rocksdb/db/write_batch.cc

2052 lines
72 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.
//
// WriteBatch::rep_ :=
// sequence: fixed64
// count: fixed32
// data: record[count]
// record :=
// kTypeValue varstring varstring
// kTypeDeletion varstring
Support for SingleDelete() Summary: This patch fixes #7460559. It introduces SingleDelete as a new database operation. This operation can be used to delete keys that were never overwritten (no put following another put of the same key). If an overwritten key is single deleted the behavior is undefined. Single deletion of a non-existent key has no effect but multiple consecutive single deletions are not allowed (see limitations). In contrast to the conventional Delete() operation, the deletion entry is removed along with the value when the two are lined up in a compaction. Note: The semantics are similar to @igor's prototype that allowed to have this behavior on the granularity of a column family ( https://reviews.facebook.net/D42093 ). This new patch, however, is more aggressive when it comes to removing tombstones: It removes the SingleDelete together with the value whenever there is no snapshot between them while the older patch only did this when the sequence number of the deletion was older than the earliest snapshot. Most of the complex additions are in the Compaction Iterator, all other changes should be relatively straightforward. The patch also includes basic support for single deletions in db_stress and db_bench. Limitations: - Not compatible with cuckoo hash tables - Single deletions cannot be used in combination with merges and normal deletions on the same key (other keys are not affected by this) - Consecutive single deletions are currently not allowed (and older version of this patch supported this so it could be resurrected if needed) Test Plan: make all check Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor Reviewed By: igor Subscribers: maykov, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 20:42:56 +02:00
// kTypeSingleDeletion varstring
// kTypeRangeDeletion varstring varstring
Support for SingleDelete() Summary: This patch fixes #7460559. It introduces SingleDelete as a new database operation. This operation can be used to delete keys that were never overwritten (no put following another put of the same key). If an overwritten key is single deleted the behavior is undefined. Single deletion of a non-existent key has no effect but multiple consecutive single deletions are not allowed (see limitations). In contrast to the conventional Delete() operation, the deletion entry is removed along with the value when the two are lined up in a compaction. Note: The semantics are similar to @igor's prototype that allowed to have this behavior on the granularity of a column family ( https://reviews.facebook.net/D42093 ). This new patch, however, is more aggressive when it comes to removing tombstones: It removes the SingleDelete together with the value whenever there is no snapshot between them while the older patch only did this when the sequence number of the deletion was older than the earliest snapshot. Most of the complex additions are in the Compaction Iterator, all other changes should be relatively straightforward. The patch also includes basic support for single deletions in db_stress and db_bench. Limitations: - Not compatible with cuckoo hash tables - Single deletions cannot be used in combination with merges and normal deletions on the same key (other keys are not affected by this) - Consecutive single deletions are currently not allowed (and older version of this patch supported this so it could be resurrected if needed) Test Plan: make all check Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor Reviewed By: igor Subscribers: maykov, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 20:42:56 +02:00
// kTypeMerge varstring varstring
// kTypeColumnFamilyValue varint32 varstring varstring
// kTypeColumnFamilyDeletion varint32 varstring
// kTypeColumnFamilySingleDeletion varint32 varstring
// kTypeColumnFamilyRangeDeletion varint32 varstring varstring
Support for SingleDelete() Summary: This patch fixes #7460559. It introduces SingleDelete as a new database operation. This operation can be used to delete keys that were never overwritten (no put following another put of the same key). If an overwritten key is single deleted the behavior is undefined. Single deletion of a non-existent key has no effect but multiple consecutive single deletions are not allowed (see limitations). In contrast to the conventional Delete() operation, the deletion entry is removed along with the value when the two are lined up in a compaction. Note: The semantics are similar to @igor's prototype that allowed to have this behavior on the granularity of a column family ( https://reviews.facebook.net/D42093 ). This new patch, however, is more aggressive when it comes to removing tombstones: It removes the SingleDelete together with the value whenever there is no snapshot between them while the older patch only did this when the sequence number of the deletion was older than the earliest snapshot. Most of the complex additions are in the Compaction Iterator, all other changes should be relatively straightforward. The patch also includes basic support for single deletions in db_stress and db_bench. Limitations: - Not compatible with cuckoo hash tables - Single deletions cannot be used in combination with merges and normal deletions on the same key (other keys are not affected by this) - Consecutive single deletions are currently not allowed (and older version of this patch supported this so it could be resurrected if needed) Test Plan: make all check Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor Reviewed By: igor Subscribers: maykov, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 20:42:56 +02:00
// kTypeColumnFamilyMerge varint32 varstring varstring
// kTypeBeginPrepareXID varstring
// kTypeEndPrepareXID
// kTypeCommitXID varstring
// kTypeRollbackXID varstring
// kTypeBeginPersistedPrepareXID varstring
// kTypeBeginUnprepareXID varstring
// kTypeNoop
// varstring :=
// len: varint32
// data: uint8[len]
#include "rocksdb/write_batch.h"
#include <map>
#include <stack>
#include <stdexcept>
#include <type_traits>
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
2015-08-15 01:59:07 +02:00
#include <vector>
#include "db/column_family.h"
#include "db/db_impl/db_impl.h"
#include "db/dbformat.h"
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
2015-08-15 01:59:07 +02:00
#include "db/flush_scheduler.h"
#include "db/memtable.h"
#include "db/merge_context.h"
#include "db/snapshot_impl.h"
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
#include "db/trim_history_scheduler.h"
#include "db/write_batch_internal.h"
#include "monitoring/perf_context_imp.h"
#include "monitoring/statistics.h"
#include "rocksdb/merge_operator.h"
refactor SavePoints (#5192) Summary: Savepoints are assumed to be used in a stack-wise fashion (only the top element should be used), so they were stored by `WriteBatch` in a member variable `save_points` using an std::stack. Conceptually this is fine, but the implementation had a few issues: - the `save_points_` instance variable was a plain pointer to a heap- allocated `SavePoints` struct. The destructor of `WriteBatch` simply deletes this pointer. However, the copy constructor of WriteBatch just copied that pointer, meaning that copying a WriteBatch with active savepoints will very likely have crashed before. Now a proper copy of the savepoints is made in the copy constructor, and not just a copy of the pointer - `save_points_` was an std::stack, which defaults to `std::deque` for the underlying container. A deque is a bit over the top here, as we only need access to the most recent savepoint (i.e. stack.top()) but never any elements at the front. std::deque is rather expensive to initialize in common environments. For example, the STL implementation shipped with GNU g++ will perform a heap allocation of more than 500 bytes to create an empty deque object. Although the `save_points_` container is created lazily by RocksDB, moving from a deque to a plain `std::vector` is much more memory-efficient. So `save_points_` is now a vector. - `save_points_` was changed from a plain pointer to an `std::unique_ptr`, making ownership more explicit. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5192 Differential Revision: D15024074 Pulled By: maysamyabandeh fbshipit-source-id: 5b128786d3789cde94e46465c9e91badd07a25d7
2019-04-20 05:30:03 +02:00
#include "util/autovector.h"
#include "util/cast_util.h"
#include "util/coding.h"
#include "util/duplicate_detector.h"
#include "util/string_util.h"
#include "util/util.h"
namespace rocksdb {
// anon namespace for file-local types
namespace {
enum ContentFlags : uint32_t {
DEFERRED = 1 << 0,
HAS_PUT = 1 << 1,
HAS_DELETE = 1 << 2,
HAS_SINGLE_DELETE = 1 << 3,
HAS_MERGE = 1 << 4,
HAS_BEGIN_PREPARE = 1 << 5,
HAS_END_PREPARE = 1 << 6,
HAS_COMMIT = 1 << 7,
HAS_ROLLBACK = 1 << 8,
HAS_DELETE_RANGE = 1 << 9,
HAS_BLOB_INDEX = 1 << 10,
HAS_BEGIN_UNPREPARE = 1 << 11,
};
struct BatchContentClassifier : public WriteBatch::Handler {
uint32_t content_flags = 0;
Status PutCF(uint32_t, const Slice&, const Slice&) override {
content_flags |= ContentFlags::HAS_PUT;
return Status::OK();
}
Status DeleteCF(uint32_t, const Slice&) override {
content_flags |= ContentFlags::HAS_DELETE;
return Status::OK();
}
Status SingleDeleteCF(uint32_t, const Slice&) override {
content_flags |= ContentFlags::HAS_SINGLE_DELETE;
return Status::OK();
}
Status DeleteRangeCF(uint32_t, const Slice&, const Slice&) override {
content_flags |= ContentFlags::HAS_DELETE_RANGE;
return Status::OK();
}
Status MergeCF(uint32_t, const Slice&, const Slice&) override {
content_flags |= ContentFlags::HAS_MERGE;
return Status::OK();
}
Status PutBlobIndexCF(uint32_t, const Slice&, const Slice&) override {
content_flags |= ContentFlags::HAS_BLOB_INDEX;
return Status::OK();
}
Status MarkBeginPrepare(bool unprepare) override {
content_flags |= ContentFlags::HAS_BEGIN_PREPARE;
if (unprepare) {
content_flags |= ContentFlags::HAS_BEGIN_UNPREPARE;
}
return Status::OK();
}
Status MarkEndPrepare(const Slice&) override {
content_flags |= ContentFlags::HAS_END_PREPARE;
return Status::OK();
}
Status MarkCommit(const Slice&) override {
content_flags |= ContentFlags::HAS_COMMIT;
return Status::OK();
}
Status MarkRollback(const Slice&) override {
content_flags |= ContentFlags::HAS_ROLLBACK;
return Status::OK();
}
};
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
class TimestampAssigner : public WriteBatch::Handler {
public:
explicit TimestampAssigner(const Slice& ts)
: timestamp_(ts), timestamps_(kEmptyTimestampList) {}
explicit TimestampAssigner(const std::vector<Slice>& ts_list)
: timestamps_(ts_list) {
SanityCheck();
}
~TimestampAssigner() override {}
Status PutCF(uint32_t, const Slice& key, const Slice&) override {
AssignTimestamp(key);
++idx_;
return Status::OK();
}
Status DeleteCF(uint32_t, const Slice& key) override {
AssignTimestamp(key);
++idx_;
return Status::OK();
}
Status SingleDeleteCF(uint32_t, const Slice& key) override {
AssignTimestamp(key);
++idx_;
return Status::OK();
}
Status DeleteRangeCF(uint32_t, const Slice& begin_key,
const Slice& end_key) override {
AssignTimestamp(begin_key);
AssignTimestamp(end_key);
++idx_;
return Status::OK();
}
Status MergeCF(uint32_t, const Slice& key, const Slice&) override {
AssignTimestamp(key);
++idx_;
return Status::OK();
}
Status PutBlobIndexCF(uint32_t, const Slice&, const Slice&) override {
// TODO (yanqin): support blob db in the future.
return Status::OK();
}
Status MarkBeginPrepare(bool) override {
// TODO (yanqin): support in the future.
return Status::OK();
}
Status MarkEndPrepare(const Slice&) override {
// TODO (yanqin): support in the future.
return Status::OK();
}
Status MarkCommit(const Slice&) override {
// TODO (yanqin): support in the future.
return Status::OK();
}
Status MarkRollback(const Slice&) override {
// TODO (yanqin): support in the future.
return Status::OK();
}
private:
void SanityCheck() const {
assert(!timestamps_.empty());
#ifndef NDEBUG
const size_t ts_sz = timestamps_[0].size();
for (size_t i = 1; i != timestamps_.size(); ++i) {
assert(ts_sz == timestamps_[i].size());
}
#endif // !NDEBUG
}
void AssignTimestamp(const Slice& key) {
assert(timestamps_.empty() || idx_ < timestamps_.size());
const Slice& ts = timestamps_.empty() ? timestamp_ : timestamps_[idx_];
size_t ts_sz = ts.size();
char* ptr = const_cast<char*>(key.data() + key.size() - ts_sz);
memcpy(ptr, ts.data(), ts_sz);
}
static const std::vector<Slice> kEmptyTimestampList;
const Slice timestamp_;
const std::vector<Slice>& timestamps_;
size_t idx_ = 0;
// No copy or move.
