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

840 lines
36 KiB

// Copyright (c) Meta Platforms, Inc. and affiliates.
//
// This source code is licensed under both the GPLv2 (found in the
// COPYING file in the root directory) and Apache 2.0 License
// (found in the LICENSE.Apache file in the root directory).
//
// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file. See the AUTHORS file for names of contributors.
#include "db/compaction/compaction_outputs.h"
#include "db/builder.h"
namespace ROCKSDB_NAMESPACE {
void CompactionOutputs::NewBuilder(const TableBuilderOptions& tboptions) {
builder_.reset(NewTableBuilder(tboptions, file_writer_.get()));
}
Status CompactionOutputs::Finish(const Status& intput_status,
const SeqnoToTimeMapping& seqno_time_mapping) {
FileMetaData* meta = GetMetaData();
assert(meta != nullptr);
Status s = intput_status;
if (s.ok()) {
std::string seqno_time_mapping_str;
seqno_time_mapping.Encode(seqno_time_mapping_str, meta->fd.smallest_seqno,
meta->fd.largest_seqno, meta->file_creation_time);
builder_->SetSeqnoTimeTableProperties(seqno_time_mapping_str,
meta->oldest_ancester_time);
s = builder_->Finish();
} else {
builder_->Abandon();
}
Status io_s = builder_->io_status();
if (s.ok()) {
s = io_s;
} else {
io_s.PermitUncheckedError();
}
const uint64_t current_bytes = builder_->FileSize();
if (s.ok()) {
meta->fd.file_size = current_bytes;
meta->marked_for_compaction = builder_->NeedCompact();
}
current_output().finished = true;
stats_.bytes_written += current_bytes;
stats_.num_output_files = outputs_.size();
return s;
}
IOStatus CompactionOutputs::WriterSyncClose(const Status& input_status,
SystemClock* clock,
Statistics* statistics,
bool use_fsync) {
IOStatus io_s;
if (input_status.ok()) {
StopWatch sw(clock, statistics, COMPACTION_OUTFILE_SYNC_MICROS);
io_s = file_writer_->Sync(use_fsync);
}
if (input_status.ok() && io_s.ok()) {
io_s = file_writer_->Close();
}
if (input_status.ok() && io_s.ok()) {
FileMetaData* meta = GetMetaData();
meta->file_checksum = file_writer_->GetFileChecksum();
meta->file_checksum_func_name = file_writer_->GetFileChecksumFuncName();
}
file_writer_.reset();
return io_s;
}
bool CompactionOutputs::UpdateFilesToCutForTTLStates(
const Slice& internal_key) {
if (!files_to_cut_for_ttl_.empty()) {
const InternalKeyComparator* icmp =
&compaction_->column_family_data()->internal_comparator();
if (cur_files_to_cut_for_ttl_ != -1) {
// Previous key is inside the range of a file
if (icmp->Compare(internal_key,
files_to_cut_for_ttl_[cur_files_to_cut_for_ttl_]
->largest.Encode()) > 0) {
next_files_to_cut_for_ttl_ = cur_files_to_cut_for_ttl_ + 1;
cur_files_to_cut_for_ttl_ = -1;
return true;
}
} else {
// Look for the key position
while (next_files_to_cut_for_ttl_ <
static_cast<int>(files_to_cut_for_ttl_.size())) {
if (icmp->Compare(internal_key,
files_to_cut_for_ttl_[next_files_to_cut_for_ttl_]
->smallest.Encode()) >= 0) {
if (icmp->Compare(internal_key,
files_to_cut_for_ttl_[next_files_to_cut_for_ttl_]
->largest.Encode()) <= 0) {
// With in the current file
cur_files_to_cut_for_ttl_ = next_files_to_cut_for_ttl_;
return true;
}
// Beyond the current file
next_files_to_cut_for_ttl_++;
} else {
// Still fall into the gap
break;
}
}
}
}
return false;
}
Align compaction output file boundaries to the next level ones (#10655) Summary: Try to align the compaction output file boundaries to the next level ones (grandparent level), to reduce the level compaction write-amplification. In level compaction, there are "wasted" data at the beginning and end of the output level files. Align the file boundary can avoid such "wasted" compaction. With this PR, it tries to align the non-bottommost level file boundaries to its next level ones. It may cut file when the file size is large enough (at least 50% of target_file_size) and not too large (2x target_file_size). db_bench shows about 12.56% compaction reduction: ``` TEST_TMPDIR=/data/dbbench2 ./db_bench --benchmarks=fillrandom,readrandom -max_background_jobs=12 -num=400000000 -target_file_size_base=33554432 # baseline: Flush(GB): cumulative 25.882, interval 7.216 Cumulative compaction: 285.90 GB write, 162.36 MB/s write, 269.68 GB read, 153.15 MB/s read, 2926.7 seconds # with this change: Flush(GB): cumulative 25.882, interval 7.753 Cumulative compaction: 249.97 GB write, 141.96 MB/s write, 233.74 GB read, 132.74 MB/s read, 2534.9 seconds ``` The compaction simulator shows a similar result (14% with 100G random data). As a side effect, with this PR, the SST file size can exceed the target_file_size, but is capped at 2x target_file_size. And there will be smaller files. Here are file size statistics when loading 100GB with the target file size 32MB: ``` baseline this_PR count 1.656000e+03 1.705000e+03 mean 3.116062e+07 3.028076e+07 std 7.145242e+06 8.046139e+06 ``` The feature is enabled by default, to revert to the old behavior disable it with `AdvancedColumnFamilyOptions.level_compaction_dynamic_file_size = false` Also includes https://github.com/facebook/rocksdb/issues/1963 to cut file before skippable grandparent file. Which is for use case like user adding 2 or more non-overlapping data range at the same time, it can reduce the overlapping of 2 datasets in the lower levels. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10655 Reviewed By: cbi42 Differential Revision: D39552321 Pulled By: jay-zhuang fbshipit-source-id: 640d15f159ab0cd973f2426cfc3af266fc8bdde2
2 years ago
size_t CompactionOutputs::UpdateGrandparentBoundaryInfo(
const Slice& internal_key) {
size_t curr_key_boundary_switched_num = 0;
const std::vector<FileMetaData*>& grandparents = compaction_->grandparents();
if (grandparents.empty()) {
return curr_key_boundary_switched_num;
}
assert(!internal_key.empty());
InternalKey ikey;
ikey.DecodeFrom(internal_key);
assert(ikey.Valid());
Align compaction output file boundaries to the next level ones (#10655) Summary: Try to align the compaction output file boundaries to the next level ones (grandparent level), to reduce the level compaction write-amplification. In level compaction, there are "wasted" data at the beginning and end of the output level files. Align the file boundary can avoid such "wasted" compaction. With this PR, it tries to align the non-bottommost level file boundaries to its next level ones. It may cut file when the file size is large enough (at least 50% of target_file_size) and not too large (2x target_file_size). db_bench shows about 12.56% compaction reduction: ``` TEST_TMPDIR=/data/dbbench2 ./db_bench --benchmarks=fillrandom,readrandom -max_background_jobs=12 -num=400000000 -target_file_size_base=33554432 # baseline: Flush(GB): cumulative 25.882, interval 7.216 Cumulative compaction: 285.90 GB write, 162.36 MB/s write, 269.68 GB read, 153.15 MB/s read, 2926.7 seconds # with this change: Flush(GB): cumulative 25.882, interval 7.753 Cumulative compaction: 249.97 GB write, 141.96 MB/s write, 233.74 GB read, 132.74 MB/s read, 2534.9 seconds ``` The compaction simulator shows a similar result (14% with 100G random data). As a side effect, with this PR, the SST file size can exceed the target_file_size, but is capped at 2x target_file_size. And there will be smaller files. Here are file size statistics when loading 100GB with the target file size 32MB: ``` baseline this_PR count 1.656000e+03 1.705000e+03 mean 3.116062e+07 3.028076e+07 std 7.145242e+06 8.046139e+06 ``` The feature is enabled by default, to revert to the old behavior disable it with `AdvancedColumnFamilyOptions.level_compaction_dynamic_file_size = false` Also includes https://github.com/facebook/rocksdb/issues/1963 to cut file before skippable grandparent file. Which is for use case like user adding 2 or more non-overlapping data range at the same time, it can reduce the overlapping of 2 datasets in the lower levels. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10655 Reviewed By: cbi42 Differential Revision: D39552321 Pulled By: jay-zhuang fbshipit-source-id: 640d15f159ab0cd973f2426cfc3af266fc8bdde2
2 years ago
const Comparator* ucmp = compaction_->column_family_data()->user_comparator();
Align compaction output file boundaries to the next level ones (#10655) Summary: Try to align the compaction output file boundaries to the next level ones (grandparent level), to reduce the level compaction write-amplification. In level compaction, there are "wasted" data at the beginning and end of the output level files. Align the file boundary can avoid such "wasted" compaction. With this PR, it tries to align the non-bottommost level file boundaries to its next level ones. It may cut file when the file size is large enough (at least 50% of target_file_size) and not too large (2x target_file_size). db_bench shows about 12.56% compaction reduction: ``` TEST_TMPDIR=/data/dbbench2 ./db_bench --benchmarks=fillrandom,readrandom -max_background_jobs=12 -num=400000000 -target_file_size_base=33554432 # baseline: Flush(GB): cumulative 25.882, interval 7.216 Cumulative compaction: 285.90 GB write, 162.36 MB/s write, 269.68 GB read, 153.15 MB/s read, 2926.7 seconds # with this change: Flush(GB): cumulative 25.882, interval 7.753 Cumulative compaction: 249.97 GB write, 141.96 MB/s write, 233.74 GB read, 132.74 MB/s read, 2534.9 seconds ``` The compaction simulator shows a similar result (14% with 100G random data). As a side effect, with this PR, the SST file size can exceed the target_file_size, but is capped at 2x target_file_size. And there will be smaller files. Here are file size statistics when loading 100GB with the target file size 32MB: ``` baseline this_PR count 1.656000e+03 1.705000e+03 mean 3.116062e+07 3.028076e+07 std 7.145242e+06 8.046139e+06 ``` The feature is enabled by default, to revert to the old behavior disable it with `AdvancedColumnFamilyOptions.level_compaction_dynamic_file_size = false` Also includes https://github.com/facebook/rocksdb/issues/1963 to cut file before skippable grandparent file. Which is for use case like user adding 2 or more non-overlapping data range at the same time, it can reduce the overlapping of 2 datasets in the lower levels. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10655 Reviewed By: cbi42 Differential Revision: D39552321 Pulled By: jay-zhuang fbshipit-source-id: 640d15f159ab0cd973f2426cfc3af266fc8bdde2
2 years ago
// Move the grandparent_index_ to the file containing the current user_key.
// If there are multiple files containing the same user_key, make sure the
// index points to the last file containing the key.
Align compaction output file boundaries to the next level ones (#10655) Summary: Try to align the compaction output file boundaries to the next level ones (grandparent level), to reduce the level compaction write-amplification. In level compaction, there are "wasted" data at the beginning and end of the output level files. Align the file boundary can avoid such "wasted" compaction. With this PR, it tries to align the non-bottommost level file boundaries to its next level ones. It may cut file when the file size is large enough (at least 50% of target_file_size) and not too large (2x target_file_size). db_bench shows about 12.56% compaction reduction: ``` TEST_TMPDIR=/data/dbbench2 ./db_bench --benchmarks=fillrandom,readrandom -max_background_jobs=12 -num=400000000 -target_file_size_base=33554432 # baseline: Flush(GB): cumulative 25.882, interval 7.216 Cumulative compaction: 285.90 GB write, 162.36 MB/s write, 269.68 GB read, 153.15 MB/s read, 2926.7 seconds # with this change: Flush(GB): cumulative 25.882, interval 7.753 Cumulative compaction: 249.97 GB write, 141.96 MB/s write, 233.74 GB read, 132.74 MB/s read, 2534.9 seconds ``` The compaction simulator shows a similar result (14% with 100G random data). As a side effect, with this PR, the SST file size can exceed the target_file_size, but is capped at 2x target_file_size. And there will be smaller files. Here are file size statistics when loading 100GB with the target file size 32MB: ``` baseline this_PR count 1.656000e+03 1.705000e+03 mean 3.116062e+07 3.028076e+07 std 7.145242e+06 8.046139e+06 ``` The feature is enabled by default, to revert to the old behavior disable it with `AdvancedColumnFamilyOptions.level_compaction_dynamic_file_size = false` Also includes https://github.com/facebook/rocksdb/issues/1963 to cut file before skippable grandparent file. Which is for use case like user adding 2 or more non-overlapping data range at the same time, it can reduce the overlapping of 2 datasets in the lower levels. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10655 Reviewed By: cbi42 Differential Revision: D39552321 Pulled By: jay-zhuang fbshipit-source-id: 640d15f159ab0cd973f2426cfc3af266fc8bdde2
2 years ago
while (grandparent_index_ < grandparents.size()) {
if (being_grandparent_gap_) {
if (sstableKeyCompare(ucmp, ikey,
grandparents[grandparent_index_]->smallest) < 0) {
Align compaction output file boundaries to the next level ones (#10655) Summary: Try to align the compaction output file boundaries to the next level ones (grandparent level), to reduce the level compaction write-amplification. In level compaction, there are "wasted" data at the beginning and end of the output level files. Align the file boundary can avoid such "wasted" compaction. With this PR, it tries to align the non-bottommost level file boundaries to its next level ones. It may cut file when the file size is large enough (at least 50% of target_file_size) and not too large (2x target_file_size). db_bench shows about 12.56% compaction reduction: ``` TEST_TMPDIR=/data/dbbench2 ./db_bench --benchmarks=fillrandom,readrandom -max_background_jobs=12 -num=400000000 -target_file_size_base=33554432 # baseline: Flush(GB): cumulative 25.882, interval 7.216 Cumulative compaction: 285.90 GB write, 162.36 MB/s write, 269.68 GB read, 153.15 MB/s read, 2926.7 seconds # with this change: Flush(GB): cumulative 25.882, interval 7.753 Cumulative compaction: 249.97 GB write, 141.96 MB/s write, 233.74 GB read, 132.74 MB/s read, 2534.9 seconds ``` The compaction simulator shows a similar result (14% with 100G random data). As a side effect, with this PR, the SST file size can exceed the target_file_size, but is capped at 2x target_file_size. And there will be smaller files. Here are file size statistics when loading 100GB with the target file size 32MB: ``` baseline this_PR count 1.656000e+03 1.705000e+03 mean 3.116062e+07 3.028076e+07 std 7.145242e+06 8.046139e+06 ``` The feature is enabled by default, to revert to the old behavior disable it with `AdvancedColumnFamilyOptions.level_compaction_dynamic_file_size = false` Also includes https://github.com/facebook/rocksdb/issues/1963 to cut file before skippable grandparent file. Which is for use case like user adding 2 or more non-overlapping data range at the same time, it can reduce the overlapping of 2 datasets in the lower levels. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10655 Reviewed By: cbi42 Differential Revision: D39552321 Pulled By: jay-zhuang fbshipit-source-id: 640d15f159ab0cd973f2426cfc3af266fc8bdde2
2 years ago
break;
}
if (seen_key_) {
curr_key_boundary_switched_num++;
grandparent_overlapped_bytes_ +=
grandparents[grandparent_index_]->fd.GetFileSize();
grandparent_boundary_switched_num_++;
}
being_grandparent_gap_ = false;
} else {
int cmp_result = sstableKeyCompare(
ucmp, ikey, grandparents[grandparent_index_]->largest);
// If it's same key, make sure grandparent_index_ is pointing to the last
// one.
