain` if they are both set to non-zero values. The new option accounts for the total memory usage of flushed immutable memtables and mutable memtable. When the total usage exceeds the limit, RocksDB may start dropping immutable memtables (which is also called trimming history), starting from the oldest one. The semantics of the old option actually works both as an upper bound and lower bound. History trimming will start if number of immutable memtables exceeds the limit, but it will never go below (limit-1) due to history trimming. In order the mimic the behavior with the new option, history trimming will stop if dropping the next immutable memtable causes the total memory usage go below the size limit. For example, assuming the size limit is set to 64MB, and there are 3 immutable memtables with sizes of 20, 30, 30. Although the total memory usage is 80MB > 64MB, dropping the oldest memtable will reduce the memory usage to 60MB < 64MB, so in this case no memtable will be dropped. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5022 Differential Revision: D14394062 Pulled By: miasantreble fbshipit-source-id: 60457a509c6af89d0993f988c9b5c2aa9e45f5c5
5 years ago
Status DBImpl::TrimMemtableHistory(WriteContext* context) {
autovector<ColumnFamilyData*> cfds;
ColumnFamilyData* tmp_cfd;
while ((tmp_cfd = trim_history_scheduler_.TakeNextColumnFamily()) !=
nullptr) {
cfds.push_back(tmp_cfd);
}
for (auto& cfd : cfds) {
autovector<MemTable*> to_delete;
bool trimmed = cfd->imm()->TrimHistory(&context->memtables_to_free_,
cfd->mem()->MemoryAllocatedBytes());
if (trimmed) {
context->superversion_context.NewSuperVersion();
assert(context->superversion_context.new_superversion.get() != nullptr);
cfd->InstallSuperVersion(&context->superversion_context, &mutex_);
Refactor trimming logic for immutable memtables (#5022) Summary: MyRocks currently sets `max_write_buffer_number_to_maintain` in order to maintain enough history for transaction conflict checking. The effectiveness of this approach depends on the size of memtables. When memtables are small, it may not keep enough history; when memtables are large, this may consume too much memory. We are proposing a new way to configure memtable list history: by limiting the memory usage of immutable memtables. The new option is `max_write_buffer_size_to_maintain` and it will take precedence over the old `max_write_buffer_number_to_maintain` if they are both set to non-zero values. The new option accounts for the total memory usage of flushed immutable memtables and mutable memtable. When the total usage exceeds the limit, RocksDB may start dropping immutable memtables (which is also called trimming history), starting from the oldest one. The semantics of the old option actually works both as an upper bound and lower bound. History trimming will start if number of immutable memtables exceeds the limit, but it will never go below (limit-1) due to history trimming. In order the mimic the behavior with the new option, history trimming will stop if dropping the next immutable memtable causes the total memory usage go below the size limit. For example, assuming the size limit is set to 64MB, and there are 3 immutable memtables with sizes of 20, 30, 30. Although the total memory usage is 80MB > 64MB, dropping the oldest memtable will reduce the memory usage to 60MB < 64MB, so in this case no memtable will be dropped. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5022 Differential Revision: D14394062 Pulled By: miasantreble fbshipit-source-id: 60457a509c6af89d0993f988c9b5c2aa9e45f5c5
5 years ago
}
if (cfd->UnrefAndTryDelete()) {
Refactor trimming logic for immutable memtables (#5022) Summary: MyRocks currently sets `max_write_buffer_number_to_maintain` in order to maintain enough history for transaction conflict checking. The effectiveness of this approach depends on the size of memtables. When memtables are small, it may not keep enough history; when memtables are large, this may consume too much memory. We are proposing a new way to configure memtable list history: by limiting the memory usage of immutable memtables. The new option is `max_write_buffer_size_to_maintain` and it will take precedence over the old `max_write_buffer_number_to_maintain` if they are both set to non-zero values. The new option accounts for the total memory usage of flushed immutable memtables and mutable memtable. When the total usage exceeds the limit, RocksDB may start dropping immutable memtables (which is also called trimming history), starting from the oldest one. The semantics of the old option actually works both as an upper bound and lower bound. History trimming will start if number of immutable memtables exceeds the limit, but it will never go below (limit-1) due to history trimming. In order the mimic the behavior with the new option, history trimming will stop if dropping the next immutable memtable causes the total memory usage go below the size limit. For example, assuming the size limit is set to 64MB, and there are 3 immutable memtables with sizes of 20, 30, 30. Although the total memory usage is 80MB > 64MB, dropping the oldest memtable will reduce the memory usage to 60MB < 64MB, so in this case no memtable will be dropped. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5022 Differential Revision: D14394062 Pulled By: miasantreble fbshipit-source-id: 60457a509c6af89d0993f988c9b5c2aa9e45f5c5
5 years ago
cfd = nullptr;
}
}
return Status::OK();
}
Status DBImpl::ScheduleFlushes(WriteContext* context) {
autovector<ColumnFamilyData*> cfds;
if (immutable_db_options_.atomic_flush) {
SelectColumnFamiliesForAtomicFlush(&cfds);
for (auto cfd : cfds) {
cfd->Ref();
}
flush_scheduler_.Clear();
} else {
ColumnFamilyData* tmp_cfd;
while ((tmp_cfd = flush_scheduler_.TakeNextColumnFamily()) != nullptr) {
cfds.push_back(tmp_cfd);
}
MaybeFlushStatsCF(&cfds);
}
Status status;
WriteThread::Writer nonmem_w;
if (two_write_queues_) {
nonmem_write_thread_.EnterUnbatched(&nonmem_w, &mutex_);
}
for (auto& cfd : cfds) {
if (!cfd->mem()->IsEmpty()) {
status = SwitchMemtable(cfd, context);
}
if (cfd->UnrefAndTryDelete()) {
cfd = nullptr;
}
if (!status.ok()) {
break;
}
}
if (two_write_queues_) {
nonmem_write_thread_.ExitUnbatched(&nonmem_w);
}
if (status.ok()) {
if (immutable_db_options_.atomic_flush) {
AssignAtomicFlushSeq(cfds);
FlushRequest flush_req;
GenerateFlushRequest(cfds, FlushReason::kWriteBufferFull, &flush_req);
SchedulePendingFlush(flush_req);
} else {
for (auto* cfd : cfds) {
FlushRequest flush_req;
GenerateFlushRequest({cfd}, FlushReason::kWriteBufferFull, &flush_req);
SchedulePendingFlush(flush_req);
}
}
MaybeScheduleFlushOrCompaction();
}
return status;
}
void DBImpl::NotifyOnMemTableSealed(ColumnFamilyData* /*cfd*/,
const MemTableInfo& mem_table_info) {
if (immutable_db_options_.listeners.size() == 0U) {
return;
}
if (shutting_down_.load(std::memory_order_acquire)) {
return;
}
mutex_.Unlock();
for (auto listener : immutable_db_options_.listeners) {
listener->OnMemTableSealed(mem_table_info);
}
mutex_.Lock();
}
// REQUIRES: mutex_ is held
// REQUIRES: this thread is currently at the front of the writer queue
// REQUIRES: this thread is currently at the front of the 2nd writer queue if
// two_write_queues_ is true (This is to simplify the reasoning.)