TimestampAssigner(const TimestampAssigner&) = delete;
TimestampAssigner(TimestampAssigner&&) = delete;
TimestampAssigner& operator=(const TimestampAssigner&) = delete;
TimestampAssigner&& operator=(TimestampAssigner&&) = delete;
};
const std::vector<Slice> TimestampAssigner::kEmptyTimestampList;
} // anon namespace
struct SavePoints {
refactor SavePoints (#5192) Summary: Savepoints are assumed to be used in a stack-wise fashion (only the top element should be used), so they were stored by `WriteBatch` in a member variable `save_points` using an std::stack. Conceptually this is fine, but the implementation had a few issues: - the `save_points_` instance variable was a plain pointer to a heap- allocated `SavePoints` struct. The destructor of `WriteBatch` simply deletes this pointer. However, the copy constructor of WriteBatch just copied that pointer, meaning that copying a WriteBatch with active savepoints will very likely have crashed before. Now a proper copy of the savepoints is made in the copy constructor, and not just a copy of the pointer - `save_points_` was an std::stack, which defaults to `std::deque` for the underlying container. A deque is a bit over the top here, as we only need access to the most recent savepoint (i.e. stack.top()) but never any elements at the front. std::deque is rather expensive to initialize in common environments. For example, the STL implementation shipped with GNU g++ will perform a heap allocation of more than 500 bytes to create an empty deque object. Although the `save_points_` container is created lazily by RocksDB, moving from a deque to a plain `std::vector` is much more memory-efficient. So `save_points_` is now a vector. - `save_points_` was changed from a plain pointer to an `std::unique_ptr`, making ownership more explicit. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5192 Differential Revision: D15024074 Pulled By: maysamyabandeh fbshipit-source-id: 5b128786d3789cde94e46465c9e91badd07a25d7
2019-04-20 05:30:03 +02:00
std::stack<SavePoint, autovector<SavePoint>> stack;
};
WriteBatch::WriteBatch(size_t reserved_bytes, size_t max_bytes)
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
: content_flags_(0), max_bytes_(max_bytes), rep_(), timestamp_size_(0) {
rep_.reserve((reserved_bytes > WriteBatchInternal::kHeader)
? reserved_bytes
: WriteBatchInternal::kHeader);
rep_.resize(WriteBatchInternal::kHeader);
}
WriteBatch::WriteBatch(size_t reserved_bytes, size_t max_bytes, size_t ts_sz)
: content_flags_(0), max_bytes_(max_bytes), rep_(), timestamp_size_(ts_sz) {
rep_.reserve((reserved_bytes > WriteBatchInternal::kHeader) ?
reserved_bytes : WriteBatchInternal::kHeader);
rep_.resize(WriteBatchInternal::kHeader);
}
WriteBatch::WriteBatch(const std::string& rep)
refactor SavePoints (#5192) Summary: Savepoints are assumed to be used in a stack-wise fashion (only the top element should be used), so they were stored by `WriteBatch` in a member variable `save_points` using an std::stack. Conceptually this is fine, but the implementation had a few issues: - the `save_points_` instance variable was a plain pointer to a heap- allocated `SavePoints` struct. The destructor of `WriteBatch` simply deletes this pointer. However, the copy constructor of WriteBatch just copied that pointer, meaning that copying a WriteBatch with active savepoints will very likely have crashed before. Now a proper copy of the savepoints is made in the copy constructor, and not just a copy of the pointer - `save_points_` was an std::stack, which defaults to `std::deque` for the underlying container. A deque is a bit over the top here, as we only need access to the most recent savepoint (i.e. stack.top()) but never any elements at the front. std::deque is rather expensive to initialize in common environments. For example, the STL implementation shipped with GNU g++ will perform a heap allocation of more than 500 bytes to create an empty deque object. Although the `save_points_` container is created lazily by RocksDB, moving from a deque to a plain `std::vector` is much more memory-efficient. So `save_points_` is now a vector. - `save_points_` was changed from a plain pointer to an `std::unique_ptr`, making ownership more explicit. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5192 Differential Revision: D15024074 Pulled By: maysamyabandeh fbshipit-source-id: 5b128786d3789cde94e46465c9e91badd07a25d7
2019-04-20 05:30:03 +02:00
: content_flags_(ContentFlags::DEFERRED),
max_bytes_(0),
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
rep_(rep),
timestamp_size_(0) {}
WriteBatch::WriteBatch(std::string&& rep)
refactor SavePoints (#5192) Summary: Savepoints are assumed to be used in a stack-wise fashion (only the top element should be used), so they were stored by `WriteBatch` in a member variable `save_points` using an std::stack. Conceptually this is fine, but the implementation had a few issues: - the `save_points_` instance variable was a plain pointer to a heap- allocated `SavePoints` struct. The destructor of `WriteBatch` simply deletes this pointer. However, the copy constructor of WriteBatch just copied that pointer, meaning that copying a WriteBatch with active savepoints will very likely have crashed before. Now a proper copy of the savepoints is made in the copy constructor, and not just a copy of the pointer - `save_points_` was an std::stack, which defaults to `std::deque` for the underlying container. A deque is a bit over the top here, as we only need access to the most recent savepoint (i.e. stack.top()) but never any elements at the front. std::deque is rather expensive to initialize in common environments. For example, the STL implementation shipped with GNU g++ will perform a heap allocation of more than 500 bytes to create an empty deque object. Although the `save_points_` container is created lazily by RocksDB, moving from a deque to a plain `std::vector` is much more memory-efficient. So `save_points_` is now a vector. - `save_points_` was changed from a plain pointer to an `std::unique_ptr`, making ownership more explicit. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5192 Differential Revision: D15024074 Pulled By: maysamyabandeh fbshipit-source-id: 5b128786d3789cde94e46465c9e91badd07a25d7
2019-04-20 05:30:03 +02:00
: content_flags_(ContentFlags::DEFERRED),
max_bytes_(0),
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
rep_(std::move(rep)),
timestamp_size_(0) {}
WriteBatch::WriteBatch(const WriteBatch& src)
refactor SavePoints (#5192) Summary: Savepoints are assumed to be used in a stack-wise fashion (only the top element should be used), so they were stored by `WriteBatch` in a member variable `save_points` using an std::stack. Conceptually this is fine, but the implementation had a few issues: - the `save_points_` instance variable was a plain pointer to a heap- allocated `SavePoints` struct. The destructor of `WriteBatch` simply deletes this pointer. However, the copy constructor of WriteBatch just copied that pointer, meaning that copying a WriteBatch with active savepoints will very likely have crashed before. Now a proper copy of the savepoints is made in the copy constructor, and not just a copy of the pointer - `save_points_` was an std::stack, which defaults to `std::deque` for the underlying container. A deque is a bit over the top here, as we only need access to the most recent savepoint (i.e. stack.top()) but never any elements at the front. std::deque is rather expensive to initialize in common environments. For example, the STL implementation shipped with GNU g++ will perform a heap allocation of more than 500 bytes to create an empty deque object. Although the `save_points_` container is created lazily by RocksDB, moving from a deque to a plain `std::vector` is much more memory-efficient. So `save_points_` is now a vector. - `save_points_` was changed from a plain pointer to an `std::unique_ptr`, making ownership more explicit. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5192 Differential Revision: D15024074 Pulled By: maysamyabandeh fbshipit-source-id: 5b128786d3789cde94e46465c9e91badd07a25d7
2019-04-20 05:30:03 +02:00
: wal_term_point_(src.wal_term_point_),
content_flags_(src.content_flags_.load(std::memory_order_relaxed)),
max_bytes_(src.max_bytes_),
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
rep_(src.rep_),
timestamp_size_(src.timestamp_size_) {
refactor SavePoints (#5192) Summary: Savepoints are assumed to be used in a stack-wise fashion (only the top element should be used), so they were stored by `WriteBatch` in a member variable `save_points` using an std::stack. Conceptually this is fine, but the implementation had a few issues: - the `save_points_` instance variable was a plain pointer to a heap- allocated `SavePoints` struct. The destructor of `WriteBatch` simply deletes this pointer. However, the copy constructor of WriteBatch just copied that pointer, meaning that copying a WriteBatch with active savepoints will very likely have crashed before. Now a proper copy of the savepoints is made in the copy constructor, and not just a copy of the pointer - `save_points_` was an std::stack, which defaults to `std::deque` for the underlying container. A deque is a bit over the top here, as we only need access to the most recent savepoint (i.e. stack.top()) but never any elements at the front. std::deque is rather expensive to initialize in common environments. For example, the STL implementation shipped with GNU g++ will perform a heap allocation of more than 500 bytes to create an empty deque object. Although the `save_points_` container is created lazily by RocksDB, moving from a deque to a plain `std::vector` is much more memory-efficient. So `save_points_` is now a vector. - `save_points_` was changed from a plain pointer to an `std::unique_ptr`, making ownership more explicit. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5192 Differential Revision: D15024074 Pulled By: maysamyabandeh fbshipit-source-id: 5b128786d3789cde94e46465c9e91badd07a25d7
2019-04-20 05:30:03 +02:00
if (src.save_points_ != nullptr) {
save_points_.reset(new SavePoints());
save_points_->stack = src.save_points_->stack;
}
}
WriteBatch::WriteBatch(WriteBatch&& src) noexcept
: save_points_(std::move(src.save_points_)),
wal_term_point_(std::move(src.wal_term_point_)),
content_flags_(src.content_flags_.load(std::memory_order_relaxed)),
max_bytes_(src.max_bytes_),
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
rep_(std::move(src.rep_)),
timestamp_size_(src.timestamp_size_) {}
WriteBatch& WriteBatch::operator=(const WriteBatch& src) {
if (&src != this) {
this->~WriteBatch();
new (this) WriteBatch(src);
}
return *this;
}
WriteBatch& WriteBatch::operator=(WriteBatch&& src) {
if (&src != this) {
this->~WriteBatch();
new (this) WriteBatch(std::move(src));
}
return *this;
}
refactor SavePoints (#5192) Summary: Savepoints are assumed to be used in a stack-wise fashion (only the top element should be used), so they were stored by `WriteBatch` in a member variable `save_points` using an std::stack. Conceptually this is fine, but the implementation had a few issues: - the `save_points_` instance variable was a plain pointer to a heap- allocated `SavePoints` struct. The destructor of `WriteBatch` simply deletes this pointer. However, the copy constructor of WriteBatch just copied that pointer, meaning that copying a WriteBatch with active savepoints will very likely have crashed before. Now a proper copy of the savepoints is made in the copy constructor, and not just a copy of the pointer - `save_points_` was an std::stack, which defaults to `std::deque` for the underlying container. A deque is a bit over the top here, as we only need access to the most recent savepoint (i.e. stack.top()) but never any elements at the front. std::deque is rather expensive to initialize in common environments. For example, the STL implementation shipped with GNU g++ will perform a heap allocation of more than 500 bytes to create an empty deque object. Although the `save_points_` container is created lazily by RocksDB, moving from a deque to a plain `std::vector` is much more memory-efficient. So `save_points_` is now a vector. - `save_points_` was changed from a plain pointer to an `std::unique_ptr`, making ownership more explicit. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5192 Differential Revision: D15024074 Pulled By: maysamyabandeh fbshipit-source-id: 5b128786d3789cde94e46465c9e91badd07a25d7
2019-04-20 05:30:03 +02:00
WriteBatch::~WriteBatch() { }
WriteBatch::Handler::~Handler() { }
void WriteBatch::Handler::LogData(const Slice& /*blob*/) {
// If the user has not specified something to do with blobs, then we ignore
// them.
}
bool WriteBatch::Handler::Continue() {
return true;
}
void WriteBatch::Clear() {
rep_.clear();
rep_.resize(WriteBatchInternal::kHeader);
content_flags_.store(0, std::memory_order_relaxed);
if (save_points_ != nullptr) {
while (!save_points_->stack.empty()) {
save_points_->stack.pop();
}
}
wal_term_point_.clear();
}
int WriteBatch::Count() const {
return WriteBatchInternal::Count(this);
}
uint32_t WriteBatch::ComputeContentFlags() const {
auto rv = content_flags_.load(std::memory_order_relaxed);
if ((rv & ContentFlags::DEFERRED) != 0) {
BatchContentClassifier classifier;
Iterate(&classifier);
rv = classifier.content_flags;
// this method is conceptually const, because it is performing a lazy
// computation that doesn't affect the abstract state of the batch.