if (cmp_result < 0 ||
(cmp_result == 0 &&
(grandparent_index_ == grandparents.size() - 1 ||
sstableKeyCompare(ucmp, ikey,
grandparents[grandparent_index_ + 1]->smallest) <
0))) {
Align compaction output file boundaries to the next level ones (#10655) Summary: Try to align the compaction output file boundaries to the next level ones (grandparent level), to reduce the level compaction write-amplification. In level compaction, there are "wasted" data at the beginning and end of the output level files. Align the file boundary can avoid such "wasted" compaction. With this PR, it tries to align the non-bottommost level file boundaries to its next level ones. It may cut file when the file size is large enough (at least 50% of target_file_size) and not too large (2x target_file_size). db_bench shows about 12.56% compaction reduction: ``` TEST_TMPDIR=/data/dbbench2 ./db_bench --benchmarks=fillrandom,readrandom -max_background_jobs=12 -num=400000000 -target_file_size_base=33554432 # baseline: Flush(GB): cumulative 25.882, interval 7.216 Cumulative compaction: 285.90 GB write, 162.36 MB/s write, 269.68 GB read, 153.15 MB/s read, 2926.7 seconds # with this change: Flush(GB): cumulative 25.882, interval 7.753 Cumulative compaction: 249.97 GB write, 141.96 MB/s write, 233.74 GB read, 132.74 MB/s read, 2534.9 seconds ``` The compaction simulator shows a similar result (14% with 100G random data). As a side effect, with this PR, the SST file size can exceed the target_file_size, but is capped at 2x target_file_size. And there will be smaller files. Here are file size statistics when loading 100GB with the target file size 32MB: ``` baseline this_PR count 1.656000e+03 1.705000e+03 mean 3.116062e+07 3.028076e+07 std 7.145242e+06 8.046139e+06 ``` The feature is enabled by default, to revert to the old behavior disable it with `AdvancedColumnFamilyOptions.level_compaction_dynamic_file_size = false` Also includes https://github.com/facebook/rocksdb/issues/1963 to cut file before skippable grandparent file. Which is for use case like user adding 2 or more non-overlapping data range at the same time, it can reduce the overlapping of 2 datasets in the lower levels. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10655 Reviewed By: cbi42 Differential Revision: D39552321 Pulled By: jay-zhuang fbshipit-source-id: 640d15f159ab0cd973f2426cfc3af266fc8bdde2
2 years ago
break;
}
if (seen_key_) {
curr_key_boundary_switched_num++;
grandparent_boundary_switched_num_++;
}
being_grandparent_gap_ = true;
grandparent_index_++;
}
}
// If the first key is in the middle of a grandparent file, adding it to the
// overlap
if (!seen_key_ && !being_grandparent_gap_) {
assert(grandparent_overlapped_bytes_ == 0);
grandparent_overlapped_bytes_ =
GetCurrentKeyGrandparentOverlappedBytes(internal_key);
Align compaction output file boundaries to the next level ones (#10655) Summary: Try to align the compaction output file boundaries to the next level ones (grandparent level), to reduce the level compaction write-amplification. In level compaction, there are "wasted" data at the beginning and end of the output level files. Align the file boundary can avoid such "wasted" compaction. With this PR, it tries to align the non-bottommost level file boundaries to its next level ones. It may cut file when the file size is large enough (at least 50% of target_file_size) and not too large (2x target_file_size). db_bench shows about 12.56% compaction reduction: ``` TEST_TMPDIR=/data/dbbench2 ./db_bench --benchmarks=fillrandom,readrandom -max_background_jobs=12 -num=400000000 -target_file_size_base=33554432 # baseline: Flush(GB): cumulative 25.882, interval 7.216 Cumulative compaction: 285.90 GB write, 162.36 MB/s write, 269.68 GB read, 153.15 MB/s read, 2926.7 seconds # with this change: Flush(GB): cumulative 25.882, interval 7.753 Cumulative compaction: 249.97 GB write, 141.96 MB/s write, 233.74 GB read, 132.74 MB/s read, 2534.9 seconds ``` The compaction simulator shows a similar result (14% with 100G random data). As a side effect, with this PR, the SST file size can exceed the target_file_size, but is capped at 2x target_file_size. And there will be smaller files. Here are file size statistics when loading 100GB with the target file size 32MB: ``` baseline this_PR count 1.656000e+03 1.705000e+03 mean 3.116062e+07 3.028076e+07 std 7.145242e+06 8.046139e+06 ``` The feature is enabled by default, to revert to the old behavior disable it with `AdvancedColumnFamilyOptions.level_compaction_dynamic_file_size = false` Also includes https://github.com/facebook/rocksdb/issues/1963 to cut file before skippable grandparent file. Which is for use case like user adding 2 or more non-overlapping data range at the same time, it can reduce the overlapping of 2 datasets in the lower levels. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10655 Reviewed By: cbi42 Differential Revision: D39552321 Pulled By: jay-zhuang fbshipit-source-id: 640d15f159ab0cd973f2426cfc3af266fc8bdde2
2 years ago
}
seen_key_ = true;
return curr_key_boundary_switched_num;
}
uint64_t CompactionOutputs::GetCurrentKeyGrandparentOverlappedBytes(
const Slice& internal_key) const {
// no overlap with any grandparent file
if (being_grandparent_gap_) {
return 0;
}
uint64_t overlapped_bytes = 0;
const std::vector<FileMetaData*>& grandparents = compaction_->grandparents();
const Comparator* ucmp = compaction_->column_family_data()->user_comparator();
InternalKey ikey;
ikey.DecodeFrom(internal_key);
#ifndef NDEBUG
// make sure the grandparent_index_ is pointing to the last files containing
// the current key.
int cmp_result =
sstableKeyCompare(ucmp, ikey, grandparents[grandparent_index_]->largest);
assert(
cmp_result < 0 ||
(cmp_result == 0 &&
(grandparent_index_ == grandparents.size() - 1 ||
sstableKeyCompare(
ucmp, ikey, grandparents[grandparent_index_ + 1]->smallest) < 0)));
assert(sstableKeyCompare(ucmp, ikey,
grandparents[grandparent_index_]->smallest) >= 0);
#endif
overlapped_bytes += grandparents[grandparent_index_]->fd.GetFileSize();
// go backwards to find all overlapped files, one key can overlap multiple
// files. In the following example, if the current output key is `c`, and one
// compaction file was cut before `c`, current `c` can overlap with 3 files:
// [a b] [c...
// [b, b] [c, c] [c, c] [c, d]
for (int64_t i = static_cast<int64_t>(grandparent_index_) - 1;
i >= 0 && sstableKeyCompare(ucmp, ikey, grandparents[i]->largest) == 0;
i--) {
overlapped_bytes += grandparents[i]->fd.GetFileSize();
}
return overlapped_bytes;
}
bool CompactionOutputs::ShouldStopBefore(const CompactionIterator& c_iter) {
assert(c_iter.Valid());
Align compaction output file boundaries to the next level ones (#10655) Summary: Try to align the compaction output file boundaries to the next level ones (grandparent level), to reduce the level compaction write-amplification. In level compaction, there are "wasted" data at the beginning and end of the output level files. Align the file boundary can avoid such "wasted" compaction. With this PR, it tries to align the non-bottommost level file boundaries to its next level ones. It may cut file when the file size is large enough (at least 50% of target_file_size) and not too large (2x target_file_size). db_bench shows about 12.56% compaction reduction: ``` TEST_TMPDIR=/data/dbbench2 ./db_bench --benchmarks=fillrandom,readrandom -max_background_jobs=12 -num=400000000 -target_file_size_base=33554432 # baseline: Flush(GB): cumulative 25.882, interval 7.216 Cumulative compaction: 285.90 GB write, 162.36 MB/s write, 269.68 GB read, 153.15 MB/s read, 2926.7 seconds # with this change: Flush(GB): cumulative 25.882, interval 7.753 Cumulative compaction: 249.97 GB write, 141.96 MB/s write, 233.74 GB read, 132.74 MB/s read, 2534.9 seconds ``` The compaction simulator shows a similar result (14% with 100G random data). As a side effect, with this PR, the SST file size can exceed the target_file_size, but is capped at 2x target_file_size. And there will be smaller files. Here are file size statistics when loading 100GB with the target file size 32MB: ``` baseline this_PR count 1.656000e+03 1.705000e+03 mean 3.116062e+07 3.028076e+07 std 7.145242e+06 8.046139e+06 ``` The feature is enabled by default, to revert to the old behavior disable it with `AdvancedColumnFamilyOptions.level_compaction_dynamic_file_size = false` Also includes https://github.com/facebook/rocksdb/issues/1963 to cut file before skippable grandparent file. Which is for use case like user adding 2 or more non-overlapping data range at the same time, it can reduce the overlapping of 2 datasets in the lower levels. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10655 Reviewed By: cbi42 Differential Revision: D39552321 Pulled By: jay-zhuang fbshipit-source-id: 640d15f159ab0cd973f2426cfc3af266fc8bdde2
2 years ago
const Slice& internal_key = c_iter.key();
const uint64_t previous_overlapped_bytes = grandparent_overlapped_bytes_;
const InternalKeyComparator* icmp =
&compaction_->column_family_data()->internal_comparator();
size_t num_grandparent_boundaries_crossed = 0;
bool should_stop_for_ttl = false;
// Always update grandparent information like overlapped file number, size
// etc., and TTL states.
// If compaction_->output_level() == 0, there is no need to update grandparent
// info, and that `grandparent` should be empty.
