2017-04-07 03:06:34 +00:00
|
|
|
// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
|
2017-07-15 23:03:42 +00:00
|
|
|
// 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).
|
2017-04-07 03:06:34 +00:00
|
|
|
//
|
|
|
|
// 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.
|
|
|
|
|
2019-05-31 18:52:59 +00:00
|
|
|
#include "db/compaction/compaction_picker_universal.h"
|
2017-04-07 03:06:34 +00:00
|
|
|
|
2023-09-12 22:53:15 +00:00
|
|
|
#include <cstdint>
|
2017-04-07 03:06:34 +00:00
|
|
|
#include <limits>
|
|
|
|
#include <queue>
|
|
|
|
#include <string>
|
|
|
|
#include <utility>
|
2021-09-29 11:01:57 +00:00
|
|
|
|
2017-04-07 03:06:34 +00:00
|
|
|
#include "db/column_family.h"
|
2019-05-30 03:44:08 +00:00
|
|
|
#include "file/filename.h"
|
2019-06-01 00:19:43 +00:00
|
|
|
#include "logging/log_buffer.h"
|
2021-09-29 11:01:57 +00:00
|
|
|
#include "logging/logging.h"
|
2023-05-17 18:27:09 +00:00
|
|
|
#include "monitoring/statistics_impl.h"
|
2019-05-31 00:39:43 +00:00
|
|
|
#include "test_util/sync_point.h"
|
2017-04-07 03:06:34 +00:00
|
|
|
#include "util/random.h"
|
|
|
|
#include "util/string_util.h"
|
|
|
|
|
2020-02-20 20:07:53 +00:00
|
|
|
namespace ROCKSDB_NAMESPACE {
|
2017-04-07 03:06:34 +00:00
|
|
|
namespace {
|
2019-09-16 17:49:20 +00:00
|
|
|
// A helper class that form universal compactions. The class is used by
|
|
|
|
// UniversalCompactionPicker::PickCompaction().
|
|
|
|
// The usage is to create the class, and get the compaction object by calling
|
|
|
|
// PickCompaction().
|
|
|
|
class UniversalCompactionBuilder {
|
|
|
|
public:
|
2020-07-23 01:31:25 +00:00
|
|
|
UniversalCompactionBuilder(
|
2021-05-05 20:59:21 +00:00
|
|
|
const ImmutableOptions& ioptions, const InternalKeyComparator* icmp,
|
2020-07-23 01:31:25 +00:00
|
|
|
const std::string& cf_name, const MutableCFOptions& mutable_cf_options,
|
|
|
|
const MutableDBOptions& mutable_db_options, VersionStorageInfo* vstorage,
|
|
|
|
UniversalCompactionPicker* picker, LogBuffer* log_buffer)
|
2019-09-16 17:49:20 +00:00
|
|
|
: ioptions_(ioptions),
|
|
|
|
icmp_(icmp),
|
|
|
|
cf_name_(cf_name),
|
|
|
|
mutable_cf_options_(mutable_cf_options),
|
2020-07-23 01:31:25 +00:00
|
|
|
mutable_db_options_(mutable_db_options),
|
2019-09-16 17:49:20 +00:00
|
|
|
vstorage_(vstorage),
|
|
|
|
picker_(picker),
|
|
|
|
log_buffer_(log_buffer) {}
|
|
|
|
|
|
|
|
// Form and return the compaction object. The caller owns return object.
|
|
|
|
Compaction* PickCompaction();
|
|
|
|
|
|
|
|
private:
|
|
|
|
struct SortedRun {
|
|
|
|
SortedRun(int _level, FileMetaData* _file, uint64_t _size,
|
|
|
|
uint64_t _compensated_file_size, bool _being_compacted)
|
|
|
|
: level(_level),
|
|
|
|
file(_file),
|
|
|
|
size(_size),
|
|
|
|
compensated_file_size(_compensated_file_size),
|
|
|
|
being_compacted(_being_compacted) {
|
|
|
|
assert(compensated_file_size > 0);
|
|
|
|
assert(level != 0 || file != nullptr);
|
|
|
|
}
|
|
|
|
|
|
|
|
void Dump(char* out_buf, size_t out_buf_size,
|
|
|
|
bool print_path = false) const;
|
|
|
|
|
|
|
|
// sorted_run_count is added into the string to print
|
|
|
|
void DumpSizeInfo(char* out_buf, size_t out_buf_size,
|
|
|
|
size_t sorted_run_count) const;
|
|
|
|
|
|
|
|
int level;
|
|
|
|
// `file` Will be null for level > 0. For level = 0, the sorted run is
|
|
|
|
// for this file.
|
|
|
|
FileMetaData* file;
|
|
|
|
// For level > 0, `size` and `compensated_file_size` are sum of sizes all
|
|
|
|
// files in the level. `being_compacted` should be the same for all files
|
|
|
|
// in a non-zero level. Use the value here.
|
|
|
|
uint64_t size;
|
|
|
|
uint64_t compensated_file_size;
|
|
|
|
bool being_compacted;
|
|
|
|
};
|
|
|
|
|
|
|
|
// Pick Universal compaction to limit read amplification
|
|
|
|
Compaction* PickCompactionToReduceSortedRuns(
|
|
|
|
unsigned int ratio, unsigned int max_number_of_files_to_compact);
|
|
|
|
|
|
|
|
// Pick Universal compaction to limit space amplification.
|
|
|
|
Compaction* PickCompactionToReduceSizeAmp();
|
|
|
|
|
Incremental Space Amp Compactions in Universal Style (#8655)
Summary:
This commit introduces incremental compaction in univeral style for space amplification. This follows the first improvement mentioned in https://rocksdb.org/blog/2021/04/12/universal-improvements.html . The implemention simply picks up files about size of max_compaction_bytes to compact and execute if the penalty is not too big. More optimizations can be done in the future, e.g. prioritizing between this compaction and other types. But for now, the feature is supposed to be functional and can often reduce frequency of full compactions, although it can introduce penalty.
In order to add cut files more efficiently so that more files from upper levels can be included, SST file cutting threshold (for current file + overlapping parent level files) is set to 1.5X of target file size. A 2MB target file size will generate files like this: https://gist.github.com/siying/29d2676fba417404f3c95e6c013c7de8 Number of files indeed increases but it is not out of control.
Two set of write benchmarks are run:
1. For ingestion rate limited scenario, we can see full compaction is mostly eliminated: https://gist.github.com/siying/959bc1186066906831cf4c808d6e0a19 . The write amp increased from 7.7 to 9.4, as expected. After applying file cutting, the number is improved to 8.9. In another benchmark, the write amp is even better with the incremental approach: https://gist.github.com/siying/d1c16c286d7c59c4d7bba718ca198163
2. For ingestion rate unlimited scenario, incremental compaction turns out to be too expensive most of the time and is not executed, as expected.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8655
Test Plan: Add unit tests to the functionality.
Reviewed By: ajkr
Differential Revision: D31787034
fbshipit-source-id: ce813e63b15a61d5a56e97bf8902a1b28e011beb
2021-10-20 17:03:03 +00:00
|
|
|
// Try to pick incremental compaction to reduce space amplification.
|
|
|
|
// It will return null if it cannot find a fanout within the threshold.
|
|
|
|
// Fanout is defined as
|
|
|
|
// total size of files to compact at output level
|
|
|
|
// --------------------------------------------------
|
|
|
|
// total size of files to compact at other levels
|
|
|
|
Compaction* PickIncrementalForReduceSizeAmp(double fanout_threshold);
|
|
|
|
|
2019-09-16 17:49:20 +00:00
|
|
|
Compaction* PickDeleteTriggeredCompaction();
|
|
|
|
|
2019-10-31 18:16:33 +00:00
|
|
|
// Form a compaction from the sorted run indicated by start_index to the
|
|
|
|
// oldest sorted run.
|
|
|
|
// The caller is responsible for making sure that those files are not in
|
|
|
|
// compaction.
|
|
|
|
Compaction* PickCompactionToOldest(size_t start_index,
|
|
|
|
CompactionReason compaction_reason);
|
|
|
|
|
2022-08-08 21:34:36 +00:00
|
|
|
Compaction* PickCompactionWithSortedRunRange(
|
|
|
|
size_t start_index, size_t end_index, CompactionReason compaction_reason);
|
|
|
|
|
2019-10-31 18:16:33 +00:00
|
|
|
// Try to pick periodic compaction. The caller should only call it
|
|
|
|
// if there is at least one file marked for periodic compaction.
|
|
|
|
// null will be returned if no such a compaction can be formed
|
|
|
|
// because some files are being compacted.
|
|
|
|
Compaction* PickPeriodicCompaction();
|
|
|
|
|
2023-09-12 22:53:15 +00:00
|
|
|
bool ShouldSkipLastSortedRunForSizeAmpCompaction() const {
|
2023-08-23 18:27:48 +00:00
|
|
|
assert(!sorted_runs_.empty());
|
|
|
|
return ioptions_.preclude_last_level_data_seconds > 0 &&
|
|
|
|
ioptions_.num_levels > 2 &&
|
|
|
|
sorted_runs_.back().level == ioptions_.num_levels - 1 &&
|
|
|
|
sorted_runs_.size() > 1;
|
|
|
|
}
|
2021-10-07 18:20:44 +00:00
|
|
|
// Used in universal compaction when the allow_trivial_move
|
2019-09-16 17:49:20 +00:00
|
|
|
// option is set. Checks whether there are any overlapping files
|
|
|
|
// in the input. Returns true if the input files are non
|
|
|
|
// overlapping.
|
|
|
|
bool IsInputFilesNonOverlapping(Compaction* c);
|
|
|
|
|
Incremental Space Amp Compactions in Universal Style (#8655)
Summary:
This commit introduces incremental compaction in univeral style for space amplification. This follows the first improvement mentioned in https://rocksdb.org/blog/2021/04/12/universal-improvements.html . The implemention simply picks up files about size of max_compaction_bytes to compact and execute if the penalty is not too big. More optimizations can be done in the future, e.g. prioritizing between this compaction and other types. But for now, the feature is supposed to be functional and can often reduce frequency of full compactions, although it can introduce penalty.
In order to add cut files more efficiently so that more files from upper levels can be included, SST file cutting threshold (for current file + overlapping parent level files) is set to 1.5X of target file size. A 2MB target file size will generate files like this: https://gist.github.com/siying/29d2676fba417404f3c95e6c013c7de8 Number of files indeed increases but it is not out of control.
Two set of write benchmarks are run:
1. For ingestion rate limited scenario, we can see full compaction is mostly eliminated: https://gist.github.com/siying/959bc1186066906831cf4c808d6e0a19 . The write amp increased from 7.7 to 9.4, as expected. After applying file cutting, the number is improved to 8.9. In another benchmark, the write amp is even better with the incremental approach: https://gist.github.com/siying/d1c16c286d7c59c4d7bba718ca198163
2. For ingestion rate unlimited scenario, incremental compaction turns out to be too expensive most of the time and is not executed, as expected.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8655
Test Plan: Add unit tests to the functionality.