// content_flags_ is marked mutable so that we can perform the
// following assignment
content_flags_.store(rv, std::memory_order_relaxed);
}
return rv;
}
void WriteBatch::MarkWalTerminationPoint() {
wal_term_point_.size = GetDataSize();
wal_term_point_.count = Count();
wal_term_point_.content_flags = content_flags_;
}
bool WriteBatch::HasPut() const {
return (ComputeContentFlags() & ContentFlags::HAS_PUT) != 0;
}
bool WriteBatch::HasDelete() const {
return (ComputeContentFlags() & ContentFlags::HAS_DELETE) != 0;
}
bool WriteBatch::HasSingleDelete() const {
return (ComputeContentFlags() & ContentFlags::HAS_SINGLE_DELETE) != 0;
}
bool WriteBatch::HasDeleteRange() const {
return (ComputeContentFlags() & ContentFlags::HAS_DELETE_RANGE) != 0;
}
bool WriteBatch::HasMerge() const {
return (ComputeContentFlags() & ContentFlags::HAS_MERGE) != 0;
}
bool ReadKeyFromWriteBatchEntry(Slice* input, Slice* key, bool cf_record) {
assert(input != nullptr && key != nullptr);
// Skip tag byte
input->remove_prefix(1);
if (cf_record) {
// Skip column_family bytes
uint32_t cf;
if (!GetVarint32(input, &cf)) {
return false;
}
}
// Extract key
return GetLengthPrefixedSlice(input, key);
}
bool WriteBatch::HasBeginPrepare() const {
return (ComputeContentFlags() & ContentFlags::HAS_BEGIN_PREPARE) != 0;
}
bool WriteBatch::HasEndPrepare() const {
return (ComputeContentFlags() & ContentFlags::HAS_END_PREPARE) != 0;
}
bool WriteBatch::HasCommit() const {
return (ComputeContentFlags() & ContentFlags::HAS_COMMIT) != 0;
}
bool WriteBatch::HasRollback() const {
return (ComputeContentFlags() & ContentFlags::HAS_ROLLBACK) != 0;
}
Status ReadRecordFromWriteBatch(Slice* input, char* tag,
uint32_t* column_family, Slice* key,
Slice* value, Slice* blob, Slice* xid) {
assert(key != nullptr && value != nullptr);
*tag = (*input)[0];
input->remove_prefix(1);
*column_family = 0; // default
switch (*tag) {
case kTypeColumnFamilyValue:
if (!GetVarint32(input, column_family)) {
return Status::Corruption("bad WriteBatch Put");
}
FALLTHROUGH_INTENDED;
case kTypeValue:
if (!GetLengthPrefixedSlice(input, key) ||
!GetLengthPrefixedSlice(input, value)) {
return Status::Corruption("bad WriteBatch Put");
}
break;
case kTypeColumnFamilyDeletion:
Support for SingleDelete() Summary: This patch fixes #7460559. It introduces SingleDelete as a new database operation. This operation can be used to delete keys that were never overwritten (no put following another put of the same key). If an overwritten key is single deleted the behavior is undefined. Single deletion of a non-existent key has no effect but multiple consecutive single deletions are not allowed (see limitations). In contrast to the conventional Delete() operation, the deletion entry is removed along with the value when the two are lined up in a compaction. Note: The semantics are similar to @igor's prototype that allowed to have this behavior on the granularity of a column family ( https://reviews.facebook.net/D42093 ). This new patch, however, is more aggressive when it comes to removing tombstones: It removes the SingleDelete together with the value whenever there is no snapshot between them while the older patch only did this when the sequence number of the deletion was older than the earliest snapshot. Most of the complex additions are in the Compaction Iterator, all other changes should be relatively straightforward. The patch also includes basic support for single deletions in db_stress and db_bench. Limitations: - Not compatible with cuckoo hash tables - Single deletions cannot be used in combination with merges and normal deletions on the same key (other keys are not affected by this) - Consecutive single deletions are currently not allowed (and older version of this patch supported this so it could be resurrected if needed) Test Plan: make all check Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor Reviewed By: igor Subscribers: maykov, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 20:42:56 +02:00
case kTypeColumnFamilySingleDeletion:
if (!GetVarint32(input, column_family)) {
return Status::Corruption("bad WriteBatch Delete");
}
FALLTHROUGH_INTENDED;
case kTypeDeletion:
Support for SingleDelete() Summary: This patch fixes #7460559. It introduces SingleDelete as a new database operation. This operation can be used to delete keys that were never overwritten (no put following another put of the same key). If an overwritten key is single deleted the behavior is undefined. Single deletion of a non-existent key has no effect but multiple consecutive single deletions are not allowed (see limitations). In contrast to the conventional Delete() operation, the deletion entry is removed along with the value when the two are lined up in a compaction. Note: The semantics are similar to @igor's prototype that allowed to have this behavior on the granularity of a column family ( https://reviews.facebook.net/D42093 ). This new patch, however, is more aggressive when it comes to removing tombstones: It removes the SingleDelete together with the value whenever there is no snapshot between them while the older patch only did this when the sequence number of the deletion was older than the earliest snapshot. Most of the complex additions are in the Compaction Iterator, all other changes should be relatively straightforward. The patch also includes basic support for single deletions in db_stress and db_bench. Limitations: - Not compatible with cuckoo hash tables - Single deletions cannot be used in combination with merges and normal deletions on the same key (other keys are not affected by this) - Consecutive single deletions are currently not allowed (and older version of this patch supported this so it could be resurrected if needed) Test Plan: make all check Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor Reviewed By: igor Subscribers: maykov, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 20:42:56 +02:00
case kTypeSingleDeletion:
if (!GetLengthPrefixedSlice(input, key)) {
return Status::Corruption("bad WriteBatch Delete");
}
break;
case kTypeColumnFamilyRangeDeletion:
if (!GetVarint32(input, column_family)) {
return Status::Corruption("bad WriteBatch DeleteRange");
}
FALLTHROUGH_INTENDED;
case kTypeRangeDeletion:
// for range delete, "key" is begin_key, "value" is end_key
if (!GetLengthPrefixedSlice(input, key) ||
!GetLengthPrefixedSlice(input, value)) {
return Status::Corruption("bad WriteBatch DeleteRange");
}
break;
case kTypeColumnFamilyMerge:
if (!GetVarint32(input, column_family)) {
return Status::Corruption("bad WriteBatch Merge");
}
FALLTHROUGH_INTENDED;
case kTypeMerge:
if (!GetLengthPrefixedSlice(input, key) ||
!GetLengthPrefixedSlice(input, value)) {
return Status::Corruption("bad WriteBatch Merge");
}
break;
case kTypeColumnFamilyBlobIndex:
if (!GetVarint32(input, column_family)) {
return Status::Corruption("bad WriteBatch BlobIndex");
}
FALLTHROUGH_INTENDED;
case kTypeBlobIndex:
if (!GetLengthPrefixedSlice(input, key) ||
!GetLengthPrefixedSlice(input, value)) {
return Status::Corruption("bad WriteBatch BlobIndex");
}
break;
case kTypeLogData:
assert(blob != nullptr);
if (!GetLengthPrefixedSlice(input, blob)) {
return Status::Corruption("bad WriteBatch Blob");
}
break;
case kTypeNoop:
case kTypeBeginPrepareXID:
// This indicates that the prepared batch is also persisted in the db.
// This is used in WritePreparedTxn
case kTypeBeginPersistedPrepareXID:
// This is used in WriteUnpreparedTxn
case kTypeBeginUnprepareXID:
break;
case kTypeEndPrepareXID:
if (!GetLengthPrefixedSlice(input, xid)) {
return Status::Corruption("bad EndPrepare XID");
}
break;
case kTypeCommitXID:
if (!GetLengthPrefixedSlice(input, xid)) {
return Status::Corruption("bad Commit XID");
}
break;
case kTypeRollbackXID:
if (!GetLengthPrefixedSlice(input, xid)) {
return Status::Corruption("bad Rollback XID");
}
break;
default:
return Status::Corruption("unknown WriteBatch tag");
}
return Status::OK();
}
Status WriteBatch::Iterate(Handler* handler) const {
if (rep_.size() < WriteBatchInternal::kHeader) {
return Status::Corruption("malformed WriteBatch (too small)");
}
return WriteBatchInternal::Iterate(this, handler, WriteBatchInternal::kHeader,
rep_.size());
}
Status WriteBatchInternal::Iterate(const WriteBatch* wb,
WriteBatch::Handler* handler, size_t begin,
size_t end) {
if (begin > wb->rep_.size() || end > wb->rep_.size() || end < begin) {
return Status::Corruption("Invalid start/end bounds for Iterate");
}
assert(begin <= end);
Slice input(wb->rep_.data() + begin, static_cast<size_t>(end - begin));
bool whole_batch =
(begin == WriteBatchInternal::kHeader) && (end == wb->rep_.size());
Slice key, value, blob, xid;
// Sometimes a sub-batch starts with a Noop. We want to exclude such Noops as
// the batch boundary symbols otherwise we would mis-count the number of
// batches. We do that by checking whether the accumulated batch is empty
// before seeing the next Noop.
bool empty_batch = true;
int found = 0;
Status s;
char tag = 0;
uint32_t column_family = 0; // default
bool last_was_try_again = false;
bool handler_continue = true;
while (((s.ok() && !input.empty()) || UNLIKELY(s.IsTryAgain()))) {
handler_continue = handler->Continue();
if (!handler_continue) {
break;
}
if (LIKELY(!s.IsTryAgain())) {
last_was_try_again = false;
tag = 0;
column_family = 0; // default
s = ReadRecordFromWriteBatch(&input, &tag, &column_family, &key, &value,
&blob, &xid);
if (!s.ok()) {
return s;
}
} else {
assert(s.IsTryAgain());
assert(!last_was_try_again); // to detect infinite loop bugs
if (UNLIKELY(last_was_try_again)) {
return Status::Corruption(
"two consecutive TryAgain in WriteBatch handler; this is either a "
"software bug or data corruption.");
}
last_was_try_again = true;
s = Status::OK();
}
switch (tag) {
case kTypeColumnFamilyValue:
case kTypeValue:
assert(wb->content_flags_.load(std::memory_order_relaxed) &
(ContentFlags::DEFERRED | ContentFlags::HAS_PUT));
s = handler->PutCF(column_family, key, value);
if (LIKELY(s.ok())) {
empty_batch = false;
found++;
}
break;
case kTypeColumnFamilyDeletion:
case kTypeDeletion:
assert(wb->content_flags_.load(std::memory_order_relaxed) &
(ContentFlags::DEFERRED | ContentFlags::HAS_DELETE));
s = handler->DeleteCF(column_family, key);
if (LIKELY(s.ok())) {
empty_batch = false;
found++;
}
break;
Support for SingleDelete() Summary: This patch fixes #7460559. It introduces SingleDelete as a new database operation. This operation can be used to delete keys that were never overwritten (no put following another put of the same key). If an overwritten key is single deleted the behavior is undefined. Single deletion of a non-existent key has no effect but multiple consecutive single deletions are not allowed (see limitations). In contrast to the conventional Delete() operation, the deletion entry is removed along with the value when the two are lined up in a compaction. Note: The semantics are similar to @igor's prototype that allowed to have this behavior on the granularity of a column family ( https://reviews.facebook.net/D42093 ). This new patch, however, is more aggressive when it comes to removing tombstones: It removes the SingleDelete together with the value whenever there is no snapshot between them while the older patch only did this when the sequence number of the deletion was older than the earliest snapshot. Most of the complex additions are in the Compaction Iterator, all other changes should be relatively straightforward. The patch also includes basic support for single deletions in db_stress and db_bench. Limitations: - Not compatible with cuckoo hash tables - Single deletions cannot be used in combination with merges and normal deletions on the same key (other keys are not affected by this) - Consecutive single deletions are currently not allowed (and older version of this patch supported this so it could be resurrected if needed) Test Plan: make all check Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor Reviewed By: igor Subscribers: maykov, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 20:42:56 +02:00
case kTypeColumnFamilySingleDeletion:
case kTypeSingleDeletion:
assert(wb->content_flags_.load(std::memory_order_relaxed) &
(ContentFlags::DEFERRED | ContentFlags::HAS_SINGLE_DELETE));
Support for SingleDelete() Summary: This patch fixes #7460559. It introduces SingleDelete as a new database operation. This operation can be used to delete keys that were never overwritten (no put following another put of the same key). If an overwritten key is single deleted the behavior is undefined. Single deletion of a non-existent key has no effect but multiple consecutive single deletions are not allowed (see limitations). In contrast to the conventional Delete() operation, the deletion entry is removed along with the value when the two are lined up in a compaction. Note: The semantics are similar to @igor's prototype that allowed to have this behavior on the granularity of a column family ( https://reviews.facebook.net/D42093 ). This new patch, however, is more aggressive when it comes to removing tombstones: It removes the SingleDelete together with the value whenever there is no snapshot between them while the older patch only did this when the sequence number of the deletion was older than the earliest snapshot. Most of the complex additions are in the Compaction Iterator, all other changes should be relatively straightforward. The patch also includes basic support for single deletions in db_stress and db_bench. Limitations: - Not compatible with cuckoo hash tables - Single deletions cannot be used in combination with merges and normal deletions on the same key (other keys are not affected by this) - Consecutive single deletions are currently not allowed (and older version of this patch supported this so it could be resurrected if needed) Test Plan: make all check Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor Reviewed By: igor Subscribers: maykov, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 20:42:56 +02:00
s = handler->SingleDeleteCF(column_family, key);
if (LIKELY(s.ok())) {
empty_batch = false;
found++;
}
Support for SingleDelete() Summary: This patch fixes #7460559. It introduces SingleDelete as a new database operation. This operation can be used to delete keys that were never overwritten (no put following another put of the same key). If an overwritten key is single deleted the behavior is undefined. Single deletion of a non-existent key has no effect but multiple consecutive single deletions are not allowed (see limitations). In contrast to the conventional Delete() operation, the deletion entry is removed along with the value when the two are lined up in a compaction. Note: The semantics are similar to @igor's prototype that allowed to have this behavior on the granularity of a column family ( https://reviews.facebook.net/D42093 ). This new patch, however, is more aggressive when it comes to removing tombstones: It removes the SingleDelete together with the value whenever there is no snapshot between them while the older patch only did this when the sequence number of the deletion was older than the earliest snapshot. Most of the complex additions are in the Compaction Iterator, all other changes should be relatively straightforward. The patch also includes basic support for single deletions in db_stress and db_bench. Limitations: - Not compatible with cuckoo hash tables - Single deletions cannot be used in combination with merges and normal deletions on the same key (other keys are not affected by this) - Consecutive single deletions are currently not allowed (and older version of this patch supported this so it could be resurrected if needed) Test Plan: make all check Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor Reviewed By: igor Subscribers: maykov, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 20:42:56 +02:00
break;
case kTypeColumnFamilyRangeDeletion:
case kTypeRangeDeletion:
assert(wb->content_flags_.load(std::memory_order_relaxed) &
(ContentFlags::DEFERRED | ContentFlags::HAS_DELETE_RANGE));
s = handler->DeleteRangeCF(column_family, key, value);
if (LIKELY(s.ok())) {
empty_batch = false;
found++;
}
break;
case kTypeColumnFamilyMerge:
case kTypeMerge:
assert(wb->content_flags_.load(std::memory_order_relaxed) &
(ContentFlags::DEFERRED | ContentFlags::HAS_MERGE));
s = handler->MergeCF(column_family, key, value);
if (LIKELY(s.ok())) {
empty_batch = false;
found++;
}
break;
case kTypeColumnFamilyBlobIndex:
case kTypeBlobIndex:
assert(wb->content_flags_.load(std::memory_order_relaxed) &
(ContentFlags::DEFERRED | ContentFlags::HAS_BLOB_INDEX));
s = handler->PutBlobIndexCF(column_family, key, value);
if (LIKELY(s.ok())) {
found++;
}
break;
case kTypeLogData:
handler->LogData(blob);
// A batch might have nothing but LogData. It is still a batch.