Status DBImpl::SwitchMemtable(ColumnFamilyData* cfd, WriteContext* context) {
mutex_.AssertHeld();
Group rocksdb.sst.read.micros stat by IOActivity flush and compaction (#11288) Summary: **Context:** The existing stat rocksdb.sst.read.micros does not reflect each of compaction and flush cases but aggregate them, which is not so helpful for us to understand IO read behavior of each of them. **Summary** - Update `StopWatch` and `RandomAccessFileReader` to record `rocksdb.sst.read.micros` and `rocksdb.file.{flush/compaction}.read.micros` - Fixed the default histogram in `RandomAccessFileReader` - New field `ReadOptions/IOOptions::io_activity`; Pass `ReadOptions` through paths under db open, flush and compaction to where we can prepare `IOOptions` and pass it to `RandomAccessFileReader` - Use `thread_status_util` for assertion in `DbStressFSWrapper` for continuous testing on we are passing correct `io_activity` under db open, flush and compaction Pull Request resolved: https://github.com/facebook/rocksdb/pull/11288 Test Plan: - **Stress test** - **Db bench 1: rocksdb.sst.read.micros COUNT ≈ sum of rocksdb.file.read.flush.micros's and rocksdb.file.read.compaction.micros's.** (without blob) - May not be exactly the same due to `HistogramStat::Add` only guarantees atomic not accuracy across threads. ``` ./db_bench -db=/dev/shm/testdb/ -statistics=true -benchmarks="fillseq" -key_size=32 -value_size=512 -num=50000 -write_buffer_size=655 -target_file_size_base=655 -disable_auto_compactions=false -compression_type=none -bloom_bits=3 (-use_plain_table=1 -prefix_size=10) ``` ``` // BlockBasedTable rocksdb.sst.read.micros P50 : 2.009374 P95 : 4.968548 P99 : 8.110362 P100 : 43.000000 COUNT : 40456 SUM : 114805 rocksdb.file.read.flush.micros P50 : 1.871841 P95 : 3.872407 P99 : 5.540541 P100 : 43.000000 COUNT : 2250 SUM : 6116 rocksdb.file.read.compaction.micros P50 : 2.023109 P95 : 5.029149 P99 : 8.196910 P100 : 26.000000 COUNT : 38206 SUM : 108689 // PlainTable Does not apply ``` - **Db bench 2: performance** **Read** SETUP: db with 900 files ``` ./db_bench -db=/dev/shm/testdb/ -benchmarks="fillseq" -key_size=32 -value_size=512 -num=50000 -write_buffer_size=655 -disable_auto_compactions=true -target_file_size_base=655 -compression_type=none ```run till convergence ``` ./db_bench -seed=1678564177044286 -use_existing_db=true -db=/dev/shm/testdb -benchmarks=readrandom[-X60] -statistics=true -num=1000000 -disable_auto_compactions=true -compression_type=none -bloom_bits=3 ``` Pre-change `readrandom [AVG 60 runs] : 21568 (± 248) ops/sec` Post-change (no regression, -0.3%) `readrandom [AVG 60 runs] : 21486 (± 236) ops/sec` **Compaction/Flush**run till convergence ``` ./db_bench -db=/dev/shm/testdb2/ -seed=1678564177044286 -benchmarks="fillseq[-X60]" -key_size=32 -value_size=512 -num=50000 -write_buffer_size=655 -disable_auto_compactions=false -target_file_size_base=655 -compression_type=none rocksdb.sst.read.micros COUNT : 33820 rocksdb.sst.read.flush.micros COUNT : 1800 rocksdb.sst.read.compaction.micros COUNT : 32020 ``` Pre-change `fillseq [AVG 46 runs] : 1391 (± 214) ops/sec; 0.7 (± 0.1) MB/sec` Post-change (no regression, ~-0.4%) `fillseq [AVG 46 runs] : 1385 (± 216) ops/sec; 0.7 (± 0.1) MB/sec` Reviewed By: ajkr Differential Revision: D44007011 Pulled By: hx235 fbshipit-source-id: a54c89e4846dfc9a135389edf3f3eedfea257132
2 years ago
// TODO: plumb Env::IOActivity
const ReadOptions read_options;
log::Writer* new_log = nullptr;
MemTable* new_mem = nullptr;
IOStatus io_s;
// Recoverable state is persisted in WAL. After memtable switch, WAL might
// be deleted, so we write the state to memtable to be persisted as well.
Status s = WriteRecoverableState();
if (!s.ok()) {
return s;
}
// Attempt to switch to a new memtable and trigger flush of old.
// Do this without holding the dbmutex lock.
assert(versions_->prev_log_number() == 0);
if (two_write_queues_) {
log_write_mutex_.Lock();
}
bool creating_new_log = !log_empty_;
if (two_write_queues_) {
log_write_mutex_.Unlock();
}
uint64_t recycle_log_number = 0;
if (creating_new_log && immutable_db_options_.recycle_log_file_num &&
!log_recycle_files_.empty()) {
recycle_log_number = log_recycle_files_.front();
}
uint64_t new_log_number =
creating_new_log ? versions_->NewFileNumber() : logfile_number_;
const MutableCFOptions mutable_cf_options = *cfd->GetLatestMutableCFOptions();
// Set memtable_info for memtable sealed callback
MemTableInfo memtable_info;
memtable_info.cf_name = cfd->GetName();
memtable_info.first_seqno = cfd->mem()->GetFirstSequenceNumber();
memtable_info.earliest_seqno = cfd->mem()->GetEarliestSequenceNumber();
memtable_info.num_entries = cfd->mem()->num_entries();
memtable_info.num_deletes = cfd->mem()->num_deletes();
// Log this later after lock release. It may be outdated, e.g., if background
// flush happens before logging, but that should be ok.
int num_imm_unflushed = cfd->imm()->NumNotFlushed();
const auto preallocate_block_size =
GetWalPreallocateBlockSize(mutable_cf_options.write_buffer_size);
mutex_.Unlock();
if (creating_new_log) {
// TODO: Write buffer size passed in should be max of all CF's instead
// of mutable_cf_options.write_buffer_size.