Reviewed By: ajkr
Differential Revision: D31787034
fbshipit-source-id: ce813e63b15a61d5a56e97bf8902a1b28e011beb
2021-10-20 17:03:03 +00:00
|
|
|
uint64_t GetMaxOverlappingBytes() const;
|
|
|
|
|
2023-09-12 22:53:15 +00:00
|
|
|
// To conditionally exclude some of the newest L0 files
|
|
|
|
// from a size amp compaction. This is to prevent a large number of L0
|
|
|
|
// files from being locked by a size amp compaction, potentially leading to
|
|
|
|
// write stop with a few more flushes.
|
|
|
|
//
|
|
|
|
// Such exclusion is based on `num_l0_input_pre_exclusion`,
|
|
|
|
// `level0_stop_writes_trigger`, `max/min_merge_width` and the pre-exclusion
|
|
|
|
// compaction score. Noted that it will not make the size amp compaction of
|
|
|
|
// interest invalid from running as a size amp compaction as long as its
|
|
|
|
// pre-exclusion compaction score satisfies the condition to run.
|
|
|
|
//
|
|
|
|
// @param `num_l0_input_pre_exclusion` Number of L0 input files prior to
|
|
|
|
// exclusion
|
|
|
|
// @param `end_index` Index of the last sorted run selected as compaction
|
|
|
|
// input. Will not be affected by this exclusion.
|
|
|
|
// @param `start_index` Index of the first input sorted run prior to
|
|
|
|
// exclusion. Will be modified as output based on the exclusion.
|
|
|
|
// @param `candidate_size` Total size of all except for the last input sorted
|
|
|
|
// runs prior to exclusion. Will be modified as output based on the exclusion.
|
|
|
|
//
|
|
|
|
// @return Number of L0 files to exclude. `start_index` and
|
|
|
|
// `candidate_size` will be modified accordingly
|
|
|
|
std::size_t MightExcludeNewL0sToReduceWriteStop(
|
|
|
|
std::size_t num_l0_input_pre_exclusion, std::size_t end_index,
|
|
|
|
std::size_t& start_index, uint64_t& candidate_size) const {
|
|
|
|
if (num_l0_input_pre_exclusion == 0) {
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
assert(start_index <= end_index && sorted_runs_.size() > end_index);
|
|
|
|
assert(mutable_cf_options_.level0_stop_writes_trigger > 0);
|
|
|
|
const std::size_t level0_stop_writes_trigger = static_cast<std::size_t>(
|
|
|
|
mutable_cf_options_.level0_stop_writes_trigger);
|
|
|
|
const std::size_t max_merge_width = static_cast<std::size_t>(
|
|
|
|
mutable_cf_options_.compaction_options_universal.max_merge_width);
|
|
|
|
const std::size_t min_merge_width = static_cast<std::size_t>(
|
|
|
|
mutable_cf_options_.compaction_options_universal.min_merge_width);
|
|
|
|
const uint64_t max_size_amplification_percent =
|
|
|
|
mutable_cf_options_.compaction_options_universal
|
|
|
|
.max_size_amplification_percent;
|
|
|
|
const uint64_t base_sr_size = sorted_runs_[end_index].size;
|
|
|
|
|
|
|
|
// Leave at least 1 L0 file and 2 input sorted runs after exclusion
|
|
|
|
const std::size_t max_num_l0_to_exclude =
|
|
|
|
std::min(num_l0_input_pre_exclusion - 1, end_index - start_index - 1);
|
|
|
|
// In universal compaction, sorted runs from non L0 levels are counted
|
|
|
|
// toward `level0_stop_writes_trigger`. Therefore we need to subtract the
|
|
|
|
// total number of sorted runs picked originally for this compaction from
|
|
|
|
// `level0_stop_writes_trigger` to calculate
|
|
|
|
// `num_extra_l0_before_write_stop`
|
|
|
|
const std::size_t num_extra_l0_before_write_stop =
|
|
|
|
level0_stop_writes_trigger -
|
|
|
|
std::min(level0_stop_writes_trigger, end_index - start_index + 1);
|
|
|
|
const std::size_t num_l0_to_exclude_for_max_merge_width =
|
|
|
|
std::min(max_merge_width -
|
|
|
|
std::min(max_merge_width, num_extra_l0_before_write_stop),
|
|
|
|
max_num_l0_to_exclude);
|
|
|
|
const std::size_t num_l0_to_exclude_for_min_merge_width =
|
|
|
|
std::min(min_merge_width -
|
|
|
|
std::min(min_merge_width, num_extra_l0_before_write_stop),
|
|
|
|
max_num_l0_to_exclude);
|
|
|
|
|
|
|
|
std::size_t num_l0_to_exclude = 0;
|
|
|
|
uint64_t candidate_size_post_exclusion = candidate_size;
|
|
|
|
|
|
|
|
for (std::size_t possible_num_l0_to_exclude =
|
|
|
|
num_l0_to_exclude_for_min_merge_width;
|
|
|
|
possible_num_l0_to_exclude <= num_l0_to_exclude_for_max_merge_width;
|
|
|
|
++possible_num_l0_to_exclude) {
|
|
|
|
uint64_t current_candidate_size = candidate_size_post_exclusion;
|
|
|
|
for (std::size_t j = num_l0_to_exclude; j < possible_num_l0_to_exclude;
|
|
|
|
++j) {
|
|
|
|
current_candidate_size -=
|
|
|
|
sorted_runs_.at(start_index + j).compensated_file_size;
|
|
|
|
}
|
|
|
|
|
|
|
|
// To ensure the compaction score before and after exclusion is similar
|
|
|
|
// so this exclusion will not make the size amp compaction of
|
|
|
|
// interest invalid from running as a size amp compaction as long as its
|
|
|
|
// pre-exclusion compaction score satisfies the condition to run.
|
|
|
|
if (current_candidate_size * 100 <
|
|
|
|
max_size_amplification_percent * base_sr_size ||
|
|
|
|
current_candidate_size < candidate_size * 9 / 10) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
num_l0_to_exclude = possible_num_l0_to_exclude;
|
|
|
|
candidate_size_post_exclusion = current_candidate_size;
|
|
|
|
}
|
|
|
|
|
|
|
|
start_index += num_l0_to_exclude;
|
|
|
|
candidate_size = candidate_size_post_exclusion;
|
|
|
|
return num_l0_to_exclude;
|
|
|
|
}
|
|
|
|
|
2021-05-05 20:59:21 +00:00
|
|
|
const ImmutableOptions& ioptions_;
|
2019-09-16 17:49:20 +00:00
|
|
|
const InternalKeyComparator* icmp_;
|
|
|
|
double score_;
|
|
|
|
std::vector<SortedRun> sorted_runs_;
|
|
|
|
const std::string& cf_name_;
|
|
|
|
const MutableCFOptions& mutable_cf_options_;
|
2020-07-23 01:31:25 +00:00
|
|
|
const MutableDBOptions& mutable_db_options_;
|
2019-09-16 17:49:20 +00:00
|
|
|
VersionStorageInfo* vstorage_;
|
|
|
|
UniversalCompactionPicker* picker_;
|
|
|
|
LogBuffer* log_buffer_;
|
|
|
|
|
2023-06-14 18:28:56 +00:00
|
|
|
static std::vector<UniversalCompactionBuilder::SortedRun> CalculateSortedRuns(
|
|
|
|
const VersionStorageInfo& vstorage, int last_level);
|
2019-09-16 17:49:20 +00:00
|
|
|
|
|
|
|
// Pick a path ID to place a newly generated file, with its estimated file
|
|
|
|
// size.
|
|
|
|
static uint32_t GetPathId(const ImmutableCFOptions& ioptions,
|
|
|
|
const MutableCFOptions& mutable_cf_options,
|
|
|
|
uint64_t file_size);
|
|
|
|
};
|
|
|
|
|
2017-04-07 03:06:34 +00:00
|
|
|
// Used in universal compaction when trivial move is enabled.
|
|
|
|
// This structure is used for the construction of min heap
|
|
|
|
// that contains the file meta data, the level of the file
|
|
|
|
// and the index of the file in that level
|
|
|
|
|
|
|
|
struct InputFileInfo {
|
2017-12-11 19:48:47 +00:00
|
|
|
InputFileInfo() : f(nullptr), level(0), index(0) {}
|
2017-04-07 03:06:34 +00:00
|
|
|
|
|
|
|
FileMetaData* f;
|
|
|
|
size_t level;
|
|
|
|
size_t index;
|
|
|
|
};
|
|
|
|
|
|
|
|
// Used in universal compaction when trivial move is enabled.
|
|
|
|
// This comparator is used for the construction of min heap
|
|
|
|
// based on the smallest key of the file.
|
|
|
|
struct SmallestKeyHeapComparator {
|
|
|
|
explicit SmallestKeyHeapComparator(const Comparator* ucmp) { ucmp_ = ucmp; }
|
|
|
|
|
|
|
|
bool operator()(InputFileInfo i1, InputFileInfo i2) const {
|
2022-09-08 20:03:07 +00:00
|
|
|
return (ucmp_->CompareWithoutTimestamp(i1.f->smallest.user_key(),
|
|
|
|
i2.f->smallest.user_key()) > 0);
|
2017-04-07 03:06:34 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
private:
|
|
|
|
const Comparator* ucmp_;
|
|
|
|
};
|
|
|
|
|
2021-09-07 18:31:12 +00:00
|
|
|
using SmallestKeyHeap =
|
|
|
|
std::priority_queue<InputFileInfo, std::vector<InputFileInfo>,
|
|
|
|
SmallestKeyHeapComparator>;
|
2017-04-07 03:06:34 +00:00
|
|
|
|
|
|
|
// This function creates the heap that is used to find if the files are
|
|
|
|
// overlapping during universal compaction when the allow_trivial_move
|
|
|
|
// is set.
|
|
|
|
SmallestKeyHeap create_level_heap(Compaction* c, const Comparator* ucmp) {
|
|
|
|
SmallestKeyHeap smallest_key_priority_q =
|
|
|
|
SmallestKeyHeap(SmallestKeyHeapComparator(ucmp));
|
|
|
|
|
|
|
|
InputFileInfo input_file;
|
|
|
|
|
|
|
|
for (size_t l = 0; l < c->num_input_levels(); l++) {
|
|
|
|
if (c->num_input_files(l) != 0) {
|
|
|
|
if (l == 0 && c->start_level() == 0) {
|
|
|
|
for (size_t i = 0; i < c->num_input_files(0); i++) {
|
|
|
|
input_file.f = c->input(0, i);
|
|
|
|
input_file.level = 0;
|
|
|
|
input_file.index = i;
|
|
|
|
smallest_key_priority_q.push(std::move(input_file));
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
input_file.f = c->input(l, 0);
|
|
|
|
input_file.level = l;
|
|
|
|
input_file.index = 0;
|
|
|
|
smallest_key_priority_q.push(std::move(input_file));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return smallest_key_priority_q;
|
|
|
|
}
|
|
|
|
|
|
|
|
#ifndef NDEBUG
|
|
|
|
// smallest_seqno and largest_seqno are set iff. `files` is not empty.
