rocksdb/db/db_impl/db_impl_write.cc

1831 lines
67 KiB
C++
Raw Normal View History

// 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 "db/db_impl/db_impl.h"
#include <cinttypes>
#include "db/error_handler.h"
#include "db/event_helpers.h"
#include "monitoring/perf_context_imp.h"
#include "options/options_helper.h"
#include "test_util/sync_point.h"
namespace rocksdb {
// Convenience methods
Status DBImpl::Put(const WriteOptions& o, ColumnFamilyHandle* column_family,
const Slice& key, const Slice& val) {
return DB::Put(o, column_family, key, val);
}
Status DBImpl::Merge(const WriteOptions& o, ColumnFamilyHandle* column_family,
const Slice& key, const Slice& val) {
auto cfh = reinterpret_cast<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::Delete(const WriteOptions& write_options,
ColumnFamilyHandle* column_family, const Slice& key) {
return DB::Delete(write_options, column_family, key);
}
Status DBImpl::SingleDelete(const WriteOptions& write_options,
ColumnFamilyHandle* column_family,
const Slice& key) {
return DB::SingleDelete(write_options, column_family, key);
}
void DBImpl::SetRecoverableStatePreReleaseCallback(
PreReleaseCallback* callback) {
recoverable_state_pre_release_callback_.reset(callback);
}
Status DBImpl::Write(const WriteOptions& write_options, WriteBatch* my_batch) {
return WriteImpl(write_options, my_batch, nullptr, nullptr);
}
#ifndef ROCKSDB_LITE
Status DBImpl::WriteWithCallback(const WriteOptions& write_options,
WriteBatch* my_batch,
WriteCallback* callback) {
return WriteImpl(write_options, my_batch, callback, nullptr);
}
#endif // ROCKSDB_LITE
// 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,
PreReleaseCallback* pre_release_callback) {
assert(!seq_per_batch_ || batch_cnt != 0);
if (my_batch == nullptr) {
return Status::Corruption("Batch is nullptr!");
}
if (tracer_) {
InstrumentedMutexLock lock(&trace_mutex_);
if (tracer_) {
tracer_->Write(my_batch);
}
}
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
2017-06-24 23:06:43 +02:00
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
2019-05-14 02:43:47 +02:00
if (immutable_db_options_.unordered_write &&
immutable_db_options_.enable_pipelined_write) {
return Status::NotSupported(
"pipelined_writes is not compatible with unordered_write");
}
// Otherwise IsLatestPersistentState optimization does not make sense
assert(!WriteBatchInternal::IsLatestPersistentState(my_batch) ||
disable_memtable);
Status status;
if (write_options.low_pri) {
status = ThrottleLowPriWritesIfNeeded(write_options, my_batch);
if (!status.ok()) {
return status;
}
}
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
2019-05-14 02:43:47 +02:00
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;
// 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 = WriteImplWALOnly(&write_thread_, write_options, my_batch, callback,
log_used, log_ref, &seq, sub_batch_cnt,
pre_release_callback, kDoAssignOrder,
kDoPublishLastSeq, disable_memtable);
if (!status.ok()) {
return status;
}
if (seq_used) {
*seq_used = seq;
}
if (!disable_memtable) {
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
2017-06-24 23:06:43 +02:00
}
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,
disable_memtable, batch_cnt, pre_release_callback);
if (!write_options.disableWAL) {
RecordTick(stats_, WRITE_WITH_WAL);
}
StopWatch write_sw(env_, immutable_db_options_.statistics.get(), 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
2019-08-23 22:54:09 +02:00
&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;
versions_->SetLastSequence(last_sequence);
MemTableInsertStatusCheck(w.status);
write_thread_.ExitAsBatchGroupFollower(&w);
}
assert(w.state == WriteThread::STATE_COMPLETED);
// STATE_COMPLETED conditional below handles exit
status = w.FinalStatus();
}
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);
// 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;
WriteThread::WriteGroup write_group;
bool in_parallel_group = false;
uint64_t last_sequence = kMaxSequenceNumber;
mutex_.Lock();
bool need_log_sync = write_options.sync;
bool need_log_dir_sync = need_log_sync && !log_dir_synced_;
if (!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
2017-06-24 23:06:43 +02:00
// 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);
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
2017-06-24 23:06:43 +02:00
status = PreprocessWrite(write_options, &need_log_sync, &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
2017-06-24 23:06:43 +02:00
}
log::Writer* log_writer = logs_.back().writer;
mutex_.Unlock();
// 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);
if (status.ok()) {