io_s = CreateWAL(new_log_number, recycle_log_number, preallocate_block_size,
&new_log);
if (s.ok()) {
s = io_s;
}
}
if (s.ok()) {
SequenceNumber seq = versions_->LastSequence();
new_mem = cfd->ConstructNewMemtable(mutable_cf_options, seq);
context->superversion_context.NewSuperVersion();
}
ROCKS_LOG_INFO(immutable_db_options_.info_log,
"[%s] New memtable created with log file: #%" PRIu64
". Immutable memtables: %d.\n",
cfd->GetName().c_str(), new_log_number, num_imm_unflushed);
Fragment memtable range tombstone in the write path (#10380) Summary: - Right now each read fragments the memtable range tombstones https://github.com/facebook/rocksdb/issues/4808. This PR explores the idea of fragmenting memtable range tombstones in the write path and reads can just read this cached fragmented tombstone without any fragmenting cost. This PR only does the caching for immutable memtable, and does so right before a memtable is added to an immutable memtable list. The fragmentation is done without holding mutex to minimize its performance impact. - db_bench is updated to print out the number of range deletions executed if there is any. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10380 Test Plan: - CI, added asserts in various places to check whether a fragmented range tombstone list should have been constructed. - Benchmark: as this PR only optimizes immutable memtable path, the number of writes in the benchmark is chosen such an immutable memtable is created and range tombstones are in that memtable. ``` single thread: ./db_bench --benchmarks=fillrandom,readrandom --writes_per_range_tombstone=1 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=500000 --reads=100000 --max_num_range_tombstones=100 multi_thread ./db_bench --benchmarks=fillrandom,readrandom --writes_per_range_tombstone=1 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=15000 --reads=20000 --threads=32 --max_num_range_tombstones=100 ``` Commit 99cdf16464a057ca44de2f747541dedf651bae9e is included in benchmark result. It was an earlier attempt where tombstones are fragmented for each write operation. Reader threads share it using a shared_ptr which would slow down multi-thread read performance as seen in benchmark results. Results are averaged over 5 runs. Single thread result: | Max # tombstones | main fillrandom micros/op | 99cdf16464a057ca44de2f747541dedf651bae9e | Post PR | main readrandom micros/op | 99cdf16464a057ca44de2f747541dedf651bae9e | Post PR | | ------------- | ------------- |------------- |------------- |------------- |------------- |------------- | | 0 |6.68 |6.57 |6.72 |4.72 |4.79 |4.54 | | 1 |6.67 |6.58 |6.62 |5.41 |4.74 |4.72 | | 10 |6.59 |6.5 |6.56 |7.83 |4.69 |4.59 | | 100 |6.62 |6.75 |6.58 |29.57 |5.04 |5.09 | | 1000 |6.54 |6.82 |6.61 |320.33 |5.22 |5.21 | 32-thread result: note that "Max # tombstones" is per thread. | Max # tombstones | main fillrandom micros/op | 99cdf16464a057ca44de2f747541dedf651bae9e | Post PR | main readrandom micros/op | 99cdf16464a057ca44de2f747541dedf651bae9e | Post PR | | ------------- | ------------- |------------- |------------- |------------- |------------- |------------- | | 0 |234.52 |260.25 |239.42 |5.06 |5.38 |5.09 | | 1 |236.46 |262.0 |231.1 |19.57 |22.14 |5.45 | | 10 |236.95 |263.84 |251.49 |151.73 |21.61 |5.73 | | 100 |268.16 |296.8 |280.13 |2308.52 |22.27 |6.57 | Reviewed By: ajkr Differential Revision: D37916564 Pulled By: cbi42 fbshipit-source-id: 05d6d2e16df26c374c57ddcca13a5bfe9d5b731e
2 years ago
// There should be no concurrent write as the thread is at the front of
// writer queue
cfd->mem()->ConstructFragmentedRangeTombstones();
mutex_.Lock();
if (recycle_log_number != 0) {
// Since renaming the file is done outside DB mutex, we need to ensure
// concurrent full purges don't delete the file while we're recycling it.
// To achieve that we hold the old log number in the recyclable list until
// after it has been renamed.
assert(log_recycle_files_.front() == recycle_log_number);
log_recycle_files_.pop_front();
}
if (s.ok() && creating_new_log) {
InstrumentedMutexLock l(&log_write_mutex_);
assert(new_log != nullptr);
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
if (!logs_.empty()) {
// Alway flush the buffer of the last log before switching to a new one
log::Writer* cur_log_writer = logs_.back().writer;
if (error_handler_.IsRecoveryInProgress()) {
// In recovery path, we force another try of writing WAL buffer.
cur_log_writer->file()->reset_seen_error();
}
io_s = cur_log_writer->WriteBuffer();
if (s.ok()) {
s = io_s;
}
if (!s.ok()) {
ROCKS_LOG_WARN(immutable_db_options_.info_log,
"[%s] Failed to switch from #%" PRIu64 " to #%" PRIu64
" WAL file\n",
cfd->GetName().c_str(), cur_log_writer->get_log_number(),
new_log_number);
}
Optimize for serial commits in 2PC Summary: Throughput: 46k tps in our sysbench settings (filling the details later) The idea is to have the simplest change that gives us a reasonable boost in 2PC throughput. Major design changes: 1. The WAL file internal buffer is not flushed after each write. Instead it is flushed before critical operations (WAL copy via fs) or when FlushWAL is called by MySQL. Flushing the WAL buffer is also protected via mutex_. 2. Use two sequence numbers: last seq, and last seq for write. Last seq is the last visible sequence number for reads. Last seq for write is the next sequence number that should be used to write to WAL/memtable. This allows to have a memtable write be in parallel to WAL writes. 3. BatchGroup is not used for writes. This means that we can have parallel writers which changes a major assumption in the code base. To accommodate for that i) allow only 1 WriteImpl that intends to write to memtable via mem_mutex_--which is fine since in 2PC almost all of the memtable writes come via group commit phase which is serial anyway, ii) make all the parts in the code base that assumed to be the only writer (via EnterUnbatched) to also acquire mem_mutex_, iii) stat updates are protected via a stat_mutex_. Note: the first commit has the approach figured out but is not clean. Submitting the PR anyway to get the early feedback on the approach. If we are ok with the approach I will go ahead with this updates: 0) Rebase with Yi's pipelining changes 1) Currently batching is disabled by default to make sure that it will be consistent with all unit tests. Will make this optional via a config. 2) A couple of unit tests are disabled. They need to be updated with the serial commit of 2PC taken into account. 3) Replacing BatchGroup with mem_mutex_ got a bit ugly as it requires releasing mutex_ beforehand (the same way EnterUnbatched does). This needs to be cleaned up. Closes https://github.com/facebook/rocksdb/pull/2345 Differential Revision: D5210732 Pulled By: maysamyabandeh fbshipit-source-id: 78653bd95a35cd1e831e555e0e57bdfd695355a4
8 years ago
}
if (s.ok()) {
logfile_number_ = new_log_number;
log_empty_ = true;
log_dir_synced_ = false;
logs_.emplace_back(logfile_number_, new_log);
alive_log_files_.push_back(LogFileNumberSize(logfile_number_));
}
}
if (!s.ok()) {
// how do we fail if we're not creating new log?