if (compaction_->output_level() > 0) {
num_grandparent_boundaries_crossed =
UpdateGrandparentBoundaryInfo(internal_key);
should_stop_for_ttl = UpdateFilesToCutForTTLStates(internal_key);
}
Align compaction output file boundaries to the next level ones (#10655) Summary: Try to align the compaction output file boundaries to the next level ones (grandparent level), to reduce the level compaction write-amplification. In level compaction, there are "wasted" data at the beginning and end of the output level files. Align the file boundary can avoid such "wasted" compaction. With this PR, it tries to align the non-bottommost level file boundaries to its next level ones. It may cut file when the file size is large enough (at least 50% of target_file_size) and not too large (2x target_file_size). db_bench shows about 12.56% compaction reduction: ``` TEST_TMPDIR=/data/dbbench2 ./db_bench --benchmarks=fillrandom,readrandom -max_background_jobs=12 -num=400000000 -target_file_size_base=33554432 # baseline: Flush(GB): cumulative 25.882, interval 7.216 Cumulative compaction: 285.90 GB write, 162.36 MB/s write, 269.68 GB read, 153.15 MB/s read, 2926.7 seconds # with this change: Flush(GB): cumulative 25.882, interval 7.753 Cumulative compaction: 249.97 GB write, 141.96 MB/s write, 233.74 GB read, 132.74 MB/s read, 2534.9 seconds ``` The compaction simulator shows a similar result (14% with 100G random data). As a side effect, with this PR, the SST file size can exceed the target_file_size, but is capped at 2x target_file_size. And there will be smaller files. Here are file size statistics when loading 100GB with the target file size 32MB: ``` baseline this_PR count 1.656000e+03 1.705000e+03 mean 3.116062e+07 3.028076e+07 std 7.145242e+06 8.046139e+06 ``` The feature is enabled by default, to revert to the old behavior disable it with `AdvancedColumnFamilyOptions.level_compaction_dynamic_file_size = false` Also includes https://github.com/facebook/rocksdb/issues/1963 to cut file before skippable grandparent file. Which is for use case like user adding 2 or more non-overlapping data range at the same time, it can reduce the overlapping of 2 datasets in the lower levels. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10655 Reviewed By: cbi42 Differential Revision: D39552321 Pulled By: jay-zhuang fbshipit-source-id: 640d15f159ab0cd973f2426cfc3af266fc8bdde2
2 years ago
if (!HasBuilder()) {
return false;
}
if (should_stop_for_ttl) {
return true;
}
// If there's user defined partitioner, check that first
Align compaction output file boundaries to the next level ones (#10655) Summary: Try to align the compaction output file boundaries to the next level ones (grandparent level), to reduce the level compaction write-amplification. In level compaction, there are "wasted" data at the beginning and end of the output level files. Align the file boundary can avoid such "wasted" compaction. With this PR, it tries to align the non-bottommost level file boundaries to its next level ones. It may cut file when the file size is large enough (at least 50% of target_file_size) and not too large (2x target_file_size). db_bench shows about 12.56% compaction reduction: ``` TEST_TMPDIR=/data/dbbench2 ./db_bench --benchmarks=fillrandom,readrandom -max_background_jobs=12 -num=400000000 -target_file_size_base=33554432 # baseline: Flush(GB): cumulative 25.882, interval 7.216 Cumulative compaction: 285.90 GB write, 162.36 MB/s write, 269.68 GB read, 153.15 MB/s read, 2926.7 seconds # with this change: Flush(GB): cumulative 25.882, interval 7.753 Cumulative compaction: 249.97 GB write, 141.96 MB/s write, 233.74 GB read, 132.74 MB/s read, 2534.9 seconds ``` The compaction simulator shows a similar result (14% with 100G random data). As a side effect, with this PR, the SST file size can exceed the target_file_size, but is capped at 2x target_file_size. And there will be smaller files. Here are file size statistics when loading 100GB with the target file size 32MB: ``` baseline this_PR count 1.656000e+03 1.705000e+03 mean 3.116062e+07 3.028076e+07 std 7.145242e+06 8.046139e+06 ``` The feature is enabled by default, to revert to the old behavior disable it with `AdvancedColumnFamilyOptions.level_compaction_dynamic_file_size = false` Also includes https://github.com/facebook/rocksdb/issues/1963 to cut file before skippable grandparent file. Which is for use case like user adding 2 or more non-overlapping data range at the same time, it can reduce the overlapping of 2 datasets in the lower levels. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10655 Reviewed By: cbi42 Differential Revision: D39552321 Pulled By: jay-zhuang fbshipit-source-id: 640d15f159ab0cd973f2426cfc3af266fc8bdde2
2 years ago
if (partitioner_ && partitioner_->ShouldPartition(PartitionerRequest(
last_key_for_partitioner_, c_iter.user_key(),
current_output_file_size_)) == kRequired) {
return true;
}
// files output to Level 0 won't be split
if (compaction_->output_level() == 0) {
return false;
}
Align compaction output file boundaries to the next level ones (#10655) Summary: Try to align the compaction output file boundaries to the next level ones (grandparent level), to reduce the level compaction write-amplification. In level compaction, there are "wasted" data at the beginning and end of the output level files. Align the file boundary can avoid such "wasted" compaction. With this PR, it tries to align the non-bottommost level file boundaries to its next level ones. It may cut file when the file size is large enough (at least 50% of target_file_size) and not too large (2x target_file_size). db_bench shows about 12.56% compaction reduction: ``` TEST_TMPDIR=/data/dbbench2 ./db_bench --benchmarks=fillrandom,readrandom -max_background_jobs=12 -num=400000000 -target_file_size_base=33554432 # baseline: Flush(GB): cumulative 25.882, interval 7.216 Cumulative compaction: 285.90 GB write, 162.36 MB/s write, 269.68 GB read, 153.15 MB/s read, 2926.7 seconds # with this change: Flush(GB): cumulative 25.882, interval 7.753 Cumulative compaction: 249.97 GB write, 141.96 MB/s write, 233.74 GB read, 132.74 MB/s read, 2534.9 seconds ``` The compaction simulator shows a similar result (14% with 100G random data). As a side effect, with this PR, the SST file size can exceed the target_file_size, but is capped at 2x target_file_size. And there will be smaller files. Here are file size statistics when loading 100GB with the target file size 32MB: ``` baseline this_PR count 1.656000e+03 1.705000e+03 mean 3.116062e+07 3.028076e+07 std 7.145242e+06 8.046139e+06 ``` The feature is enabled by default, to revert to the old behavior disable it with `AdvancedColumnFamilyOptions.level_compaction_dynamic_file_size = false` Also includes https://github.com/facebook/rocksdb/issues/1963 to cut file before skippable grandparent file. Which is for use case like user adding 2 or more non-overlapping data range at the same time, it can reduce the overlapping of 2 datasets in the lower levels. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10655 Reviewed By: cbi42 Differential Revision: D39552321 Pulled By: jay-zhuang fbshipit-source-id: 640d15f159ab0cd973f2426cfc3af266fc8bdde2
2 years ago
// reach the max file size
if (current_output_file_size_ >= compaction_->max_output_file_size()) {
return true;
}
// Check if it needs to split for RoundRobin
// Invalid local_output_split_key indicates that we do not need to split
if (local_output_split_key_ != nullptr && !is_split_) {
// Split occurs when the next key is larger than/equal to the cursor
if (icmp->Compare(internal_key, local_output_split_key_->Encode()) >= 0) {
is_split_ = true;
return true;
}
}
Align compaction output file boundaries to the next level ones (#10655) Summary: Try to align the compaction output file boundaries to the next level ones (grandparent level), to reduce the level compaction write-amplification. In level compaction, there are "wasted" data at the beginning and end of the output level files. Align the file boundary can avoid such "wasted" compaction. With this PR, it tries to align the non-bottommost level file boundaries to its next level ones. It may cut file when the file size is large enough (at least 50% of target_file_size) and not too large (2x target_file_size). db_bench shows about 12.56% compaction reduction: ``` TEST_TMPDIR=/data/dbbench2 ./db_bench --benchmarks=fillrandom,readrandom -max_background_jobs=12 -num=400000000 -target_file_size_base=33554432 # baseline: Flush(GB): cumulative 25.882, interval 7.216 Cumulative compaction: 285.90 GB write, 162.36 MB/s write, 269.68 GB read, 153.15 MB/s read, 2926.7 seconds # with this change: Flush(GB): cumulative 25.882, interval 7.753 Cumulative compaction: 249.97 GB write, 141.96 MB/s write, 233.74 GB read, 132.74 MB/s read, 2534.9 seconds ``` The compaction simulator shows a similar result (14% with 100G random data). As a side effect, with this PR, the SST file size can exceed the target_file_size, but is capped at 2x target_file_size. And there will be smaller files. Here are file size statistics when loading 100GB with the target file size 32MB: ``` baseline this_PR count 1.656000e+03 1.705000e+03 mean 3.116062e+07 3.028076e+07 std 7.145242e+06 8.046139e+06 ``` The feature is enabled by default, to revert to the old behavior disable it with `AdvancedColumnFamilyOptions.level_compaction_dynamic_file_size = false` Also includes https://github.com/facebook/rocksdb/issues/1963 to cut file before skippable grandparent file. Which is for use case like user adding 2 or more non-overlapping data range at the same time, it can reduce the overlapping of 2 datasets in the lower levels. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10655 Reviewed By: cbi42 Differential Revision: D39552321 Pulled By: jay-zhuang fbshipit-source-id: 640d15f159ab0cd973f2426cfc3af266fc8bdde2
2 years ago
// only check if the current key is going to cross the grandparents file
// boundary (either the file beginning or ending).
if (num_grandparent_boundaries_crossed > 0) {
// Cut the file before the current key if the size of the current output
// file + its overlapped grandparent files is bigger than
// max_compaction_bytes. Which is to prevent future bigger than
// max_compaction_bytes compaction from the current output level.
if (grandparent_overlapped_bytes_ + current_output_file_size_ >
compaction_->max_compaction_bytes()) {
return true;
}
Align compaction output file boundaries to the next level ones (#10655) Summary: Try to align the compaction output file boundaries to the next level ones (grandparent level), to reduce the level compaction write-amplification. In level compaction, there are "wasted" data at the beginning and end of the output level files. Align the file boundary can avoid such "wasted" compaction. With this PR, it tries to align the non-bottommost level file boundaries to its next level ones. It may cut file when the file size is large enough (at least 50% of target_file_size) and not too large (2x target_file_size). db_bench shows about 12.56% compaction reduction: ``` TEST_TMPDIR=/data/dbbench2 ./db_bench --benchmarks=fillrandom,readrandom -max_background_jobs=12 -num=400000000 -target_file_size_base=33554432 # baseline: Flush(GB): cumulative 25.882, interval 7.216 Cumulative compaction: 285.90 GB write, 162.36 MB/s write, 269.68 GB read, 153.15 MB/s read, 2926.7 seconds # with this change: Flush(GB): cumulative 25.882, interval 7.753 Cumulative compaction: 249.97 GB write, 141.96 MB/s write, 233.74 GB read, 132.74 MB/s read, 2534.9 seconds ``` The compaction simulator shows a similar result (14% with 100G random data). As a side effect, with this PR, the SST file size can exceed the target_file_size, but is capped at 2x target_file_size. And there will be smaller files. Here are file size statistics when loading 100GB with the target file size 32MB: ``` baseline this_PR count 1.656000e+03 1.705000e+03 mean 3.116062e+07 3.028076e+07 std 7.145242e+06 8.046139e+06 ``` The feature is enabled by default, to revert to the old behavior disable it with `AdvancedColumnFamilyOptions.level_compaction_dynamic_file_size = false` Also includes https://github.com/facebook/rocksdb/issues/1963 to cut file before skippable grandparent file. Which is for use case like user adding 2 or more non-overlapping data range at the same time, it can reduce the overlapping of 2 datasets in the lower levels. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10655 Reviewed By: cbi42 Differential Revision: D39552321 Pulled By: jay-zhuang fbshipit-source-id: 640d15f159ab0cd973f2426cfc3af266fc8bdde2
2 years ago
// Cut the file if including the key is going to add a skippable file on
// the grandparent level AND its size is reasonably big (1/8 of target file
// size). For example, if it's compacting the files L0 + L1:
// L0: [1, 21]
// L1: [3, 23]
// L2: [2, 4] [11, 15] [22, 24]
// Without this break, it will output as:
// L1: [1,3, 21,23]
// With this break, it will output as (assuming [11, 15] at L2 is bigger
// than 1/8 of target size):
// L1: [1,3] [21,23]
// Then for the future compactions, [11,15] won't be included.
// For random datasets (either evenly distributed or skewed), it rarely
// triggers this condition, but if the user is adding 2 different datasets
// without any overlap, it may likely happen.
// More details, check PR #1963
const size_t num_skippable_boundaries_crossed =
being_grandparent_gap_ ? 2 : 3;
if (compaction_->immutable_options()->compaction_style ==
kCompactionStyleLevel &&
compaction_->immutable_options()->level_compaction_dynamic_file_size &&
num_grandparent_boundaries_crossed >=
num_skippable_boundaries_crossed &&
grandparent_overlapped_bytes_ - previous_overlapped_bytes >
compaction_->target_output_file_size() / 8) {
return true;
}
// Pre-cut the output file if it's reaching a certain size AND it's at the
// boundary of a grandparent file. It can reduce the future compaction size,
// the cost is having smaller files.
// The pre-cut size threshold is based on how many grandparent boundaries
// it has seen before. Basically, if it has seen no boundary at all, then it
// will pre-cut at 50% target file size. Every boundary it has seen
// increases the threshold by 5%, max at 90%, which it will always cut.
// The idea is based on if it has seen more boundaries before, it will more
// likely to see another boundary (file cutting opportunity) before the
// target file size. The test shows it can generate larger files than a
// static threshold like 75% and has a similar write amplification
// improvement.