|
|
|
|
void GetSmallestLargestSeqno(const std::vector<FileMetaData*>& files,
|
|
|
|
SequenceNumber* smallest_seqno,
|
|
|
|
SequenceNumber* largest_seqno) {
|
|
|
|
bool is_first = true;
|
|
|
|
for (FileMetaData* f : files) {
|
2018-07-27 23:00:26 +00:00
|
|
|
assert(f->fd.smallest_seqno <= f->fd.largest_seqno);
|
2017-04-07 03:06:34 +00:00
|
|
|
if (is_first) {
|
|
|
|
is_first = false;
|
2018-07-27 23:00:26 +00:00
|
|
|
*smallest_seqno = f->fd.smallest_seqno;
|
|
|
|
*largest_seqno = f->fd.largest_seqno;
|
2017-04-07 03:06:34 +00:00
|
|
|
} else {
|
2018-07-27 23:00:26 +00:00
|
|
|
if (f->fd.smallest_seqno < *smallest_seqno) {
|
|
|
|
*smallest_seqno = f->fd.smallest_seqno;
|
2017-04-07 03:06:34 +00:00
|
|
|
}
|
2018-07-27 23:00:26 +00:00
|
|
|
if (f->fd.largest_seqno > *largest_seqno) {
|
|
|
|
*largest_seqno = f->fd.largest_seqno;
|
2017-04-07 03:06:34 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
} // namespace
|
|
|
|
|
|
|
|
// Algorithm that checks to see if there are any overlapping
|
|
|
|
// files in the input
|
2019-09-16 17:49:20 +00:00
|
|
|
bool UniversalCompactionBuilder::IsInputFilesNonOverlapping(Compaction* c) {
|
2017-04-07 03:06:34 +00:00
|
|
|
auto comparator = icmp_->user_comparator();
|
|
|
|
int first_iter = 1;
|
|
|
|
|
|
|
|
InputFileInfo prev, curr, next;
|
|
|
|
|
|
|
|
SmallestKeyHeap smallest_key_priority_q =
|
|
|
|
create_level_heap(c, icmp_->user_comparator());
|
|
|
|
|
|
|
|
while (!smallest_key_priority_q.empty()) {
|
|
|
|
curr = smallest_key_priority_q.top();
|
|
|
|
smallest_key_priority_q.pop();
|
|
|
|
|
|
|
|
if (first_iter) {
|
|
|
|
prev = curr;
|
|
|
|
first_iter = 0;
|
|
|
|
} else {
|
2022-09-08 20:03:07 +00:00
|
|
|
if (comparator->CompareWithoutTimestamp(
|
|
|
|
prev.f->largest.user_key(), curr.f->smallest.user_key()) >= 0) {
|
2017-04-07 03:06:34 +00:00
|
|
|
// found overlapping files, return false
|
|
|
|
return false;
|
|
|
|
}
|
2022-09-08 20:03:07 +00:00
|
|
|
assert(comparator->CompareWithoutTimestamp(
|
|
|
|
curr.f->largest.user_key(), prev.f->largest.user_key()) > 0);
|
2017-04-07 03:06:34 +00:00
|
|
|
prev = curr;
|
|
|
|
}
|
|
|
|
|
|
|
|
next.f = nullptr;
|
|
|
|
|
2019-12-11 19:25:33 +00:00
|
|
|
if (c->level(curr.level) != 0 &&
|
|
|
|
curr.index < c->num_input_files(curr.level) - 1) {
|
2017-04-07 03:06:34 +00:00
|
|
|
next.f = c->input(curr.level, curr.index + 1);
|
|
|
|
next.level = curr.level;
|
|
|
|
next.index = curr.index + 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (next.f) {
|
|
|
|
smallest_key_priority_q.push(std::move(next));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool UniversalCompactionPicker::NeedsCompaction(
|
|
|
|
const VersionStorageInfo* vstorage) const {
|
|
|
|
const int kLevel0 = 0;
|
2018-05-29 22:42:14 +00:00
|
|
|
if (vstorage->CompactionScore(kLevel0) >= 1) {
|
|
|
|
return true;
|
|
|
|
}
|
2019-10-31 18:16:33 +00:00
|
|
|
if (!vstorage->FilesMarkedForPeriodicCompaction().empty()) {
|
|
|
|
return true;
|
|
|
|
}
|
2018-05-29 22:42:14 +00:00
|
|
|
if (!vstorage->FilesMarkedForCompaction().empty()) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
return false;
|
2017-04-07 03:06:34 +00:00
|
|
|
}
|
|
|
|
|
2019-09-16 17:49:20 +00:00
|
|
|
Compaction* UniversalCompactionPicker::PickCompaction(
|
|
|
|
const std::string& cf_name, const MutableCFOptions& mutable_cf_options,
|
2020-07-23 01:31:25 +00:00
|
|
|
const MutableDBOptions& mutable_db_options, VersionStorageInfo* vstorage,
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2022-12-13 21:29:37 +00:00
|
|
|
LogBuffer* log_buffer) {
|
2019-09-16 17:49:20 +00:00
|
|
|
UniversalCompactionBuilder builder(ioptions_, icmp_, cf_name,
|
2020-07-23 01:31:25 +00:00
|
|
|
mutable_cf_options, mutable_db_options,
|
|
|
|
vstorage, this, log_buffer);
|
2019-09-16 17:49:20 +00:00
|
|
|
return builder.PickCompaction();
|
|
|
|
}
|
|
|
|
|
|
|
|
void UniversalCompactionBuilder::SortedRun::Dump(char* out_buf,
|
|
|
|
size_t out_buf_size,
|
|
|
|
bool print_path) const {
|
2017-04-07 03:06:34 +00:00
|
|
|
if (level == 0) {
|
|
|
|
assert(file != nullptr);
|
|
|
|
if (file->fd.GetPathId() == 0 || !print_path) {
|
|
|
|
snprintf(out_buf, out_buf_size, "file %" PRIu64, file->fd.GetNumber());
|
|
|
|
} else {
|
2022-10-26 19:35:12 +00:00
|
|
|
snprintf(out_buf, out_buf_size,
|
|
|
|
"file %" PRIu64
|
|
|
|
"(path "
|
|
|
|
"%" PRIu32 ")",
|
2017-04-07 03:06:34 +00:00
|
|
|
file->fd.GetNumber(), file->fd.GetPathId());
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
snprintf(out_buf, out_buf_size, "level %d", level);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2019-09-16 17:49:20 +00:00
|
|
|
void UniversalCompactionBuilder::SortedRun::DumpSizeInfo(
|
2017-04-07 03:06:34 +00:00
|
|
|
char* out_buf, size_t out_buf_size, size_t sorted_run_count) const {
|
|
|
|
if (level == 0) {
|
|
|
|
assert(file != nullptr);
|
|
|
|
snprintf(out_buf, out_buf_size,
|
|
|
|
"file %" PRIu64 "[%" ROCKSDB_PRIszt
|
|
|
|
"] "
|
|
|
|
"with size %" PRIu64 " (compensated size %" PRIu64 ")",
|
|
|
|
file->fd.GetNumber(), sorted_run_count, file->fd.GetFileSize(),
|
|
|
|
file->compensated_file_size);
|
|
|
|
} else {
|
|
|
|
snprintf(out_buf, out_buf_size,
|
|
|
|
"level %d[%" ROCKSDB_PRIszt
|
|
|
|
"] "
|
|
|
|
"with size %" PRIu64 " (compensated size %" PRIu64 ")",
|
|
|
|
level, sorted_run_count, size, compensated_file_size);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2019-09-16 17:49:20 +00:00
|
|
|
std::vector<UniversalCompactionBuilder::SortedRun>
|
|
|
|
UniversalCompactionBuilder::CalculateSortedRuns(
|
2023-06-14 18:28:56 +00:00
|
|
|
const VersionStorageInfo& vstorage, int last_level) {
|
2019-09-16 17:49:20 +00:00
|
|
|
std::vector<UniversalCompactionBuilder::SortedRun> ret;
|
2017-04-07 03:06:34 +00:00
|
|
|
for (FileMetaData* f : vstorage.LevelFiles(0)) {
|
|
|
|
ret.emplace_back(0, f, f->fd.GetFileSize(), f->compensated_file_size,
|
|
|
|
f->being_compacted);
|
|
|
|
}
|
2023-06-14 18:28:56 +00:00
|
|
|
for (int level = 1; level <= last_level; level++) {
|
2017-04-07 03:06:34 +00:00
|
|
|
uint64_t total_compensated_size = 0U;
|
|
|
|
uint64_t total_size = 0U;
|
|
|
|
bool being_compacted = false;
|
|
|
|
for (FileMetaData* f : vstorage.LevelFiles(level)) {
|
|
|
|
total_compensated_size += f->compensated_file_size;
|
|
|
|
total_size += f->fd.GetFileSize();
|
2020-05-08 00:30:41 +00:00
|
|
|
// Size amp, read amp and periodic compactions always include all files
|
|
|
|
// for a non-zero level. However, a delete triggered compaction and
|
|
|
|
// a trivial move might pick a subset of files in a sorted run. So
|
|
|
|
// always check all files in a sorted run and mark the entire run as
|
|
|
|
// being compacted if one or more files are being compacted
|
|
|
|
if (f->being_compacted) {
|
2017-04-07 03:06:34 +00:00
|
|
|
being_compacted = f->being_compacted;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (total_compensated_size > 0) {
|
|
|
|
ret.emplace_back(level, nullptr, total_size, total_compensated_size,
|
|
|
|
being_compacted);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Universal style of compaction. Pick files that are contiguous in
|
|
|
|
// time-range to compact.
|
2019-09-16 17:49:20 +00:00
|
|
|
Compaction* UniversalCompactionBuilder::PickCompaction() {
|
2017-04-07 03:06:34 +00:00
|
|
|
const int kLevel0 = 0;
|
2019-09-16 17:49:20 +00:00
|
|
|
score_ = vstorage_->CompactionScore(kLevel0);
|
2023-06-14 18:28:56 +00:00
|
|
|
int max_output_level =
|
|
|
|
vstorage_->MaxOutputLevel(ioptions_.allow_ingest_behind);
|
|
|
|
sorted_runs_ = CalculateSortedRuns(*vstorage_, max_output_level);
|
2019-09-16 17:49:20 +00:00
|
|
|
|
|
|
|
if (sorted_runs_.size() == 0 ||
|
2019-10-31 18:16:33 +00:00
|
|
|
(vstorage_->FilesMarkedForPeriodicCompaction().empty() &&
|
|
|
|
vstorage_->FilesMarkedForCompaction().empty() &&
|
2019-09-16 17:49:20 +00:00
|
|
|
sorted_runs_.size() < (unsigned int)mutable_cf_options_
|
|
|
|
.level0_file_num_compaction_trigger)) {
|
|
|
|
ROCKS_LOG_BUFFER(log_buffer_, "[%s] Universal: nothing to do\n",
|
|
|
|
cf_name_.c_str());
|
|
|
|
TEST_SYNC_POINT_CALLBACK(
|
|
|
|
"UniversalCompactionBuilder::PickCompaction:Return", nullptr);
|
2017-04-07 03:06:34 +00:00
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
VersionStorageInfo::LevelSummaryStorage tmp;
|
|
|
|
ROCKS_LOG_BUFFER_MAX_SZ(
|
2019-09-16 17:49:20 +00:00
|
|
|
log_buffer_, 3072,
|
2020-04-28 22:23:33 +00:00
|
|
|
"[%s] Universal: sorted runs: %" ROCKSDB_PRIszt " files: %s\n",
|
2019-09-16 17:49:20 +00:00
|
|
|
cf_name_.c_str(), sorted_runs_.size(), vstorage_->LevelSummary(&tmp));
|
2017-04-07 03:06:34 +00:00
|
|
|
|
2018-05-29 22:42:14 +00:00
|
|
|
Compaction* c = nullptr;
|
2019-10-31 18:16:33 +00:00
|
|
|
// Periodic compaction has higher priority than other type of compaction
|
|
|
|
// because it's a hard requirement.
|
|
|
|
if (!vstorage_->FilesMarkedForPeriodicCompaction().empty()) {
|
|
|
|
// Always need to do a full compaction for periodic compaction.
|
|
|
|
c = PickPeriodicCompaction();
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2022-12-13 21:29:37 +00:00
|
|
|
TEST_SYNC_POINT_CALLBACK("PostPickPeriodicCompaction", c);
|
2019-10-31 18:16:33 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// Check for size amplification.
|
|
|
|
if (c == nullptr &&
|
|
|
|
sorted_runs_.size() >=
|
|
|
|
static_cast<size_t>(
|
|
|
|
mutable_cf_options_.level0_file_num_compaction_trigger)) {
|
2019-09-16 17:49:20 +00:00
|
|
|
if ((c = PickCompactionToReduceSizeAmp()) != nullptr) {
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2022-12-13 21:29:37 +00:00
|
|
|
TEST_SYNC_POINT("PickCompactionToReduceSizeAmpReturnNonnullptr");
|
2019-09-16 17:49:20 +00:00
|
|
|
ROCKS_LOG_BUFFER(log_buffer_, "[%s] Universal: compacting for size amp\n",
|
|
|
|
cf_name_.c_str());
|
2017-04-07 03:06:34 +00:00
|
|
|
} else {
|
2018-05-29 22:42:14 +00:00
|
|
|
// Size amplification is within limits. Try reducing read
|
|
|
|
// amplification while maintaining file size ratios.
|
|
|
|
unsigned int ratio =
|
2019-09-16 17:49:20 +00:00
|
|
|
mutable_cf_options_.compaction_options_universal.size_ratio;
|
2018-05-29 22:42:14 +00:00
|
|
|
|
2019-09-16 17:49:20 +00:00
|
|
|
if ((c = PickCompactionToReduceSortedRuns(ratio, UINT_MAX)) != nullptr) {
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2022-12-13 21:29:37 +00:00
|
|
|
TEST_SYNC_POINT("PickCompactionToReduceSortedRunsReturnNonnullptr");
|
2019-09-16 17:49:20 +00:00
|
|
|
ROCKS_LOG_BUFFER(log_buffer_,
|
2018-05-29 22:42:14 +00:00
|
|
|
"[%s] Universal: compacting for size ratio\n",
|
2019-09-16 17:49:20 +00:00
|
|
|
cf_name_.c_str());
|
2018-05-29 22:42:14 +00:00
|
|
|
} else {
|
|
|
|
// Size amplification and file size ratios are within configured limits.