assert(creating_new_log);
delete new_mem;
delete new_log;
context->superversion_context.new_superversion.reset();
// We may have lost data from the WritableFileBuffer in-memory buffer for
// the current log, so treat it as a fatal error and set bg_error
if (!io_s.ok()) {
error_handler_.SetBGError(io_s, BackgroundErrorReason::kMemTable);
} else {
error_handler_.SetBGError(s, BackgroundErrorReason::kMemTable);
}
// Read back bg_error in order to get the right severity
s = error_handler_.GetBGError();
return s;
}
Track WAL obsoletion when updating empty CF's log number (#7781) Summary: In the write path, there is an optimization: when a new WAL is created during SwitchMemtable, we update the internal log number of the empty column families to the new WAL. `FindObsoleteFiles` marks a WAL as obsolete if the WAL's log number is less than `VersionSet::MinLogNumberWithUnflushedData`. After updating the empty column families' internal log number, `VersionSet::MinLogNumberWithUnflushedData` might change, so some WALs might become obsolete to be purged from disk. For example, consider there are 3 column families: 0, 1, 2: 1. initially, all the column families' log number is 1; 2. write some data to cf0, and flush cf0, but the flush is pending; 3. now a new WAL 2 is created; 4. write data to cf1 and WAL 2, now cf0's log number is 1, cf1's log number is 2, cf2's log number is 2 (because cf1 and cf2 are empty, so their log numbers will be set to the highest log number); 5. now cf0's flush hasn't finished, flush cf1, a new WAL 3 is created, and cf1's flush finishes, now cf0's log number is 1, cf1's log number is 3, cf2's log number is 3, since WAL 1 still contains data for the unflushed cf0, no WAL can be deleted from disk; 6. now cf0's flush finishes, cf0's log number is 2 (because when cf0 was switching memtable, WAL 3 does not exist yet), cf1's log number is 3, cf2's log number is 3, so WAL 1 can be purged from disk now, but WAL 2 still cannot because `MinLogNumberToKeep()` is 2; 7. write data to cf2 and WAL 3, because cf0 is empty, its log number is updated to 3, so now cf0's log number is 3, cf1's log number is 3, cf2's log number is 3; 8. now if the background threads want to purge obsolete files from disk, WAL 2 can be purged because `MinLogNumberToKeep()` is 3. But there are only two flush results written to MANIFEST: the first is for flushing cf1, and the `MinLogNumberToKeep` is 1, the second is for flushing cf0, and the `MinLogNumberToKeep` is 2. So without this PR, if the DB crashes at this point and try to recover, `WalSet` will still expect WAL 2 to exist. When WAL tracking is enabled, we assume WALs will only become obsolete after a flush result is written to MANIFEST in `MemtableList::TryInstallMemtableFlushResults` (or its atomic flush counterpart). The above situation breaks this assumption. This PR tracks WAL obsoletion if necessary before updating the empty column families' log numbers. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7781 Test Plan: watch existing tests and stress tests to pass. `make -j48 blackbox_crash_test` on devserver Reviewed By: ltamasi Differential Revision: D25631695 Pulled By: cheng-chang fbshipit-source-id: ca7fff967bdb42204b84226063d909893bc0a4ec
4 years ago
bool empty_cf_updated = false;
if (immutable_db_options_.track_and_verify_wals_in_manifest &&
!immutable_db_options_.allow_2pc && creating_new_log) {
// In non-2pc mode, WALs become obsolete if they do not contain unflushed
// data. Updating the empty CF's log number might cause some WALs to become
// obsolete. So we should track the WAL obsoletion event before actually
// updating the empty CF's log number.
uint64_t min_wal_number_to_keep =
versions_->PreComputeMinLogNumberWithUnflushedData(logfile_number_);
if (min_wal_number_to_keep >
versions_->GetWalSet().GetMinWalNumberToKeep()) {
// Get a snapshot of the empty column families.
// LogAndApply may release and reacquire db
// mutex, during that period, column family may become empty (e.g. its
// flush succeeds), then it affects the computed min_log_number_to_keep,
// so we take a snapshot for consistency of column family data
// status. If a column family becomes non-empty afterwards, its active log
// should still be the created new log, so the min_log_number_to_keep is
// not affected.
autovector<ColumnFamilyData*> empty_cfs;
for (auto cf : *versions_->GetColumnFamilySet()) {
if (cf->IsEmpty()) {
empty_cfs.push_back(cf);
}
}
VersionEdit wal_deletion;
wal_deletion.DeleteWalsBefore(min_wal_number_to_keep);
Group rocksdb.sst.read.micros stat by IOActivity flush and compaction (#11288) Summary: **Context:** The existing stat rocksdb.sst.read.micros does not reflect each of compaction and flush cases but aggregate them, which is not so helpful for us to understand IO read behavior of each of them. **Summary** - Update `StopWatch` and `RandomAccessFileReader` to record `rocksdb.sst.read.micros` and `rocksdb.file.{flush/compaction}.read.micros` - Fixed the default histogram in `RandomAccessFileReader` - New field `ReadOptions/IOOptions::io_activity`; Pass `ReadOptions` through paths under db open, flush and compaction to where we can prepare `IOOptions` and pass it to `RandomAccessFileReader` - Use `thread_status_util` for assertion in `DbStressFSWrapper` for continuous testing on we are passing correct `io_activity` under db open, flush and compaction Pull Request resolved: https://github.com/facebook/rocksdb/pull/11288 Test Plan: - **Stress test** - **Db bench 1: rocksdb.sst.read.micros COUNT ≈ sum of rocksdb.file.read.flush.micros's and rocksdb.file.read.compaction.micros's.