// 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) {
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
2017-06-24 23:06:43 +02:00
if (status.ok() && !write_options.disableWAL) {
PERF_TIMER_GUARD(write_wal_time);
status = WriteToWAL(write_group, log_writer, log_used, need_log_sync,
need_log_dir_sync, last_sequence + 1);
}
} 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
2017-06-24 23:06:43 +02:00
// wal_write_mutex_ to ensure ordered events in WAL
status = 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
2017-06-24 23:06:43 +02:00
} else {
// Otherwise we inc seq number for memtable writes
last_sequence = versions_->FetchAddLastAllocatedSequence(seq_inc);
}
}
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
2017-06-24 23:06:43 +02:00
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 = 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
2019-08-23 22:54:09 +02:00
&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
2019-08-23 22:54:09 +02:00
&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 (!w.CallbackFailed()) {
WriteStatusCheck(status);
}
if (need_log_sync) {
mutex_.Lock();
MarkLogsSynced(logfile_number_, need_log_dir_sync, status);
mutex_.Unlock();
// 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
2017-06-24 23:06:43 +02:00
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
2017-06-24 23:06:43 +02:00
if (status.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
2017-06-24 23:06:43 +02:00
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(env_, immutable_db_options_.statistics.get(), DB_WRITE);
WriteContext write_context;
WriteThread::Writer w(write_options, my_batch, callback, log_ref,
disable_memtable);
write_thread_.JoinBatchGroup(&w);
if (w.state == WriteThread::STATE_GROUP_LEADER) {
WriteThread::WriteGroup wal_write_group;
if (w.callback && !w.callback->AllowWriteBatching()) {
write_thread_.WaitForMemTableWriters();
}
mutex_.Lock();
bool need_log_sync = !write_options.disableWAL && write_options.sync;
bool need_log_dir_sync = need_log_sync && !log_dir_synced_;
// PreprocessWrite does its own perf timing.
PERF_TIMER_STOP(write_pre_and_post_process_time);
w.status = PreprocessWrite(write_options, &need_log_sync, &write_context);
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
2017-06-24 23:06:43 +02:00
log::Writer* log_writer = logs_.back().writer;
mutex_.Unlock();
// 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()) {
SequenceNumber next_sequence = current_sequence;
for (auto writer : wal_write_group) {
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);
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);
}
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
2017-06-24 23:06:43 +02:00
w.status = WriteToWAL(wal_write_group, log_writer, log_used,
need_log_sync, need_log_dir_sync, current_sequence);
}
if (!w.CallbackFailed()) {
WriteStatusCheck(w.status);
}
if (need_log_sync) {
mutex_.Lock();
MarkLogsSynced(logfile_number_, need_log_dir_sync, w.status);
mutex_.Unlock();
}
write_thread_.ExitAsBatchGroupLeader(wal_write_group, w.status);
}
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
2019-08-23 22:54:09 +02:00
&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);
}
}
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
2019-08-23 22:54:09 +02:00
&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
2017-06-24 23:06:43 +02:00
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
2019-05-14 02:43:47 +02:00
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(env_, immutable_db_options_.statistics.get(), DB_WRITE);
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
2019-05-14 02:43:47 +02:00
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
2019-08-23 22:54:09 +02:00
&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
2019-05-14 02:43:47 +02:00
WriteStatusCheck(w.status);
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
2019-05-14 02:43:47 +02:00
switch_cv_.notify_all();
}
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
2019-05-14 02:43:47 +02:00
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
2017-06-24 23:06:43 +02:00
Status status;
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
2019-05-14 02:43:47 +02:00
disable_memtable, sub_batch_cnt, pre_release_callback);
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
2017-06-24 23:06:43 +02:00
RecordTick(stats_, WRITE_WITH_WAL);
StopWatch write_sw(env_, immutable_db_options_.statistics.get(), 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
2019-05-14 02:43:47 +02:00
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
2017-06-24 23:06:43 +02:00