|
|
|
|
// If max read amplification is exceeding configured limits, then force
|
|
|
|
// compaction without looking at filesize ratios and try to reduce
|
|
|
|
// the number of files to fewer than level0_file_num_compaction_trigger.
|
|
|
|
// This is guaranteed by NeedsCompaction()
|
2019-09-16 17:49:20 +00:00
|
|
|
assert(sorted_runs_.size() >=
|
2018-05-29 22:42:14 +00:00
|
|
|
static_cast<size_t>(
|
2019-09-16 17:49:20 +00:00
|
|
|
mutable_cf_options_.level0_file_num_compaction_trigger));
|
2018-05-29 22:42:14 +00:00
|
|
|
// Get the total number of sorted runs that are not being compacted
|
|
|
|
int num_sr_not_compacted = 0;
|
2019-09-16 17:49:20 +00:00
|
|
|
for (size_t i = 0; i < sorted_runs_.size(); i++) {
|
|
|
|
if (sorted_runs_[i].being_compacted == false) {
|
2018-05-29 22:42:14 +00:00
|
|
|
num_sr_not_compacted++;
|
|
|
|
}
|
2017-04-07 03:06:34 +00:00
|
|
|
}
|
|
|
|
|
2018-05-29 22:42:14 +00:00
|
|
|
// The number of sorted runs that are not being compacted is greater
|
|
|
|
// than the maximum allowed number of sorted runs
|
|
|
|
if (num_sr_not_compacted >
|
2019-09-16 17:49:20 +00:00
|
|
|
mutable_cf_options_.level0_file_num_compaction_trigger) {
|
2018-05-29 22:42:14 +00:00
|
|
|
unsigned int num_files =
|
|
|
|
num_sr_not_compacted -
|
2019-09-16 17:49:20 +00:00
|
|
|
mutable_cf_options_.level0_file_num_compaction_trigger + 1;
|
|
|
|
if ((c = PickCompactionToReduceSortedRuns(UINT_MAX, num_files)) !=
|
|
|
|
nullptr) {
|
|
|
|
ROCKS_LOG_BUFFER(log_buffer_,
|
2018-05-29 22:42:14 +00:00
|
|
|
"[%s] Universal: compacting for file num -- %u\n",
|
2019-09-16 17:49:20 +00:00
|
|
|
cf_name_.c_str(), num_files);
|
2018-05-29 22:42:14 +00:00
|
|
|
}
|
2017-04-07 03:06:34 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2018-05-29 22:42:14 +00:00
|
|
|
|
|
|
|
if (c == nullptr) {
|
2019-09-16 17:49:20 +00:00
|
|
|
if ((c = PickDeleteTriggeredCompaction()) != nullptr) {
|
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2022-12-13 21:29:37 +00:00
|
|
|
TEST_SYNC_POINT("PickDeleteTriggeredCompactionReturnNonnullptr");
|
2019-09-16 17:49:20 +00:00
|
|
|
ROCKS_LOG_BUFFER(log_buffer_,
|
2018-05-29 22:42:14 +00:00
|
|
|
"[%s] Universal: delete triggered compaction\n",
|
2019-09-16 17:49:20 +00:00
|
|
|
cf_name_.c_str());
|
2018-05-29 22:42:14 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-04-07 03:06:34 +00:00
|
|
|
if (c == nullptr) {
|
2019-09-16 17:49:20 +00:00
|
|
|
TEST_SYNC_POINT_CALLBACK(
|
|
|
|
"UniversalCompactionBuilder::PickCompaction:Return", nullptr);
|
2017-04-07 03:06:34 +00:00
|
|
|
return nullptr;
|
|
|
|
}
|
2023-06-14 18:28:56 +00:00
|
|
|
assert(c->output_level() <=
|
|
|
|
vstorage_->MaxOutputLevel(ioptions_.allow_ingest_behind));
|
2017-04-07 03:06:34 +00:00
|
|
|
|
2019-09-16 17:49:20 +00:00
|
|
|
if (mutable_cf_options_.compaction_options_universal.allow_trivial_move ==
|
2019-10-31 18:16:33 +00:00
|
|
|
true &&
|
|
|
|
c->compaction_reason() != CompactionReason::kPeriodicCompaction) {
|
2017-04-07 03:06:34 +00:00
|
|
|
c->set_is_trivial_move(IsInputFilesNonOverlapping(c));
|
|
|
|
}
|
|
|
|
|
|
|
|
// validate that all the chosen files of L0 are non overlapping in time
|
|
|
|
#ifndef NDEBUG
|
|
|
|
bool is_first = true;
|
|
|
|
|
|
|
|
size_t level_index = 0U;
|
|
|
|
if (c->start_level() == 0) {
|
|
|
|
for (auto f : *c->inputs(0)) {
|
2018-07-27 23:00:26 +00:00
|
|
|
assert(f->fd.smallest_seqno <= f->fd.largest_seqno);
|
2017-04-07 03:06:34 +00:00
|
|
|
if (is_first) {
|
|
|
|
is_first = false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
level_index = 1U;
|
|
|
|
}
|
|
|
|
for (; level_index < c->num_input_levels(); level_index++) {
|
|
|
|
if (c->num_input_files(level_index) != 0) {
|
|
|
|
SequenceNumber smallest_seqno = 0U;
|
|
|
|
SequenceNumber largest_seqno = 0U;
|
|
|
|
GetSmallestLargestSeqno(*(c->inputs(level_index)), &smallest_seqno,
|
|
|
|
&largest_seqno);
|
|
|
|
if (is_first) {
|
|
|
|
is_first = false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
// update statistics
|
2021-11-30 23:10:12 +00:00
|
|
|
size_t num_files = 0;
|
|
|
|
for (auto& each_level : *c->inputs()) {
|
|
|
|
num_files += each_level.files.size();
|
|
|
|
}
|
|
|
|
RecordInHistogram(ioptions_.stats, NUM_FILES_IN_SINGLE_COMPACTION, num_files);
|
2017-04-07 03:06:34 +00:00
|
|
|
|
2019-09-16 17:49:20 +00:00
|
|
|
picker_->RegisterCompaction(c);
|
|
|
|
vstorage_->ComputeCompactionScore(ioptions_, mutable_cf_options_);
|
2017-04-07 03:06:34 +00:00
|
|
|
|
2019-09-16 17:49:20 +00:00
|
|
|
TEST_SYNC_POINT_CALLBACK("UniversalCompactionBuilder::PickCompaction:Return",
|
2017-04-07 03:06:34 +00:00
|
|
|
c);
|
|
|
|
return c;
|
|
|
|
}
|
|
|
|
|
2019-09-16 17:49:20 +00:00
|
|
|
uint32_t UniversalCompactionBuilder::GetPathId(
|
2017-12-11 21:12:12 +00:00
|
|
|
const ImmutableCFOptions& ioptions,
|
|
|
|
const MutableCFOptions& mutable_cf_options, uint64_t file_size) {
|
2017-04-07 03:06:34 +00:00
|
|
|
// Two conditions need to be satisfied:
|
|
|
|
// (1) the target path needs to be able to hold the file's size
|
|
|
|
// (2) Total size left in this and previous paths need to be not
|
|
|
|
// smaller than expected future file size before this new file is
|
|
|
|
// compacted, which is estimated based on size_ratio.
|
|
|
|
// For example, if now we are compacting files of size (1, 1, 2, 4, 8),
|
|
|
|
// we will make sure the target file, probably with size of 16, will be
|
|
|
|
// placed in a path so that eventually when new files are generated and
|
|
|
|
// compacted to (1, 1, 2, 4, 8, 16), all those files can be stored in or
|
|
|
|
// before the path we chose.
|
|
|
|
//
|
|
|
|
// TODO(sdong): now the case of multiple column families is not
|
|
|
|
// considered in this algorithm. So the target size can be violated in
|
|
|
|
// that case. We need to improve it.
|
|
|
|
uint64_t accumulated_size = 0;
|
|
|
|
uint64_t future_size =
|
2017-12-11 21:12:12 +00:00
|
|
|
file_size *
|
|
|
|
(100 - mutable_cf_options.compaction_options_universal.size_ratio) / 100;
|
2017-04-07 03:06:34 +00:00
|
|
|
uint32_t p = 0;
|
2018-04-06 02:49:06 +00:00
|
|
|
assert(!ioptions.cf_paths.empty());
|
|
|
|
for (; p < ioptions.cf_paths.size() - 1; p++) {
|
|
|
|
uint64_t target_size = ioptions.cf_paths[p].target_size;
|
2017-04-07 03:06:34 +00:00
|
|
|
if (target_size > file_size &&
|
|
|
|
accumulated_size + (target_size - file_size) > future_size) {
|
|
|
|
return p;
|
|
|
|
}
|
|
|
|
accumulated_size += target_size;
|
|
|
|
}
|
|
|
|
return p;
|
|
|
|
}
|
|
|
|
|
|
|
|
//
|
|
|
|
// Consider compaction files based on their size differences with
|
|
|
|
// the next file in time order.
|
|
|
|
//
|
2019-09-16 17:49:20 +00:00
|
|
|
Compaction* UniversalCompactionBuilder::PickCompactionToReduceSortedRuns(
|
|
|
|
unsigned int ratio, unsigned int max_number_of_files_to_compact) {
|
2017-04-07 03:06:34 +00:00
|
|
|
unsigned int min_merge_width =
|
2019-09-16 17:49:20 +00:00
|
|
|
mutable_cf_options_.compaction_options_universal.min_merge_width;
|
2017-04-07 03:06:34 +00:00
|
|
|
unsigned int max_merge_width =
|
2019-09-16 17:49:20 +00:00
|
|
|
mutable_cf_options_.compaction_options_universal.max_merge_width;
|
2017-04-07 03:06:34 +00:00
|
|
|
|
|
|
|
const SortedRun* sr = nullptr;
|
|
|
|
bool done = false;
|
|
|
|
size_t start_index = 0;
|
|
|
|
unsigned int candidate_count = 0;
|
|
|
|
|
|
|
|
unsigned int max_files_to_compact =
|
|
|
|
std::min(max_merge_width, max_number_of_files_to_compact);
|
|
|
|
min_merge_width = std::max(min_merge_width, 2U);
|
|
|
|
|
|
|
|
// Caller checks the size before executing this function. This invariant is
|
|
|
|
// important because otherwise we may have a possible integer underflow when
|
|
|
|
// dealing with unsigned types.
|
2019-09-16 17:49:20 +00:00
|
|
|
assert(sorted_runs_.size() > 0);
|
2017-04-07 03:06:34 +00:00
|
|
|
|
|
|
|
// Considers a candidate file only if it is smaller than the
|
|
|
|
// total size accumulated so far.
|
2019-09-16 17:49:20 +00:00
|
|
|
for (size_t loop = 0; loop < sorted_runs_.size(); loop++) {
|
2017-04-07 03:06:34 +00:00
|
|
|
candidate_count = 0;
|
|
|
|
|
|
|
|
// Skip files that are already being compacted
|
2019-09-16 17:49:20 +00:00
|
|
|
for (sr = nullptr; loop < sorted_runs_.size(); loop++) {
|
|
|
|
sr = &sorted_runs_[loop];
|
2017-04-07 03:06:34 +00:00
|
|
|
|
|
|
|
if (!sr->being_compacted) {
|
|
|
|
candidate_count = 1;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
char file_num_buf[kFormatFileNumberBufSize];
|
|
|
|
sr->Dump(file_num_buf, sizeof(file_num_buf));
|
2019-09-16 17:49:20 +00:00
|
|
|
ROCKS_LOG_BUFFER(log_buffer_,
|
2017-04-07 03:06:34 +00:00
|
|
|
"[%s] Universal: %s"
|
|
|
|
"[%d] being compacted, skipping",
|
2019-09-16 17:49:20 +00:00
|
|
|
cf_name_.c_str(), file_num_buf, loop);
|
2017-04-07 03:06:34 +00:00
|
|
|
|
|
|
|
sr = nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
// This file is not being compacted. Consider it as the
|
|
|
|
// first candidate to be compacted.