** (without blob) - May not be exactly the same due to `HistogramStat::Add` only guarantees atomic not accuracy across threads. ``` ./db_bench -db=/dev/shm/testdb/ -statistics=true -benchmarks="fillseq" -key_size=32 -value_size=512 -num=50000 -write_buffer_size=655 -target_file_size_base=655 -disable_auto_compactions=false -compression_type=none -bloom_bits=3 (-use_plain_table=1 -prefix_size=10) ``` ``` // BlockBasedTable rocksdb.sst.read.micros P50 : 2.009374 P95 : 4.968548 P99 : 8.110362 P100 : 43.000000 COUNT : 40456 SUM : 114805 rocksdb.file.read.flush.micros P50 : 1.871841 P95 : 3.872407 P99 : 5.540541 P100 : 43.000000 COUNT : 2250 SUM : 6116 rocksdb.file.read.compaction.micros P50 : 2.023109 P95 : 5.029149 P99 : 8.196910 P100 : 26.000000 COUNT : 38206 SUM : 108689 // PlainTable Does not apply ``` - **Db bench 2: performance** **Read** SETUP: db with 900 files ``` ./db_bench -db=/dev/shm/testdb/ -benchmarks="fillseq" -key_size=32 -value_size=512 -num=50000 -write_buffer_size=655 -disable_auto_compactions=true -target_file_size_base=655 -compression_type=none ```run till convergence ``` ./db_bench -seed=1678564177044286 -use_existing_db=true -db=/dev/shm/testdb -benchmarks=readrandom[-X60] -statistics=true -num=1000000 -disable_auto_compactions=true -compression_type=none -bloom_bits=3 ``` Pre-change `readrandom [AVG 60 runs] : 21568 (± 248) ops/sec` Post-change (no regression, -0.3%) `readrandom [AVG 60 runs] : 21486 (± 236) ops/sec` **Compaction/Flush**run till convergence ``` ./db_bench -db=/dev/shm/testdb2/ -seed=1678564177044286 -benchmarks="fillseq[-X60]" -key_size=32 -value_size=512 -num=50000 -write_buffer_size=655 -disable_auto_compactions=false -target_file_size_base=655 -compression_type=none rocksdb.sst.read.micros COUNT : 33820 rocksdb.sst.read.flush.micros COUNT : 1800 rocksdb.sst.read.compaction.micros COUNT : 32020 ``` Pre-change `fillseq [AVG 46 runs] : 1391 (± 214) ops/sec; 0.7 (± 0.1) MB/sec` Post-change (no regression, ~-0.4%) `fillseq [AVG 46 runs] : 1385 (± 216) ops/sec; 0.7 (± 0.1) MB/sec` Reviewed By: ajkr Differential Revision: D44007011 Pulled By: hx235 fbshipit-source-id: a54c89e4846dfc9a135389edf3f3eedfea257132
2 years ago
s = versions_->LogAndApplyToDefaultColumnFamily(
read_options, &wal_deletion, &mutex_, directories_.GetDbDir());
Track WAL obsoletion when updating empty CF's log number (#7781) Summary: In the write path, there is an optimization: when a new WAL is created during SwitchMemtable, we update the internal log number of the empty column families to the new WAL. `FindObsoleteFiles` marks a WAL as obsolete if the WAL's log number is less than `VersionSet::MinLogNumberWithUnflushedData`. After updating the empty column families' internal log number, `VersionSet::MinLogNumberWithUnflushedData` might change, so some WALs might become obsolete to be purged from disk. For example, consider there are 3 column families: 0, 1, 2: 1. initially, all the column families' log number is 1; 2. write some data to cf0, and flush cf0, but the flush is pending; 3. now a new WAL 2 is created; 4. write data to cf1 and WAL 2, now cf0's log number is 1, cf1's log number is 2, cf2's log number is 2 (because cf1 and cf2 are empty, so their log numbers will be set to the highest log number); 5. now cf0's flush hasn't finished, flush cf1, a new WAL 3 is created, and cf1's flush finishes, now cf0's log number is 1, cf1's log number is 3, cf2's log number is 3, since WAL 1 still contains data for the unflushed cf0, no WAL can be deleted from disk; 6. now cf0's flush finishes, cf0's log number is 2 (because when cf0 was switching memtable, WAL 3 does not exist yet), cf1's log number is 3, cf2's log number is 3, so WAL 1 can be purged from disk now, but WAL 2 still cannot because `MinLogNumberToKeep()` is 2; 7. write data to cf2 and WAL 3, because cf0 is empty, its log number is updated to 3, so now cf0's log number is 3, cf1's log number is 3, cf2's log number is 3; 8. now if the background threads want to purge obsolete files from disk, WAL 2 can be purged because `MinLogNumberToKeep()` is 3. But there are only two flush results written to MANIFEST: the first is for flushing cf1, and the `MinLogNumberToKeep` is 1, the second is for flushing cf0, and the `MinLogNumberToKeep` is 2. So without this PR, if the DB crashes at this point and try to recover, `WalSet` will still expect WAL 2 to exist. When WAL tracking is enabled, we assume WALs will only become obsolete after a flush result is written to MANIFEST in `MemtableList::TryInstallMemtableFlushResults` (or its atomic flush counterpart). The above situation breaks this assumption. This PR tracks WAL obsoletion if necessary before updating the empty column families' log numbers. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7781 Test Plan: watch existing tests and stress tests to pass. `make -j48 blackbox_crash_test` on devserver Reviewed By: ltamasi Differential Revision: D25631695 Pulled By: cheng-chang fbshipit-source-id: ca7fff967bdb42204b84226063d909893bc0a4ec
4 years ago
if (!s.ok() && versions_->io_status().IsIOError()) {
s = error_handler_.SetBGError(versions_->io_status(),
BackgroundErrorReason::kManifestWrite);
}
if (!s.ok()) {
return s;
}
for (auto cf : empty_cfs) {
if (cf->IsEmpty()) {
cf->SetLogNumber(logfile_number_);
Memtable "MemPurge" prototype (#8454) Summary: Implement an experimental feature called "MemPurge", which consists in purging "garbage" bytes out of a memtable and reuse the memtable struct instead of making it immutable and eventually flushing its content to storage. The prototype is by default deactivated and is not intended for use. It is intended for correctness and validation testing. At the moment, the "MemPurge" feature can be switched on by using the `options.experimental_allow_mempurge` flag. For this early stage, when the allow_mempurge flag is set to `true`, all the flush operations will be rerouted to perform a MemPurge. This is a temporary design decision that will give us the time to explore meaningful heuristics to use MemPurge at the right time for relevant workloads . Moreover, the current MemPurge operation only supports `Puts`, `Deletes`, `DeleteRange` operations, and handles `Iterators` as well as `CompactionFilter`s that are invoked at flush time . Three unit tests are added to `db_flush_test.cc` to test if MemPurge works correctly (and checks that the previously mentioned operations are fully supported thoroughly tested). One noticeable design decision is the timing of the MemPurge operation in the memtable workflow: for this prototype, the mempurge happens when the memtable is switched (and usually made immutable). This is an inefficient process because it implies that the entirety of the MemPurge operation happens while holding the db_mutex. Future commits will make the MemPurge operation a background task (akin to the regular flush operation) and aim at drastically enhancing the performance of this operation. The MemPurge is also not fully "WAL-compatible" yet, but when the WAL is full, or when the regular MemPurge operation fails (or when the purged memtable still needs to be flushed), a regular flush operation takes place. Later commits will also correct these behaviors. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8454 Reviewed By: anand1976 Differential Revision: D29433971 Pulled By: bjlemaire fbshipit-source-id: 6af48213554e35048a7e03816955100a80a26dc5
4 years ago
// MEMPURGE: No need to change this, because new adds
// should still receive new sequence numbers.