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
2017-06-24 23:06:43 +02:00
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
2019-05-14 02:43:47 +02:00
if (publish_last_seq == kDoPublishLastSeq) {
// 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()) {
InstrumentedMutexLock l(&mutex_);
bool need_log_sync = false;
status = PreprocessWrite(write_options, &need_log_sync, &write_context);
WriteStatusCheck(status);
}
if (!status.ok()) {
WriteThread::WriteGroup write_group;
write_thread->EnterAsBatchGroupLeader(&w, &write_group);
write_thread->ExitAsBatchGroupLeader(write_group, status);
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
2017-06-24 23:06:43 +02:00
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
2019-05-14 02:43:47 +02:00
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
2017-06-24 23:06:43 +02:00
// Note: no need to update last_batch_group_size_ here since the batch writes
// to WAL only
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
2017-06-24 23:06:43 +02:00
for (auto* writer : write_group) {
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
2017-06-24 23:06:43 +02:00
}
}
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
2017-06-24 23:06:43 +02:00
// 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
2017-06-24 23:06:43 +02:00
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
2017-06-24 23:06:43 +02:00
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
2017-06-24 23:06:43 +02:00
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
2017-06-24 23:06:43 +02:00
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
2017-06-24 23:06:43 +02:00
// 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
2019-05-14 02:43:47 +02:00
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
2019-05-14 02:43:47 +02:00
assert(writer->batch_cnt || !seq_per_batch_);
if (!writer->CallbackFailed()) {
total_batch_cnt += writer->batch_cnt;
}
}
seq_inc = total_batch_cnt;
}
if (!write_options.disableWAL) {
status =
ConcurrentWriteToWAL(write_group, log_used, &last_sequence, seq_inc);
} 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
2019-05-14 02:43:47 +02:00
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
2019-05-14 02:43:47 +02:00
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
2019-05-14 02:43:47 +02:00
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
2017-06-24 23:06:43 +02:00
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
2017-06-24 23:06:43 +02:00
// 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
2017-06-24 23:06:43 +02:00
}
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
2019-05-14 02:43:47 +02:00
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
2017-06-24 23:06:43 +02:00
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
2017-06-24 23:06:43 +02:00
return status;
}
void DBImpl::WriteStatusCheck(const Status& status) {
// Is setting bg_error_ enough here? This will at least stop
// compaction and fail any further writes.
if (immutable_db_options_.paranoid_checks && !status.ok() &&
!status.IsBusy() && !status.IsIncomplete()) {
mutex_.Lock();
error_handler_.SetBGError(status, BackgroundErrorReason::kWriteCallback);
mutex_.Unlock();
}
}
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());
error_handler_.SetBGError(status, BackgroundErrorReason::kMemTable);
mutex_.Unlock();
}
}
Status DBImpl::PreprocessWrite(const WriteOptions& write_options,
bool* need_log_sync,
WriteContext* write_context) {
mutex_.AssertHeld();
assert(write_context != nullptr && need_log_sync != 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
2018-09-15 22:36:19 +02:00
if (error_handler_.IsDBStopped()) {
status = error_handler_.GetBGError();
}
PERF_TIMER_GUARD(write_scheduling_flushes_compactions_time);
assert(!single_column_family_mode_ ||
versions_->GetColumnFamilySet()->NumberOfColumnFamilies() == 1);
if (UNLIKELY(status.ok() && !single_column_family_mode_ &&
total_log_size_ > GetMaxTotalWalSize())) {
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
2019-05-14 02:43:47 +02:00
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.
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
2019-05-14 02:43:47 +02:00
WaitForPendingWrites();
status = HandleWriteBufferFull(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
2019-08-23 22:54:09 +02:00
if (UNLIKELY(status.ok() && !trim_history_scheduler_.Empty())) {
status = TrimMemtableHistory(write_context);
}
if (UNLIKELY(status.ok() && !flush_scheduler_.Empty())) {
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
2019-05-14 02:43:47 +02:00
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.