|
|
|
|
uint64_t candidate_size = sr != nullptr ? sr->compensated_file_size : 0;
|
|
|
|
if (sr != nullptr) {
|
|
|
|
char file_num_buf[kFormatFileNumberBufSize];
|
|
|
|
sr->Dump(file_num_buf, sizeof(file_num_buf), true);
|
2019-09-16 17:49:20 +00:00
|
|
|
ROCKS_LOG_BUFFER(log_buffer_,
|
|
|
|
"[%s] Universal: Possible candidate %s[%d].",
|
|
|
|
cf_name_.c_str(), file_num_buf, loop);
|
2017-04-07 03:06:34 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// Check if the succeeding files need compaction.
|
|
|
|
for (size_t i = loop + 1;
|
2019-09-16 17:49:20 +00:00
|
|
|
candidate_count < max_files_to_compact && i < sorted_runs_.size();
|
2017-04-07 03:06:34 +00:00
|
|
|
i++) {
|
2019-09-16 17:49:20 +00:00
|
|
|
const SortedRun* succeeding_sr = &sorted_runs_[i];
|
2017-04-07 03:06:34 +00:00
|
|
|
if (succeeding_sr->being_compacted) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
// Pick files if the total/last candidate file size (increased by the
|
|
|
|
// specified ratio) is still larger than the next candidate file.
|
|
|
|
// candidate_size is the total size of files picked so far with the
|
|
|
|
// default kCompactionStopStyleTotalSize; with
|
|
|
|
// kCompactionStopStyleSimilarSize, it's simply the size of the last
|
|
|
|
// picked file.
|
|
|
|
double sz = candidate_size * (100.0 + ratio) / 100.0;
|
|
|
|
if (sz < static_cast<double>(succeeding_sr->size)) {
|
|
|
|
break;
|
|
|
|
}
|
2019-09-16 17:49:20 +00:00
|
|
|
if (mutable_cf_options_.compaction_options_universal.stop_style ==
|
2017-04-07 03:06:34 +00:00
|
|
|
kCompactionStopStyleSimilarSize) {
|
|
|
|
// Similar-size stopping rule: also check the last picked file isn't
|
|
|
|
// far larger than the next candidate file.
|
|
|
|
sz = (succeeding_sr->size * (100.0 + ratio)) / 100.0;
|
|
|
|
if (sz < static_cast<double>(candidate_size)) {
|
|
|
|
// If the small file we've encountered begins a run of similar-size
|
|
|
|
// files, we'll pick them up on a future iteration of the outer
|
|
|
|
// loop. If it's some lonely straggler, it'll eventually get picked
|
|
|
|
// by the last-resort read amp strategy which disregards size ratios.
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
candidate_size = succeeding_sr->compensated_file_size;
|
|
|
|
} else { // default kCompactionStopStyleTotalSize
|
|
|
|
candidate_size += succeeding_sr->compensated_file_size;
|
|
|
|
}
|
|
|
|
candidate_count++;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Found a series of consecutive files that need compaction.
|
|
|
|
if (candidate_count >= (unsigned int)min_merge_width) {
|
|
|
|
start_index = loop;
|
|
|
|
done = true;
|
|
|
|
break;
|
|
|
|
} else {
|
|
|
|
for (size_t i = loop;
|
2019-09-16 17:49:20 +00:00
|
|
|
i < loop + candidate_count && i < sorted_runs_.size(); i++) {
|
|
|
|
const SortedRun* skipping_sr = &sorted_runs_[i];
|
2017-04-07 03:06:34 +00:00
|
|
|
char file_num_buf[256];
|
|
|
|
skipping_sr->DumpSizeInfo(file_num_buf, sizeof(file_num_buf), loop);
|
2019-09-16 17:49:20 +00:00
|
|
|
ROCKS_LOG_BUFFER(log_buffer_, "[%s] Universal: Skipping %s",
|
|
|
|
cf_name_.c_str(), file_num_buf);
|
2017-04-07 03:06:34 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (!done || candidate_count <= 1) {
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
size_t first_index_after = start_index + candidate_count;
|
|
|
|
// Compression is enabled if files compacted earlier already reached
|
|
|
|
// size ratio of compression.
|
|
|
|
bool enable_compression = true;
|
|
|
|
int ratio_to_compress =
|
2019-09-16 17:49:20 +00:00
|
|
|
mutable_cf_options_.compaction_options_universal.compression_size_percent;
|
2017-04-07 03:06:34 +00:00
|
|
|
if (ratio_to_compress >= 0) {
|
|
|
|
uint64_t total_size = 0;
|
2019-09-16 17:49:20 +00:00
|
|
|
for (auto& sorted_run : sorted_runs_) {
|
2017-04-07 03:06:34 +00:00
|
|
|
total_size += sorted_run.compensated_file_size;
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t older_file_size = 0;
|
2019-09-16 17:49:20 +00:00
|
|
|
for (size_t i = sorted_runs_.size() - 1; i >= first_index_after; i--) {
|
|
|
|
older_file_size += sorted_runs_[i].size;
|
2017-04-07 03:06:34 +00:00
|
|
|
if (older_file_size * 100L >= total_size * (long)ratio_to_compress) {
|
|
|
|
enable_compression = false;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t estimated_total_size = 0;
|
|
|
|
for (unsigned int i = 0; i < first_index_after; i++) {
|
2019-09-16 17:49:20 +00:00
|
|
|
estimated_total_size += sorted_runs_[i].size;
|
2017-04-07 03:06:34 +00:00
|
|
|
}
|
2017-12-11 21:12:12 +00:00
|
|
|
uint32_t path_id =
|
2019-09-16 17:49:20 +00:00
|
|
|
GetPathId(ioptions_, mutable_cf_options_, estimated_total_size);
|
|
|
|
int start_level = sorted_runs_[start_index].level;
|
2017-04-07 03:06:34 +00:00
|
|
|
int output_level;
|
2023-06-14 18:28:56 +00:00
|
|
|
// last level is reserved for the files ingested behind
|
|
|
|
int max_output_level =
|
|
|
|
vstorage_->MaxOutputLevel(ioptions_.allow_ingest_behind);
|
2019-09-16 17:49:20 +00:00
|
|
|
if (first_index_after == sorted_runs_.size()) {
|
2023-06-14 18:28:56 +00:00
|
|
|
output_level = max_output_level;
|
2019-09-16 17:49:20 +00:00
|
|
|
} else if (sorted_runs_[first_index_after].level == 0) {
|
2017-04-07 03:06:34 +00:00
|
|
|
output_level = 0;
|
|
|
|
} else {
|
2019-09-16 17:49:20 +00:00
|
|
|
output_level = sorted_runs_[first_index_after].level - 1;
|
2017-04-07 03:06:34 +00:00
|
|
|
}
|
|
|
|
|
2023-06-14 18:28:56 +00:00
|
|
|
std::vector<CompactionInputFiles> inputs(max_output_level + 1);
|
2017-04-07 03:06:34 +00:00
|
|
|
for (size_t i = 0; i < inputs.size(); ++i) {
|
|
|
|
inputs[i].level = start_level + static_cast<int>(i);
|
|
|
|
}
|
|
|
|
for (size_t i = start_index; i < first_index_after; i++) {
|
2019-09-16 17:49:20 +00:00
|
|
|
auto& picking_sr = sorted_runs_[i];
|
2017-04-07 03:06:34 +00:00
|
|
|
if (picking_sr.level == 0) {
|
|
|
|
FileMetaData* picking_file = picking_sr.file;
|
|
|
|
inputs[0].files.push_back(picking_file);
|
|
|
|
} else {
|
|
|
|
auto& files = inputs[picking_sr.level - start_level].files;
|
2019-09-16 17:49:20 +00:00
|
|
|
for (auto* f : vstorage_->LevelFiles(picking_sr.level)) {
|
2017-04-07 03:06:34 +00:00
|
|
|
files.push_back(f);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
char file_num_buf[256];
|
|
|
|
picking_sr.DumpSizeInfo(file_num_buf, sizeof(file_num_buf), i);
|
2019-09-16 17:49:20 +00:00
|
|
|
ROCKS_LOG_BUFFER(log_buffer_, "[%s] Universal: Picking %s",
|
|
|
|
cf_name_.c_str(), file_num_buf);
|
2017-04-07 03:06:34 +00:00
|
|
|
}
|
|
|
|
|
Incremental Space Amp Compactions in Universal Style (#8655)
Summary:
This commit introduces incremental compaction in univeral style for space amplification. This follows the first improvement mentioned in https://rocksdb.org/blog/2021/04/12/universal-improvements.html . The implemention simply picks up files about size of max_compaction_bytes to compact and execute if the penalty is not too big. More optimizations can be done in the future, e.g. prioritizing between this compaction and other types. But for now, the feature is supposed to be functional and can often reduce frequency of full compactions, although it can introduce penalty.
In order to add cut files more efficiently so that more files from upper levels can be included, SST file cutting threshold (for current file + overlapping parent level files) is set to 1.5X of target file size. A 2MB target file size will generate files like this: https://gist.github.com/siying/29d2676fba417404f3c95e6c013c7de8 Number of files indeed increases but it is not out of control.
Two set of write benchmarks are run:
1. For ingestion rate limited scenario, we can see full compaction is mostly eliminated: https://gist.github.com/siying/959bc1186066906831cf4c808d6e0a19 . The write amp increased from 7.7 to 9.4, as expected. After applying file cutting, the number is improved to 8.9. In another benchmark, the write amp is even better with the incremental approach: https://gist.github.com/siying/d1c16c286d7c59c4d7bba718ca198163
2. For ingestion rate unlimited scenario, incremental compaction turns out to be too expensive most of the time and is not executed, as expected.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8655
Test Plan: Add unit tests to the functionality.
Reviewed By: ajkr
Differential Revision: D31787034
fbshipit-source-id: ce813e63b15a61d5a56e97bf8902a1b28e011beb
2021-10-20 17:03:03 +00:00
|
|
|
std::vector<FileMetaData*> grandparents;
|
|
|
|
// Include grandparents for potential file cutting in incremental
|
|
|
|
// mode. It is for aligning file cutting boundaries across levels,
|
|
|
|
// so that subsequent compactions can pick files with aligned
|
|
|
|
// buffer.
|
|
|
|
// Single files are only picked up in incremental mode, so that
|
|
|
|
// there is no need for full range.
|
|
|
|
if (mutable_cf_options_.compaction_options_universal.incremental &&
|
|
|
|
first_index_after < sorted_runs_.size() &&
|
|
|
|
sorted_runs_[first_index_after].level > 1) {
|
|
|
|
grandparents = vstorage_->LevelFiles(sorted_runs_[first_index_after].level);
|
|
|
|
}
|
|
|
|
|
2022-10-22 15:57:38 +00:00
|
|
|
if (output_level != 0 &&
|
|
|
|
picker_->FilesRangeOverlapWithCompaction(
|
|
|
|
inputs, output_level,
|
|
|
|
Compaction::EvaluatePenultimateLevel(vstorage_, ioptions_,
|
|
|
|
start_level, output_level))) {
|
|
|
|
return nullptr;
|
|
|
|
}
|
2017-04-07 03:06:34 +00:00
|
|
|
CompactionReason compaction_reason;
|
|
|
|
if (max_number_of_files_to_compact == UINT_MAX) {
|
|
|
|
compaction_reason = CompactionReason::kUniversalSizeRatio;
|
2018-01-26 19:01:54 +00:00
|
|
|
} else {
|
|
|
|
compaction_reason = CompactionReason::kUniversalSortedRunNum;
|
2017-04-07 03:06:34 +00:00
|
|
|
}
|
2022-03-08 02:06:19 +00:00
|
|
|
return new Compaction(vstorage_, ioptions_, mutable_cf_options_,
|
|
|
|
mutable_db_options_, std::move(inputs), output_level,
|
|
|
|
MaxFileSizeForLevel(mutable_cf_options_, output_level,
|
|
|
|
kCompactionStyleUniversal),
|
|
|
|
GetMaxOverlappingBytes(), path_id,
|
|
|
|
GetCompressionType(vstorage_, mutable_cf_options_,
|
2022-08-23 21:58:02 +00:00
|
|
|
output_level, 1, enable_compression),
|
2022-03-08 02:06:19 +00:00
|
|
|
GetCompressionOptions(mutable_cf_options_, vstorage_,
|
2022-08-23 21:58:02 +00:00
|
|
|
output_level, enable_compression),
|
2022-03-08 02:06:19 +00:00
|
|
|
Temperature::kUnknown,
|
|
|
|
/* max_subcompactions */ 0, grandparents,
|
2022-03-12 00:13:23 +00:00
|
|
|
/* is manual */ false, /* trim_ts */ "", score_,
|
2022-07-01 01:04:23 +00:00
|
|
|
false /* deletion_compaction */,
|
|
|
|
/* l0_files_might_overlap */ true, compaction_reason);
|
2017-04-07 03:06:34 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// Look at overall size amplification. If size amplification
|
2021-03-26 04:17:17 +00:00
|
|
|
// exceeds the configured value, then do a compaction
|
2023-08-23 18:27:48 +00:00
|
|
|
// on longest span of candidate files without conflict with other compactions
|
|
|
|
// ending at the earliest base file (overriding configured values of file-size
|
|
|
|
// ratios, min_merge_width and max_merge_width).