Track WAL obsoletion when updating empty CF's log number (#7781) Summary: In the write path, there is an optimization: when a new WAL is created during SwitchMemtable, we update the internal log number of the empty column families to the new WAL. `FindObsoleteFiles` marks a WAL as obsolete if the WAL's log number is less than `VersionSet::MinLogNumberWithUnflushedData`. After updating the empty column families' internal log number, `VersionSet::MinLogNumberWithUnflushedData` might change, so some WALs might become obsolete to be purged from disk. For example, consider there are 3 column families: 0, 1, 2: 1. initially, all the column families' log number is 1; 2. write some data to cf0, and flush cf0, but the flush is pending; 3. now a new WAL 2 is created; 4. write data to cf1 and WAL 2, now cf0's log number is 1, cf1's log number is 2, cf2's log number is 2 (because cf1 and cf2 are empty, so their log numbers will be set to the highest log number); 5. now cf0's flush hasn't finished, flush cf1, a new WAL 3 is created, and cf1's flush finishes, now cf0's log number is 1, cf1's log number is 3, cf2's log number is 3, since WAL 1 still contains data for the unflushed cf0, no WAL can be deleted from disk; 6. now cf0's flush finishes, cf0's log number is 2 (because when cf0 was switching memtable, WAL 3 does not exist yet), cf1's log number is 3, cf2's log number is 3, so WAL 1 can be purged from disk now, but WAL 2 still cannot because `MinLogNumberToKeep()` is 2; 7. write data to cf2 and WAL 3, because cf0 is empty, its log number is updated to 3, so now cf0's log number is 3, cf1's log number is 3, cf2's log number is 3; 8. now if the background threads want to purge obsolete files from disk, WAL 2 can be purged because `MinLogNumberToKeep()` is 3. But there are only two flush results written to MANIFEST: the first is for flushing cf1, and the `MinLogNumberToKeep` is 1, the second is for flushing cf0, and the `MinLogNumberToKeep` is 2. So without this PR, if the DB crashes at this point and try to recover, `WalSet` will still expect WAL 2 to exist. When WAL tracking is enabled, we assume WALs will only become obsolete after a flush result is written to MANIFEST in `MemtableList::TryInstallMemtableFlushResults` (or its atomic flush counterpart). The above situation breaks this assumption. This PR tracks WAL obsoletion if necessary before updating the empty column families' log numbers. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7781 Test Plan: watch existing tests and stress tests to pass. `make -j48 blackbox_crash_test` on devserver Reviewed By: ltamasi Differential Revision: D25631695 Pulled By: cheng-chang fbshipit-source-id: ca7fff967bdb42204b84226063d909893bc0a4ec
4 years ago
cf->mem()->SetCreationSeq(versions_->LastSequence());
} // cf may become non-empty.
}
empty_cf_updated = true;
}
}
if (!empty_cf_updated) {
for (auto cf : *versions_->GetColumnFamilySet()) {
// all this is just optimization to delete logs that
// are no longer needed -- if CF is empty, that means it
// doesn't need that particular log to stay alive, so we just
// advance the log number. no need to persist this in the manifest
if (cf->IsEmpty()) {
if (creating_new_log) {
cf->SetLogNumber(logfile_number_);
}
cf->mem()->SetCreationSeq(versions_->LastSequence());
}
}
}
cfd->mem()->SetNextLogNumber(logfile_number_);
assert(new_mem != nullptr);
Make mempurge a background process (equivalent to in-memory compaction). (#8505) Summary: In https://github.com/facebook/rocksdb/issues/8454, I introduced a new process baptized `MemPurge` (memtable garbage collection). This new PR is built upon this past mempurge prototype. In this PR, I made the `mempurge` process a background task, which provides superior performance since the mempurge process does not cling on the db_mutex anymore, and addresses severe restrictions from the past iteration (including a scenario where the past mempurge was failling, when a memtable was mempurged but was still referred to by an iterator/snapshot/...). Now the mempurge process ressembles an in-memory compaction process: the stack of immutable memtables is filtered out, and the useful payload is used to populate an output memtable. If the output memtable is filled at more than 60% capacity (arbitrary heuristic) the mempurge process is aborted and a regular flush process takes place, else the output memtable is kept in the immutable memtable stack. Note that adding this output memtable to the `imm()` memtable stack does not trigger another flush process, so that the flush thread can go to sleep at the end of a successful mempurge. MemPurge is activated by making the `experimental_allow_mempurge` flag `true`. When activated, the `MemPurge` process will always happen when the flush reason is `kWriteBufferFull`. The 3 unit tests confirm that this process supports `Put`, `Get`, `Delete`, `DeleteRange` operators and is compatible with `Iterators` and `CompactionFilters`. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8505 Reviewed By: pdillinger Differential Revision: D29619283 Pulled By: bjlemaire fbshipit-source-id: 8a99bee76b63a8211bff1a00e0ae32360aaece95
4 years ago
cfd->imm()->Add(cfd->mem(), &context->memtables_to_free_);
new_mem->Ref();
cfd->SetMemtable(new_mem);
InstallSuperVersionAndScheduleWork(cfd, &context->superversion_context,
Make mempurge a background process (equivalent to in-memory compaction). (#8505) Summary: In https://github.com/facebook/rocksdb/issues/8454, I introduced a new process baptized `MemPurge` (memtable garbage collection). This new PR is built upon this past mempurge prototype. In this PR, I made the `mempurge` process a background task, which provides superior performance since the mempurge process does not cling on the db_mutex anymore, and addresses severe restrictions from the past iteration (including a scenario where the past mempurge was failling, when a memtable was mempurged but was still referred to by an iterator/snapshot/...). Now the mempurge process ressembles an in-memory compaction process: the stack of immutable memtables is filtered out, and the useful payload is used to populate an output memtable. If the output memtable is filled at more than 60% capacity (arbitrary heuristic) the mempurge process is aborted and a regular flush process takes place, else the output memtable is kept in the immutable memtable stack. Note that adding this output memtable to the `imm()` memtable stack does not trigger another flush process, so that the flush thread can go to sleep at the end of a successful mempurge. MemPurge is activated by making the `experimental_allow_mempurge` flag `true`. When activated, the `MemPurge` process will always happen when the flush reason is `kWriteBufferFull`. The 3 unit tests confirm that this process supports `Put`, `Get`, `Delete`, `DeleteRange` operators and is compatible with `Iterators` and `CompactionFilters`. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8505 Reviewed By: pdillinger Differential Revision: D29619283 Pulled By: bjlemaire fbshipit-source-id: 8a99bee76b63a8211bff1a00e0ae32360aaece95
4 years ago
mutable_cf_options);
