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

607 lines
24 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;
}
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;
}
// TODO: here it uses the internal comparator but Compaction picker uses user
// comparator to get a clean cut with `GetOverlappingInputs()`, which means
// this function may cut files still be overlapped in compaction, for example
// current function can generate L1 files like:
// L1: [2-21] [22-30]
// L2: [1-10] [21-30]
// Because L1 `21` has higher seq_number, which is smaller than `21` on L2,
// it cuts in the first file. But for compaction picker L1 [2-21] file
// overlaps with both files on L2. Ideally it should cut to
// L1: [2-20] [21-30]
const InternalKeyComparator* icmp =
&compaction_->column_family_data()->internal_comparator();
while (grandparent_index_ < grandparents.size()) {
if (being_grandparent_gap_) {
if (icmp->Compare(internal_key,
grandparents[grandparent_index_]->smallest.Encode()) <
0) {
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 {
if (icmp->Compare(internal_key,
grandparents[grandparent_index_]->largest.Encode()) <=
0) {
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_ =
grandparents[grandparent_index_]->fd.GetFileSize();
}
seen_key_ = true;
return curr_key_boundary_switched_num;
}
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
// always update grandparent information like overlapped file number, size
// etc.
const Slice& internal_key = c_iter.key();
const uint64_t previous_overlapped_bytes = grandparent_overlapped_bytes_;
size_t num_grandparent_boundaries_crossed =
UpdateGrandparentBoundaryInfo(internal_key);
if (!HasBuilder()) {
return false;
}
// 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;
}
const InternalKeyComparator* icmp =
&compaction_->column_family_data()->internal_comparator();
// 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_ >
((compaction_->target_output_file_size() + 99) / 100) *
(50 + std::min(grandparent_boundary_switched_num_ * 5,
size_t{40}))) {
return true;
}
}
// check ttl file boundaries if there's any
if (!files_to_cut_for_ttl_.empty()) {
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;
}
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
const std::vector<FileMetaData*>& grandparents =
compaction_->grandparents();
grandparent_overlapped_bytes_ =
being_grandparent_gap_
? 0
: grandparents[grandparent_index_]->fd.GetFileSize();
grandparent_boundary_switched_num_ = 0;
}
// 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;
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();
}
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_);
if ((has_overlapping_endpoints && cmp < 0) ||
(!has_overlapping_endpoints && cmp <= 0)) {
// 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.
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);
InternalKey smallest_candidate = std::move(kv.first);
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.
//
// When lower_bound is chosen by a subcompaction, 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.
//
// 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
if (lower_bound_from_sub_compact) {
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 {
smallest_candidate = InternalKey(*lower_bound, 0, kTypeRangeDeletion);
}
}
InternalKey largest_candidate = tombstone.SerializeEndKey();
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);
}
}
#ifndef NDEBUG
SequenceNumber smallest_ikey_seqnum = kMaxSequenceNumber;
if (meta.smallest.size() > 0) {
smallest_ikey_seqnum = GetInternalKeySeqno(meta.smallest.Encode());
}
#endif
meta.UpdateBoundariesForRange(smallest_candidate, largest_candidate,
tombstone.seq_, icmp);
// The smallest key in a file is used for range tombstone truncation, so
// it cannot have a seqnum of 0 (unless the smallest data key in a file
// has a seqnum of 0). Otherwise, the truncated tombstone may expose
// deleted keys at lower levels.
assert(smallest_ikey_seqnum == 0 ||
ExtractInternalKeyFooter(meta.smallest.Encode()) !=
PackSequenceAndType(0, kTypeRangeDeletion));
}
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