|
2019-09-16 17:49:20 +00:00
|
|
|
Compaction* UniversalCompactionBuilder::PickCompactionToReduceSizeAmp() {
|
|
|
|
assert(!sorted_runs_.empty());
|
2023-08-23 18:27:48 +00:00
|
|
|
|
|
|
|
const size_t end_index = ShouldSkipLastSortedRunForSizeAmpCompaction()
|
|
|
|
? sorted_runs_.size() - 2
|
|
|
|
: sorted_runs_.size() - 1;
|
|
|
|
if (sorted_runs_[end_index].being_compacted) {
|
2018-06-26 17:28:41 +00:00
|
|
|
return nullptr;
|
|
|
|
}
|
2023-09-12 17:00:57 +00:00
|
|
|
const uint64_t base_sr_size = sorted_runs_[end_index].size;
|
2023-08-23 18:27:48 +00:00
|
|
|
size_t start_index = end_index;
|
2023-09-12 17:00:57 +00:00
|
|
|
uint64_t candidate_size = 0;
|
2023-09-12 22:53:15 +00:00
|
|
|
size_t num_l0_files = 0;
|
2018-06-26 17:28:41 +00:00
|
|
|
|
2023-08-23 18:27:48 +00:00
|
|
|
// Get longest span (i.e, [start_index, end_index]) of available sorted runs
|
|
|
|
while (start_index > 0) {
|
|
|
|
const SortedRun* sr = &sorted_runs_[start_index - 1];
|
|
|
|
if (sr->being_compacted) {
|
|
|
|
char file_num_buf[kFormatFileNumberBufSize];
|
|
|
|
sr->Dump(file_num_buf, sizeof(file_num_buf), true);
|
|
|
|
ROCKS_LOG_BUFFER(
|
|
|
|
log_buffer_,
|
|
|
|
"[%s] Universal: stopping at sorted run undergoing compaction: "
|
|
|
|
"%s[%" ROCKSDB_PRIszt "]",
|
|
|
|
cf_name_.c_str(), file_num_buf, start_index - 1);
|
2017-04-07 03:06:34 +00:00
|
|
|
break;
|
|
|
|
}
|
2023-08-23 18:27:48 +00:00
|
|
|
candidate_size += sr->compensated_file_size;
|
2023-09-12 22:53:15 +00:00
|
|
|
num_l0_files += sr->level == 0 ? 1 : 0;
|
2023-08-23 18:27:48 +00:00
|
|
|
--start_index;
|
2017-04-07 03:06:34 +00:00
|
|
|
}
|
|
|
|
|
2023-08-23 18:27:48 +00:00
|
|
|
if (start_index == end_index) {
|
|
|
|
return nullptr;
|
2017-04-07 03:06:34 +00:00
|
|
|
}
|
2023-08-23 18:27:48 +00:00
|
|
|
|
2023-09-12 22:53:15 +00:00
|
|
|
{
|
|
|
|
const size_t num_l0_to_exclude = MightExcludeNewL0sToReduceWriteStop(
|
|
|
|
num_l0_files, end_index, start_index, candidate_size);
|
|
|
|
ROCKS_LOG_BUFFER(log_buffer_,
|
|
|
|
"[%s] Universal: Excluding %" ROCKSDB_PRIszt
|
|
|
|
" latest L0 files to reduce potential write stop "
|
|
|
|
"triggered by `level0_stop_writes_trigger`",
|
|
|
|
cf_name_.c_str(), num_l0_to_exclude);
|
|
|
|
}
|
|
|
|
|
2017-04-07 03:06:34 +00:00
|
|
|
{
|
|
|
|
char file_num_buf[kFormatFileNumberBufSize];
|
2023-08-23 18:27:48 +00:00
|
|
|
sorted_runs_[start_index].Dump(file_num_buf, sizeof(file_num_buf), true);
|
2017-04-07 03:06:34 +00:00
|
|
|
ROCKS_LOG_BUFFER(
|
2019-09-16 17:49:20 +00:00
|
|
|
log_buffer_,
|
2017-04-07 03:06:34 +00:00
|
|
|
"[%s] Universal: First candidate %s[%" ROCKSDB_PRIszt "] %s",
|
2019-09-16 17:49:20 +00:00
|
|
|
cf_name_.c_str(), file_num_buf, start_index, " to reduce size amp.\n");
|
2017-04-07 03:06:34 +00:00
|
|
|
}
|
|
|
|
|
2023-08-23 18:27:48 +00:00
|
|
|
// percentage flexibility while reducing size amplification
|
|
|
|
const uint64_t ratio = mutable_cf_options_.compaction_options_universal
|
|
|
|
.max_size_amplification_percent;
|
2017-04-07 03:06:34 +00:00
|
|
|
|
|
|
|
// size amplification = percentage of additional size
|
2022-08-08 21:34:36 +00:00
|
|
|
if (candidate_size * 100 < ratio * base_sr_size) {
|
2017-04-07 03:06:34 +00:00
|
|
|
ROCKS_LOG_BUFFER(
|
2019-09-16 17:49:20 +00:00
|
|
|
log_buffer_,
|
2017-04-07 03:06:34 +00:00
|
|
|
"[%s] Universal: size amp not needed. newer-files-total-size %" PRIu64
|
|
|
|
" earliest-file-size %" PRIu64,
|
2022-08-08 21:34:36 +00:00
|
|
|
cf_name_.c_str(), candidate_size, base_sr_size);
|
2017-04-07 03:06:34 +00:00
|
|
|
return nullptr;
|
|
|
|
} else {
|
|
|
|
ROCKS_LOG_BUFFER(
|
2019-09-16 17:49:20 +00:00
|
|
|
log_buffer_,
|
2017-04-07 03:06:34 +00:00
|
|
|
"[%s] Universal: size amp needed. newer-files-total-size %" PRIu64
|
|
|
|
" earliest-file-size %" PRIu64,
|
2022-08-08 21:34:36 +00:00
|
|
|
cf_name_.c_str(), candidate_size, base_sr_size);
|
2017-04-07 03:06:34 +00:00
|
|
|
}
|
Incremental Space Amp Compactions in Universal Style (#8655)
Summary:
This commit introduces incremental compaction in univeral style for space amplification. This follows the first improvement mentioned in https://rocksdb.org/blog/2021/04/12/universal-improvements.html . The implemention simply picks up files about size of max_compaction_bytes to compact and execute if the penalty is not too big. More optimizations can be done in the future, e.g. prioritizing between this compaction and other types. But for now, the feature is supposed to be functional and can often reduce frequency of full compactions, although it can introduce penalty.
In order to add cut files more efficiently so that more files from upper levels can be included, SST file cutting threshold (for current file + overlapping parent level files) is set to 1.5X of target file size. A 2MB target file size will generate files like this: https://gist.github.com/siying/29d2676fba417404f3c95e6c013c7de8 Number of files indeed increases but it is not out of control.
Two set of write benchmarks are run:
1. For ingestion rate limited scenario, we can see full compaction is mostly eliminated: https://gist.github.com/siying/959bc1186066906831cf4c808d6e0a19 . The write amp increased from 7.7 to 9.4, as expected. After applying file cutting, the number is improved to 8.9. In another benchmark, the write amp is even better with the incremental approach: https://gist.github.com/siying/d1c16c286d7c59c4d7bba718ca198163
2. For ingestion rate unlimited scenario, incremental compaction turns out to be too expensive most of the time and is not executed, as expected.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8655
Test Plan: Add unit tests to the functionality.
Reviewed By: ajkr
Differential Revision: D31787034
fbshipit-source-id: ce813e63b15a61d5a56e97bf8902a1b28e011beb
2021-10-20 17:03:03 +00:00
|
|
|
// Since incremental compaction can't include more than second last
|
|
|
|
// level, it can introduce penalty, compared to full compaction. We
|
|
|
|
// hard code the pentalty to be 80%. If we end up with a compaction
|
|
|
|
// fanout higher than 80% of full level compactions, we fall back
|
|
|
|
// to full level compaction.
|
|
|
|
// The 80% threshold is arbitrary and can be adjusted or made
|
|
|
|
// configurable in the future.
|
|
|
|
// This also prevent the case when compaction falls behind and we
|
|
|
|
// need to compact more levels for compactions to catch up.
|
|
|
|
if (mutable_cf_options_.compaction_options_universal.incremental) {
|
2022-08-08 21:34:36 +00:00
|
|
|
double fanout_threshold = static_cast<double>(base_sr_size) /
|
Incremental Space Amp Compactions in Universal Style (#8655)
Summary:
This commit introduces incremental compaction in univeral style for space amplification. This follows the first improvement mentioned in https://rocksdb.org/blog/2021/04/12/universal-improvements.html . The implemention simply picks up files about size of max_compaction_bytes to compact and execute if the penalty is not too big. More optimizations can be done in the future, e.g. prioritizing between this compaction and other types. But for now, the feature is supposed to be functional and can often reduce frequency of full compactions, although it can introduce penalty.
In order to add cut files more efficiently so that more files from upper levels can be included, SST file cutting threshold (for current file + overlapping parent level files) is set to 1.5X of target file size. A 2MB target file size will generate files like this: https://gist.github.com/siying/29d2676fba417404f3c95e6c013c7de8 Number of files indeed increases but it is not out of control.
Two set of write benchmarks are run:
1. For ingestion rate limited scenario, we can see full compaction is mostly eliminated: https://gist.github.com/siying/959bc1186066906831cf4c808d6e0a19 . The write amp increased from 7.7 to 9.4, as expected. After applying file cutting, the number is improved to 8.9. In another benchmark, the write amp is even better with the incremental approach: https://gist.github.com/siying/d1c16c286d7c59c4d7bba718ca198163
2. For ingestion rate unlimited scenario, incremental compaction turns out to be too expensive most of the time and is not executed, as expected.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8655
Test Plan: Add unit tests to the functionality.
Reviewed By: ajkr
Differential Revision: D31787034
fbshipit-source-id: ce813e63b15a61d5a56e97bf8902a1b28e011beb
2021-10-20 17:03:03 +00:00
|
|
|
static_cast<double>(candidate_size) * 1.8;
|
|
|
|
Compaction* picked = PickIncrementalForReduceSizeAmp(fanout_threshold);
|
|
|
|
if (picked != nullptr) {
|
|
|
|
// As the feature is still incremental, picking incremental compaction
|
|
|
|
// might fail and we will fall bck to compacting full level.
|
|
|
|
return picked;
|
|
|
|
}
|
|
|
|
}
|
2022-08-08 21:34:36 +00:00
|
|
|
return PickCompactionWithSortedRunRange(
|
2023-08-23 18:27:48 +00:00
|
|
|
start_index, end_index, CompactionReason::kUniversalSizeAmplification);
|
2017-04-07 03:06:34 +00:00
|
|
|
}
|
2018-05-29 22:42:14 +00:00
|
|
|
|
Incremental Space Amp Compactions in Universal Style (#8655)
Summary:
This commit introduces incremental compaction in univeral style for space amplification. This follows the first improvement mentioned in https://rocksdb.org/blog/2021/04/12/universal-improvements.html . The implemention simply picks up files about size of max_compaction_bytes to compact and execute if the penalty is not too big. More optimizations can be done in the future, e.g. prioritizing between this compaction and other types. But for now, the feature is supposed to be functional and can often reduce frequency of full compactions, although it can introduce penalty.