Memtable "MemPurge" prototype (#8454) Summary: Implement an experimental feature called "MemPurge", which consists in purging "garbage" bytes out of a memtable and reuse the memtable struct instead of making it immutable and eventually flushing its content to storage. The prototype is by default deactivated and is not intended for use. It is intended for correctness and validation testing. At the moment, the "MemPurge" feature can be switched on by using the `options.experimental_allow_mempurge` flag. For this early stage, when the allow_mempurge flag is set to `true`, all the flush operations will be rerouted to perform a MemPurge. This is a temporary design decision that will give us the time to explore meaningful heuristics to use MemPurge at the right time for relevant workloads . Moreover, the current MemPurge operation only supports `Puts`, `Deletes`, `DeleteRange` operations, and handles `Iterators` as well as `CompactionFilter`s that are invoked at flush time . Three unit tests are added to `db_flush_test.cc` to test if MemPurge works correctly (and checks that the previously mentioned operations are fully supported thoroughly tested). One noticeable design decision is the timing of the MemPurge operation in the memtable workflow: for this prototype, the mempurge happens when the memtable is switched (and usually made immutable). This is an inefficient process because it implies that the entirety of the MemPurge operation happens while holding the db_mutex. Future commits will make the MemPurge operation a background task (akin to the regular flush operation) and aim at drastically enhancing the performance of this operation. The MemPurge is also not fully "WAL-compatible" yet, but when the WAL is full, or when the regular MemPurge operation fails (or when the purged memtable still needs to be flushed), a regular flush operation takes place. Later commits will also correct these behaviors. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8454 Reviewed By: anand1976 Differential Revision: D29433971 Pulled By: bjlemaire fbshipit-source-id: 6af48213554e35048a7e03816955100a80a26dc5
4 years ago
// Notify client that memtable is sealed, now that we have successfully
// installed a new memtable
NotifyOnMemTableSealed(cfd, memtable_info);
// It is possible that we got here without checking the value of i_os, but
// that is okay. If we did, it most likely means that s was already an error.
// In any case, ignore any unchecked error for i_os here.
io_s.PermitUncheckedError();
return s;
}
size_t DBImpl::GetWalPreallocateBlockSize(uint64_t write_buffer_size) const {
mutex_.AssertHeld();
size_t bsize =
static_cast<size_t>(write_buffer_size / 10 + write_buffer_size);
// Some users might set very high write_buffer_size and rely on
// max_total_wal_size or other parameters to control the WAL size.
if (mutable_db_options_.max_total_wal_size > 0) {
bsize = std::min<size_t>(
bsize, static_cast<size_t>(mutable_db_options_.max_total_wal_size));
}
if (immutable_db_options_.db_write_buffer_size > 0) {
bsize = std::min<size_t>(bsize, immutable_db_options_.db_write_buffer_size);
}
if (immutable_db_options_.write_buffer_manager &&
immutable_db_options_.write_buffer_manager->enabled()) {
bsize = std::min<size_t>(
bsize, immutable_db_options_.write_buffer_manager->buffer_size());
}
return bsize;
}
// Default implementations of convenience methods that subclasses of DB
// can call if they wish
Status DB::Put(const WriteOptions& opt, ColumnFamilyHandle* column_family,
const Slice& key, const Slice& value) {
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
// Pre-allocate size of write batch conservatively.
// 8 bytes are taken by header, 4 bytes for count, 1 byte for type,
// and we allocate 11 extra bytes for key length, as well as value length.
WriteBatch batch(key.size() + value.size() + 24, 0 /* max_bytes */,
opt.protection_bytes_per_key, 0 /* default_cf_ts_sz */);
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
Status s = batch.Put(column_family, key, value);
if (!s.ok()) {
return s;
Add support for timestamp in Get/Put (#5079) Summary: It's useful to be able to (optionally) associate key-value pairs with user-provided timestamps. This PR is an early effort towards this goal and continues the work of facebook#4942. A suite of new unit tests exist in DBBasicTestWithTimestampWithParam. Support for timestamp requires the user to provide timestamp as a slice in `ReadOptions` and `WriteOptions`. All timestamps of the same database must share the same length, format, etc. The format of the timestamp is the same throughout the same database, and the user is responsible for providing a comparator function (Comparator) to order the <key, timestamp> tuples. Once created, the format and length of the timestamp cannot change (at least for now). Test plan (on devserver): ``` $COMPILE_WITH_ASAN=1 make -j32 all $./db_basic_test --gtest_filter=Timestamp/DBBasicTestWithTimestampWithParam.PutAndGet/* $make check ``` All tests must pass. We also run the following db_bench tests to verify whether there is regression on Get/Put while timestamp is not enabled. ``` $TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillseq,readrandom -num=1000000 $TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=1000000 ``` Repeat for 6 times for both versions. Results are as follows: ``` | | readrandom | fillrandom | | master | 16.77 MB/s | 47.05 MB/s | | PR5079 | 16.44 MB/s | 47.03 MB/s | ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/5079 Differential Revision: D15132946 Pulled By: riversand963 fbshipit-source-id: 833a0d657eac21182f0f206c910a6438154c742c
6 years ago
}
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
return Write(opt, &batch);
}
Status DB::Put(const WriteOptions& opt, ColumnFamilyHandle* column_family,
const Slice& key, const Slice& ts, const Slice& value) {
ColumnFamilyHandle* default_cf = DefaultColumnFamily();
assert(default_cf);
const Comparator* const default_cf_ucmp = default_cf->GetComparator();
assert(default_cf_ucmp);
WriteBatch batch(0 /* reserved_bytes */, 0 /* max_bytes */,
opt.protection_bytes_per_key,
default_cf_ucmp->timestamp_size());
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
Status s = batch.Put(column_family, key, ts, value);
if (!s.ok()) {
return s;
}
return Write(opt, &batch);
}
Status DB::PutEntity(const WriteOptions& options,
ColumnFamilyHandle* column_family, const Slice& key,
const WideColumns& columns) {
const ColumnFamilyHandle* const default_cf = DefaultColumnFamily();
assert(default_cf);
const Comparator* const default_cf_ucmp = default_cf->GetComparator();
assert(default_cf_ucmp);
WriteBatch batch(/* reserved_bytes */ 0, /* max_bytes */ 0,
options.protection_bytes_per_key,
default_cf_ucmp->timestamp_size());
const Status s = batch.PutEntity(column_family, key, columns);
if (!s.ok()) {
return s;
}
return Write(options, &batch);
}
Status DB::Delete(const WriteOptions& opt, ColumnFamilyHandle* column_family,
const Slice& key) {
WriteBatch batch(0 /* reserved_bytes */, 0 /* max_bytes */,
opt.protection_bytes_per_key, 0 /* default_cf_ts_sz */);
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
Status s = batch.Delete(column_family, key);