empty_batch = false;
break;
case kTypeBeginPrepareXID:
assert(wb->content_flags_.load(std::memory_order_relaxed) &
(ContentFlags::DEFERRED | ContentFlags::HAS_BEGIN_PREPARE));
handler->MarkBeginPrepare();
empty_batch = false;
if (!handler->WriteAfterCommit()) {
s = Status::NotSupported(
"WriteCommitted txn tag when write_after_commit_ is disabled (in "
"WritePrepared/WriteUnprepared mode). If it is not due to "
"corruption, the WAL must be emptied before changing the "
"WritePolicy.");
}
if (handler->WriteBeforePrepare()) {
s = Status::NotSupported(
"WriteCommitted txn tag when write_before_prepare_ is enabled "
"(in WriteUnprepared mode). If it is not due to corruption, the "
"WAL must be emptied before changing the WritePolicy.");
}
break;
case kTypeBeginPersistedPrepareXID:
assert(wb->content_flags_.load(std::memory_order_relaxed) &
(ContentFlags::DEFERRED | ContentFlags::HAS_BEGIN_PREPARE));
handler->MarkBeginPrepare();
empty_batch = false;
if (handler->WriteAfterCommit()) {
s = Status::NotSupported(
"WritePrepared/WriteUnprepared txn tag when write_after_commit_ "
"is enabled (in default WriteCommitted mode). If it is not due "
"to corruption, the WAL must be emptied before changing the "
"WritePolicy.");
}
break;
case kTypeBeginUnprepareXID:
assert(wb->content_flags_.load(std::memory_order_relaxed) &
(ContentFlags::DEFERRED | ContentFlags::HAS_BEGIN_UNPREPARE));
handler->MarkBeginPrepare(true /* unprepared */);
empty_batch = false;
if (handler->WriteAfterCommit()) {
s = Status::NotSupported(
"WriteUnprepared txn tag when write_after_commit_ is enabled (in "
"default WriteCommitted mode). If it is not due to corruption, "
"the WAL must be emptied before changing the WritePolicy.");
}
if (!handler->WriteBeforePrepare()) {
s = Status::NotSupported(
"WriteUnprepared txn tag when write_before_prepare_ is disabled "
"(in WriteCommitted/WritePrepared mode). If it is not due to "
"corruption, the WAL must be emptied before changing the "
"WritePolicy.");
}
break;
case kTypeEndPrepareXID:
assert(wb->content_flags_.load(std::memory_order_relaxed) &
(ContentFlags::DEFERRED | ContentFlags::HAS_END_PREPARE));
handler->MarkEndPrepare(xid);
empty_batch = true;
break;
case kTypeCommitXID:
assert(wb->content_flags_.load(std::memory_order_relaxed) &
(ContentFlags::DEFERRED | ContentFlags::HAS_COMMIT));
handler->MarkCommit(xid);
empty_batch = true;
break;
case kTypeRollbackXID:
assert(wb->content_flags_.load(std::memory_order_relaxed) &
(ContentFlags::DEFERRED | ContentFlags::HAS_ROLLBACK));
handler->MarkRollback(xid);
empty_batch = true;
break;
case kTypeNoop:
handler->MarkNoop(empty_batch);
empty_batch = true;
break;
default:
return Status::Corruption("unknown WriteBatch tag");
}
}
if (!s.ok()) {
return s;
}
if (handler_continue && whole_batch &&
found != WriteBatchInternal::Count(wb)) {
return Status::Corruption("WriteBatch has wrong count");
} else {
return Status::OK();
}
}
bool WriteBatchInternal::IsLatestPersistentState(const WriteBatch* b) {
return b->is_latest_persistent_state_;
}
void WriteBatchInternal::SetAsLastestPersistentState(WriteBatch* b) {
b->is_latest_persistent_state_ = true;
}
int WriteBatchInternal::Count(const WriteBatch* b) {
return DecodeFixed32(b->rep_.data() + 8);
}
void WriteBatchInternal::SetCount(WriteBatch* b, int n) {
EncodeFixed32(&b->rep_[8], n);
}
SequenceNumber WriteBatchInternal::Sequence(const WriteBatch* b) {
return SequenceNumber(DecodeFixed64(b->rep_.data()));
}
void WriteBatchInternal::SetSequence(WriteBatch* b, SequenceNumber seq) {
EncodeFixed64(&b->rep_[0], seq);
}
size_t WriteBatchInternal::GetFirstOffset(WriteBatch* /*b*/) {
return WriteBatchInternal::kHeader;
}
Status WriteBatchInternal::Put(WriteBatch* b, uint32_t column_family_id,
const Slice& key, const Slice& value) {
if (key.size() > size_t{port::kMaxUint32}) {
return Status::InvalidArgument("key is too large");
}
if (value.size() > size_t{port::kMaxUint32}) {
return Status::InvalidArgument("value is too large");
}
LocalSavePoint save(b);
WriteBatchInternal::SetCount(b, WriteBatchInternal::Count(b) + 1);
if (column_family_id == 0) {
b->rep_.push_back(static_cast<char>(kTypeValue));
} else {
b->rep_.push_back(static_cast<char>(kTypeColumnFamilyValue));
PutVarint32(&b->rep_, column_family_id);
}
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
if (0 == b->timestamp_size_) {
PutLengthPrefixedSlice(&b->rep_, key);
} else {
PutVarint32(&b->rep_,
static_cast<uint32_t>(key.size() + b->timestamp_size_));
b->rep_.append(key.data(), key.size());
b->rep_.append(b->timestamp_size_, '\0');
}
PutLengthPrefixedSlice(&b->rep_, value);
b->content_flags_.store(
b->content_flags_.load(std::memory_order_relaxed) | ContentFlags::HAS_PUT,
std::memory_order_relaxed);
return save.commit();
}
Status WriteBatch::Put(ColumnFamilyHandle* column_family, const Slice& key,
const Slice& value) {
return WriteBatchInternal::Put(this, GetColumnFamilyID(column_family), key,
value);
}
Status WriteBatchInternal::CheckSlicePartsLength(const SliceParts& key,
const SliceParts& value) {
size_t total_key_bytes = 0;
for (int i = 0; i < key.num_parts; ++i) {
total_key_bytes += key.parts[i].size();
}
if (total_key_bytes >= size_t{port::kMaxUint32}) {
return Status::InvalidArgument("key is too large");
}
size_t total_value_bytes = 0;
for (int i = 0; i < value.num_parts; ++i) {
total_value_bytes += value.parts[i].size();
}
if (total_value_bytes >= size_t{port::kMaxUint32}) {
return Status::InvalidArgument("value is too large");
}
return Status::OK();
}
Status WriteBatchInternal::Put(WriteBatch* b, uint32_t column_family_id,
const SliceParts& key, const SliceParts& value) {
Status s = CheckSlicePartsLength(key, value);
if (!s.ok()) {
return s;
}
LocalSavePoint save(b);
WriteBatchInternal::SetCount(b, WriteBatchInternal::Count(b) + 1);
if (column_family_id == 0) {
b->rep_.push_back(static_cast<char>(kTypeValue));
} else {
b->rep_.push_back(static_cast<char>(kTypeColumnFamilyValue));
PutVarint32(&b->rep_, column_family_id);
}
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
if (0 == b->timestamp_size_) {
PutLengthPrefixedSliceParts(&b->rep_, key);
} else {
PutLengthPrefixedSlicePartsWithPadding(&b->rep_, key, b->timestamp_size_);
}
PutLengthPrefixedSliceParts(&b->rep_, value);
b->content_flags_.store(
b->content_flags_.load(std::memory_order_relaxed) | ContentFlags::HAS_PUT,
std::memory_order_relaxed);
return save.commit();
}
Status WriteBatch::Put(ColumnFamilyHandle* column_family, const SliceParts& key,
const SliceParts& value) {
return WriteBatchInternal::Put(this, GetColumnFamilyID(column_family), key,
value);
}
Status WriteBatchInternal::InsertNoop(WriteBatch* b) {
b->rep_.push_back(static_cast<char>(kTypeNoop));
return Status::OK();
}
Status WriteBatchInternal::MarkEndPrepare(WriteBatch* b, const Slice& xid,
bool write_after_commit,
bool unprepared_batch) {
// a manually constructed batch can only contain one prepare section
assert(b->rep_[12] == static_cast<char>(kTypeNoop));
// all savepoints up to this point are cleared
if (b->save_points_ != nullptr) {
while (!b->save_points_->stack.empty()) {
b->save_points_->stack.pop();
}
}
// rewrite noop as begin marker
b->rep_[12] = static_cast<char>(
write_after_commit ? kTypeBeginPrepareXID
: (unprepared_batch ? kTypeBeginUnprepareXID
: kTypeBeginPersistedPrepareXID));
b->rep_.push_back(static_cast<char>(kTypeEndPrepareXID));
PutLengthPrefixedSlice(&b->rep_, xid);
b->content_flags_.store(b->content_flags_.load(std::memory_order_relaxed) |
ContentFlags::HAS_END_PREPARE |
ContentFlags::HAS_BEGIN_PREPARE,
std::memory_order_relaxed);
if (unprepared_batch) {
b->content_flags_.store(b->content_flags_.load(std::memory_order_relaxed) |
ContentFlags::HAS_BEGIN_UNPREPARE,
std::memory_order_relaxed);
}
return Status::OK();
}
Status WriteBatchInternal::MarkCommit(WriteBatch* b, const Slice& xid) {
b->rep_.push_back(static_cast<char>(kTypeCommitXID));
PutLengthPrefixedSlice(&b->rep_, xid);
b->content_flags_.store(b->content_flags_.load(std::memory_order_relaxed) |
ContentFlags::HAS_COMMIT,
std::memory_order_relaxed);
return Status::OK();
}
Status WriteBatchInternal::MarkRollback(WriteBatch* b, const Slice& xid) {
b->rep_.push_back(static_cast<char>(kTypeRollbackXID));
PutLengthPrefixedSlice(&b->rep_, xid);
b->content_flags_.store(b->content_flags_.load(std::memory_order_relaxed) |
ContentFlags::HAS_ROLLBACK,
std::memory_order_relaxed);
return Status::OK();
}
Status WriteBatchInternal::Delete(WriteBatch* b, uint32_t column_family_id,
const Slice& key) {
LocalSavePoint save(b);
WriteBatchInternal::SetCount(b, WriteBatchInternal::Count(b) + 1);
if (column_family_id == 0) {
b->rep_.push_back(static_cast<char>(kTypeDeletion));
} else {
b->rep_.push_back(static_cast<char>(kTypeColumnFamilyDeletion));
PutVarint32(&b->rep_, column_family_id);
}
PutLengthPrefixedSlice(&b->rep_, key);
b->content_flags_.store(b->content_flags_.load(std::memory_order_relaxed) |
ContentFlags::HAS_DELETE,
std::memory_order_relaxed);
return save.commit();
}
Status WriteBatch::Delete(ColumnFamilyHandle* column_family, const Slice& key) {
return WriteBatchInternal::Delete(this, GetColumnFamilyID(column_family),
key);
}
Status WriteBatchInternal::Delete(WriteBatch* b, uint32_t column_family_id,
const SliceParts& key) {
LocalSavePoint save(b);
WriteBatchInternal::SetCount(b, WriteBatchInternal::Count(b) + 1);
if (column_family_id == 0) {
b->rep_.push_back(static_cast<char>(kTypeDeletion));
} else {
b->rep_.push_back(static_cast<char>(kTypeColumnFamilyDeletion));
PutVarint32(&b->rep_, column_family_id);
}
PutLengthPrefixedSliceParts(&b->rep_, key);
b->content_flags_.store(b->content_flags_.load(std::memory_order_relaxed) |
ContentFlags::HAS_DELETE,
std::memory_order_relaxed);
return save.commit();
}
Status WriteBatch::Delete(ColumnFamilyHandle* column_family,
const SliceParts& key) {
return WriteBatchInternal::Delete(this, GetColumnFamilyID(column_family),
key);
}
Status WriteBatchInternal::SingleDelete(WriteBatch* b,
uint32_t column_family_id,
const Slice& key) {
LocalSavePoint save(b);
Support for SingleDelete() Summary: This patch fixes #7460559. It introduces SingleDelete as a new database operation. This operation can be used to delete keys that were never overwritten (no put following another put of the same key). If an overwritten key is single deleted the behavior is undefined. Single deletion of a non-existent key has no effect but multiple consecutive single deletions are not allowed (see limitations). In contrast to the conventional Delete() operation, the deletion entry is removed along with the value when the two are lined up in a compaction. Note: The semantics are similar to @igor's prototype that allowed to have this behavior on the granularity of a column family ( https://reviews.facebook.net/D42093 ). This new patch, however, is more aggressive when it comes to removing tombstones: It removes the SingleDelete together with the value whenever there is no snapshot between them while the older patch only did this when the sequence number of the deletion was older than the earliest snapshot. Most of the complex additions are in the Compaction Iterator, all other changes should be relatively straightforward. The patch also includes basic support for single deletions in db_stress and db_bench. Limitations: - Not compatible with cuckoo hash tables - Single deletions cannot be used in combination with merges and normal deletions on the same key (other keys are not affected by this) - Consecutive single deletions are currently not allowed (and older version of this patch supported this so it could be resurrected if needed) Test Plan: make all check Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor Reviewed By: igor Subscribers: maykov, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 20:42:56 +02:00
WriteBatchInternal::SetCount(b, WriteBatchInternal::Count(b) + 1);
if (column_family_id == 0) {
b->rep_.push_back(static_cast<char>(kTypeSingleDeletion));
} else {
b->rep_.push_back(static_cast<char>(kTypeColumnFamilySingleDeletion));
PutVarint32(&b->rep_, column_family_id);
}
PutLengthPrefixedSlice(&b->rep_, key);
b->content_flags_.store(b->content_flags_.load(std::memory_order_relaxed) |
ContentFlags::HAS_SINGLE_DELETE,
std::memory_order_relaxed);
return save.commit();
Support for SingleDelete() Summary: This patch fixes #7460559. It introduces SingleDelete as a new database operation. This operation can be used to delete keys that were never overwritten (no put following another put of the same key). If an overwritten key is single deleted the behavior is undefined. Single deletion of a non-existent key has no effect but multiple consecutive single deletions are not allowed (see limitations). In contrast to the conventional Delete() operation, the deletion entry is removed along with the value when the two are lined up in a compaction. Note: The semantics are similar to @igor's prototype that allowed to have this behavior on the granularity of a column family ( https://reviews.facebook.net/D42093 ). This new patch, however, is more aggressive when it comes to removing tombstones: It removes the SingleDelete together with the value whenever there is no snapshot between them while the older patch only did this when the sequence number of the deletion was older than the earliest snapshot. Most of the complex additions are in the Compaction Iterator, all other changes should be relatively straightforward. The patch also includes basic support for single deletions in db_stress and db_bench. Limitations: - Not compatible with cuckoo hash tables - Single deletions cannot be used in combination with merges and normal deletions on the same key (other keys are not affected by this) - Consecutive single deletions are currently not allowed (and older version of this patch supported this so it could be resurrected if needed) Test Plan: make all check Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor Reviewed By: igor Subscribers: maykov, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 20:42:56 +02:00
}
Status WriteBatch::SingleDelete(ColumnFamilyHandle* column_family,
const Slice& key) {
return WriteBatchInternal::SingleDelete(
this, GetColumnFamilyID(column_family), key);
Support for SingleDelete() Summary: This patch fixes #7460559. It introduces SingleDelete as a new database operation. This operation can be used to delete keys that were never overwritten (no put following another put of the same key). If an overwritten key is single deleted the behavior is undefined. Single deletion of a non-existent key has no effect but multiple consecutive single deletions are not allowed (see limitations). In contrast to the conventional Delete() operation, the deletion entry is removed along with the value when the two are lined up in a compaction. Note: The semantics are similar to @igor's prototype that allowed to have this behavior on the granularity of a column family ( https://reviews.facebook.net/D42093 ). This new patch, however, is more aggressive when it comes to removing tombstones: It removes the SingleDelete together with the value whenever there is no snapshot between them while the older patch only did this when the sequence number of the deletion was older than the earliest snapshot. Most of the complex additions are in the Compaction Iterator, all other changes should be relatively straightforward. The patch also includes basic support for single deletions in db_stress and db_bench. Limitations: - Not compatible with cuckoo hash tables - Single deletions cannot be used in combination with merges and normal deletions on the same key (other keys are not affected by this) - Consecutive single deletions are currently not allowed (and older version of this patch supported this so it could be resurrected if needed) Test Plan: make all check Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor Reviewed By: igor Subscribers: maykov, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 20:42:56 +02:00