status = DelayWrite(last_batch_group_size_, write_options);
PERF_TIMER_START(write_pre_and_post_process_time);
}
if (status.ok() && *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().getting_synced) {
log_sync_cv_.Wait();
}
for (auto& log : logs_) {
assert(!log.getting_synced);
// 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.getting_synced = true;
}
} else {
*need_log_sync = false;
}
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
2017-06-24 23:06:43 +02:00
WriteBatch* DBImpl::MergeBatch(const WriteThread::WriteGroup& write_group,
WriteBatch* tmp_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
2017-06-24 23:06:43 +02:00
assert(write_with_wal != nullptr);
assert(tmp_batch != nullptr);
assert(*to_be_cached_state == nullptr);
WriteBatch* merged_batch = 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
2017-06-24 23:06:43 +02:00
*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
2017-06-24 23:06:43 +02:00
*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
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
2017-06-24 23:06:43 +02:00
merged_batch = tmp_batch;
for (auto writer : write_group) {
if (!writer->CallbackFailed()) {
WriteBatchInternal::Append(merged_batch, writer->batch,
/*WAL_only*/ true);
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
2017-06-24 23:06:43 +02:00
(*write_with_wal)++;
}
}
}
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
2017-06-24 23:06:43 +02:00
return merged_batch;
}
// When two_write_queues_ is disabled, this function is called from the only
// write thread. Otherwise this must be called holding log_write_mutex_.
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
2017-06-24 23:06:43 +02:00
Status DBImpl::WriteToWAL(const WriteBatch& merged_batch,
log::Writer* log_writer, uint64_t* log_used,
uint64_t* log_size) {
assert(log_size != nullptr);
Slice log_entry = WriteBatchInternal::Contents(&merged_batch);
*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();
}
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
2017-06-24 23:06:43 +02:00
Status status = log_writer->AddRecord(log_entry);
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
2017-06-24 23:06:43 +02:00
if (log_used != nullptr) {
*log_used = logfile_number_;
}
total_log_size_ += log_entry.size();
// TODO(myabandeh): it might be unsafe to access alive_log_files_.back() here
// since alive_log_files_ might be modified concurrently
alive_log_files_.back().AddSize(log_entry.size());
log_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
2017-06-24 23:06:43 +02:00
return status;
}
Status 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) {
Status status;
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
2017-06-24 23:06:43 +02:00
size_t write_with_wal = 0;
WriteBatch* to_be_cached_state = nullptr;
WriteBatch* merged_batch = MergeBatch(write_group, &tmp_batch_,
&write_with_wal, &to_be_cached_state);
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
2017-06-24 23:06:43 +02:00
WriteBatchInternal::SetSequence(merged_batch, sequence);
uint64_t log_size;
status = WriteToWAL(*merged_batch, log_writer, log_used, &log_size);
if (to_be_cached_state) {
cached_recoverable_state_ = *to_be_cached_state;
cached_recoverable_state_empty_ = false;
}
if (status.ok() && need_log_sync) {
StopWatch sw(env_, 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.
for (auto& log : logs_) {
status = log.writer->file()->Sync(immutable_db_options_.use_fsync);
if (!status.ok()) {
break;
}
}
if (status.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.
status = directories_.GetWalDir()->Fsync();
}
}
if (merged_batch == &tmp_batch_) {
tmp_batch_.Clear();
}
if (status.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 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
2017-06-24 23:06:43 +02:00
Status DBImpl::ConcurrentWriteToWAL(const WriteThread::WriteGroup& write_group,
uint64_t* log_used,
SequenceNumber* last_sequence,
size_t 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
2017-06-24 23:06:43 +02:00
Status status;
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
2017-06-24 23:06:43 +02:00
WriteBatch tmp_batch;
size_t write_with_wal = 0;
WriteBatch* 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
2017-06-24 23:06:43 +02:00
WriteBatch* merged_batch =
MergeBatch(write_group, &tmp_batch, &write_with_wal, &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
2017-06-24 23:06:43 +02:00
// 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
2017-06-24 23:06:43 +02:00
auto sequence = *last_sequence + 1;
WriteBatchInternal::SetSequence(merged_batch, sequence);
log::Writer* log_writer = logs_.back().writer;
uint64_t log_size;
status = WriteToWAL(*merged_batch, log_writer, log_used, &log_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
2017-06-24 23:06:43 +02:00
log_write_mutex_.Unlock();
if (status.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
2017-06-24 23:06:43 +02:00
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
2017-06-24 23:06:43 +02:00
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
2017-06-24 23:06:43 +02:00
RecordTick(stats_, WRITE_WITH_WAL, write_with_wal);
}
return status;
}
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
2019-08-23 22:54:09 +02:00
&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(
autovector<ColumnFamilyData*>* cfds) {
for (ColumnFamilyData* cfd : *versions_->GetColumnFamilySet()) {
if (cfd->IsDropped()) {
continue;
}
if (cfd->imm()->NumNotFlushed() != 0 || !cfd->mem()->IsEmpty() ||
!cached_recoverable_state_empty_.load()) {
cfds->push_back(cfd);
}
}
}
// 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
2018-05-04 00:35:11 +02:00
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
2018-05-04 00:35:11 +02:00
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
2018-05-04 00:35:11 +02:00
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
2018-05-04 00:35:11 +02:00
// 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
2018-05-04 00:35:11 +02:00
} 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
2018-05-04 00:35:11 +02:00
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
2018-05-04 00:35:11 +02:00
}
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
2018-05-04 00:35:11 +02:00
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->Unref();
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();
}
FlushRequest flush_req;
GenerateFlushRequest(cfds, &flush_req);
SchedulePendingFlush(flush_req, FlushReason::kWriteBufferManager);
MaybeScheduleFlushOrCompaction();
}
return status;
}
Status DBImpl::HandleWriteBufferFull(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.