In order to add cut files more efficiently so that more files from upper levels can be included, SST file cutting threshold (for current file + overlapping parent level files) is set to 1.5X of target file size. A 2MB target file size will generate files like this: https://gist.github.com/siying/29d2676fba417404f3c95e6c013c7de8 Number of files indeed increases but it is not out of control.
Two set of write benchmarks are run:
1. For ingestion rate limited scenario, we can see full compaction is mostly eliminated: https://gist.github.com/siying/959bc1186066906831cf4c808d6e0a19 . The write amp increased from 7.7 to 9.4, as expected. After applying file cutting, the number is improved to 8.9. In another benchmark, the write amp is even better with the incremental approach: https://gist.github.com/siying/d1c16c286d7c59c4d7bba718ca198163
2. For ingestion rate unlimited scenario, incremental compaction turns out to be too expensive most of the time and is not executed, as expected.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8655
Test Plan: Add unit tests to the functionality.
Reviewed By: ajkr
Differential Revision: D31787034
fbshipit-source-id: ce813e63b15a61d5a56e97bf8902a1b28e011beb
2021-10-20 17:03:03 +00:00
|
|
|
Compaction* UniversalCompactionBuilder::PickIncrementalForReduceSizeAmp(
|
|
|
|
double fanout_threshold) {
|
|
|
|
// Try find all potential compactions with total size just over
|
|
|
|
// options.max_compaction_size / 2, and take the one with the lowest
|
|
|
|
// fanout (defined in declaration of the function).
|
|
|
|
// This is done by having a sliding window of the files at the second
|
|
|
|
// lowest level, and keep expanding while finding overlapping in the
|
|
|
|
// last level. Once total size exceeds the size threshold, calculate
|
|
|
|
// the fanout value. And then shrinking from the small side of the
|
|
|
|
// window. Keep doing it until the end.
|
|
|
|
// Finally, we try to include upper level files if they fall into
|
|
|
|
// the range.
|
|
|
|
//
|
|
|
|
// Note that it is a similar problem as leveled compaction's
|
|
|
|
// kMinOverlappingRatio priority, but instead of picking single files
|
|
|
|
// we expand to a target compaction size. The reason is that in
|
|
|
|
// leveled compaction, actual fanout value tends to high, e.g. 10, so
|
|
|
|
// even with single file in down merging level, the extra size
|
|
|
|
// compacted in boundary files is at a lower ratio. But here users
|
|
|
|
// often have size of second last level size to be 1/4, 1/3 or even
|
|
|
|
// 1/2 of the bottommost level, so picking single file in second most
|
|
|
|
// level will cause significant waste, which is not desirable.
|
|
|
|
//
|
|
|
|
// This algorithm has lots of room to improve to pick more efficient
|
|
|
|
// compactions.
|
|
|
|
assert(sorted_runs_.size() >= 2);
|
|
|
|
int second_last_level = sorted_runs_[sorted_runs_.size() - 2].level;
|
|
|
|
if (second_last_level == 0) {
|
|
|
|
// Can't split Level 0.
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
int output_level = sorted_runs_.back().level;
|
|
|
|
const std::vector<FileMetaData*>& bottom_files =
|
|
|
|
vstorage_->LevelFiles(output_level);
|
|
|
|
const std::vector<FileMetaData*>& files =
|
|
|
|
vstorage_->LevelFiles(second_last_level);
|
|
|
|
assert(!bottom_files.empty());
|
|
|
|
assert(!files.empty());
|
|
|
|
|
|
|
|
// std::unordered_map<uint64_t, uint64_t> file_to_order;
|
|
|
|
|
|
|
|
int picked_start_idx = 0;
|
|
|
|
int picked_end_idx = 0;
|
|
|
|
double picked_fanout = fanout_threshold;
|
|
|
|
|
|
|
|
// Use half target compaction bytes as anchor to stop growing second most
|
|
|
|
// level files, and reserve growing space for more overlapping bottom level,
|
|
|
|
// clean cut, files from other levels, etc.
|
|
|
|
uint64_t comp_thres_size = mutable_cf_options_.max_compaction_bytes / 2;
|
|
|
|
int start_idx = 0;
|
|
|
|
int bottom_end_idx = 0;
|
|
|
|
int bottom_start_idx = 0;
|
|
|
|
uint64_t non_bottom_size = 0;
|
|
|
|
uint64_t bottom_size = 0;
|
|
|
|
bool end_bottom_size_counted = false;
|
|
|
|
for (int end_idx = 0; end_idx < static_cast<int>(files.size()); end_idx++) {
|
|
|
|
FileMetaData* end_file = files[end_idx];
|
|
|
|
|
|
|
|
// Include bottom most level files smaller than the current second
|
|
|
|
// last level file.
|
|
|
|
int num_skipped = 0;
|
|
|
|
while (bottom_end_idx < static_cast<int>(bottom_files.size()) &&
|
|
|
|
icmp_->Compare(bottom_files[bottom_end_idx]->largest,
|
|
|
|
end_file->smallest) < 0) {
|
|
|
|
if (!end_bottom_size_counted) {
|
|
|
|
bottom_size += bottom_files[bottom_end_idx]->fd.file_size;
|
|
|
|
}
|
|
|
|
bottom_end_idx++;
|
|
|
|
end_bottom_size_counted = false;
|
|
|
|
num_skipped++;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (num_skipped > 1) {
|
|
|
|
// At least a file in the bottom most level falls into the file gap. No
|
|
|
|
// reason to include the file. We cut the range and start a new sliding
|
|
|
|
// window.
|
|
|
|
start_idx = end_idx;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (start_idx == end_idx) {
|
|
|
|
// new sliding window.
|
|
|
|
non_bottom_size = 0;
|
|
|
|
bottom_size = 0;
|
|
|
|
bottom_start_idx = bottom_end_idx;
|
|
|
|
end_bottom_size_counted = false;
|
|
|
|
}
|
|
|
|
|
|
|
|
non_bottom_size += end_file->fd.file_size;
|
|
|
|
|
|
|
|
// Include all overlapping files in bottom level.
|
|
|
|
while (bottom_end_idx < static_cast<int>(bottom_files.size()) &&
|
|
|
|
icmp_->Compare(bottom_files[bottom_end_idx]->smallest,
|
|
|
|
end_file->largest) < 0) {
|
|
|
|
if (!end_bottom_size_counted) {
|
|
|
|
bottom_size += bottom_files[bottom_end_idx]->fd.file_size;
|
|
|
|
end_bottom_size_counted = true;
|
|
|
|
}
|
|
|
|
if (icmp_->Compare(bottom_files[bottom_end_idx]->largest,
|
|
|
|
end_file->largest) > 0) {
|
|
|
|
// next level file cross large boundary of current file.
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
bottom_end_idx++;
|
|
|
|
end_bottom_size_counted = false;
|
|
|
|
}
|
|
|
|
|
|
|
|
if ((non_bottom_size + bottom_size > comp_thres_size ||
|
|
|
|
end_idx == static_cast<int>(files.size()) - 1) &&
|
|
|
|
non_bottom_size > 0) { // Do we alow 0 size file at all?
|
|
|
|
// If it is a better compaction, remember it in picked* variables.
|
|
|
|
double fanout = static_cast<double>(bottom_size) /
|
|
|
|
static_cast<double>(non_bottom_size);
|
|
|
|
if (fanout < picked_fanout) {
|
|
|
|
picked_start_idx = start_idx;
|
|
|
|
picked_end_idx = end_idx;
|
|
|
|
picked_fanout = fanout;
|
|
|
|
}
|
|
|
|
// Shrink from the start end to under comp_thres_size
|
|
|
|
while (non_bottom_size + bottom_size > comp_thres_size &&
|
|
|
|
start_idx <= end_idx) {
|
|
|
|
non_bottom_size -= files[start_idx]->fd.file_size;
|
|
|
|
start_idx++;
|
|
|
|
if (start_idx < static_cast<int>(files.size())) {
|
|
|
|
while (bottom_start_idx <= bottom_end_idx &&
|
|
|
|
icmp_->Compare(bottom_files[bottom_start_idx]->largest,
|
|
|
|
files[start_idx]->smallest) < 0) {
|
|
|
|
bottom_size -= bottom_files[bottom_start_idx]->fd.file_size;
|
|
|
|
bottom_start_idx++;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (picked_fanout >= fanout_threshold) {
|
|
|
|
assert(picked_fanout == fanout_threshold);
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
std::vector<CompactionInputFiles> inputs;
|
|
|
|
CompactionInputFiles bottom_level_inputs;
|
|
|
|
CompactionInputFiles second_last_level_inputs;
|
|
|
|
second_last_level_inputs.level = second_last_level;
|
|
|
|
bottom_level_inputs.level = output_level;
|
|
|
|
for (int i = picked_start_idx; i <= picked_end_idx; i++) {
|
|
|
|
if (files[i]->being_compacted) {
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
second_last_level_inputs.files.push_back(files[i]);
|
|
|
|
}
|
|
|
|
assert(!second_last_level_inputs.empty());
|
|
|
|
if (!picker_->ExpandInputsToCleanCut(cf_name_, vstorage_,
|
|
|
|
&second_last_level_inputs,
|
|
|
|
/*next_smallest=*/nullptr)) {
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
// We might be able to avoid this binary search if we save and expand
|
|
|
|
// from bottom_start_idx and bottom_end_idx, but for now, we use
|
|
|
|
// SetupOtherInputs() for simplicity.
|
|
|
|
int parent_index = -1; // Create and use bottom_start_idx?
|
2024-02-03 01:09:42 +00:00
|
|
|
if (!picker_->SetupOtherInputs(cf_name_, vstorage_, &second_last_level_inputs,
|
Incremental Space Amp Compactions in Universal Style (#8655)
Summary:
This commit introduces incremental compaction in univeral style for space amplification. This follows the first improvement mentioned in https://rocksdb.org/blog/2021/04/12/universal-improvements.html . The implemention simply picks up files about size of max_compaction_bytes to compact and execute if the penalty is not too big. More optimizations can be done in the future, e.g. prioritizing between this compaction and other types. But for now, the feature is supposed to be functional and can often reduce frequency of full compactions, although it can introduce penalty.
In order to add cut files more efficiently so that more files from upper levels can be included, SST file cutting threshold (for current file + overlapping parent level files) is set to 1.5X of target file size. A 2MB target file size will generate files like this: https://gist.github.com/siying/29d2676fba417404f3c95e6c013c7de8 Number of files indeed increases but it is not out of control.
Two set of write benchmarks are run:
1. For ingestion rate limited scenario, we can see full compaction is mostly eliminated: https://gist.github.com/siying/959bc1186066906831cf4c808d6e0a19 . The write amp increased from 7.7 to 9.4, as expected. After applying file cutting, the number is improved to 8.9. In another benchmark, the write amp is even better with the incremental approach: https://gist.github.com/siying/d1c16c286d7c59c4d7bba718ca198163
2. For ingestion rate unlimited scenario, incremental compaction turns out to be too expensive most of the time and is not executed, as expected.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8655
Test Plan: Add unit tests to the functionality.