}
Status WriteBatchInternal::SingleDelete(WriteBatch* b,
uint32_t column_family_id,
const SliceParts& key) {
LocalSavePoint save(b);
Support for SingleDelete() Summary: This patch fixes #7460559. It introduces SingleDelete as a new database operation. This operation can be used to delete keys that were never overwritten (no put following another put of the same key). If an overwritten key is single deleted the behavior is undefined. Single deletion of a non-existent key has no effect but multiple consecutive single deletions are not allowed (see limitations). In contrast to the conventional Delete() operation, the deletion entry is removed along with the value when the two are lined up in a compaction. Note: The semantics are similar to @igor's prototype that allowed to have this behavior on the granularity of a column family ( https://reviews.facebook.net/D42093 ). This new patch, however, is more aggressive when it comes to removing tombstones: It removes the SingleDelete together with the value whenever there is no snapshot between them while the older patch only did this when the sequence number of the deletion was older than the earliest snapshot. Most of the complex additions are in the Compaction Iterator, all other changes should be relatively straightforward. The patch also includes basic support for single deletions in db_stress and db_bench. Limitations: - Not compatible with cuckoo hash tables - Single deletions cannot be used in combination with merges and normal deletions on the same key (other keys are not affected by this) - Consecutive single deletions are currently not allowed (and older version of this patch supported this so it could be resurrected if needed) Test Plan: make all check Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor Reviewed By: igor Subscribers: maykov, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 20:42:56 +02:00
WriteBatchInternal::SetCount(b, WriteBatchInternal::Count(b) + 1);
if (column_family_id == 0) {
b->rep_.push_back(static_cast<char>(kTypeSingleDeletion));
} else {
b->rep_.push_back(static_cast<char>(kTypeColumnFamilySingleDeletion));
PutVarint32(&b->rep_, column_family_id);
}
PutLengthPrefixedSliceParts(&b->rep_, key);
b->content_flags_.store(b->content_flags_.load(std::memory_order_relaxed) |
ContentFlags::HAS_SINGLE_DELETE,
std::memory_order_relaxed);
return save.commit();
Support for SingleDelete() Summary: This patch fixes #7460559. It introduces SingleDelete as a new database operation. This operation can be used to delete keys that were never overwritten (no put following another put of the same key). If an overwritten key is single deleted the behavior is undefined. Single deletion of a non-existent key has no effect but multiple consecutive single deletions are not allowed (see limitations). In contrast to the conventional Delete() operation, the deletion entry is removed along with the value when the two are lined up in a compaction. Note: The semantics are similar to @igor's prototype that allowed to have this behavior on the granularity of a column family ( https://reviews.facebook.net/D42093 ). This new patch, however, is more aggressive when it comes to removing tombstones: It removes the SingleDelete together with the value whenever there is no snapshot between them while the older patch only did this when the sequence number of the deletion was older than the earliest snapshot. Most of the complex additions are in the Compaction Iterator, all other changes should be relatively straightforward. The patch also includes basic support for single deletions in db_stress and db_bench. Limitations: - Not compatible with cuckoo hash tables - Single deletions cannot be used in combination with merges and normal deletions on the same key (other keys are not affected by this) - Consecutive single deletions are currently not allowed (and older version of this patch supported this so it could be resurrected if needed) Test Plan: make all check Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor Reviewed By: igor Subscribers: maykov, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 20:42:56 +02:00
}
Status WriteBatch::SingleDelete(ColumnFamilyHandle* column_family,
const SliceParts& key) {
return WriteBatchInternal::SingleDelete(
this, GetColumnFamilyID(column_family), key);
Support for SingleDelete() Summary: This patch fixes #7460559. It introduces SingleDelete as a new database operation. This operation can be used to delete keys that were never overwritten (no put following another put of the same key). If an overwritten key is single deleted the behavior is undefined. Single deletion of a non-existent key has no effect but multiple consecutive single deletions are not allowed (see limitations). In contrast to the conventional Delete() operation, the deletion entry is removed along with the value when the two are lined up in a compaction. Note: The semantics are similar to @igor's prototype that allowed to have this behavior on the granularity of a column family ( https://reviews.facebook.net/D42093 ). This new patch, however, is more aggressive when it comes to removing tombstones: It removes the SingleDelete together with the value whenever there is no snapshot between them while the older patch only did this when the sequence number of the deletion was older than the earliest snapshot. Most of the complex additions are in the Compaction Iterator, all other changes should be relatively straightforward. The patch also includes basic support for single deletions in db_stress and db_bench. Limitations: - Not compatible with cuckoo hash tables - Single deletions cannot be used in combination with merges and normal deletions on the same key (other keys are not affected by this) - Consecutive single deletions are currently not allowed (and older version of this patch supported this so it could be resurrected if needed) Test Plan: make all check Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor Reviewed By: igor Subscribers: maykov, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 20:42:56 +02:00
}
Status WriteBatchInternal::DeleteRange(WriteBatch* b, uint32_t column_family_id,
const Slice& begin_key,
const Slice& end_key) {
LocalSavePoint save(b);
WriteBatchInternal::SetCount(b, WriteBatchInternal::Count(b) + 1);
if (column_family_id == 0) {
b->rep_.push_back(static_cast<char>(kTypeRangeDeletion));
} else {
b->rep_.push_back(static_cast<char>(kTypeColumnFamilyRangeDeletion));
PutVarint32(&b->rep_, column_family_id);
}
PutLengthPrefixedSlice(&b->rep_, begin_key);
PutLengthPrefixedSlice(&b->rep_, end_key);
b->content_flags_.store(b->content_flags_.load(std::memory_order_relaxed) |
ContentFlags::HAS_DELETE_RANGE,
std::memory_order_relaxed);
return save.commit();
}
Status WriteBatch::DeleteRange(ColumnFamilyHandle* column_family,
const Slice& begin_key, const Slice& end_key) {
return WriteBatchInternal::DeleteRange(this, GetColumnFamilyID(column_family),
begin_key, end_key);
}
Status WriteBatchInternal::DeleteRange(WriteBatch* b, uint32_t column_family_id,
const SliceParts& begin_key,
const SliceParts& end_key) {
LocalSavePoint save(b);
WriteBatchInternal::SetCount(b, WriteBatchInternal::Count(b) + 1);
if (column_family_id == 0) {
b->rep_.push_back(static_cast<char>(kTypeRangeDeletion));
} else {
b->rep_.push_back(static_cast<char>(kTypeColumnFamilyRangeDeletion));
PutVarint32(&b->rep_, column_family_id);
}
PutLengthPrefixedSliceParts(&b->rep_, begin_key);
PutLengthPrefixedSliceParts(&b->rep_, end_key);
b->content_flags_.store(b->content_flags_.load(std::memory_order_relaxed) |
ContentFlags::HAS_DELETE_RANGE,
std::memory_order_relaxed);
return save.commit();
}
Status WriteBatch::DeleteRange(ColumnFamilyHandle* column_family,
const SliceParts& begin_key,
const SliceParts& end_key) {
return WriteBatchInternal::DeleteRange(this, GetColumnFamilyID(column_family),
begin_key, end_key);
}
Status WriteBatchInternal::Merge(WriteBatch* b, uint32_t column_family_id,
const Slice& key, const Slice& value) {
if (key.size() > size_t{port::kMaxUint32}) {
return Status::InvalidArgument("key is too large");
}
if (value.size() > size_t{port::kMaxUint32}) {
return Status::InvalidArgument("value is too large");
}
LocalSavePoint save(b);
WriteBatchInternal::SetCount(b, WriteBatchInternal::Count(b) + 1);
if (column_family_id == 0) {
b->rep_.push_back(static_cast<char>(kTypeMerge));
} else {
b->rep_.push_back(static_cast<char>(kTypeColumnFamilyMerge));
PutVarint32(&b->rep_, column_family_id);
}
PutLengthPrefixedSlice(&b->rep_, key);
PutLengthPrefixedSlice(&b->rep_, value);
b->content_flags_.store(b->content_flags_.load(std::memory_order_relaxed) |
ContentFlags::HAS_MERGE,
std::memory_order_relaxed);
return save.commit();
}
Status WriteBatch::Merge(ColumnFamilyHandle* column_family, const Slice& key,
const Slice& value) {
return WriteBatchInternal::Merge(this, GetColumnFamilyID(column_family), key,
value);
}
Status WriteBatchInternal::Merge(WriteBatch* b, uint32_t column_family_id,
const SliceParts& key,
const SliceParts& value) {
Status s = CheckSlicePartsLength(key, value);
if (!s.ok()) {
return s;
}
LocalSavePoint save(b);
WriteBatchInternal::SetCount(b, WriteBatchInternal::Count(b) + 1);
if (column_family_id == 0) {
b->rep_.push_back(static_cast<char>(kTypeMerge));
} else {
b->rep_.push_back(static_cast<char>(kTypeColumnFamilyMerge));
PutVarint32(&b->rep_, column_family_id);
}
PutLengthPrefixedSliceParts(&b->rep_, key);
PutLengthPrefixedSliceParts(&b->rep_, value);
b->content_flags_.store(b->content_flags_.load(std::memory_order_relaxed) |
ContentFlags::HAS_MERGE,
std::memory_order_relaxed);
return save.commit();
}
Status WriteBatch::Merge(ColumnFamilyHandle* column_family,
const SliceParts& key, const SliceParts& value) {
return WriteBatchInternal::Merge(this, GetColumnFamilyID(column_family), key,
value);
}
Status WriteBatchInternal::PutBlobIndex(WriteBatch* b,
uint32_t column_family_id,
const Slice& key, const Slice& value) {
LocalSavePoint save(b);
WriteBatchInternal::SetCount(b, WriteBatchInternal::Count(b) + 1);
if (column_family_id == 0) {
b->rep_.push_back(static_cast<char>(kTypeBlobIndex));
} else {
b->rep_.push_back(static_cast<char>(kTypeColumnFamilyBlobIndex));
PutVarint32(&b->rep_, column_family_id);
}
PutLengthPrefixedSlice(&b->rep_, key);
PutLengthPrefixedSlice(&b->rep_, value);
b->content_flags_.store(b->content_flags_.load(std::memory_order_relaxed) |
ContentFlags::HAS_BLOB_INDEX,
std::memory_order_relaxed);
return save.commit();
}
Status WriteBatch::PutLogData(const Slice& blob) {
LocalSavePoint save(this);
rep_.push_back(static_cast<char>(kTypeLogData));
PutLengthPrefixedSlice(&rep_, blob);
return save.commit();
}
void WriteBatch::SetSavePoint() {
if (save_points_ == nullptr) {
refactor SavePoints (#5192) Summary: Savepoints are assumed to be used in a stack-wise fashion (only the top element should be used), so they were stored by `WriteBatch` in a member variable `save_points` using an std::stack. Conceptually this is fine, but the implementation had a few issues: - the `save_points_` instance variable was a plain pointer to a heap- allocated `SavePoints` struct. The destructor of `WriteBatch` simply deletes this pointer. However, the copy constructor of WriteBatch just copied that pointer, meaning that copying a WriteBatch with active savepoints will very likely have crashed before. Now a proper copy of the savepoints is made in the copy constructor, and not just a copy of the pointer - `save_points_` was an std::stack, which defaults to `std::deque` for the underlying container. A deque is a bit over the top here, as we only need access to the most recent savepoint (i.e. stack.top()) but never any elements at the front. std::deque is rather expensive to initialize in common environments. For example, the STL implementation shipped with GNU g++ will perform a heap allocation of more than 500 bytes to create an empty deque object. Although the `save_points_` container is created lazily by RocksDB, moving from a deque to a plain `std::vector` is much more memory-efficient. So `save_points_` is now a vector. - `save_points_` was changed from a plain pointer to an `std::unique_ptr`, making ownership more explicit. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5192 Differential Revision: D15024074 Pulled By: maysamyabandeh fbshipit-source-id: 5b128786d3789cde94e46465c9e91badd07a25d7
2019-04-20 05:30:03 +02:00
save_points_.reset(new SavePoints());
}
// Record length and count of current batch of writes.
save_points_->stack.push(SavePoint(
GetDataSize(), Count(), content_flags_.load(std::memory_order_relaxed)));
}
Status WriteBatch::RollbackToSavePoint() {
if (save_points_ == nullptr || save_points_->stack.size() == 0) {
return Status::NotFound();
}
// Pop the most recent savepoint off the stack
SavePoint savepoint = save_points_->stack.top();
save_points_->stack.pop();
assert(savepoint.size <= rep_.size());
assert(savepoint.count <= Count());
if (savepoint.size == rep_.size()) {
// No changes to rollback
} else if (savepoint.size == 0) {
// Rollback everything
Clear();
} else {
rep_.resize(savepoint.size);
WriteBatchInternal::SetCount(this, savepoint.count);
content_flags_.store(savepoint.content_flags, std::memory_order_relaxed);
}
return Status::OK();
}
Status WriteBatch::PopSavePoint() {
if (save_points_ == nullptr || save_points_->stack.size() == 0) {
return Status::NotFound();
}
// Pop the most recent savepoint off the stack
save_points_->stack.pop();
return Status::OK();
}
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
Status WriteBatch::AssignTimestamp(const Slice& ts) {
TimestampAssigner ts_assigner(ts);
return Iterate(&ts_assigner);
}
Status WriteBatch::AssignTimestamps(const std::vector<Slice>& ts_list) {
TimestampAssigner ts_assigner(ts_list);
return Iterate(&ts_assigner);
}
class MemTableInserter : public WriteBatch::Handler {
SequenceNumber sequence_;
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
2015-08-15 01:59:07 +02:00
ColumnFamilyMemTables* const cf_mems_;
FlushScheduler* const 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
TrimHistoryScheduler* const trim_history_scheduler_;
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
2015-08-15 01:59:07 +02:00
const bool ignore_missing_column_families_;
const uint64_t recovering_log_number_;
// log number that all Memtables inserted into should reference
uint64_t log_number_ref_;
DBImpl* db_;
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
2015-08-15 01:59:07 +02:00
const bool concurrent_memtable_writes_;
bool post_info_created_;
bool* has_valid_writes_;
// On some (!) platforms just default creating
// a map is too expensive in the Write() path as they
// cause memory allocations though unused.