if (compaction_->immutable_options()->compaction_style ==
kCompactionStyleLevel &&
compaction_->immutable_options()->level_compaction_dynamic_file_size &&
current_output_file_size_ >=
Align compaction output file boundaries to the next level ones (#10655) Summary: Try to align the compaction output file boundaries to the next level ones (grandparent level), to reduce the level compaction write-amplification. In level compaction, there are "wasted" data at the beginning and end of the output level files. Align the file boundary can avoid such "wasted" compaction. With this PR, it tries to align the non-bottommost level file boundaries to its next level ones. It may cut file when the file size is large enough (at least 50% of target_file_size) and not too large (2x target_file_size). db_bench shows about 12.56% compaction reduction: ``` TEST_TMPDIR=/data/dbbench2 ./db_bench --benchmarks=fillrandom,readrandom -max_background_jobs=12 -num=400000000 -target_file_size_base=33554432 # baseline: Flush(GB): cumulative 25.882, interval 7.216 Cumulative compaction: 285.90 GB write, 162.36 MB/s write, 269.68 GB read, 153.15 MB/s read, 2926.7 seconds # with this change: Flush(GB): cumulative 25.882, interval 7.753 Cumulative compaction: 249.97 GB write, 141.96 MB/s write, 233.74 GB read, 132.74 MB/s read, 2534.9 seconds ``` The compaction simulator shows a similar result (14% with 100G random data). As a side effect, with this PR, the SST file size can exceed the target_file_size, but is capped at 2x target_file_size. And there will be smaller files. Here are file size statistics when loading 100GB with the target file size 32MB: ``` baseline this_PR count 1.656000e+03 1.705000e+03 mean 3.116062e+07 3.028076e+07 std 7.145242e+06 8.046139e+06 ``` The feature is enabled by default, to revert to the old behavior disable it with `AdvancedColumnFamilyOptions.level_compaction_dynamic_file_size = false` Also includes https://github.com/facebook/rocksdb/issues/1963 to cut file before skippable grandparent file. Which is for use case like user adding 2 or more non-overlapping data range at the same time, it can reduce the overlapping of 2 datasets in the lower levels. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10655 Reviewed By: cbi42 Differential Revision: D39552321 Pulled By: jay-zhuang fbshipit-source-id: 640d15f159ab0cd973f2426cfc3af266fc8bdde2
2 years ago
((compaction_->target_output_file_size() + 99) / 100) *
(50 + std::min(grandparent_boundary_switched_num_ * 5,
size_t{40}))) {
return true;
}
}
return false;
}
Status CompactionOutputs::AddToOutput(
const CompactionIterator& c_iter,
const CompactionFileOpenFunc& open_file_func,
const CompactionFileCloseFunc& close_file_func) {
Status s;
const Slice& key = c_iter.key();
if (ShouldStopBefore(c_iter) && HasBuilder()) {
s = close_file_func(*this, c_iter.InputStatus(), key);
if (!s.ok()) {
return s;
}
Align compaction output file boundaries to the next level ones (#10655) Summary: Try to align the compaction output file boundaries to the next level ones (grandparent level), to reduce the level compaction write-amplification. In level compaction, there are "wasted" data at the beginning and end of the output level files. Align the file boundary can avoid such "wasted" compaction. With this PR, it tries to align the non-bottommost level file boundaries to its next level ones. It may cut file when the file size is large enough (at least 50% of target_file_size) and not too large (2x target_file_size). db_bench shows about 12.56% compaction reduction: ``` TEST_TMPDIR=/data/dbbench2 ./db_bench --benchmarks=fillrandom,readrandom -max_background_jobs=12 -num=400000000 -target_file_size_base=33554432 # baseline: Flush(GB): cumulative 25.882, interval 7.216 Cumulative compaction: 285.90 GB write, 162.36 MB/s write, 269.68 GB read, 153.15 MB/s read, 2926.7 seconds # with this change: Flush(GB): cumulative 25.882, interval 7.753 Cumulative compaction: 249.97 GB write, 141.96 MB/s write, 233.74 GB read, 132.74 MB/s read, 2534.9 seconds ``` The compaction simulator shows a similar result (14% with 100G random data). As a side effect, with this PR, the SST file size can exceed the target_file_size, but is capped at 2x target_file_size. And there will be smaller files. Here are file size statistics when loading 100GB with the target file size 32MB: ``` baseline this_PR count 1.656000e+03 1.705000e+03 mean 3.116062e+07 3.028076e+07 std 7.145242e+06 8.046139e+06 ``` The feature is enabled by default, to revert to the old behavior disable it with `AdvancedColumnFamilyOptions.level_compaction_dynamic_file_size = false` Also includes https://github.com/facebook/rocksdb/issues/1963 to cut file before skippable grandparent file. Which is for use case like user adding 2 or more non-overlapping data range at the same time, it can reduce the overlapping of 2 datasets in the lower levels. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10655 Reviewed By: cbi42 Differential Revision: D39552321 Pulled By: jay-zhuang fbshipit-source-id: 640d15f159ab0cd973f2426cfc3af266fc8bdde2
2 years ago
// reset grandparent information
grandparent_boundary_switched_num_ = 0;
grandparent_overlapped_bytes_ =
GetCurrentKeyGrandparentOverlappedBytes(key);
}
// Open output file if necessary
if (!HasBuilder()) {
s = open_file_func(*this);
if (!s.ok()) {
return s;
}
}
assert(builder_ != nullptr);
const Slice& value = c_iter.value();
Align compaction output file boundaries to the next level ones (#10655) Summary: Try to align the compaction output file boundaries to the next level ones (grandparent level), to reduce the level compaction write-amplification. In level compaction, there are "wasted" data at the beginning and end of the output level files. Align the file boundary can avoid such "wasted" compaction. With this PR, it tries to align the non-bottommost level file boundaries to its next level ones. It may cut file when the file size is large enough (at least 50% of target_file_size) and not too large (2x target_file_size). db_bench shows about 12.56% compaction reduction: ``` TEST_TMPDIR=/data/dbbench2 ./db_bench --benchmarks=fillrandom,readrandom -max_background_jobs=12 -num=400000000 -target_file_size_base=33554432 # baseline: Flush(GB): cumulative 25.882, interval 7.216 Cumulative compaction: 285.90 GB write, 162.36 MB/s write, 269.68 GB read, 153.15 MB/s read, 2926.7 seconds # with this change: Flush(GB): cumulative 25.882, interval 7.753 Cumulative compaction: 249.97 GB write, 141.96 MB/s write, 233.74 GB read, 132.74 MB/s read, 2534.9 seconds ``` The compaction simulator shows a similar result (14% with 100G random data). As a side effect, with this PR, the SST file size can exceed the target_file_size, but is capped at 2x target_file_size. And there will be smaller files. Here are file size statistics when loading 100GB with the target file size 32MB: ``` baseline this_PR count 1.656000e+03 1.705000e+03 mean 3.116062e+07 3.028076e+07 std 7.145242e+06 8.046139e+06 ``` The feature is enabled by default, to revert to the old behavior disable it with `AdvancedColumnFamilyOptions.level_compaction_dynamic_file_size = false` Also includes https://github.com/facebook/rocksdb/issues/1963 to cut file before skippable grandparent file. Which is for use case like user adding 2 or more non-overlapping data range at the same time, it can reduce the overlapping of 2 datasets in the lower levels. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10655 Reviewed By: cbi42 Differential Revision: D39552321 Pulled By: jay-zhuang fbshipit-source-id: 640d15f159ab0cd973f2426cfc3af266fc8bdde2
2 years ago
s = current_output().validator.Add(key, value);
if (!s.ok()) {
return s;
}
builder_->Add(key, value);
stats_.num_output_records++;
current_output_file_size_ = builder_->EstimatedFileSize();
if (blob_garbage_meter_) {
s = blob_garbage_meter_->ProcessOutFlow(key, value);
}
if (!s.ok()) {
return s;
}
const ParsedInternalKey& ikey = c_iter.ikey();
s = current_output().meta.UpdateBoundaries(key, value, ikey.sequence,
ikey.type);
if (partitioner_) {
last_key_for_partitioner_.assign(c_iter.user_key().data_,
c_iter.user_key().size_);
}
return s;
}
Status CompactionOutputs::AddRangeDels(
User-defined timestamp support for `DeleteRange()` (#10661) Summary: Add user-defined timestamp support for range deletion. The new API is `DeleteRange(opt, cf, begin_key, end_key, ts)`. Most of the change is to update the comparator to compare without timestamp. Other than that, major changes are - internal range tombstone data structures (`FragmentedRangeTombstoneList`, `RangeTombstone`, etc.) to store timestamps. - Garbage collection of range tombstones and range tombstone covered keys during compaction. - Get()/MultiGet() to return the timestamp of a range tombstone when needed. - Get/Iterator with range tombstones bounded by readoptions.timestamp. - timestamp crash test now issues DeleteRange by default. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10661 Test Plan: - Added unit test: `make check` - Stress test: `python3 tools/db_crashtest.py --enable_ts whitebox --readpercent=57 --prefixpercent=4 --writepercent=25 -delpercent=5 --iterpercent=5 --delrangepercent=4` - Ran `db_bench` to measure regression when timestamp is not enabled. The tests are for write (with some range deletion) and iterate with DB fitting in memory: `./db_bench--benchmarks=fillrandom,seekrandom --writes_per_range_tombstone=200 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=500000 --reads=500000 --seek_nexts=10 --disable_auto_compactions -disable_wal=true --max_num_range_tombstones=1000`. Did not see consistent regression in no timestamp case. | micros/op | fillrandom | seekrandom | | --- | --- | --- | |main| 2.58 |10.96| |PR 10661| 2.68 |10.63| Reviewed By: riversand963 Differential Revision: D39441192 Pulled By: cbi42 fbshipit-source-id: f05aca3c41605caf110daf0ff405919f300ddec2
2 years ago
const Slice* comp_start_user_key, const Slice* comp_end_user_key,
CompactionIterationStats& range_del_out_stats, bool bottommost_level,
const InternalKeyComparator& icmp, SequenceNumber earliest_snapshot,
User-defined timestamp support for `DeleteRange()` (#10661) Summary: Add user-defined timestamp support for range deletion. The new API is `DeleteRange(opt, cf, begin_key, end_key, ts)`. Most of the change is to update the comparator to compare without timestamp. Other than that, major changes are - internal range tombstone data structures (`FragmentedRangeTombstoneList`, `RangeTombstone`, etc.) to store timestamps. - Garbage collection of range tombstones and range tombstone covered keys during compaction. - Get()/MultiGet() to return the timestamp of a range tombstone when needed. - Get/Iterator with range tombstones bounded by readoptions.timestamp. - timestamp crash test now issues DeleteRange by default. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10661 Test Plan: - Added unit test: `make check` - Stress test: `python3 tools/db_crashtest.py --enable_ts whitebox --readpercent=57 --prefixpercent=4 --writepercent=25 -delpercent=5 --iterpercent=5 --delrangepercent=4` - Ran `db_bench` to measure regression when timestamp is not enabled. The tests are for write (with some range deletion) and iterate with DB fitting in memory: `./db_bench--benchmarks=fillrandom,seekrandom --writes_per_range_tombstone=200 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=500000 --reads=500000 --seek_nexts=10 --disable_auto_compactions -disable_wal=true --max_num_range_tombstones=1000`. Did not see consistent regression in no timestamp case. | micros/op | fillrandom | seekrandom | | --- | --- | --- | |main| 2.58 |10.96| |PR 10661| 2.68 |10.63| Reviewed By: riversand963 Differential Revision: D39441192 Pulled By: cbi42 fbshipit-source-id: f05aca3c41605caf110daf0ff405919f300ddec2
2 years ago
const Slice& next_table_min_key, const std::string& full_history_ts_low) {
assert(HasRangeDel());
FileMetaData& meta = current_output().meta;
const Comparator* ucmp = icmp.user_comparator();
Slice lower_bound_guard, upper_bound_guard;
std::string smallest_user_key;
const Slice *lower_bound, *upper_bound;
bool lower_bound_from_sub_compact = false;
// The following example does not happen since
// CompactionOutput::ShouldStopBefore() always return false for the first
// point key. But we should consider removing this dependency. Suppose for the
// first compaction output file,
// - next_table_min_key.user_key == comp_start_user_key
// - no point key is in the output file
// - there is a range tombstone @seqno to be added that covers
// comp_start_user_key
// Then meta.smallest will be set to comp_start_user_key@seqno
// and meta.largest will be set to comp_start_user_key@kMaxSequenceNumber
// which violates the assumption that meta.smallest should be <= meta.largest.
size_t output_size = outputs_.size();
if (output_size == 1) {
// For the first output table, include range tombstones before the min
// key but after the subcompaction boundary.
User-defined timestamp support for `DeleteRange()` (#10661) Summary: Add user-defined timestamp support for range deletion. The new API is `DeleteRange(opt, cf, begin_key, end_key, ts)`. Most of the change is to update the comparator to compare without timestamp. Other than that, major changes are - internal range tombstone data structures (`FragmentedRangeTombstoneList`, `RangeTombstone`, etc.) to store timestamps. - Garbage collection of range tombstones and range tombstone covered keys during compaction. - Get()/MultiGet() to return the timestamp of a range tombstone when needed. - Get/Iterator with range tombstones bounded by readoptions.timestamp. - timestamp crash test now issues DeleteRange by default. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10661 Test Plan: - Added unit test: `make check` - Stress test: `python3 tools/db_crashtest.py --enable_ts whitebox --readpercent=57 --prefixpercent=4 --writepercent=25 -delpercent=5 --iterpercent=5 --delrangepercent=4` - Ran `db_bench` to measure regression when timestamp is not enabled. The tests are for write (with some range deletion) and iterate with DB fitting in memory: `./db_bench--benchmarks=fillrandom,seekrandom --writes_per_range_tombstone=200 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=500000 --reads=500000 --seek_nexts=10 --disable_auto_compactions -disable_wal=true --max_num_range_tombstones=1000`. Did not see consistent regression in no timestamp case. | micros/op | fillrandom | seekrandom | | --- | --- | --- | |main| 2.58 |10.96| |PR 10661| 2.68 |10.63| Reviewed By: riversand963 Differential Revision: D39441192 Pulled By: cbi42 fbshipit-source-id: f05aca3c41605caf110daf0ff405919f300ddec2
2 years ago
lower_bound = comp_start_user_key;
lower_bound_from_sub_compact = true;
} else if (meta.smallest.size() > 0) {
// For subsequent output tables, only include range tombstones from min
// key onwards since the previous file was extended to contain range
// tombstones falling before min key.
smallest_user_key = meta.smallest.user_key().ToString(false /*hex*/);
lower_bound_guard = Slice(smallest_user_key);
lower_bound = &lower_bound_guard;
} else {
lower_bound = nullptr;
}
if (!next_table_min_key.empty()) {
// This may be the last file in the subcompaction in some cases, so we
// need to compare the end key of subcompaction with the next file start
// key. When the end key is chosen by the subcompaction, we know that
// it must be the biggest key in output file. Therefore, it is safe to
// use the smaller key as the upper bound of the output file, to ensure
// that there is no overlapping between different output files.
upper_bound_guard = ExtractUserKey(next_table_min_key);
User-defined timestamp support for `DeleteRange()` (#10661) Summary: Add user-defined timestamp support for range deletion. The new API is `DeleteRange(opt, cf, begin_key, end_key, ts)`. Most of the change is to update the comparator to compare without timestamp. Other than that, major changes are - internal range tombstone data structures (`FragmentedRangeTombstoneList`, `RangeTombstone`, etc.) to store timestamps. - Garbage collection of range tombstones and range tombstone covered keys during compaction. - Get()/MultiGet() to return the timestamp of a range tombstone when needed. - Get/Iterator with range tombstones bounded by readoptions.timestamp. - timestamp crash test now issues DeleteRange by default. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10661 Test Plan: - Added unit test: `make check` - Stress test: `python3 tools/db_crashtest.py --enable_ts whitebox --readpercent=57 --prefixpercent=4 --writepercent=25 -delpercent=5 --iterpercent=5 --delrangepercent=4` - Ran `db_bench` to measure regression when timestamp is not enabled. The tests are for write (with some range deletion) and iterate with DB fitting in memory: `./db_bench--benchmarks=fillrandom,seekrandom --writes_per_range_tombstone=200 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=500000 --reads=500000 --seek_nexts=10 --disable_auto_compactions -disable_wal=true --max_num_range_tombstones=1000`. Did not see consistent regression in no timestamp case. | micros/op | fillrandom | seekrandom | | --- | --- | --- | |main| 2.58 |10.96| |PR 10661| 2.68 |10.63| Reviewed By: riversand963 Differential Revision: D39441192 Pulled By: cbi42 fbshipit-source-id: f05aca3c41605caf110daf0ff405919f300ddec2
2 years ago
if (comp_end_user_key != nullptr &&
ucmp->CompareWithoutTimestamp(upper_bound_guard, *comp_end_user_key) >=
0) {
upper_bound = comp_end_user_key;
} else {
upper_bound = &upper_bound_guard;
}
} else {
// This is the last file in the subcompaction, so extend until the
// subcompaction ends.