ROCKS_LOG_INFO(
immutable_db_options_.info_log,
"Flushing column family with largest mem table size. Write buffer is "
"using %" ROCKSDB_PRIszt " bytes out of a total of %" ROCKSDB_PRIszt ".",
write_buffer_manager_->memory_usage(),
write_buffer_manager_->buffer_size());
// 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;
}
if (!cfd->mem()->IsEmpty()) {
// We only consider active mem table, hoping immutable memtable is
// already in the process of flushing.
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);
}
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->Unref();
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();
}
FlushRequest flush_req;
GenerateFlushRequest(cfds, &flush_req);
SchedulePendingFlush(flush_req, FlushReason::kWriteBufferFull);
MaybeScheduleFlushOrCompaction();
}
return status;
}
uint64_t DBImpl::GetMaxTotalWalSize() const {
mutex_.AssertHeld();
return mutable_db_options_.max_total_wal_size == 0
? 4 * max_total_in_memory_state_
: mutable_db_options_.max_total_wal_size;
}
// REQUIRES: mutex_ is held
// REQUIRES: this thread is currently at the front of the writer queue
Status DBImpl::DelayWrite(uint64_t num_bytes,
const WriteOptions& write_options) {
uint64_t time_delayed = 0;
bool delayed = false;
{
StopWatch sw(env_, stats_, WRITE_STALL, &time_delayed);
uint64_t delay = write_controller_.GetDelay(env_, num_bytes);
if (delay > 0) {
if (write_options.no_slowdown) {
return Status::Incomplete("Write stall");
}
TEST_SYNC_POINT("DBImpl::DelayWrite:Sleep");
// Notify write_thread_ about the stall so it can setup a barrier and
// fail any pending writers with no_slowdown
write_thread_.BeginWriteStall();
TEST_SYNC_POINT("DBImpl::DelayWrite:BeginWriteStallDone");
mutex_.Unlock();
// We will delay the write until we have slept for delay ms or
// we don't need a delay anymore
const uint64_t kDelayInterval = 1000;
uint64_t stall_end = sw.start_time() + delay;
while (write_controller_.NeedsDelay()) {
if (env_->NowMicros() >= stall_end) {
// We already delayed this write `delay` microseconds
break;
}
delayed = true;
// Sleep for 0.001 seconds
env_->SleepForMicroseconds(kDelayInterval);
}
mutex_.Lock();
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
2018-09-15 22:36:19 +02:00
// 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()) {
if (write_options.no_slowdown) {
return Status::Incomplete("Write stall");
}
delayed = true;
// Notify write_thread_ about the stall so it can setup a barrier and
// fail any pending writers with no_slowdown
write_thread_.BeginWriteStall();
TEST_SYNC_POINT("DBImpl::DelayWrite:Wait");
bg_cv_.Wait();
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
2018-09-15 22:36:19 +02:00
// 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 writes are still stopped, it means we bailed due to a background
// error
s = Status::Incomplete(error_handler_.GetBGError().ToString());
}
if (error_handler_.IsDBStopped()) {
s = error_handler_.GetBGError();
}
return s;
}
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();
} 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
2019-08-23 22:54:09 +02:00
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;
cfd->imm()->TrimHistory(&to_delete, cfd->mem()->ApproximateMemoryUsage());
for (auto m : to_delete) {
delete m;
}
context->superversion_context.NewSuperVersion();
assert(context->superversion_context.new_superversion.get() != nullptr);
cfd->InstallSuperVersion(&context->superversion_context, &mutex_);
if (cfd->Unref()) {
delete cfd;
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->Unref()) {
delete cfd;
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, &flush_req);
SchedulePendingFlush(flush_req, FlushReason::kWriteBufferFull);
MaybeScheduleFlushOrCompaction();
}
return status;
}
#ifndef ROCKSDB_LITE
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;
}
for (auto listener : immutable_db_options_.listeners) {
listener->OnMemTableSealed(mem_table_info);
}
}
#endif // ROCKSDB_LITE
// 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();
WriteThread::Writer nonmem_w;
std::unique_ptr<WritableFile> lfile;
log::Writer* new_log = nullptr;
MemTable* new_mem = nullptr;
// 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();
log_recycle_files_.pop_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
#ifndef ROCKSDB_LITE
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();
#endif // ROCKSDB_LITE
// 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.