// Make creation optional but do not incur
// std::unique_ptr additional allocation
using MemPostInfoMap = std::map<MemTable*, MemTablePostProcessInfo>;
using PostMapType = std::aligned_storage<sizeof(MemPostInfoMap)>::type;
PostMapType mem_post_info_map_;
// current recovered transaction we are rebuilding (recovery)
WriteBatch* rebuilding_trx_;
SequenceNumber rebuilding_trx_seq_;
// Increase seq number once per each write batch. Otherwise increase it once
// per key.
bool seq_per_batch_;
// Whether the memtable write will be done only after the commit
bool write_after_commit_;
// Whether memtable write can be done before prepare
bool write_before_prepare_;
// Whether this batch was unprepared or not
bool unprepared_batch_;
using DupDetector = std::aligned_storage<sizeof(DuplicateDetector)>::type;
DupDetector duplicate_detector_;
bool dup_dectector_on_;
MemPostInfoMap& GetPostMap() {
assert(concurrent_memtable_writes_);
if(!post_info_created_) {
new (&mem_post_info_map_) MemPostInfoMap();
post_info_created_ = true;
}
return *reinterpret_cast<MemPostInfoMap*>(&mem_post_info_map_);
}
bool IsDuplicateKeySeq(uint32_t column_family_id, const Slice& key) {
assert(!write_after_commit_);
assert(rebuilding_trx_ != nullptr);
if (!dup_dectector_on_) {
new (&duplicate_detector_) DuplicateDetector(db_);
dup_dectector_on_ = true;
}
return reinterpret_cast<DuplicateDetector*>
(&duplicate_detector_)->IsDuplicateKeySeq(column_family_id, key, sequence_);
}
protected:
bool WriteBeforePrepare() const override { return write_before_prepare_; }
bool WriteAfterCommit() const override { return write_after_commit_; }
public:
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
2015-08-15 01:59:07 +02:00
// cf_mems should not be shared with concurrent inserters
MemTableInserter(SequenceNumber _sequence, ColumnFamilyMemTables* cf_mems,
FlushScheduler* 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
TrimHistoryScheduler* trim_history_scheduler,
bool ignore_missing_column_families,
uint64_t recovering_log_number, DB* db,
bool concurrent_memtable_writes,
bool* has_valid_writes = nullptr, bool seq_per_batch = false,
bool batch_per_txn = true)
: sequence_(_sequence),
cf_mems_(cf_mems),
flush_scheduler_(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_(trim_history_scheduler),
ignore_missing_column_families_(ignore_missing_column_families),
recovering_log_number_(recovering_log_number),
log_number_ref_(0),
db_(static_cast_with_check<DBImpl, DB>(db)),
concurrent_memtable_writes_(concurrent_memtable_writes),
post_info_created_(false),
has_valid_writes_(has_valid_writes),
rebuilding_trx_(nullptr),
rebuilding_trx_seq_(0),
seq_per_batch_(seq_per_batch),
// Write after commit currently uses one seq per key (instead of per
// batch). So seq_per_batch being false indicates write_after_commit
// approach.
write_after_commit_(!seq_per_batch),
// WriteUnprepared can write WriteBatches per transaction, so
// batch_per_txn being false indicates write_before_prepare.
write_before_prepare_(!batch_per_txn),
unprepared_batch_(false),
duplicate_detector_(),
dup_dectector_on_(false) {
assert(cf_mems_);
}
~MemTableInserter() override {
if (dup_dectector_on_) {
reinterpret_cast<DuplicateDetector*>
(&duplicate_detector_)->~DuplicateDetector();
}
if (post_info_created_) {
reinterpret_cast<MemPostInfoMap*>
(&mem_post_info_map_)->~MemPostInfoMap();
}
delete rebuilding_trx_;
}
MemTableInserter(const MemTableInserter&) = delete;
MemTableInserter& operator=(const MemTableInserter&) = delete;
// The batch seq is regularly restarted; In normal mode it is set when
// MemTableInserter is constructed in the write thread and in recovery mode it
// is set when a batch, which is tagged with seq, is read from the WAL.
// Within a sequenced batch, which could be a merge of multiple batches, we
// have two policies to advance the seq: i) seq_per_key (default) and ii)
// seq_per_batch. To implement the latter we need to mark the boundary between
// the individual batches. The approach is this: 1) Use the terminating
// markers to indicate the boundary (kTypeEndPrepareXID, kTypeCommitXID,
// kTypeRollbackXID) 2) Terminate a batch with kTypeNoop in the absence of a
// natural boundary marker.
void MaybeAdvanceSeq(bool batch_boundry = false) {
if (batch_boundry == seq_per_batch_) {
sequence_++;
}
}
void set_log_number_ref(uint64_t log) { log_number_ref_ = log; }
SequenceNumber sequence() const { return sequence_; }
void PostProcess() {
assert(concurrent_memtable_writes_);
// If post info was not created there is nothing
// to process and no need to create on demand
if(post_info_created_) {
for (auto& pair : GetPostMap()) {
pair.first->BatchPostProcess(pair.second);
}
}
}
bool SeekToColumnFamily(uint32_t column_family_id, Status* s) {
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
2015-08-15 01:59:07 +02:00
// If we are in a concurrent mode, it is the caller's responsibility
// to clone the original ColumnFamilyMemTables so that each thread
// has its own instance. Otherwise, it must be guaranteed that there
// is no concurrent access
bool found = cf_mems_->Seek(column_family_id);
if (!found) {
if (ignore_missing_column_families_) {
*s = Status::OK();
} else {
*s = Status::InvalidArgument(
"Invalid column family specified in write batch");
}
return false;
}
if (recovering_log_number_ != 0 &&
recovering_log_number_ < cf_mems_->GetLogNumber()) {
// This is true only in recovery environment (recovering_log_number_ is
// always 0 in
// non-recovery, regular write code-path)
// * If recovering_log_number_ < cf_mems_->GetLogNumber(), this means that
// column
// family already contains updates from this log. We can't apply updates
// twice because of update-in-place or merge workloads -- ignore the
// update
*s = Status::OK();
return false;
}
if (has_valid_writes_ != nullptr) {
*has_valid_writes_ = true;
}
if (log_number_ref_ > 0) {
cf_mems_->GetMemTable()->RefLogContainingPrepSection(log_number_ref_);
}
return true;
}
Status PutCFImpl(uint32_t column_family_id, const Slice& key,
const Slice& value, ValueType value_type) {
// optimize for non-recovery mode
if (UNLIKELY(write_after_commit_ && rebuilding_trx_ != nullptr)) {
WriteBatchInternal::Put(rebuilding_trx_, column_family_id, key, value);
return Status::OK();
// else insert the values to the memtable right away
}
Status seek_status;
if (UNLIKELY(!SeekToColumnFamily(column_family_id, &seek_status))) {
bool batch_boundry = false;
if (rebuilding_trx_ != nullptr) {
assert(!write_after_commit_);
// The CF is probably flushed and hence no need for insert but we still
// need to keep track of the keys for upcoming rollback/commit.
WriteBatchInternal::Put(rebuilding_trx_, column_family_id, key, value);
batch_boundry = IsDuplicateKeySeq(column_family_id, key);
}
MaybeAdvanceSeq(batch_boundry);
return seek_status;
}
Status ret_status;
MemTable* mem = cf_mems_->GetMemTable();
auto* moptions = mem->GetImmutableMemTableOptions();
// inplace_update_support is inconsistent with snapshots, and therefore with
// any kind of transactions including the ones that use seq_per_batch
assert(!seq_per_batch_ || !moptions->inplace_update_support);
if (!moptions->inplace_update_support) {
bool mem_res =
mem->Add(sequence_, value_type, key, value,
concurrent_memtable_writes_, get_post_process_info(mem));
if (UNLIKELY(!mem_res)) {
assert(seq_per_batch_);
ret_status = Status::TryAgain("key+seq exists");
const bool BATCH_BOUNDRY = true;
MaybeAdvanceSeq(BATCH_BOUNDRY);
}
} else if (moptions->inplace_callback == nullptr) {
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
2015-08-15 01:59:07 +02:00
assert(!concurrent_memtable_writes_);
mem->Update(sequence_, key, value);
} else {
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
2015-08-15 01:59:07 +02:00
assert(!concurrent_memtable_writes_);
if (mem->UpdateCallback(sequence_, key, value)) {
} else {
// key not found in memtable. Do sst get, update, add
SnapshotImpl read_from_snapshot;
read_from_snapshot.number_ = sequence_;
ReadOptions ropts;
// it's going to be overwritten for sure, so no point caching data block
// containing the old version
ropts.fill_cache = false;
ropts.snapshot = &read_from_snapshot;
std::string prev_value;
std::string merged_value;
auto cf_handle = cf_mems_->GetColumnFamilyHandle();
Status s = Status::NotSupported();
if (db_ != nullptr && recovering_log_number_ == 0) {
if (cf_handle == nullptr) {
cf_handle = db_->DefaultColumnFamily();
}
s = db_->Get(ropts, cf_handle, key, &prev_value);
}
char* prev_buffer = const_cast<char*>(prev_value.c_str());
uint32_t prev_size = static_cast<uint32_t>(prev_value.size());
auto status = moptions->inplace_callback(s.ok() ? prev_buffer : nullptr,
s.ok() ? &prev_size : nullptr,
value, &merged_value);
if (status == UpdateStatus::UPDATED_INPLACE) {
// prev_value is updated in-place with final value.
bool mem_res __attribute__((__unused__));
mem_res = mem->Add(
sequence_, value_type, key, Slice(prev_buffer, prev_size));
assert(mem_res);
RecordTick(moptions->statistics, NUMBER_KEYS_WRITTEN);
} else if (status == UpdateStatus::UPDATED) {
// merged_value contains the final value.
bool mem_res __attribute__((__unused__));
mem_res =
mem->Add(sequence_, value_type, key, Slice(merged_value));
assert(mem_res);
RecordTick(moptions->statistics, NUMBER_KEYS_WRITTEN);
}
}
}
// optimize for non-recovery mode
if (UNLIKELY(!ret_status.IsTryAgain() && rebuilding_trx_ != nullptr)) {
assert(!write_after_commit_);
// If the ret_status is TryAgain then let the next try to add the ky to
// the rebuilding transaction object.
WriteBatchInternal::Put(rebuilding_trx_, column_family_id, key, value);
}
// Since all Puts are logged in transaction logs (if enabled), always bump
// sequence number. Even if the update eventually fails and does not result
// in memtable add/update.
MaybeAdvanceSeq();
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
2015-08-15 01:59:07 +02:00
CheckMemtableFull();
return ret_status;
}
Status PutCF(uint32_t column_family_id, const Slice& key,
const Slice& value) override {
return PutCFImpl(column_family_id, key, value, kTypeValue);
}
Status DeleteImpl(uint32_t /*column_family_id*/, const Slice& key,
const Slice& value, ValueType delete_type) {
Status ret_status;
Support for SingleDelete() Summary: This patch fixes #7460559. It introduces SingleDelete as a new database operation. This operation can be used to delete keys that were never overwritten (no put following another put of the same key). If an overwritten key is single deleted the behavior is undefined. Single deletion of a non-existent key has no effect but multiple consecutive single deletions are not allowed (see limitations). In contrast to the conventional Delete() operation, the deletion entry is removed along with the value when the two are lined up in a compaction. Note: The semantics are similar to @igor's prototype that allowed to have this behavior on the granularity of a column family ( https://reviews.facebook.net/D42093 ). This new patch, however, is more aggressive when it comes to removing tombstones: It removes the SingleDelete together with the value whenever there is no snapshot between them while the older patch only did this when the sequence number of the deletion was older than the earliest snapshot. Most of the complex additions are in the Compaction Iterator, all other changes should be relatively straightforward. The patch also includes basic support for single deletions in db_stress and db_bench. Limitations: - Not compatible with cuckoo hash tables - Single deletions cannot be used in combination with merges and normal deletions on the same key (other keys are not affected by this) - Consecutive single deletions are currently not allowed (and older version of this patch supported this so it could be resurrected if needed) Test Plan: make all check Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor Reviewed By: igor Subscribers: maykov, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 20:42:56 +02:00
MemTable* mem = cf_mems_->GetMemTable();
bool mem_res =
mem->Add(sequence_, delete_type, key, value,
concurrent_memtable_writes_, get_post_process_info(mem));
if (UNLIKELY(!mem_res)) {
assert(seq_per_batch_);
ret_status = Status::TryAgain("key+seq exists");
const bool BATCH_BOUNDRY = true;
MaybeAdvanceSeq(BATCH_BOUNDRY);
}
MaybeAdvanceSeq();
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
2015-08-15 01:59:07 +02:00
CheckMemtableFull();
return ret_status;
Support for SingleDelete() Summary: This patch fixes #7460559. It introduces SingleDelete as a new database operation. This operation can be used to delete keys that were never overwritten (no put following another put of the same key). If an overwritten key is single deleted the behavior is undefined. Single deletion of a non-existent key has no effect but multiple consecutive single deletions are not allowed (see limitations). In contrast to the conventional Delete() operation, the deletion entry is removed along with the value when the two are lined up in a compaction. Note: The semantics are similar to @igor's prototype that allowed to have this behavior on the granularity of a column family ( https://reviews.facebook.net/D42093 ). This new patch, however, is more aggressive when it comes to removing tombstones: It removes the SingleDelete together with the value whenever there is no snapshot between them while the older patch only did this when the sequence number of the deletion was older than the earliest snapshot. Most of the complex additions are in the Compaction Iterator, all other changes should be relatively straightforward. The patch also includes basic support for single deletions in db_stress and db_bench. Limitations: - Not compatible with cuckoo hash tables - Single deletions cannot be used in combination with merges and normal deletions on the same key (other keys are not affected by this) - Consecutive single deletions are currently not allowed (and older version of this patch supported this so it could be resurrected if needed) Test Plan: make all check Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor Reviewed By: igor Subscribers: maykov, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 20:42:56 +02:00
}
Status DeleteCF(uint32_t column_family_id, const Slice& key) override {
// optimize for non-recovery mode
if (UNLIKELY(write_after_commit_ && rebuilding_trx_ != nullptr)) {
WriteBatchInternal::Delete(rebuilding_trx_, column_family_id, key);
return Status::OK();
// else insert the values to the memtable right away
}
Status seek_status;
if (UNLIKELY(!SeekToColumnFamily(column_family_id, &seek_status))) {
bool batch_boundry = false;
if (rebuilding_trx_ != nullptr) {
assert(!write_after_commit_);
// The CF is probably flushed and hence no need for insert but we still
// need to keep track of the keys for upcoming rollback/commit.