User-defined timestamp support for `DeleteRange()` (#10661) Summary: Add user-defined timestamp support for range deletion. The new API is `DeleteRange(opt, cf, begin_key, end_key, ts)`. Most of the change is to update the comparator to compare without timestamp. Other than that, major changes are - internal range tombstone data structures (`FragmentedRangeTombstoneList`, `RangeTombstone`, etc.) to store timestamps. - Garbage collection of range tombstones and range tombstone covered keys during compaction. - Get()/MultiGet() to return the timestamp of a range tombstone when needed. - Get/Iterator with range tombstones bounded by readoptions.timestamp. - timestamp crash test now issues DeleteRange by default. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10661 Test Plan: - Added unit test: `make check` - Stress test: `python3 tools/db_crashtest.py --enable_ts whitebox --readpercent=57 --prefixpercent=4 --writepercent=25 -delpercent=5 --iterpercent=5 --delrangepercent=4` - Ran `db_bench` to measure regression when timestamp is not enabled. The tests are for write (with some range deletion) and iterate with DB fitting in memory: `./db_bench--benchmarks=fillrandom,seekrandom --writes_per_range_tombstone=200 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=500000 --reads=500000 --seek_nexts=10 --disable_auto_compactions -disable_wal=true --max_num_range_tombstones=1000`. Did not see consistent regression in no timestamp case. | micros/op | fillrandom | seekrandom | | --- | --- | --- | |main| 2.58 |10.96| |PR 10661| 2.68 |10.63| Reviewed By: riversand963 Differential Revision: D39441192 Pulled By: cbi42 fbshipit-source-id: f05aca3c41605caf110daf0ff405919f300ddec2
2 years ago
upper_bound = comp_end_user_key;
}
bool has_overlapping_endpoints;
if (upper_bound != nullptr && meta.largest.size() > 0) {
User-defined timestamp support for `DeleteRange()` (#10661) Summary: Add user-defined timestamp support for range deletion. The new API is `DeleteRange(opt, cf, begin_key, end_key, ts)`. Most of the change is to update the comparator to compare without timestamp. Other than that, major changes are - internal range tombstone data structures (`FragmentedRangeTombstoneList`, `RangeTombstone`, etc.) to store timestamps. - Garbage collection of range tombstones and range tombstone covered keys during compaction. - Get()/MultiGet() to return the timestamp of a range tombstone when needed. - Get/Iterator with range tombstones bounded by readoptions.timestamp. - timestamp crash test now issues DeleteRange by default. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10661 Test Plan: - Added unit test: `make check` - Stress test: `python3 tools/db_crashtest.py --enable_ts whitebox --readpercent=57 --prefixpercent=4 --writepercent=25 -delpercent=5 --iterpercent=5 --delrangepercent=4` - Ran `db_bench` to measure regression when timestamp is not enabled. The tests are for write (with some range deletion) and iterate with DB fitting in memory: `./db_bench--benchmarks=fillrandom,seekrandom --writes_per_range_tombstone=200 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=500000 --reads=500000 --seek_nexts=10 --disable_auto_compactions -disable_wal=true --max_num_range_tombstones=1000`. Did not see consistent regression in no timestamp case. | micros/op | fillrandom | seekrandom | | --- | --- | --- | |main| 2.58 |10.96| |PR 10661| 2.68 |10.63| Reviewed By: riversand963 Differential Revision: D39441192 Pulled By: cbi42 fbshipit-source-id: f05aca3c41605caf110daf0ff405919f300ddec2
2 years ago
has_overlapping_endpoints = ucmp->CompareWithoutTimestamp(
meta.largest.user_key(), *upper_bound) == 0;
} else {
has_overlapping_endpoints = false;
}
// The end key of the subcompaction must be bigger or equal to the upper
// bound. If the end of subcompaction is null or the upper bound is null,
// it means that this file is the last file in the compaction. So there
// will be no overlapping between this file and others.
User-defined timestamp support for `DeleteRange()` (#10661) Summary: Add user-defined timestamp support for range deletion. The new API is `DeleteRange(opt, cf, begin_key, end_key, ts)`. Most of the change is to update the comparator to compare without timestamp. Other than that, major changes are - internal range tombstone data structures (`FragmentedRangeTombstoneList`, `RangeTombstone`, etc.) to store timestamps. - Garbage collection of range tombstones and range tombstone covered keys during compaction. - Get()/MultiGet() to return the timestamp of a range tombstone when needed. - Get/Iterator with range tombstones bounded by readoptions.timestamp. - timestamp crash test now issues DeleteRange by default. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10661 Test Plan: - Added unit test: `make check` - Stress test: `python3 tools/db_crashtest.py --enable_ts whitebox --readpercent=57 --prefixpercent=4 --writepercent=25 -delpercent=5 --iterpercent=5 --delrangepercent=4` - Ran `db_bench` to measure regression when timestamp is not enabled. The tests are for write (with some range deletion) and iterate with DB fitting in memory: `./db_bench--benchmarks=fillrandom,seekrandom --writes_per_range_tombstone=200 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=500000 --reads=500000 --seek_nexts=10 --disable_auto_compactions -disable_wal=true --max_num_range_tombstones=1000`. Did not see consistent regression in no timestamp case. | micros/op | fillrandom | seekrandom | | --- | --- | --- | |main| 2.58 |10.96| |PR 10661| 2.68 |10.63| Reviewed By: riversand963 Differential Revision: D39441192 Pulled By: cbi42 fbshipit-source-id: f05aca3c41605caf110daf0ff405919f300ddec2
2 years ago
assert(comp_end_user_key == nullptr || upper_bound == nullptr ||
ucmp->CompareWithoutTimestamp(*upper_bound, *comp_end_user_key) <= 0);
auto it = range_del_agg_->NewIterator(lower_bound, upper_bound,
has_overlapping_endpoints);
// Position the range tombstone output iterator. There may be tombstone
// fragments that are entirely out of range, so make sure that we do not
// include those.
if (lower_bound != nullptr) {
it->Seek(*lower_bound);
} else {
it->SeekToFirst();
}
Slice last_tombstone_start_user_key{};
for (; it->Valid(); it->Next()) {
auto tombstone = it->Tombstone();
if (upper_bound != nullptr) {
User-defined timestamp support for `DeleteRange()` (#10661) Summary: Add user-defined timestamp support for range deletion. The new API is `DeleteRange(opt, cf, begin_key, end_key, ts)`. Most of the change is to update the comparator to compare without timestamp. Other than that, major changes are - internal range tombstone data structures (`FragmentedRangeTombstoneList`, `RangeTombstone`, etc.) to store timestamps. - Garbage collection of range tombstones and range tombstone covered keys during compaction. - Get()/MultiGet() to return the timestamp of a range tombstone when needed. - Get/Iterator with range tombstones bounded by readoptions.timestamp. - timestamp crash test now issues DeleteRange by default. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10661 Test Plan: - Added unit test: `make check` - Stress test: `python3 tools/db_crashtest.py --enable_ts whitebox --readpercent=57 --prefixpercent=4 --writepercent=25 -delpercent=5 --iterpercent=5 --delrangepercent=4` - Ran `db_bench` to measure regression when timestamp is not enabled. The tests are for write (with some range deletion) and iterate with DB fitting in memory: `./db_bench--benchmarks=fillrandom,seekrandom --writes_per_range_tombstone=200 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=500000 --reads=500000 --seek_nexts=10 --disable_auto_compactions -disable_wal=true --max_num_range_tombstones=1000`. Did not see consistent regression in no timestamp case. | micros/op | fillrandom | seekrandom | | --- | --- | --- | |main| 2.58 |10.96| |PR 10661| 2.68 |10.63| Reviewed By: riversand963 Differential Revision: D39441192 Pulled By: cbi42 fbshipit-source-id: f05aca3c41605caf110daf0ff405919f300ddec2
2 years ago
int cmp =
ucmp->CompareWithoutTimestamp(*upper_bound, tombstone.start_key_);
// Tombstones starting after upper_bound only need to be included in
// the next table.
// If the current SST ends before upper_bound, i.e.,
// `has_overlapping_endpoints == false`, we can also skip over range
// tombstones that start exactly at upper_bound. Such range
// tombstones will be included in the next file and are not relevant
// to the point keys or endpoints of the current file.
// If the current SST ends at the same user key at upper_bound,
// i.e., `has_overlapping_endpoints == true`, AND the tombstone has
// the same start key as upper_bound, i.e., cmp == 0, then
// the tombstone is relevant only if the tombstone's sequence number
// is no larger than this file's largest key's sequence number. This
// is because the upper bound to truncate this file's range tombstone
// will be meta.largest in this case, and any tombstone that starts after
// it will not be relevant.
if (cmp < 0) {
break;
} else if (cmp == 0) {
if (!has_overlapping_endpoints ||
tombstone.seq_ < GetInternalKeySeqno(meta.largest.Encode())) {
break;
}
}
}
User-defined timestamp support for `DeleteRange()` (#10661) Summary: Add user-defined timestamp support for range deletion. The new API is `DeleteRange(opt, cf, begin_key, end_key, ts)`. Most of the change is to update the comparator to compare without timestamp. Other than that, major changes are - internal range tombstone data structures (`FragmentedRangeTombstoneList`, `RangeTombstone`, etc.) to store timestamps. - Garbage collection of range tombstones and range tombstone covered keys during compaction. - Get()/MultiGet() to return the timestamp of a range tombstone when needed. - Get/Iterator with range tombstones bounded by readoptions.timestamp. - timestamp crash test now issues DeleteRange by default. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10661 Test Plan: - Added unit test: `make check` - Stress test: `python3 tools/db_crashtest.py --enable_ts whitebox --readpercent=57 --prefixpercent=4 --writepercent=25 -delpercent=5 --iterpercent=5 --delrangepercent=4` - Ran `db_bench` to measure regression when timestamp is not enabled. The tests are for write (with some range deletion) and iterate with DB fitting in memory: `./db_bench--benchmarks=fillrandom,seekrandom --writes_per_range_tombstone=200 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=500000 --reads=500000 --seek_nexts=10 --disable_auto_compactions -disable_wal=true --max_num_range_tombstones=1000`. Did not see consistent regression in no timestamp case. | micros/op | fillrandom | seekrandom | | --- | --- | --- | |main| 2.58 |10.96| |PR 10661| 2.68 |10.63| Reviewed By: riversand963 Differential Revision: D39441192 Pulled By: cbi42 fbshipit-source-id: f05aca3c41605caf110daf0ff405919f300ddec2
2 years ago
const size_t ts_sz = ucmp->timestamp_size();
// Garbage collection for range tombstones.
// If user-defined timestamp is enabled, range tombstones are dropped if
// they are at bottommost_level, below full_history_ts_low and not visible
// in any snapshot. trim_ts_ is passed to the constructor for
// range_del_agg_, and range_del_agg_ internally drops tombstones above
// trim_ts_.
if (bottommost_level && tombstone.seq_ <= earliest_snapshot &&
(ts_sz == 0 ||
(!full_history_ts_low.empty() &&
ucmp->CompareTimestamp(tombstone.ts_, full_history_ts_low) < 0))) {
// TODO(andrewkr): tombstones that span multiple output files are
// counted for each compaction output file, so lots of double
// counting.
range_del_out_stats.num_range_del_drop_obsolete++;
range_del_out_stats.num_record_drop_obsolete++;
continue;
}
auto kv = tombstone.Serialize();
assert(lower_bound == nullptr ||
User-defined timestamp support for `DeleteRange()` (#10661) Summary: Add user-defined timestamp support for range deletion. The new API is `DeleteRange(opt, cf, begin_key, end_key, ts)`. Most of the change is to update the comparator to compare without timestamp. Other than that, major changes are - internal range tombstone data structures (`FragmentedRangeTombstoneList`, `RangeTombstone`, etc.) to store timestamps. - Garbage collection of range tombstones and range tombstone covered keys during compaction. - Get()/MultiGet() to return the timestamp of a range tombstone when needed. - Get/Iterator with range tombstones bounded by readoptions.timestamp. - timestamp crash test now issues DeleteRange by default. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10661 Test Plan: - Added unit test: `make check` - Stress test: `python3 tools/db_crashtest.py --enable_ts whitebox --readpercent=57 --prefixpercent=4 --writepercent=25 -delpercent=5 --iterpercent=5 --delrangepercent=4` - Ran `db_bench` to measure regression when timestamp is not enabled. The tests are for write (with some range deletion) and iterate with DB fitting in memory: `./db_bench--benchmarks=fillrandom,seekrandom --writes_per_range_tombstone=200 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=500000 --reads=500000 --seek_nexts=10 --disable_auto_compactions -disable_wal=true --max_num_range_tombstones=1000`. Did not see consistent regression in no timestamp case. | micros/op | fillrandom | seekrandom | | --- | --- | --- | |main| 2.58 |10.96| |PR 10661| 2.68 |10.63| Reviewed By: riversand963 Differential Revision: D39441192 Pulled By: cbi42 fbshipit-source-id: f05aca3c41605caf110daf0ff405919f300ddec2
2 years ago
ucmp->CompareWithoutTimestamp(*lower_bound, kv.second) < 0);
// Range tombstone is not supported by output validator yet.