if (!s.ok()) {
return s;
}
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
return Write(opt, &batch);
}
Status DB::Delete(const WriteOptions& opt, ColumnFamilyHandle* column_family,
const Slice& key, const Slice& ts) {
ColumnFamilyHandle* default_cf = DefaultColumnFamily();
assert(default_cf);
const Comparator* const default_cf_ucmp = default_cf->GetComparator();
assert(default_cf_ucmp);
WriteBatch batch(0 /* reserved_bytes */, 0 /* max_bytes */,
opt.protection_bytes_per_key,
default_cf_ucmp->timestamp_size());
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
Status s = batch.Delete(column_family, key, ts);
if (!s.ok()) {
return s;
}
return Write(opt, &batch);
}
Status DB::SingleDelete(const WriteOptions& opt,
ColumnFamilyHandle* column_family, const Slice& key) {
WriteBatch batch(0 /* reserved_bytes */, 0 /* max_bytes */,
opt.protection_bytes_per_key, 0 /* default_cf_ts_sz */);
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
Status s = batch.SingleDelete(column_family, key);
if (!s.ok()) {
return s;
}
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
return Write(opt, &batch);
}
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
Status DB::SingleDelete(const WriteOptions& opt,
ColumnFamilyHandle* column_family, const Slice& key,
const Slice& ts) {
ColumnFamilyHandle* default_cf = DefaultColumnFamily();
assert(default_cf);
const Comparator* const default_cf_ucmp = default_cf->GetComparator();
assert(default_cf_ucmp);
WriteBatch batch(0 /* reserved_bytes */, 0 /* max_bytes */,
opt.protection_bytes_per_key,
default_cf_ucmp->timestamp_size());
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
Status s = batch.SingleDelete(column_family, key, ts);
if (!s.ok()) {
return s;
}
Revise APIs related to user-defined timestamp (#8946) Summary: ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`. Namely, `WriteOptions` should not include information about "what-to-write", but should just include information about "how-to-write". According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore, this PR removes `WriteOptions::timestamp` for compliance. After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and `SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity made me reconsider doing it in another PR (maybe). For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take extra `timestamp` information when writing to `WriteBatch`es. These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list. Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to `WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps allocated already and multiple timestamps can be updated. The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp size of the default column family. This will be used to allocate space when calling APIs that do not specify a column family handle. Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing some assertions about timestamp to returning Status code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946 Test Plan: make check ./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8 ./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0 Make sure there is no perf regression by running the following ``` ./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom ``` Before this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s ``` After this PR ``` DB path: [/dev/shm/rocksdb] fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s ``` Reviewed By: ltamasi Differential Revision: D33721359 Pulled By: riversand963 fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
return Write(opt, &batch);
}
Status DB::DeleteRange(const WriteOptions& opt,
ColumnFamilyHandle* column_family,
const Slice& begin_key, const Slice& end_key) {
WriteBatch batch(0 /* reserved_bytes */, 0 /* max_bytes */,
opt.protection_bytes_per_key, 0 /* default_cf_ts_sz */);
Status s = batch.DeleteRange(column_family, begin_key, end_key);
if (!s.ok()) {
return s;
}
return Write(opt, &batch);
}
User-defined timestamp support for `DeleteRange()` (#10661) Summary: Add user-defined timestamp support for range deletion. The new API is `DeleteRange(opt, cf, begin_key, end_key, ts)`. Most of the change is to update the comparator to compare without timestamp. Other than that, major changes are - internal range tombstone data structures (`FragmentedRangeTombstoneList`, `RangeTombstone`, etc.) to store timestamps. - Garbage collection of range tombstones and range tombstone covered keys during compaction. - Get()/MultiGet() to return the timestamp of a range tombstone when needed. - Get/Iterator with range tombstones bounded by readoptions.timestamp. - timestamp crash test now issues DeleteRange by default. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10661 Test Plan: - Added unit test: `make check` - Stress test: `python3 tools/db_crashtest.py --enable_ts whitebox --readpercent=57 --prefixpercent=4 --writepercent=25 -delpercent=5 --iterpercent=5 --delrangepercent=4` - Ran `db_bench` to measure regression when timestamp is not enabled. The tests are for write (with some range deletion) and iterate with DB fitting in memory: `./db_bench--benchmarks=fillrandom,seekrandom --writes_per_range_tombstone=200 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=500000 --reads=500000 --seek_nexts=10 --disable_auto_compactions -disable_wal=true --max_num_range_tombstones=1000`. Did not see consistent regression in no timestamp case. | micros/op | fillrandom | seekrandom | | --- | --- | --- | |main| 2.58 |10.96| |PR 10661| 2.68 |10.63| Reviewed By: riversand963 Differential Revision: D39441192 Pulled By: cbi42 fbshipit-source-id: f05aca3c41605caf110daf0ff405919f300ddec2
2 years ago
Status DB::DeleteRange(const WriteOptions& opt,
ColumnFamilyHandle* column_family,
const Slice& begin_key, const Slice& end_key,
const Slice& ts) {
ColumnFamilyHandle* default_cf = DefaultColumnFamily();
assert(default_cf);
const Comparator* const default_cf_ucmp = default_cf->GetComparator();
assert(default_cf_ucmp);
WriteBatch batch(0 /* reserved_bytes */, 0 /* max_bytes */,
opt.protection_bytes_per_key,
default_cf_ucmp->timestamp_size());
Status s = batch.DeleteRange(column_family, begin_key, end_key, ts);
if (!s.ok()) {
return s;
}
return Write(opt, &batch);
}
Status DB::Merge(const WriteOptions& opt, ColumnFamilyHandle* column_family,
const Slice& key, const Slice& value) {
WriteBatch batch(0 /* reserved_bytes */, 0 /* max_bytes */,
opt.protection_bytes_per_key, 0 /* default_cf_ts_sz */);
Status s = batch.Merge(column_family, key, value);
if (!s.ok()) {
return s;
}
return Write(opt, &batch);
}
Status DB::Merge(const WriteOptions& opt, ColumnFamilyHandle* column_family,
const Slice& key, const Slice& ts, const Slice& value) {
ColumnFamilyHandle* default_cf = DefaultColumnFamily();
assert(default_cf);
const Comparator* const default_cf_ucmp = default_cf->GetComparator();
assert(default_cf_ucmp);
WriteBatch batch(0 /* reserved_bytes */, 0 /* max_bytes */,
opt.protection_bytes_per_key,
default_cf_ucmp->timestamp_size());
Status s = batch.Merge(column_family, key, ts, value);
if (!s.ok()) {
return s;
}
return Write(opt, &batch);
}
} // namespace ROCKSDB_NAMESPACE