WriteBatchInternal::Delete(rebuilding_trx_, column_family_id, key);
batch_boundry = IsDuplicateKeySeq(column_family_id, key);
}
MaybeAdvanceSeq(batch_boundry);
return seek_status;
}
auto ret_status = DeleteImpl(column_family_id, key, Slice(), kTypeDeletion);
// optimize for non-recovery mode
if (UNLIKELY(!ret_status.IsTryAgain() && rebuilding_trx_ != nullptr)) {
assert(!write_after_commit_);
// If the ret_status is TryAgain then let the next try to add the ky to
// the rebuilding transaction object.
WriteBatchInternal::Delete(rebuilding_trx_, column_family_id, key);
}
return ret_status;
}
Status SingleDeleteCF(uint32_t column_family_id, const Slice& key) override {
// optimize for non-recovery mode
if (UNLIKELY(write_after_commit_ && rebuilding_trx_ != nullptr)) {
WriteBatchInternal::SingleDelete(rebuilding_trx_, column_family_id, key);
return Status::OK();
// else insert the values to the memtable right away
}
Status seek_status;
if (UNLIKELY(!SeekToColumnFamily(column_family_id, &seek_status))) {
bool batch_boundry = false;
if (rebuilding_trx_ != nullptr) {
assert(!write_after_commit_);
// The CF is probably flushed and hence no need for insert but we still
// need to keep track of the keys for upcoming rollback/commit.
WriteBatchInternal::SingleDelete(rebuilding_trx_, column_family_id,
key);
batch_boundry = IsDuplicateKeySeq(column_family_id, key);
}
MaybeAdvanceSeq(batch_boundry);
return seek_status;
}
auto ret_status =
DeleteImpl(column_family_id, key, Slice(), kTypeSingleDeletion);
// optimize for non-recovery mode
if (UNLIKELY(!ret_status.IsTryAgain() && rebuilding_trx_ != nullptr)) {
assert(!write_after_commit_);
// If the ret_status is TryAgain then let the next try to add the ky to
// the rebuilding transaction object.
WriteBatchInternal::SingleDelete(rebuilding_trx_, column_family_id, key);
}
return ret_status;
}
Status DeleteRangeCF(uint32_t column_family_id, const Slice& begin_key,
const Slice& end_key) override {
// optimize for non-recovery mode
if (UNLIKELY(write_after_commit_ && rebuilding_trx_ != nullptr)) {
WriteBatchInternal::DeleteRange(rebuilding_trx_, column_family_id,
begin_key, end_key);
return Status::OK();
// else insert the values to the memtable right away
}
Status seek_status;
if (UNLIKELY(!SeekToColumnFamily(column_family_id, &seek_status))) {
bool batch_boundry = false;
if (rebuilding_trx_ != nullptr) {
assert(!write_after_commit_);
// The CF is probably flushed and hence no need for insert but we still
// need to keep track of the keys for upcoming rollback/commit.
WriteBatchInternal::DeleteRange(rebuilding_trx_, column_family_id,
begin_key, end_key);
// TODO(myabandeh): when transactional DeleteRange support is added,
// check if end_key must also be added.
batch_boundry = IsDuplicateKeySeq(column_family_id, begin_key);
}
MaybeAdvanceSeq(batch_boundry);
return seek_status;
}
if (db_ != nullptr) {
auto cf_handle = cf_mems_->GetColumnFamilyHandle();
if (cf_handle == nullptr) {
cf_handle = db_->DefaultColumnFamily();
}
auto* cfd = reinterpret_cast<ColumnFamilyHandleImpl*>(cf_handle)->cfd();
if (!cfd->is_delete_range_supported()) {
return Status::NotSupported(
std::string("DeleteRange not supported for table type ") +
cfd->ioptions()->table_factory->Name() + " in CF " +
cfd->GetName());
}
}
auto ret_status =
DeleteImpl(column_family_id, begin_key, end_key, kTypeRangeDeletion);
// optimize for non-recovery mode
if (UNLIKELY(!ret_status.IsTryAgain() && rebuilding_trx_ != nullptr)) {
assert(!write_after_commit_);
// If the ret_status is TryAgain then let the next try to add the ky to
// the rebuilding transaction object.
WriteBatchInternal::DeleteRange(rebuilding_trx_, column_family_id,
begin_key, end_key);
}
return ret_status;
Support for SingleDelete() Summary: This patch fixes #7460559. It introduces SingleDelete as a new database operation. This operation can be used to delete keys that were never overwritten (no put following another put of the same key). If an overwritten key is single deleted the behavior is undefined. Single deletion of a non-existent key has no effect but multiple consecutive single deletions are not allowed (see limitations). In contrast to the conventional Delete() operation, the deletion entry is removed along with the value when the two are lined up in a compaction. Note: The semantics are similar to @igor's prototype that allowed to have this behavior on the granularity of a column family ( https://reviews.facebook.net/D42093 ). This new patch, however, is more aggressive when it comes to removing tombstones: It removes the SingleDelete together with the value whenever there is no snapshot between them while the older patch only did this when the sequence number of the deletion was older than the earliest snapshot. Most of the complex additions are in the Compaction Iterator, all other changes should be relatively straightforward. The patch also includes basic support for single deletions in db_stress and db_bench. Limitations: - Not compatible with cuckoo hash tables - Single deletions cannot be used in combination with merges and normal deletions on the same key (other keys are not affected by this) - Consecutive single deletions are currently not allowed (and older version of this patch supported this so it could be resurrected if needed) Test Plan: make all check Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor Reviewed By: igor Subscribers: maykov, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 20:42:56 +02:00
}
Status MergeCF(uint32_t column_family_id, const Slice& key,
const Slice& value) override {
// optimize for non-recovery mode
if (UNLIKELY(write_after_commit_ && rebuilding_trx_ != nullptr)) {
WriteBatchInternal::Merge(rebuilding_trx_, column_family_id, key, value);
return Status::OK();
// else insert the values to the memtable right away
}
Status seek_status;
if (UNLIKELY(!SeekToColumnFamily(column_family_id, &seek_status))) {
bool batch_boundry = false;
if (rebuilding_trx_ != nullptr) {
assert(!write_after_commit_);
// The CF is probably flushed and hence no need for insert but we still
// need to keep track of the keys for upcoming rollback/commit.
WriteBatchInternal::Merge(rebuilding_trx_, column_family_id, key,
value);
batch_boundry = IsDuplicateKeySeq(column_family_id, key);
}
MaybeAdvanceSeq(batch_boundry);
return seek_status;
}
Status ret_status;
MemTable* mem = cf_mems_->GetMemTable();
auto* moptions = mem->GetImmutableMemTableOptions();
bool perform_merge = false;
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(!concurrent_memtable_writes_ ||
moptions->max_successive_merges == 0);
// If we pass DB through and options.max_successive_merges is hit
// during recovery, Get() will be issued which will try to acquire
// DB mutex and cause deadlock, as DB mutex is already held.
// So we disable merge in recovery
if (moptions->max_successive_merges > 0 && db_ != nullptr &&
recovering_log_number_ == 0) {
Unordered Writes (#5218) Summary: Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees. The patch is prepared based on an original by siying Existing unit tests are extended to include unordered_write option. Benchmark Results: ``` TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1 ``` With WAL - Vanilla RocksDB: 78.6 MB/s - WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x) - unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees) Without WAL - Vanilla RocksDB: 111.3 MB/s - WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x) - unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees) - WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x) Limitations: - The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218 Differential Revision: D15219029 Pulled By: maysamyabandeh fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
2019-05-14 02:43:47 +02:00
assert(!concurrent_memtable_writes_);
LookupKey lkey(key, sequence_);
// Count the number of successive merges at the head
// of the key in the memtable
size_t num_merges = mem->CountSuccessiveMergeEntries(lkey);
if (num_merges >= moptions->max_successive_merges) {
perform_merge = true;
}
}
if (perform_merge) {
// 1) Get the existing value
std::string get_value;
// Pass in the sequence number so that we also include previous merge
// operations in the same batch.
SnapshotImpl read_from_snapshot;
read_from_snapshot.number_ = sequence_;
ReadOptions read_options;
read_options.snapshot = &read_from_snapshot;
auto cf_handle = cf_mems_->GetColumnFamilyHandle();
if (cf_handle == nullptr) {
cf_handle = db_->DefaultColumnFamily();
}
db_->Get(read_options, cf_handle, key, &get_value);
Slice get_value_slice = Slice(get_value);
// 2) Apply this merge
auto merge_operator = moptions->merge_operator;
assert(merge_operator);
std::string new_value;
Status merge_status = MergeHelper::TimedFullMerge(
Introduce FullMergeV2 (eliminate memcpy from merge operators) Summary: This diff update the code to pin the merge operator operands while the merge operation is done, so that we can eliminate the memcpy cost, to do that we need a new public API for FullMerge that replace the std::deque<std::string> with std::vector<Slice> This diff is stacked on top of D56493 and D56511 In this diff we - Update FullMergeV2 arguments to be encapsulated in MergeOperationInput and MergeOperationOutput which will make it easier to add new arguments in the future - Replace std::deque<std::string> with std::vector<Slice> to pass operands - Replace MergeContext std::deque with std::vector (based on a simple benchmark I ran https://gist.github.com/IslamAbdelRahman/78fc86c9ab9f52b1df791e58943fb187) - Allow FullMergeV2 output to be an existing operand ``` [Everything in Memtable | 10K operands | 10 KB each | 1 operand per key] DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=10000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000 [FullMergeV2] readseq : 0.607 micros/op 1648235 ops/sec; 16121.2 MB/s readseq : 0.478 micros/op 2091546 ops/sec; 20457.2 MB/s readseq : 0.252 micros/op 3972081 ops/sec; 38850.5 MB/s readseq : 0.237 micros/op 4218328 ops/sec; 41259.0 MB/s readseq : 0.247 micros/op 4043927 ops/sec; 39553.2 MB/s [master] readseq : 3.935 micros/op 254140 ops/sec; 2485.7 MB/s readseq : 3.722 micros/op 268657 ops/sec; 2627.7 MB/s readseq : 3.149 micros/op 317605 ops/sec; 3106.5 MB/s readseq : 3.125 micros/op 320024 ops/sec; 3130.1 MB/s readseq : 4.075 micros/op 245374 ops/sec; 2400.0 MB/s ``` ``` [Everything in Memtable | 10K operands | 10 KB each | 10 operand per key] DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=1000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000 [FullMergeV2] readseq : 3.472 micros/op 288018 ops/sec; 2817.1 MB/s readseq : 2.304 micros/op 434027 ops/sec; 4245.2 MB/s readseq : 1.163 micros/op 859845 ops/sec; 8410.0 MB/s readseq : 1.192 micros/op 838926 ops/sec; 8205.4 MB/s readseq : 1.250 micros/op 800000 ops/sec; 7824.7 MB/s [master] readseq : 24.025 micros/op 41623 ops/sec; 407.1 MB/s readseq : 18.489 micros/op 54086 ops/sec; 529.0 MB/s readseq : 18.693 micros/op 53495 ops/sec; 523.2 MB/s readseq : 23.621 micros/op 42335 ops/sec; 414.1 MB/s readseq : 18.775 micros/op 53262 ops/sec; 521.0 MB/s ``` ``` [Everything in Block cache | 10K operands | 10 KB each | 1 operand per key] [FullMergeV2] $ DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions readseq : 14.741 micros/op 67837 ops/sec; 663.5 MB/s readseq : 1.029 micros/op 971446 ops/sec; 9501.6 MB/s readseq : 0.974 micros/op 1026229 ops/sec; 10037.4 MB/s readseq : 0.965 micros/op 1036080 ops/sec; 10133.8 MB/s readseq : 0.943 micros/op 1060657 ops/sec; 10374.2 MB/s [master] readseq : 16.735 micros/op 59755 ops/sec; 584.5 MB/s readseq : 3.029 micros/op 330151 ops/sec; 3229.2 MB/s readseq : 3.136 micros/op 318883 ops/sec; 3119.0 MB/s readseq : 3.065 micros/op 326245 ops/sec; 3191.0 MB/s readseq : 3.014 micros/op 331813 ops/sec; 3245.4 MB/s ``` ``` [Everything in Block cache | 10K operands | 10 KB each | 10 operand per key] DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10-operands-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions [FullMergeV2] readseq : 24.325 micros/op 41109 ops/sec; 402.1 MB/s readseq : 1.470 micros/op 680272 ops/sec; 6653.7 MB/s readseq : 1.231 micros/op 812347 ops/sec; 7945.5 MB/s readseq : 1.091 micros/op 916590 ops/sec; 8965.1 MB/s readseq : 1.109 micros/op 901713 ops/sec; 8819.6 MB/s [master] readseq : 27.257 micros/op 36687 ops/sec; 358.8 MB/s readseq : 4.443 micros/op 225073 ops/sec; 2201.4 MB/s readseq : 5.830 micros/op 171526 ops/sec; 1677.7 MB/s readseq : 4.173 micros/op 239635 ops/sec; 2343.8 MB/s readseq : 4.150 micros/op 240963 ops/sec; 2356.8 MB/s ``` Test Plan: COMPILE_WITH_ASAN=1 make check -j64 Reviewers: yhchiang, andrewkr, sdong Reviewed By: sdong Subscribers: lovro, andrewkr, dhruba Differential Revision: https://reviews.facebook.net/D57075
2016-07-20 18:49:03 +02:00
merge_operator, key, &get_value_slice, {value}, &new_value,
moptions->info_log, moptions->statistics, Env::Default());
if (!merge_status.ok()) {
// Failed to merge!