builder_->Add(kv.first.Encode(), kv.second);
Include estimated bytes deleted by range tombstones in compensated file size (#10734) Summary: compensate file sizes in compaction picking so files with range tombstones are preferred, such that they get compacted down earlier as they tend to delete a lot of data. This PR adds a `compensated_range_deletion_size` field in FileMeta that is computed during Flush/Compaction and persisted in MANIFEST. This value is added to `compensated_file_size` which will be used for compaction picking. Currently, for a file in level L, `compensated_range_deletion_size` is set to the estimated bytes deleted by range tombstone of this file in all levels > L. This helps to reduce space amp when data in older levels are covered by range tombstones in level L. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10734 Test Plan: - Added unit tests. - benchmark to check if the above definition `compensated_range_deletion_size` is reducing space amp as intended, without affecting write amp too much. The experiment set up favorable for this optimization: large range tombstone issued infrequently. Command used: ``` ./db_bench -benchmarks=fillrandom,waitforcompaction,stats,levelstats -use_existing_db=false -avoid_flush_during_recovery=true -write_buffer_size=33554432 -level_compaction_dynamic_level_bytes=true -max_background_jobs=8 -max_bytes_for_level_base=134217728 -target_file_size_base=33554432 -writes_per_range_tombstone=500000 -range_tombstone_width=5000000 -num=50000000 -benchmark_write_rate_limit=8388608 -threads=16 -duration=1800 --max_num_range_tombstones=1000000000 ``` In this experiment, each thread wrote 16 range tombstones over the duration of 30 minutes, each range tombstone has width 5M that is the 10% of the key space width. Results shows this PR generates a smaller DB size. Compaction stats from this PR: ``` Level Files Size Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) Moved(GB) W-Amp Rd(MB/s) Wr(MB/s) Comp(sec) CompMergeCPU(sec) Comp(cnt) Avg(sec) KeyIn KeyDrop Rblob(GB) Wblob(GB) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ L0 2/0 31.54 MB 0.5 0.0 0.0 0.0 8.4 8.4 0.0 1.0 0.0 63.4 135.56 110.94 544 0.249 0 0 0.0 0.0 L4 3/0 96.55 MB 0.8 18.5 6.7 11.8 18.4 6.6 0.0 2.7 65.3 64.9 290.08 284.03 108 2.686 284M 1957K 0.0 0.0 L5 15/0 404.41 MB 1.0 19.1 7.7 11.4 18.8 7.4 0.3 2.5 66.6 65.7 292.93 285.34 220 1.332 293M 3808K 0.0 0.0 L6 143/0 4.12 GB 0.0 45.0 7.5 37.5 41.6 4.1 0.0 5.5 71.2 65.9 647.00 632.66 251 2.578 739M 47M 0.0 0.0 Sum 163/0 4.64 GB 0.0 82.6 21.9 60.7 87.2 26.5 0.3 10.4 61.9 65.4 1365.58 1312.97 1123 1.216 1318M 52M 0.0 0.0 ``` Compaction stats from main: ``` Level Files Size Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) Moved(GB) W-Amp Rd(MB/s) Wr(MB/s) Comp(sec) CompMergeCPU(sec) Comp(cnt) Avg(sec) KeyIn KeyDrop Rblob(GB) Wblob(GB) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ L0 0/0 0.00 KB 0.0 0.0 0.0 0.0 8.4 8.4 0.0 1.0 0.0 60.5 142.12 115.89 569 0.250 0 0 0.0 0.0 L4 3/0 85.68 MB 1.0 17.7 6.8 10.9 17.6 6.7 0.0 2.6 62.7 62.3 289.05 281.79 112 2.581 272M 2309K 0.0 0.0 L5 11/0 293.73 MB 1.0 18.8 7.5 11.2 18.5 7.2 0.5 2.5 64.9 63.9 296.07 288.50 220 1.346 288M 4365K 0.0 0.0 L6 130/0 3.94 GB 0.0 51.5 7.6 43.9 47.9 3.9 0.0 6.3 67.2 62.4 784.95 765.92 258 3.042 848M 51M 0.0 0.0 Sum 144/0 4.31 GB 0.0 88.0 21.9 66.0 92.3 26.3 0.5 11.0 59.6 62.5 1512.19 1452.09 1159 1.305 1409M 58M 0.0 0.0``` Reviewed By: ajkr Differential Revision: D39834713 Pulled By: cbi42 fbshipit-source-id: fe9341040b8704a8fbb10cad5cf5c43e962c7e6b
2 years ago
InternalKey tombstone_start = std::move(kv.first);
InternalKey smallest_candidate{tombstone_start};
if (lower_bound != nullptr &&
User-defined timestamp support for `DeleteRange()` (#10661) Summary: Add user-defined timestamp support for range deletion. The new API is `DeleteRange(opt, cf, begin_key, end_key, ts)`. Most of the change is to update the comparator to compare without timestamp. Other than that, major changes are - internal range tombstone data structures (`FragmentedRangeTombstoneList`, `RangeTombstone`, etc.) to store timestamps. - Garbage collection of range tombstones and range tombstone covered keys during compaction. - Get()/MultiGet() to return the timestamp of a range tombstone when needed. - Get/Iterator with range tombstones bounded by readoptions.timestamp. - timestamp crash test now issues DeleteRange by default. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10661 Test Plan: - Added unit test: `make check` - Stress test: `python3 tools/db_crashtest.py --enable_ts whitebox --readpercent=57 --prefixpercent=4 --writepercent=25 -delpercent=5 --iterpercent=5 --delrangepercent=4` - Ran `db_bench` to measure regression when timestamp is not enabled. The tests are for write (with some range deletion) and iterate with DB fitting in memory: `./db_bench--benchmarks=fillrandom,seekrandom --writes_per_range_tombstone=200 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=500000 --reads=500000 --seek_nexts=10 --disable_auto_compactions -disable_wal=true --max_num_range_tombstones=1000`. Did not see consistent regression in no timestamp case. | micros/op | fillrandom | seekrandom | | --- | --- | --- | |main| 2.58 |10.96| |PR 10661| 2.68 |10.63| Reviewed By: riversand963 Differential Revision: D39441192 Pulled By: cbi42 fbshipit-source-id: f05aca3c41605caf110daf0ff405919f300ddec2
2 years ago
ucmp->CompareWithoutTimestamp(smallest_candidate.user_key(),
*lower_bound) <= 0) {
// Pretend the smallest key has the same user key as lower_bound
// (the max key in the previous table or subcompaction) in order for
// files to appear key-space partitioned.
User-defined timestamp support for `DeleteRange()` (#10661) Summary: Add user-defined timestamp support for range deletion. The new API is `DeleteRange(opt, cf, begin_key, end_key, ts)`. Most of the change is to update the comparator to compare without timestamp. Other than that, major changes are - internal range tombstone data structures (`FragmentedRangeTombstoneList`, `RangeTombstone`, etc.) to store timestamps. - Garbage collection of range tombstones and range tombstone covered keys during compaction. - Get()/MultiGet() to return the timestamp of a range tombstone when needed. - Get/Iterator with range tombstones bounded by readoptions.timestamp. - timestamp crash test now issues DeleteRange by default. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10661 Test Plan: - Added unit test: `make check` - Stress test: `python3 tools/db_crashtest.py --enable_ts whitebox --readpercent=57 --prefixpercent=4 --writepercent=25 -delpercent=5 --iterpercent=5 --delrangepercent=4` - Ran `db_bench` to measure regression when timestamp is not enabled. The tests are for write (with some range deletion) and iterate with DB fitting in memory: `./db_bench--benchmarks=fillrandom,seekrandom --writes_per_range_tombstone=200 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=500000 --reads=500000 --seek_nexts=10 --disable_auto_compactions -disable_wal=true --max_num_range_tombstones=1000`. Did not see consistent regression in no timestamp case. | micros/op | fillrandom | seekrandom | | --- | --- | --- | |main| 2.58 |10.96| |PR 10661| 2.68 |10.63| Reviewed By: riversand963 Differential Revision: D39441192 Pulled By: cbi42 fbshipit-source-id: f05aca3c41605caf110daf0ff405919f300ddec2
2 years ago
if (lower_bound_from_sub_compact) {
// When lower_bound is chosen by a subcompaction
// (lower_bound_from_sub_compact), we know that subcompactions over
// smaller keys cannot contain any keys at lower_bound. We also know
// that smaller subcompactions exist, because otherwise the
// subcompaction woud be unbounded on the left. As a result, we know
// that no other files on the output level will contain actual keys at
// lower_bound (an output file may have a largest key of
// lower_bound@kMaxSequenceNumber, but this only indicates a large range
// tombstone was truncated). Therefore, it is safe to use the
// tombstone's sequence number, to ensure that keys at lower_bound at
// lower levels are covered by truncated tombstones.
User-defined timestamp support for `DeleteRange()` (#10661) Summary: Add user-defined timestamp support for range deletion. The new API is `DeleteRange(opt, cf, begin_key, end_key, ts)`. Most of the change is to update the comparator to compare without timestamp. Other than that, major changes are - internal range tombstone data structures (`FragmentedRangeTombstoneList`, `RangeTombstone`, etc.) to store timestamps. - Garbage collection of range tombstones and range tombstone covered keys during compaction. - Get()/MultiGet() to return the timestamp of a range tombstone when needed. - Get/Iterator with range tombstones bounded by readoptions.timestamp. - timestamp crash test now issues DeleteRange by default. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10661 Test Plan: - Added unit test: `make check` - Stress test: `python3 tools/db_crashtest.py --enable_ts whitebox --readpercent=57 --prefixpercent=4 --writepercent=25 -delpercent=5 --iterpercent=5 --delrangepercent=4` - Ran `db_bench` to measure regression when timestamp is not enabled. The tests are for write (with some range deletion) and iterate with DB fitting in memory: `./db_bench--benchmarks=fillrandom,seekrandom --writes_per_range_tombstone=200 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=500000 --reads=500000 --seek_nexts=10 --disable_auto_compactions -disable_wal=true --max_num_range_tombstones=1000`. Did not see consistent regression in no timestamp case. | micros/op | fillrandom | seekrandom | | --- | --- | --- | |main| 2.58 |10.96| |PR 10661| 2.68 |10.63| Reviewed By: riversand963 Differential Revision: D39441192 Pulled By: cbi42 fbshipit-source-id: f05aca3c41605caf110daf0ff405919f300ddec2
2 years ago
if (ts_sz) {
assert(tombstone.ts_.size() == ts_sz);
smallest_candidate = InternalKey(*lower_bound, tombstone.seq_,
kTypeRangeDeletion, tombstone.ts_);
} else {
smallest_candidate =
InternalKey(*lower_bound, tombstone.seq_, kTypeRangeDeletion);
}
} else {
// If lower_bound was chosen by the smallest data key in the file,
// choose lowest seqnum so this file's smallest internal key comes
// after the previous file's largest. The fake seqnum is OK because
// the read path's file-picking code only considers user key.
User-defined timestamp support for `DeleteRange()` (#10661) Summary: Add user-defined timestamp support for range deletion. The new API is `DeleteRange(opt, cf, begin_key, end_key, ts)`. Most of the change is to update the comparator to compare without timestamp. Other than that, major changes are - internal range tombstone data structures (`FragmentedRangeTombstoneList`, `RangeTombstone`, etc.) to store timestamps. - Garbage collection of range tombstones and range tombstone covered keys during compaction. - Get()/MultiGet() to return the timestamp of a range tombstone when needed. - Get/Iterator with range tombstones bounded by readoptions.timestamp. - timestamp crash test now issues DeleteRange by default. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10661 Test Plan: - Added unit test: `make check` - Stress test: `python3 tools/db_crashtest.py --enable_ts whitebox --readpercent=57 --prefixpercent=4 --writepercent=25 -delpercent=5 --iterpercent=5 --delrangepercent=4` - Ran `db_bench` to measure regression when timestamp is not enabled. The tests are for write (with some range deletion) and iterate with DB fitting in memory: `./db_bench--benchmarks=fillrandom,seekrandom --writes_per_range_tombstone=200 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=500000 --reads=500000 --seek_nexts=10 --disable_auto_compactions -disable_wal=true --max_num_range_tombstones=1000`. Did not see consistent regression in no timestamp case. | micros/op | fillrandom | seekrandom | | --- | --- | --- | |main| 2.58 |10.96| |PR 10661| 2.68 |10.63| Reviewed By: riversand963 Differential Revision: D39441192 Pulled By: cbi42 fbshipit-source-id: f05aca3c41605caf110daf0ff405919f300ddec2
2 years ago
smallest_candidate = InternalKey(*lower_bound, 0, kTypeRangeDeletion);
}
}
Include estimated bytes deleted by range tombstones in compensated file size (#10734) Summary: compensate file sizes in compaction picking so files with range tombstones are preferred, such that they get compacted down earlier as they tend to delete a lot of data. This PR adds a `compensated_range_deletion_size` field in FileMeta that is computed during Flush/Compaction and persisted in MANIFEST. This value is added to `compensated_file_size` which will be used for compaction picking. Currently, for a file in level L, `compensated_range_deletion_size` is set to the estimated bytes deleted by range tombstone of this file in all levels > L. This helps to reduce space amp when data in older levels are covered by range tombstones in level L. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10734 Test Plan: - Added unit tests. - benchmark to check if the above definition `compensated_range_deletion_size` is reducing space amp as intended, without affecting write amp too much. The experiment set up favorable for this optimization: large range tombstone issued infrequently. Command used: ``` ./db_bench -benchmarks=fillrandom,waitforcompaction,stats,levelstats -use_existing_db=false -avoid_flush_during_recovery=true -write_buffer_size=33554432 -level_compaction_dynamic_level_bytes=true -max_background_jobs=8 -max_bytes_for_level_base=134217728 -target_file_size_base=33554432 -writes_per_range_tombstone=500000 -range_tombstone_width=5000000 -num=50000000 -benchmark_write_rate_limit=8388608 -threads=16 -duration=1800 --max_num_range_tombstones=1000000000 ``` In this experiment, each thread wrote 16 range tombstones over the duration of 30 minutes, each range tombstone has width 5M that is the 10% of the key space width. Results shows this PR generates a smaller DB size. Compaction stats from this PR: ``` Level Files Size Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) Moved(GB) W-Amp Rd(MB/s) Wr(MB/s) Comp(sec) CompMergeCPU(sec) Comp(cnt) Avg(sec) KeyIn KeyDrop Rblob(GB) Wblob(GB) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ L0 2/0 31.54 MB 0.5 0.0 0.0 0.0 8.4 8.4 0.0 1.0 0.0 63.4 135.56 110.94 544 0.249 0 0 0.0 0.0 L4 3/0 96.55 MB 0.8 18.5 6.7 11.8 18.4 6.6 0.0 2.7 65.3 64.9 290.08 284.03 108 2.686 284M 1957K 0.0 0.0 L5 15/0 404.41 MB 1.0 19.1 7.7 11.4 18.8 7.4 0.3 2.5 66.6 65.7 292.93 285.34 220 1.332 293M 3808K 0.0 0.0 L6 143/0 4.12 GB 0.0 45.0 7.5 37.5 41.6 4.1 0.0 5.5 71.2 65.9 647.00 632.66 251 2.578 739M 47M 0.0 0.0 Sum 163/0 4.64 GB 0.0 82.6 21.9 60.7 87.2 26.5 0.3 10.4 61.9 65.4 1365.58 1312.97 1123 1.216 1318M 52M 0.0 0.0 ``` Compaction stats from main: ``` Level Files Size Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) Moved(GB) W-Amp Rd(MB/s) Wr(MB/s) Comp(sec) CompMergeCPU(sec) Comp(cnt) Avg(sec) KeyIn KeyDrop Rblob(GB) Wblob(GB) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ L0 0/0 0.00 KB 0.0 0.0 0.0 0.0 8.4 8.4 0.0 1.0 0.0 60.5 142.12 115.89 569 0.250 0 0 0.0 0.0 L4 3/0 85.68 MB 1.0 17.7 6.8 10.9 17.6 6.7 0.0 2.6 62.7 62.3 289.05 281.79 112 2.581 272M 2309K 0.0 0.0 L5 11/0 293.73 MB 1.0 18.8 7.5 11.2 18.5 7.2 0.5 2.5 64.9 63.9 296.07 288.50 220 1.346 288M 4365K 0.0 0.0 L6 130/0 3.94 GB 0.0 51.5 7.6 43.9 47.9 3.9 0.0 6.3 67.2 62.4 784.95 765.92 258 3.042 848M 51M 0.0 0.0 Sum 144/0 4.31 GB 0.0 88.0 21.9 66.0 92.3 26.3 0.5 11.0 59.6 62.5 1512.19 1452.09 1159 1.305 1409M 58M 0.0 0.0``` Reviewed By: ajkr Differential Revision: D39834713 Pulled By: cbi42 fbshipit-source-id: fe9341040b8704a8fbb10cad5cf5c43e962c7e6b
2 years ago
InternalKey tombstone_end = tombstone.SerializeEndKey();
InternalKey largest_candidate{tombstone_end};
if (upper_bound != nullptr &&
User-defined timestamp support for `DeleteRange()` (#10661) Summary: Add user-defined timestamp support for range deletion. The new API is `DeleteRange(opt, cf, begin_key, end_key, ts)`. Most of the change is to update the comparator to compare without timestamp. Other than that, major changes are - internal range tombstone data structures (`FragmentedRangeTombstoneList`, `RangeTombstone`, etc.) to store timestamps. - Garbage collection of range tombstones and range tombstone covered keys during compaction. - Get()/MultiGet() to return the timestamp of a range tombstone when needed. - Get/Iterator with range tombstones bounded by readoptions.timestamp. - timestamp crash test now issues DeleteRange by default. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10661 Test Plan: - Added unit test: `make check` - Stress test: `python3 tools/db_crashtest.py --enable_ts whitebox --readpercent=57 --prefixpercent=4 --writepercent=25 -delpercent=5 --iterpercent=5 --delrangepercent=4` - Ran `db_bench` to measure regression when timestamp is not enabled. The tests are for write (with some range deletion) and iterate with DB fitting in memory: `./db_bench--benchmarks=fillrandom,seekrandom --writes_per_range_tombstone=200 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=500000 --reads=500000 --seek_nexts=10 --disable_auto_compactions -disable_wal=true --max_num_range_tombstones=1000`. Did not see consistent regression in no timestamp case. | micros/op | fillrandom | seekrandom | | --- | --- | --- | |main| 2.58 |10.96| |PR 10661| 2.68 |10.63| Reviewed By: riversand963 Differential Revision: D39441192 Pulled By: cbi42 fbshipit-source-id: f05aca3c41605caf110daf0ff405919f300ddec2
2 years ago
ucmp->CompareWithoutTimestamp(*upper_bound,
largest_candidate.user_key()) <= 0) {
// Pretend the largest key has the same user key as upper_bound (the
// min key in the following table or subcompaction) in order for files
// to appear key-space partitioned.