s = CreateWAL(new_log_number, recycle_log_number, preallocate_block_size,
&new_log);
}
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);
mutex_.Lock();
if (s.ok() && creating_new_log) {
log_write_mutex_.Lock();
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
2017-06-24 23:06:43 +02:00
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;
s = cur_log_writer->WriteBuffer();
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
2017-06-24 23:06:43 +02:00
}
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_));
}
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
2017-06-24 23:06:43 +02:00
log_write_mutex_.Unlock();
}
if (!s.ok()) {
// how do we fail if we're not creating new log?
assert(creating_new_log);
if (new_mem) {
delete new_mem;
}
if (new_log) {
delete new_log;
}
SuperVersion* new_superversion =
context->superversion_context.new_superversion.release();
if (new_superversion != nullptr) {
delete new_superversion;
}
// 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
error_handler_.SetBGError(s, BackgroundErrorReason::kMemTable);
// Read back bg_error in order to get the right severity
s = error_handler_.GetBGError();
return s;
}
for (auto loop_cfd : *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 (loop_cfd->mem()->GetFirstSequenceNumber() == 0 &&
loop_cfd->imm()->NumNotFlushed() == 0) {
if (creating_new_log) {
loop_cfd->SetLogNumber(logfile_number_);
}
loop_cfd->mem()->SetCreationSeq(versions_->LastSequence());
}
}
cfd->mem()->SetNextLogNumber(logfile_number_);
cfd->imm()->Add(cfd->mem(), &context->memtables_to_free_);
new_mem->Ref();
cfd->SetMemtable(new_mem);
InstallSuperVersionAndScheduleWork(cfd, &context->superversion_context,
mutable_cf_options);
#ifndef ROCKSDB_LITE
mutex_.Unlock();
// Notify client that memtable is sealed, now that we have successfully
// installed a new memtable
NotifyOnMemTableSealed(cfd, memtable_info);
mutex_.Lock();
#endif // ROCKSDB_LITE
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) {
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
2019-06-06 08:07:28 +02:00
if (nullptr == opt.timestamp) {
// 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);
Status s = batch.Put(column_family, key, value);
if (!s.ok()) {
return s;
}
return Write(opt, &batch);
}
Avoid user key copying for Get/Put/Write with user-timestamp (#5502) Summary: In previous https://github.com/facebook/rocksdb/issues/5079, we added user-specified timestamp to `DB::Get()` and `DB::Put()`. Limitation is that these two functions may cause extra memory allocation and key copy. The reason is that `WriteBatch` does not allocate extra memory for timestamps because it is not aware of timestamp size, and we did not provide an API to assign/update timestamp of each key within a `WriteBatch`. We address these issues in this PR by doing the following. 1. Add a `timestamp_size_` to `WriteBatch` so that `WriteBatch` can take timestamps into account when calling `WriteBatch::Put`, `WriteBatch::Delete`, etc. 2. Add APIs `WriteBatch::AssignTimestamp` and `WriteBatch::AssignTimestamps` so that application can assign/update timestamps for each key in a `WriteBatch`. 3. Avoid key copy in `GetImpl` by adding new constructor to `LookupKey`. Test plan (on devserver): ``` $make clean && COMPILE_WITH_ASAN=1 make -j32 all $./db_basic_test --gtest_filter=Timestamp/DBBasicTestWithTimestampWithParam.PutAndGet/* $make check ``` If the API extension looks good, I will add more unit tests. Some simple benchmark using db_bench. ``` $rm -rf /dev/shm/dbbench/* && TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillseq,readrandom -num=1000000 $rm -rf /dev/shm/dbbench/* && TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=1000000 -disable_wal=true ``` Master is at a78503bd6c80a3c4137df1962a972fe406b4d90b. ``` | | readrandom | fillrandom | | master | 15.53 MB/s | 25.97 MB/s | | PR5502 | 16.70 MB/s | 25.80 MB/s | ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/5502 Differential Revision: D16340894 Pulled By: riversand963 fbshipit-source-id: 51132cf792be07d1efc3ac33f5768c4ee2608bb8