// Store the delta in memtable
perform_merge = false;
} else {
// 3) Add value to 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
assert(!concurrent_memtable_writes_);
bool mem_res = mem->Add(sequence_, kTypeValue, key, new_value);
if (UNLIKELY(!mem_res)) {
assert(seq_per_batch_);
ret_status = Status::TryAgain("key+seq exists");
const bool BATCH_BOUNDRY = true;
MaybeAdvanceSeq(BATCH_BOUNDRY);
}
}
}
if (!perform_merge) {
// Add merge operator to 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
bool mem_res =
mem->Add(sequence_, kTypeMerge, key, value,
concurrent_memtable_writes_, get_post_process_info(mem));
if (UNLIKELY(!mem_res)) {
assert(seq_per_batch_);
ret_status = Status::TryAgain("key+seq exists");
const bool BATCH_BOUNDRY = true;
MaybeAdvanceSeq(BATCH_BOUNDRY);
}
}
// optimize for non-recovery mode
if (UNLIKELY(!ret_status.IsTryAgain() && rebuilding_trx_ != nullptr)) {
assert(!write_after_commit_);
// If the ret_status is TryAgain then let the next try to add the ky to
// the rebuilding transaction object.
WriteBatchInternal::Merge(rebuilding_trx_, column_family_id, key, value);
}
MaybeAdvanceSeq();
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
2015-08-15 01:59:07 +02:00
CheckMemtableFull();
return ret_status;
}
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
2015-08-15 01:59:07 +02:00
Status PutBlobIndexCF(uint32_t column_family_id, const Slice& key,
const Slice& value) override {
// Same as PutCF except for value type.
return PutCFImpl(column_family_id, key, value, kTypeBlobIndex);
}
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
2015-08-15 01:59:07 +02:00
void CheckMemtableFull() {
if (flush_scheduler_ != nullptr) {
auto* cfd = cf_mems_->current();
assert(cfd != nullptr);
if (cfd->mem()->ShouldScheduleFlush() &&
cfd->mem()->MarkFlushScheduled()) {
// MarkFlushScheduled only returns true if we are the one that
// should take action, so no need to dedup further
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_->ScheduleWork(cfd);
}
}
// check if memtable_list size exceeds max_write_buffer_size_to_maintain
if (trim_history_scheduler_ != nullptr) {
auto* cfd = cf_mems_->current();
assert(cfd != nullptr);
if (cfd->ioptions()->max_write_buffer_size_to_maintain > 0 &&
cfd->mem()->ApproximateMemoryUsageFast() +
cfd->imm()->ApproximateMemoryUsageExcludingLast() >=
static_cast<size_t>(
cfd->ioptions()->max_write_buffer_size_to_maintain) &&
cfd->imm()->MarkTrimHistoryNeeded()) {
trim_history_scheduler_->ScheduleWork(cfd);
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
2015-08-15 01:59:07 +02:00
}
}
}
// The write batch handler calls MarkBeginPrepare with unprepare set to true
// if it encounters the kTypeBeginUnprepareXID marker.
Status MarkBeginPrepare(bool unprepare) override {
assert(rebuilding_trx_ == nullptr);
assert(db_);
if (recovering_log_number_ != 0) {
// during recovery we rebuild a hollow transaction
// from all encountered prepare sections of the wal
if (db_->allow_2pc() == false) {
return Status::NotSupported(
"WAL contains prepared transactions. Open with "
"TransactionDB::Open().");
}
// we are now iterating through a prepared section
rebuilding_trx_ = new WriteBatch();
rebuilding_trx_seq_ = sequence_;
// We only call MarkBeginPrepare once per batch, and unprepared_batch_
// is initialized to false by default.
assert(!unprepared_batch_);
unprepared_batch_ = unprepare;
if (has_valid_writes_ != nullptr) {
*has_valid_writes_ = true;
}
}
return Status::OK();
}
Status MarkEndPrepare(const Slice& name) override {
assert(db_);
assert((rebuilding_trx_ != nullptr) == (recovering_log_number_ != 0));
if (recovering_log_number_ != 0) {
assert(db_->allow_2pc());
size_t batch_cnt =
write_after_commit_
? 0 // 0 will disable further checks
: static_cast<size_t>(sequence_ - rebuilding_trx_seq_ + 1);
db_->InsertRecoveredTransaction(recovering_log_number_, name.ToString(),
rebuilding_trx_, rebuilding_trx_seq_,
batch_cnt, unprepared_batch_);
rebuilding_trx_ = nullptr;
} else {
assert(rebuilding_trx_ == nullptr);
}
const bool batch_boundry = true;
MaybeAdvanceSeq(batch_boundry);
return Status::OK();
}
Status MarkNoop(bool empty_batch) override {
// A hack in pessimistic transaction could result into a noop at the start
// of the write batch, that should be ignored.
if (!empty_batch) {
// In the absence of Prepare markers, a kTypeNoop tag indicates the end of
// a batch. This happens when write batch commits skipping the prepare
// phase.
const bool batch_boundry = true;
MaybeAdvanceSeq(batch_boundry);
}
return Status::OK();
}
Status MarkCommit(const Slice& name) override {
assert(db_);
Status s;
if (recovering_log_number_ != 0) {
// in recovery when we encounter a commit marker
// we lookup this transaction in our set of rebuilt transactions
// and commit.
auto trx = db_->GetRecoveredTransaction(name.ToString());
// the log containing the prepared section may have
// been released in the last incarnation because the
// data was flushed to L0
if (trx != nullptr) {
// at this point individual CF lognumbers will prevent
// duplicate re-insertion of values.
assert(log_number_ref_ == 0);
if (write_after_commit_) {
// write_after_commit_ can only have one batch in trx.
assert(trx->batches_.size() == 1);
const auto& batch_info = trx->batches_.begin()->second;
// all inserts must reference this trx log number
log_number_ref_ = batch_info.log_number_;
s = batch_info.batch_->Iterate(this);
log_number_ref_ = 0;
}
// else the values are already inserted before the commit
if (s.ok()) {
db_->DeleteRecoveredTransaction(name.ToString());
}
if (has_valid_writes_ != nullptr) {
*has_valid_writes_ = true;
}
}
} else {
// When writes are not delayed until commit, there is no disconnect
// between a memtable write and the WAL that supports it. So the commit
// need not reference any log as the only log to which it depends.
assert(!write_after_commit_ || log_number_ref_ > 0);
}
const bool batch_boundry = true;
MaybeAdvanceSeq(batch_boundry);
return s;
}
Status MarkRollback(const Slice& name) override {
assert(db_);
if (recovering_log_number_ != 0) {
auto trx = db_->GetRecoveredTransaction(name.ToString());
// the log containing the transactions prep section
// may have been released in the previous incarnation
// because we knew it had been rolled back
if (trx != nullptr) {
db_->DeleteRecoveredTransaction(name.ToString());
}
} else {
// in non recovery we simply ignore this tag
}
const bool batch_boundry = true;
MaybeAdvanceSeq(batch_boundry);
return Status::OK();
}
private:
MemTablePostProcessInfo* get_post_process_info(MemTable* mem) {
if (!concurrent_memtable_writes_) {
// No need to batch counters locally if we don't use concurrent mode.
return nullptr;
}
return &GetPostMap()[mem];
}
};
2015-01-06 21:44:21 +01:00
// This function can only be called in these conditions:
// 1) During Recovery()
// 2) During Write(), in a single-threaded write thread
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
2015-08-15 01:59:07 +02:00
// 3) During Write(), in a concurrent context where memtables has been cloned
// The reason is that it calls memtables->Seek(), which has a stateful cache
Status WriteBatchInternal::InsertInto(
WriteThread::WriteGroup& write_group, SequenceNumber sequence,
ColumnFamilyMemTables* memtables, FlushScheduler* 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
TrimHistoryScheduler* trim_history_scheduler,
bool ignore_missing_column_families, uint64_t recovery_log_number, DB* db,
bool concurrent_memtable_writes, bool seq_per_batch, bool batch_per_txn) {
MemTableInserter inserter(
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
sequence, memtables, flush_scheduler, trim_history_scheduler,
ignore_missing_column_families, recovery_log_number, db,
concurrent_memtable_writes, nullptr /*has_valid_writes*/, seq_per_batch,
batch_per_txn);
for (auto w : write_group) {
if (w->CallbackFailed()) {
continue;
}
w->sequence = inserter.sequence();
if (!w->ShouldWriteToMemtable()) {
// In seq_per_batch_ mode this advances the seq by one.
inserter.MaybeAdvanceSeq(true);
continue;
}
SetSequence(w->batch, inserter.sequence());
inserter.set_log_number_ref(w->log_ref);
w->status = w->batch->Iterate(&inserter);
if (!w->status.ok()) {
return w->status;
}
assert(!seq_per_batch || w->batch_cnt != 0);
assert(!seq_per_batch || inserter.sequence() - w->sequence == w->batch_cnt);
}
return Status::OK();
}
Status WriteBatchInternal::InsertInto(
WriteThread::Writer* writer, SequenceNumber sequence,
ColumnFamilyMemTables* memtables, FlushScheduler* 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
TrimHistoryScheduler* trim_history_scheduler,
bool ignore_missing_column_families, uint64_t log_number, DB* db,
bool concurrent_memtable_writes, bool seq_per_batch, size_t batch_cnt,
bool batch_per_txn) {
#ifdef NDEBUG
(void)batch_cnt;
#endif
assert(writer->ShouldWriteToMemtable());
MemTableInserter inserter(
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
sequence, memtables, flush_scheduler, trim_history_scheduler,
ignore_missing_column_families, log_number, db,
concurrent_memtable_writes, nullptr /*has_valid_writes*/, seq_per_batch,
batch_per_txn);
SetSequence(writer->batch, sequence);
inserter.set_log_number_ref(writer->log_ref);
Status s = writer->batch->Iterate(&inserter);
assert(!seq_per_batch || batch_cnt != 0);
assert(!seq_per_batch || inserter.sequence() - sequence == batch_cnt);
if (concurrent_memtable_writes) {
inserter.PostProcess();
}
return s;
}
Status WriteBatchInternal::InsertInto(
const WriteBatch* batch, ColumnFamilyMemTables* memtables,
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
FlushScheduler* flush_scheduler,
TrimHistoryScheduler* trim_history_scheduler,
bool ignore_missing_column_families, uint64_t log_number, DB* db,
bool concurrent_memtable_writes, SequenceNumber* next_seq,
bool* has_valid_writes, bool seq_per_batch, bool batch_per_txn) {
MemTableInserter inserter(Sequence(batch), 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,
ignore_missing_column_families, log_number, db,
concurrent_memtable_writes, has_valid_writes,
seq_per_batch, batch_per_txn);
Status s = batch->Iterate(&inserter);
if (next_seq != nullptr) {
*next_seq = inserter.sequence();
}
if (concurrent_memtable_writes) {
inserter.PostProcess();
}
return s;
}
Status WriteBatchInternal::SetContents(WriteBatch* b, const Slice& contents) {
assert(contents.size() >= WriteBatchInternal::kHeader);
b->rep_.assign(contents.data(), contents.size());
b->content_flags_.store(ContentFlags::DEFERRED, std::memory_order_relaxed);
return Status::OK();
}
Status WriteBatchInternal::Append(WriteBatch* dst, const WriteBatch* src,
const bool wal_only) {
size_t src_len;
int src_count;
uint32_t src_flags;
const SavePoint& batch_end = src->GetWalTerminationPoint();
if (wal_only && !batch_end.is_cleared()) {
src_len = batch_end.size - WriteBatchInternal::kHeader;
src_count = batch_end.count;
src_flags = batch_end.content_flags;
} else {
src_len = src->rep_.size() - WriteBatchInternal::kHeader;
src_count = Count(src);
src_flags = src->content_flags_.load(std::memory_order_relaxed);
}
SetCount(dst, Count(dst) + src_count);
assert(src->rep_.size() >= WriteBatchInternal::kHeader);
dst->rep_.append(src->rep_.data() + WriteBatchInternal::kHeader, src_len);
dst->content_flags_.store(
dst->content_flags_.load(std::memory_order_relaxed) | src_flags,
std::memory_order_relaxed);
return Status::OK();
}
size_t WriteBatchInternal::AppendedByteSize(size_t leftByteSize,
size_t rightByteSize) {
if (leftByteSize == 0 || rightByteSize == 0) {
return leftByteSize + rightByteSize;
} else {
return leftByteSize + rightByteSize - WriteBatchInternal::kHeader;
}
}
} // namespace rocksdb