//
// Choose highest seqnum so this file's largest internal key comes
// before the next file's/subcompaction's smallest. The fake seqnum is
// OK because the read path's file-picking code only considers the
// user key portion.
//
// Note Seek() also creates InternalKey with (user_key,
// kMaxSequenceNumber), but with kTypeDeletion (0x7) instead of
// kTypeRangeDeletion (0xF), so the range tombstone comes before the
// Seek() key in InternalKey's ordering. So Seek() will look in the
User-defined timestamp support for `DeleteRange()` (#10661) Summary: Add user-defined timestamp support for range deletion. The new API is `DeleteRange(opt, cf, begin_key, end_key, ts)`. Most of the change is to update the comparator to compare without timestamp. Other than that, major changes are - internal range tombstone data structures (`FragmentedRangeTombstoneList`, `RangeTombstone`, etc.) to store timestamps. - Garbage collection of range tombstones and range tombstone covered keys during compaction. - Get()/MultiGet() to return the timestamp of a range tombstone when needed. - Get/Iterator with range tombstones bounded by readoptions.timestamp. - timestamp crash test now issues DeleteRange by default. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10661 Test Plan: - Added unit test: `make check` - Stress test: `python3 tools/db_crashtest.py --enable_ts whitebox --readpercent=57 --prefixpercent=4 --writepercent=25 -delpercent=5 --iterpercent=5 --delrangepercent=4` - Ran `db_bench` to measure regression when timestamp is not enabled. The tests are for write (with some range deletion) and iterate with DB fitting in memory: `./db_bench--benchmarks=fillrandom,seekrandom --writes_per_range_tombstone=200 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=500000 --reads=500000 --seek_nexts=10 --disable_auto_compactions -disable_wal=true --max_num_range_tombstones=1000`. Did not see consistent regression in no timestamp case. | micros/op | fillrandom | seekrandom | | --- | --- | --- | |main| 2.58 |10.96| |PR 10661| 2.68 |10.63| Reviewed By: riversand963 Differential Revision: D39441192 Pulled By: cbi42 fbshipit-source-id: f05aca3c41605caf110daf0ff405919f300ddec2
2 years ago
// next file for the user key
if (ts_sz) {
static constexpr char kTsMax[] = "\xff\xff\xff\xff\xff\xff\xff\xff\xff";
if (ts_sz <= strlen(kTsMax)) {
largest_candidate =
InternalKey(*upper_bound, kMaxSequenceNumber, kTypeRangeDeletion,
Slice(kTsMax, ts_sz));
} else {
largest_candidate =
InternalKey(*upper_bound, kMaxSequenceNumber, kTypeRangeDeletion,
std::string(ts_sz, '\xff'));
}
} else {
largest_candidate =
InternalKey(*upper_bound, kMaxSequenceNumber, kTypeRangeDeletion);
}
}
meta.UpdateBoundariesForRange(smallest_candidate, largest_candidate,
tombstone.seq_, icmp);
Include estimated bytes deleted by range tombstones in compensated file size (#10734) Summary: compensate file sizes in compaction picking so files with range tombstones are preferred, such that they get compacted down earlier as they tend to delete a lot of data. This PR adds a `compensated_range_deletion_size` field in FileMeta that is computed during Flush/Compaction and persisted in MANIFEST. This value is added to `compensated_file_size` which will be used for compaction picking. Currently, for a file in level L, `compensated_range_deletion_size` is set to the estimated bytes deleted by range tombstone of this file in all levels > L. This helps to reduce space amp when data in older levels are covered by range tombstones in level L. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10734 Test Plan: - Added unit tests. - benchmark to check if the above definition `compensated_range_deletion_size` is reducing space amp as intended, without affecting write amp too much. The experiment set up favorable for this optimization: large range tombstone issued infrequently. Command used: ``` ./db_bench -benchmarks=fillrandom,waitforcompaction,stats,levelstats -use_existing_db=false -avoid_flush_during_recovery=true -write_buffer_size=33554432 -level_compaction_dynamic_level_bytes=true -max_background_jobs=8 -max_bytes_for_level_base=134217728 -target_file_size_base=33554432 -writes_per_range_tombstone=500000 -range_tombstone_width=5000000 -num=50000000 -benchmark_write_rate_limit=8388608 -threads=16 -duration=1800 --max_num_range_tombstones=1000000000 ``` In this experiment, each thread wrote 16 range tombstones over the duration of 30 minutes, each range tombstone has width 5M that is the 10% of the key space width. Results shows this PR generates a smaller DB size. Compaction stats from this PR: ``` Level Files Size Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) Moved(GB) W-Amp Rd(MB/s) Wr(MB/s) Comp(sec) CompMergeCPU(sec) Comp(cnt) Avg(sec) KeyIn KeyDrop Rblob(GB) Wblob(GB) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ L0 2/0 31.54 MB 0.5 0.0 0.0 0.0 8.4 8.4 0.0 1.0 0.0 63.4 135.56 110.94 544 0.249 0 0 0.0 0.0 L4 3/0 96.55 MB 0.8 18.5 6.7 11.8 18.4 6.6 0.0 2.7 65.3 64.9 290.08 284.03 108 2.686 284M 1957K 0.0 0.0 L5 15/0 404.41 MB 1.0 19.1 7.7 11.4 18.8 7.4 0.3 2.5 66.6 65.7 292.93 285.34 220 1.332 293M 3808K 0.0 0.0 L6 143/0 4.12 GB 0.0 45.0 7.5 37.5 41.6 4.1 0.0 5.5 71.2 65.9 647.00 632.66 251 2.578 739M 47M 0.0 0.0 Sum 163/0 4.64 GB 0.0 82.6 21.9 60.7 87.2 26.5 0.3 10.4 61.9 65.4 1365.58 1312.97 1123 1.216 1318M 52M 0.0 0.0 ``` Compaction stats from main: ``` Level Files Size Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) Moved(GB) W-Amp Rd(MB/s) Wr(MB/s) Comp(sec) CompMergeCPU(sec) Comp(cnt) Avg(sec) KeyIn KeyDrop Rblob(GB) Wblob(GB) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ L0 0/0 0.00 KB 0.0 0.0 0.0 0.0 8.4 8.4 0.0 1.0 0.0 60.5 142.12 115.89 569 0.250 0 0 0.0 0.0 L4 3/0 85.68 MB 1.0 17.7 6.8 10.9 17.6 6.7 0.0 2.6 62.7 62.3 289.05 281.79 112 2.581 272M 2309K 0.0 0.0 L5 11/0 293.73 MB 1.0 18.8 7.5 11.2 18.5 7.2 0.5 2.5 64.9 63.9 296.07 288.50 220 1.346 288M 4365K 0.0 0.0 L6 130/0 3.94 GB 0.0 51.5 7.6 43.9 47.9 3.9 0.0 6.3 67.2 62.4 784.95 765.92 258 3.042 848M 51M 0.0 0.0 Sum 144/0 4.31 GB 0.0 88.0 21.9 66.0 92.3 26.3 0.5 11.0 59.6 62.5 1512.19 1452.09 1159 1.305 1409M 58M 0.0 0.0``` Reviewed By: ajkr Differential Revision: D39834713 Pulled By: cbi42 fbshipit-source-id: fe9341040b8704a8fbb10cad5cf5c43e962c7e6b
2 years ago
if (!bottommost_level) {
bool start_user_key_changed =
last_tombstone_start_user_key.empty() ||
ucmp->CompareWithoutTimestamp(last_tombstone_start_user_key,
it->start_key()) < 0;
last_tombstone_start_user_key = it->start_key();
Include estimated bytes deleted by range tombstones in compensated file size (#10734) Summary: compensate file sizes in compaction picking so files with range tombstones are preferred, such that they get compacted down earlier as they tend to delete a lot of data. This PR adds a `compensated_range_deletion_size` field in FileMeta that is computed during Flush/Compaction and persisted in MANIFEST. This value is added to `compensated_file_size` which will be used for compaction picking. Currently, for a file in level L, `compensated_range_deletion_size` is set to the estimated bytes deleted by range tombstone of this file in all levels > L. This helps to reduce space amp when data in older levels are covered by range tombstones in level L. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10734 Test Plan: - Added unit tests. - benchmark to check if the above definition `compensated_range_deletion_size` is reducing space amp as intended, without affecting write amp too much. The experiment set up favorable for this optimization: large range tombstone issued infrequently. Command used: ``` ./db_bench -benchmarks=fillrandom,waitforcompaction,stats,levelstats -use_existing_db=false -avoid_flush_during_recovery=true -write_buffer_size=33554432 -level_compaction_dynamic_level_bytes=true -max_background_jobs=8 -max_bytes_for_level_base=134217728 -target_file_size_base=33554432 -writes_per_range_tombstone=500000 -range_tombstone_width=5000000 -num=50000000 -benchmark_write_rate_limit=8388608 -threads=16 -duration=1800 --max_num_range_tombstones=1000000000 ``` In this experiment, each thread wrote 16 range tombstones over the duration of 30 minutes, each range tombstone has width 5M that is the 10% of the key space width. Results shows this PR generates a smaller DB size. Compaction stats from this PR: ``` Level Files Size Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) Moved(GB) W-Amp Rd(MB/s) Wr(MB/s) Comp(sec) CompMergeCPU(sec) Comp(cnt) Avg(sec) KeyIn KeyDrop Rblob(GB) Wblob(GB) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ L0 2/0 31.54 MB 0.5 0.0 0.0 0.0 8.4 8.4 0.0 1.0 0.0 63.4 135.56 110.94 544 0.249 0 0 0.0 0.0 L4 3/0 96.55 MB 0.8 18.5 6.7 11.8 18.4 6.6 0.0 2.7 65.3 64.9 290.08 284.03 108 2.686 284M 1957K 0.0 0.0 L5 15/0 404.41 MB 1.0 19.1 7.7 11.4 18.8 7.4 0.3 2.5 66.6 65.7 292.93 285.34 220 1.332 293M 3808K 0.0 0.0 L6 143/0 4.12 GB 0.0 45.0 7.5 37.5 41.6 4.1 0.0 5.5 71.2 65.9 647.00 632.66 251 2.578 739M 47M 0.0 0.0 Sum 163/0 4.64 GB 0.0 82.6 21.9 60.7 87.2 26.5 0.3 10.4 61.9 65.4 1365.58 1312.97 1123 1.216 1318M 52M 0.0 0.0 ``` Compaction stats from main: ``` Level Files Size Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) Moved(GB) W-Amp Rd(MB/s) Wr(MB/s) Comp(sec) CompMergeCPU(sec) Comp(cnt) Avg(sec) KeyIn KeyDrop Rblob(GB) Wblob(GB) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ L0 0/0 0.00 KB 0.0 0.0 0.0 0.0 8.4 8.4 0.0 1.0 0.0 60.5 142.12 115.89 569 0.250 0 0 0.0 0.0 L4 3/0 85.68 MB 1.0 17.7 6.8 10.9 17.6 6.7 0.0 2.6 62.7 62.3 289.05 281.79 112 2.581 272M 2309K 0.0 0.0 L5 11/0 293.73 MB 1.0 18.8 7.5 11.2 18.5 7.2 0.5 2.5 64.9 63.9 296.07 288.50 220 1.346 288M 4365K 0.0 0.0 L6 130/0 3.94 GB 0.0 51.5 7.6 43.9 47.9 3.9 0.0 6.3 67.2 62.4 784.95 765.92 258 3.042 848M 51M 0.0 0.0 Sum 144/0 4.31 GB 0.0 88.0 21.9 66.0 92.3 26.3 0.5 11.0 59.6 62.5 1512.19 1452.09 1159 1.305 1409M 58M 0.0 0.0``` Reviewed By: ajkr Differential Revision: D39834713 Pulled By: cbi42 fbshipit-source-id: fe9341040b8704a8fbb10cad5cf5c43e962c7e6b
2 years ago
// Range tombstones are truncated at file boundaries
if (icmp.Compare(tombstone_start, meta.smallest) < 0) {
tombstone_start = meta.smallest;
}
if (icmp.Compare(tombstone_end, meta.largest) > 0) {
tombstone_end = meta.largest;
}
// this assertion validates invariant (2) in the comment below.
assert(icmp.Compare(tombstone_start, tombstone_end) <= 0);
if (start_user_key_changed) {
// if tombstone_start >= tombstone_end, then either no key range is
// covered, or that they have the same user key. If they have the same
// user key, then the internal key range should only be within this
// level, and no keys from older levels is covered.