2019-07-26 00:23:46 +02:00
const Slice* ts = opt.timestamp;
assert(nullptr != ts);
size_t ts_sz = ts->size();
WriteBatch batch(key.size() + ts_sz + value.size() + 24, /*max_bytes=*/0,
ts_sz);
Status s = batch.Put(column_family, key, value);
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
2019-06-06 08:07:28 +02:00
if (!s.ok()) {
return s;
}
Avoid user key copying for Get/Put/Write with user-timestamp (#5502) Summary: In previous https://github.com/facebook/rocksdb/issues/5079, we added user-specified timestamp to `DB::Get()` and `DB::Put()`. Limitation is that these two functions may cause extra memory allocation and key copy. The reason is that `WriteBatch` does not allocate extra memory for timestamps because it is not aware of timestamp size, and we did not provide an API to assign/update timestamp of each key within a `WriteBatch`. We address these issues in this PR by doing the following. 1. Add a `timestamp_size_` to `WriteBatch` so that `WriteBatch` can take timestamps into account when calling `WriteBatch::Put`, `WriteBatch::Delete`, etc. 2. Add APIs `WriteBatch::AssignTimestamp` and `WriteBatch::AssignTimestamps` so that application can assign/update timestamps for each key in a `WriteBatch`. 3. Avoid key copy in `GetImpl` by adding new constructor to `LookupKey`. Test plan (on devserver): ``` $make clean && COMPILE_WITH_ASAN=1 make -j32 all $./db_basic_test --gtest_filter=Timestamp/DBBasicTestWithTimestampWithParam.PutAndGet/* $make check ``` If the API extension looks good, I will add more unit tests. Some simple benchmark using db_bench. ``` $rm -rf /dev/shm/dbbench/* && TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillseq,readrandom -num=1000000 $rm -rf /dev/shm/dbbench/* && TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=1000000 -disable_wal=true ``` Master is at a78503bd6c80a3c4137df1962a972fe406b4d90b. ``` | | readrandom | fillrandom | | master | 15.53 MB/s | 25.97 MB/s | | PR5502 | 16.70 MB/s | 25.80 MB/s | ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/5502 Differential Revision: D16340894 Pulled By: riversand963 fbshipit-source-id: 51132cf792be07d1efc3ac33f5768c4ee2608bb8
2019-07-26 00:23:46 +02:00
s = batch.AssignTimestamp(*ts);
if (!s.ok()) {
return s;
}
return Write(opt, &batch);
}
Status DB::Delete(const WriteOptions& opt, ColumnFamilyHandle* column_family,
const Slice& key) {
WriteBatch batch;
batch.Delete(column_family, key);
return Write(opt, &batch);
}
Status DB::SingleDelete(const WriteOptions& opt,
ColumnFamilyHandle* column_family, const Slice& key) {
WriteBatch batch;
batch.SingleDelete(column_family, key);
return Write(opt, &batch);
}
Status DB::DeleteRange(const WriteOptions& opt,
ColumnFamilyHandle* column_family,
const Slice& begin_key, const Slice& end_key) {
WriteBatch batch;
batch.DeleteRange(column_family, begin_key, end_key);
return Write(opt, &batch);
}
Status DB::Merge(const WriteOptions& opt, ColumnFamilyHandle* column_family,
const Slice& key, const Slice& value) {
WriteBatch batch;
Status s = batch.Merge(column_family, key, value);
if (!s.ok()) {
return s;
}
return Write(opt, &batch);
}
} // namespace rocksdb