if (ucmp->CompareWithoutTimestamp(tombstone_start.user_key(),
tombstone_end.user_key()) < 0) {
SizeApproximationOptions approx_opts;
approx_opts.files_size_error_margin = 0.1;
auto approximate_covered_size =
compaction_->input_version()->version_set()->ApproximateSize(
approx_opts, compaction_->input_version(),
tombstone_start.Encode(), tombstone_end.Encode(),
compaction_->output_level() + 1 /* start_level */,
-1 /* end_level */, kCompaction);
meta.compensated_range_deletion_size += approximate_covered_size;
}
}
Include estimated bytes deleted by range tombstones in compensated file size (#10734) Summary: compensate file sizes in compaction picking so files with range tombstones are preferred, such that they get compacted down earlier as they tend to delete a lot of data. This PR adds a `compensated_range_deletion_size` field in FileMeta that is computed during Flush/Compaction and persisted in MANIFEST. This value is added to `compensated_file_size` which will be used for compaction picking. Currently, for a file in level L, `compensated_range_deletion_size` is set to the estimated bytes deleted by range tombstone of this file in all levels > L. This helps to reduce space amp when data in older levels are covered by range tombstones in level L. Pull Request resolved: https://github.com/facebook/rocksdb/pull/10734 Test Plan: - Added unit tests. - benchmark to check if the above definition `compensated_range_deletion_size` is reducing space amp as intended, without affecting write amp too much. The experiment set up favorable for this optimization: large range tombstone issued infrequently. Command used: ``` ./db_bench -benchmarks=fillrandom,waitforcompaction,stats,levelstats -use_existing_db=false -avoid_flush_during_recovery=true -write_buffer_size=33554432 -level_compaction_dynamic_level_bytes=true -max_background_jobs=8 -max_bytes_for_level_base=134217728 -target_file_size_base=33554432 -writes_per_range_tombstone=500000 -range_tombstone_width=5000000 -num=50000000 -benchmark_write_rate_limit=8388608 -threads=16 -duration=1800 --max_num_range_tombstones=1000000000 ``` In this experiment, each thread wrote 16 range tombstones over the duration of 30 minutes, each range tombstone has width 5M that is the 10% of the key space width. Results shows this PR generates a smaller DB size. Compaction stats from this PR: ``` Level Files Size Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) Moved(GB) W-Amp Rd(MB/s) Wr(MB/s) Comp(sec) CompMergeCPU(sec) Comp(cnt) Avg(sec) KeyIn KeyDrop Rblob(GB) Wblob(GB) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ L0 2/0 31.54 MB 0.5 0.0 0.0 0.0 8.4 8.4 0.0 1.0 0.0 63.4 135.56 110.94 544 0.249 0 0 0.0 0.0 L4 3/0 96.55 MB 0.8 18.5 6.7 11.8 18.4 6.6 0.0 2.7 65.3 64.9 290.08 284.03 108 2.686 284M 1957K 0.0 0.0 L5 15/0 404.41 MB 1.0 19.1 7.7 11.4 18.8 7.4 0.3 2.5 66.6 65.7 292.93 285.34 220 1.332 293M 3808K 0.0 0.0 L6 143/0 4.12 GB 0.0 45.0 7.5 37.5 41.6 4.1 0.0 5.5 71.2 65.9 647.00 632.66 251 2.578 739M 47M 0.0 0.0 Sum 163/0 4.64 GB 0.0 82.6 21.9 60.7 87.2 26.5 0.3 10.4 61.9 65.4 1365.58 1312.97 1123 1.216 1318M 52M 0.0 0.0 ``` Compaction stats from main: ``` Level Files Size Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) Moved(GB) W-Amp Rd(MB/s) Wr(MB/s) Comp(sec) CompMergeCPU(sec) Comp(cnt) Avg(sec) KeyIn KeyDrop Rblob(GB) Wblob(GB) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ L0 0/0 0.00 KB 0.0 0.0 0.0 0.0 8.4 8.4 0.0 1.0 0.0 60.5 142.12 115.89 569 0.250 0 0 0.0 0.0 L4 3/0 85.68 MB 1.0 17.7 6.8 10.9 17.6 6.7 0.0 2.6 62.7 62.3 289.05 281.79 112 2.581 272M 2309K 0.0 0.0 L5 11/0 293.73 MB 1.0 18.8 7.5 11.2 18.5 7.2 0.5 2.5 64.9 63.9 296.07 288.50 220 1.346 288M 4365K 0.0 0.0 L6 130/0 3.94 GB 0.0 51.5 7.6 43.9 47.9 3.9 0.0 6.3 67.2 62.4 784.95 765.92 258 3.042 848M 51M 0.0 0.0 Sum 144/0 4.31 GB 0.0 88.0 21.9 66.0 92.3 26.3 0.5 11.0 59.6 62.5 1512.19 1452.09 1159 1.305 1409M 58M 0.0 0.0``` Reviewed By: ajkr Differential Revision: D39834713 Pulled By: cbi42 fbshipit-source-id: fe9341040b8704a8fbb10cad5cf5c43e962c7e6b
2 years ago
}
// TODO: show invariants that ensure all necessary range tombstones are
// added
// and that file boundaries ensure no coverage is lost.
// Each range tombstone with internal key range [tombstone_start,
// tombstone_end] is being added to the current compaction output file here.
// The range tombstone is going to be truncated at range [meta.smallest,
// meta.largest] during reading/scanning. We should maintain invariants
// (1) meta.smallest <= meta.largest and,
// (2) [tombstone_start, tombstone_end] and [meta.smallest, meta.largest]
// overlaps, as there is no point adding range tombstone with a range
// outside the file's range.
// Since `tombstone_end` is always some user_key@kMaxSeqno, it is okay to
// use either open or closed range. Using closed range here to make
// reasoning easier, and it is more consistent with an ongoing work that
// tries to simplify this method.
//
// There are two cases:
// Case 1. Output file has no point key:
// First we show this case only happens when the entire compaction output
// is range tombstone only. This is true if CompactionIterator does not
// emit any point key. Suppose CompactionIterator emits some point key.
// Based on the assumption that CompactionOutputs::ShouldStopBefore()
// always return false for the first point key, the first compaction
// output file always contains a point key. Each new compaction output
// file is created if there is a point key for which ShouldStopBefore()
// returns true, and the point key would be added to the new compaction
// output file. So each new compaction file always contains a point key.
// So Case 1 only happens when CompactionIterator does not emit any
// point key.
//
// To show (1) meta.smallest <= meta.largest:
// Since the compaction output is range tombstone only, `lower_bound` and
// `upper_bound` are either null or comp_start/end_user_key respectively.
// According to how UpdateBoundariesForRange() is implemented, it blindly
// updates meta.smallest and meta.largest to smallest_candidate and
// largest_candidate the first time it is called. Subsequently, it
// compares input parameter with meta.smallest and meta.largest and only
// updates them when input is smaller/larger. So we only need to show
// smallest_candidate <= largest_candidate the first time
// UpdateBoundariesForRange() is called. Here we show something stronger
// that smallest_candidate.user_key < largest_candidate.user_key always
// hold for Case 1.
// We assume comp_start_user_key < comp_end_user_key, if provided. We
// assume that tombstone_start < tombstone_end. This assumption is based
// on that each fragment in FragmentedTombstoneList has
// start_key < end_key (user_key) and that
// FragmentedTombstoneIterator::Tombstone() returns the pair
// (start_key@tombstone_seqno with op_type kTypeRangeDeletion, end_key).
// The logic in this loop sets smallest_candidate to
// max(tombstone_start.user_key, comp_start_user_key)@tombstone.seq_ with
// op_type kTypeRangeDeletion, largest_candidate to
// min(tombstone_end.user_key, comp_end_user_key)@kMaxSequenceNumber with
// op_type kTypeRangeDeletion. When a bound is null, there is no
// truncation on that end. To show that smallest_candidate.user_key <
// largest_candidate.user_key, it suffices to show
// tombstone_start.user_key < comp_end_user_key (if not null) AND
// comp_start_user_key (if not null) < tombstone_end.user_key.
// Since the file has no point key, `has_overlapping_endpoints` is false.
// In the first sanity check of this for-loop, we compare
// tombstone_start.user_key against upper_bound = comp_end_user_key,
// and only proceed if tombstone_start.user_key < comp_end_user_key.
// We assume FragmentedTombstoneIterator::Seek(k) lands
// on a tombstone with end_key > k. So the call it->Seek(*lower_bound)
// above implies compact_start_user_key < tombstone_end.user_key.
//
// To show (2) [tombstone_start, tombstone_end] and [meta.smallest,
// meta.largest] overlaps (after the call to UpdateBoundariesForRange()):
// In the proof for (1) we have shown that
// smallest_candidate <= largest_candidate. Since tombstone_start <=
// smallest_candidate <= largest_candidate <= tombstone_end, for (2) to
// hold, it suffices to show that [smallest_candidate, largest_candidate]
// overlaps with [meta.smallest, meta.largest]. too.
// Given meta.smallest <= meta.largest shown above, we need to show
// that it is impossible to have largest_candidate < meta.smallest or
// meta.largest < smallest_candidate. If the above
// meta.UpdateBoundariesForRange(smallest_candidate, largest_candidate)
// updates meta.largest or meta.smallest, then the two ranges overlap.
// So we assume meta.UpdateBoundariesForRange(smallest_candidate,
// largest_candidate) did not update meta.smallest nor meta.largest, which
// means meta.smallest < smallest_candidate and largest_candidate <
// meta.largest.
//
// Case 2. Output file has >= 1 point key. This means meta.smallest and
// meta.largest are not empty when AddRangeDels() is called.
// To show (1) meta.smallest <= meta.largest:
// Assume meta.smallest <= meta.largest when AddRangeDels() is called,
// this follow from how UpdateBoundariesForRange() is implemented where it
// takes min or max to update meta.smallest or meta.largest.
//
// To show (2) [tombstone_start, tombstone_end] and [meta.smallest,
// meta.largest] overlaps (after the call to UpdateBoundariesForRange()):
// When smallest_candidate <= largest_candidate, the proof in Case 1
// applies, so we only need to show (2) holds when smallest_candidate >
// largest_candidate. When both bounds are either null or from
// subcompaction boundary, the proof in Case 1 applies, so we only need to
// show (2) holds when at least one bound is from a point key (either
// meta.smallest for lower bound or next_table_min_key for upper bound).
//
// Suppose lower bound is meta.smallest.user_key. The call
// it->Seek(*lower_bound) implies tombstone_end.user_key >
// meta.smallest.user_key. We have smallest_candidate.user_key =
// max(tombstone_start.user_key, meta.smallest.user_key). For
// smallest_candidate to be > largest_candidate, we need
// largest_candidate.user_key = upper_bound = smallest_candidate.user_key,
// where tombstone_end is truncated to largest_candidate.
// Subcase 1:
// Suppose largest_candidate.user_key = comp_end_user_key (there is no
// next point key). Subcompaction ensures any point key from this
// subcompaction has a user_key < comp_end_user_key, so 1)
// meta.smallest.user_key < comp_end_user_key, 2)
// `has_overlapping_endpoints` is false, and the first if condition in
// this for-loop ensures tombstone_start.user_key < comp_end_user_key. So
// smallest_candidate.user_key < largest_candidate.user_key. This case
// cannot happen when smallest > largest_candidate.
// Subcase 2:
// Suppose largest_candidate.user_key = next_table_min_key.user_key.
// The first if condition in this for-loop together with
// smallest_candidate.user_key = next_table_min_key.user_key =
// upper_bound implies `has_overlapping_endpoints` is true (so meta
// largest.user_key = upper_bound) and
// tombstone.seq_ < meta.largest.seqno. So
// tombstone_start < meta.largest < tombstone_end.
//
// Suppose lower bound is comp_start_user_key and upper_bound is
// next_table_min_key. The call it->Seek(*lower_bound) implies we have
// tombstone_end_key.user_key > comp_start_user_key. So
// tombstone_end_key.user_key > smallest_candidate.user_key. For
// smallest_candidate to be > largest_candidate, we need
// tombstone_start.user_key = largest_candidate.user_key = upper_bound =
// next_table_min_key.user_key. This means `has_overlapping_endpoints` is
// true (so meta.largest.user_key = upper_bound) and tombstone.seq_ <
// meta.largest.seqno. So tombstone_start < meta.largest < tombstone_end.
}
return Status::OK();
}
void CompactionOutputs::FillFilesToCutForTtl() {
if (compaction_->immutable_options()->compaction_style !=
kCompactionStyleLevel ||
compaction_->immutable_options()->compaction_pri !=
kMinOverlappingRatio ||
compaction_->mutable_cf_options()->ttl == 0 ||
compaction_->num_input_levels() < 2 || compaction_->bottommost_level()) {
return;
}
// We define new file with the oldest ancestor time to be younger than 1/4
// TTL, and an old one to be older than 1/2 TTL time.
int64_t temp_current_time;
auto get_time_status =
compaction_->immutable_options()->clock->GetCurrentTime(
&temp_current_time);
if (!get_time_status.ok()) {
return;
}
auto current_time = static_cast<uint64_t>(temp_current_time);
if (current_time < compaction_->mutable_cf_options()->ttl) {
return;
}
uint64_t old_age_thres =
current_time - compaction_->mutable_cf_options()->ttl / 2;
const std::vector<FileMetaData*>& olevel =
*(compaction_->inputs(compaction_->num_input_levels() - 1));
for (FileMetaData* file : olevel) {
// Worth filtering out by start and end?
uint64_t oldest_ancester_time = file->TryGetOldestAncesterTime();
// We put old files if they are not too small to prevent a flood
// of small files.
if (oldest_ancester_time < old_age_thres &&
file->fd.GetFileSize() >
compaction_->mutable_cf_options()->target_file_size_base / 2) {
files_to_cut_for_ttl_.push_back(file);
}
}
}
CompactionOutputs::CompactionOutputs(const Compaction* compaction,
const bool is_penultimate_level)
: compaction_(compaction), is_penultimate_level_(is_penultimate_level) {
partitioner_ = compaction->output_level() == 0
? nullptr
: compaction->CreateSstPartitioner();
if (compaction->output_level() != 0) {
FillFilesToCutForTtl();
}
}
} // namespace ROCKSDB_NAMESPACE