Reviewed By: ajkr
Differential Revision: D31787034
fbshipit-source-id: ce813e63b15a61d5a56e97bf8902a1b28e011beb
2021-10-20 17:03:03 +00:00
|
|
|
&bottom_level_inputs, &parent_index,
|
|
|
|
/*base_index=*/-1)) {
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Try to include files in upper levels if they fall into the range.
|
|
|
|
// Since we need to go from lower level up and this is in the reverse
|
|
|
|
// order, compared to level order, we first write to an reversed
|
|
|
|
// data structure and finally copy them to compaction inputs.
|
|
|
|
InternalKey smallest, largest;
|
|
|
|
picker_->GetRange(second_last_level_inputs, &smallest, &largest);
|
|
|
|
std::vector<CompactionInputFiles> inputs_reverse;
|
|
|
|
for (auto it = ++(++sorted_runs_.rbegin()); it != sorted_runs_.rend(); it++) {
|
|
|
|
SortedRun& sr = *it;
|
|
|
|
if (sr.level == 0) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
std::vector<FileMetaData*> level_inputs;
|
|
|
|
vstorage_->GetCleanInputsWithinInterval(sr.level, &smallest, &largest,
|
|
|
|
&level_inputs);
|
|
|
|
if (!level_inputs.empty()) {
|
|
|
|
inputs_reverse.push_back({});
|
|
|
|
inputs_reverse.back().level = sr.level;
|
|
|
|
inputs_reverse.back().files = level_inputs;
|
|
|
|
picker_->GetRange(inputs_reverse.back(), &smallest, &largest);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
for (auto it = inputs_reverse.rbegin(); it != inputs_reverse.rend(); it++) {
|
|
|
|
inputs.push_back(*it);
|
|
|
|
}
|
|
|
|
|
|
|
|
inputs.push_back(second_last_level_inputs);
|
|
|
|
inputs.push_back(bottom_level_inputs);
|
|
|
|
|
2022-10-22 15:57:38 +00:00
|
|
|
int start_level = Compaction::kInvalidLevel;
|
|
|
|
for (const auto& in : inputs) {
|
|
|
|
if (!in.empty()) {
|
|
|
|
// inputs should already be sorted by level
|
|
|
|
start_level = in.level;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// intra L0 compactions outputs could have overlap
|
|
|
|
if (output_level != 0 &&
|
|
|
|
picker_->FilesRangeOverlapWithCompaction(
|
|
|
|
inputs, output_level,
|
|
|
|
Compaction::EvaluatePenultimateLevel(vstorage_, ioptions_,
|
|
|
|
start_level, output_level))) {
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
|
Incremental Space Amp Compactions in Universal Style (#8655)
Summary:
This commit introduces incremental compaction in univeral style for space amplification. This follows the first improvement mentioned in https://rocksdb.org/blog/2021/04/12/universal-improvements.html . The implemention simply picks up files about size of max_compaction_bytes to compact and execute if the penalty is not too big. More optimizations can be done in the future, e.g. prioritizing between this compaction and other types. But for now, the feature is supposed to be functional and can often reduce frequency of full compactions, although it can introduce penalty.
In order to add cut files more efficiently so that more files from upper levels can be included, SST file cutting threshold (for current file + overlapping parent level files) is set to 1.5X of target file size. A 2MB target file size will generate files like this: https://gist.github.com/siying/29d2676fba417404f3c95e6c013c7de8 Number of files indeed increases but it is not out of control.
Two set of write benchmarks are run:
1. For ingestion rate limited scenario, we can see full compaction is mostly eliminated: https://gist.github.com/siying/959bc1186066906831cf4c808d6e0a19 . The write amp increased from 7.7 to 9.4, as expected. After applying file cutting, the number is improved to 8.9. In another benchmark, the write amp is even better with the incremental approach: https://gist.github.com/siying/d1c16c286d7c59c4d7bba718ca198163
2. For ingestion rate unlimited scenario, incremental compaction turns out to be too expensive most of the time and is not executed, as expected.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8655
Test Plan: Add unit tests to the functionality.
Reviewed By: ajkr
Differential Revision: D31787034
fbshipit-source-id: ce813e63b15a61d5a56e97bf8902a1b28e011beb
2021-10-20 17:03:03 +00:00
|
|
|
// TODO support multi paths?
|
|
|
|
uint32_t path_id = 0;
|
|
|
|
return new Compaction(
|
|
|
|
vstorage_, ioptions_, mutable_cf_options_, mutable_db_options_,
|
|
|
|
std::move(inputs), output_level,
|
|
|
|
MaxFileSizeForLevel(mutable_cf_options_, output_level,
|
|
|
|
kCompactionStyleUniversal),
|
|
|
|
GetMaxOverlappingBytes(), path_id,
|
2022-03-08 02:06:19 +00:00
|
|
|
GetCompressionType(vstorage_, mutable_cf_options_, output_level, 1,
|
|
|
|
true /* enable_compression */),
|
Incremental Space Amp Compactions in Universal Style (#8655)
Summary:
This commit introduces incremental compaction in univeral style for space amplification. This follows the first improvement mentioned in https://rocksdb.org/blog/2021/04/12/universal-improvements.html . The implemention simply picks up files about size of max_compaction_bytes to compact and execute if the penalty is not too big. More optimizations can be done in the future, e.g. prioritizing between this compaction and other types. But for now, the feature is supposed to be functional and can often reduce frequency of full compactions, although it can introduce penalty.
In order to add cut files more efficiently so that more files from upper levels can be included, SST file cutting threshold (for current file + overlapping parent level files) is set to 1.5X of target file size. A 2MB target file size will generate files like this: https://gist.github.com/siying/29d2676fba417404f3c95e6c013c7de8 Number of files indeed increases but it is not out of control.
Two set of write benchmarks are run:
1. For ingestion rate limited scenario, we can see full compaction is mostly eliminated: https://gist.github.com/siying/959bc1186066906831cf4c808d6e0a19 . The write amp increased from 7.7 to 9.4, as expected. After applying file cutting, the number is improved to 8.9. In another benchmark, the write amp is even better with the incremental approach: https://gist.github.com/siying/d1c16c286d7c59c4d7bba718ca198163
2. For ingestion rate unlimited scenario, incremental compaction turns out to be too expensive most of the time and is not executed, as expected.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8655
Test Plan: Add unit tests to the functionality.
Reviewed By: ajkr
Differential Revision: D31787034
fbshipit-source-id: ce813e63b15a61d5a56e97bf8902a1b28e011beb
2021-10-20 17:03:03 +00:00
|
|
|
GetCompressionOptions(mutable_cf_options_, vstorage_, output_level,
|
|
|
|
true /* enable_compression */),
|
|
|
|
Temperature::kUnknown,
|
|
|
|
/* max_subcompactions */ 0, /* grandparents */ {}, /* is manual */ false,
|
2022-03-12 00:13:23 +00:00
|
|
|
/* trim_ts */ "", score_, false /* deletion_compaction */,
|
2022-07-01 01:04:23 +00:00
|
|
|
/* l0_files_might_overlap */ true,
|
Incremental Space Amp Compactions in Universal Style (#8655)
Summary:
This commit introduces incremental compaction in univeral style for space amplification. This follows the first improvement mentioned in https://rocksdb.org/blog/2021/04/12/universal-improvements.html . The implemention simply picks up files about size of max_compaction_bytes to compact and execute if the penalty is not too big. More optimizations can be done in the future, e.g. prioritizing between this compaction and other types. But for now, the feature is supposed to be functional and can often reduce frequency of full compactions, although it can introduce penalty.
In order to add cut files more efficiently so that more files from upper levels can be included, SST file cutting threshold (for current file + overlapping parent level files) is set to 1.5X of target file size. A 2MB target file size will generate files like this: https://gist.github.com/siying/29d2676fba417404f3c95e6c013c7de8 Number of files indeed increases but it is not out of control.
Two set of write benchmarks are run:
1. For ingestion rate limited scenario, we can see full compaction is mostly eliminated: https://gist.github.com/siying/959bc1186066906831cf4c808d6e0a19 . The write amp increased from 7.7 to 9.4, as expected. After applying file cutting, the number is improved to 8.9. In another benchmark, the write amp is even better with the incremental approach: https://gist.github.com/siying/d1c16c286d7c59c4d7bba718ca198163
2. For ingestion rate unlimited scenario, incremental compaction turns out to be too expensive most of the time and is not executed, as expected.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8655
Test Plan: Add unit tests to the functionality.
Reviewed By: ajkr
Differential Revision: D31787034
fbshipit-source-id: ce813e63b15a61d5a56e97bf8902a1b28e011beb
2021-10-20 17:03:03 +00:00
|
|
|
CompactionReason::kUniversalSizeAmplification);
|
|
|
|
}
|
|
|
|
|
2018-05-29 22:42:14 +00:00
|
|
|
// Pick files marked for compaction. Typically, files are marked by
|
|
|
|
// CompactOnDeleteCollector due to the presence of tombstones.
|
2019-09-16 17:49:20 +00:00
|
|
|
Compaction* UniversalCompactionBuilder::PickDeleteTriggeredCompaction() {
|
2018-05-29 22:42:14 +00:00
|
|
|
CompactionInputFiles start_level_inputs;
|
|
|
|
int output_level;
|
|
|
|
std::vector<CompactionInputFiles> inputs;
|
Incremental Space Amp Compactions in Universal Style (#8655)
Summary:
This commit introduces incremental compaction in univeral style for space amplification. This follows the first improvement mentioned in https://rocksdb.org/blog/2021/04/12/universal-improvements.html . The implemention simply picks up files about size of max_compaction_bytes to compact and execute if the penalty is not too big. More optimizations can be done in the future, e.g. prioritizing between this compaction and other types. But for now, the feature is supposed to be functional and can often reduce frequency of full compactions, although it can introduce penalty.
In order to add cut files more efficiently so that more files from upper levels can be included, SST file cutting threshold (for current file + overlapping parent level files) is set to 1.5X of target file size. A 2MB target file size will generate files like this: https://gist.github.com/siying/29d2676fba417404f3c95e6c013c7de8 Number of files indeed increases but it is not out of control.
Two set of write benchmarks are run:
1. For ingestion rate limited scenario, we can see full compaction is mostly eliminated: https://gist.github.com/siying/959bc1186066906831cf4c808d6e0a19 . The write amp increased from 7.7 to 9.4, as expected. After applying file cutting, the number is improved to 8.9. In another benchmark, the write amp is even better with the incremental approach: https://gist.github.com/siying/d1c16c286d7c59c4d7bba718ca198163
2. For ingestion rate unlimited scenario, incremental compaction turns out to be too expensive most of the time and is not executed, as expected.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8655
Test Plan: Add unit tests to the functionality.
Reviewed By: ajkr
Differential Revision: D31787034
fbshipit-source-id: ce813e63b15a61d5a56e97bf8902a1b28e011beb
2021-10-20 17:03:03 +00:00
|
|
|
std::vector<FileMetaData*> grandparents;
|
2018-05-29 22:42:14 +00:00
|
|
|
|
2019-09-16 17:49:20 +00:00
|
|
|
if (vstorage_->num_levels() == 1) {
|
2018-05-29 22:42:14 +00:00
|
|
|
// This is single level universal. Since we're basically trying to reclaim
|
|
|
|
// space by processing files marked for compaction due to high tombstone
|
|
|
|
// density, let's do the same thing as compaction to reduce size amp which
|
|
|
|
// has the same goals.
|
2020-08-10 19:17:46 +00:00
|
|
|
int start_index = -1;
|
2018-05-29 22:42:14 +00:00
|
|
|
|
|
|
|
start_level_inputs.level = 0;
|
|
|
|
start_level_inputs.files.clear();
|
|
|
|
output_level = 0;
|
2020-08-10 19:17:46 +00:00
|
|
|
// Find the first file marked for compaction. Ignore the last file
|
|
|
|
for (size_t loop = 0; loop + 1 < sorted_runs_.size(); loop++) {
|
|
|
|
SortedRun* sr = &sorted_runs_[loop];
|
|
|
|
if (sr->being_compacted) {
|
|
|
|
continue;
|
2018-05-29 22:42:14 +00:00
|
|
|
}
|
2020-08-10 19:17:46 +00:00
|
|
|
FileMetaData* f = vstorage_->LevelFiles(0)[loop];
|
|
|
|
if (f->marked_for_compaction) {
|
2018-05-29 22:42:14 +00:00
|
|
|
start_level_inputs.files.push_back(f);
|
2020-08-10 19:17:46 +00:00
|
|
|
start_index =
|
|
|
|
static_cast<int>(loop); // Consider this as the first candidate.
|
|
|
|
break;
|
2018-05-29 22:42:14 +00:00
|
|
|
}
|
|
|
|
}
|
2020-08-10 19:17:46 +00:00
|
|
|
if (start_index < 0) {
|
|
|
|
// Either no file marked, or they're already being compacted
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
for (size_t loop = start_index + 1; loop < sorted_runs_.size(); loop++) {
|
|
|
|
SortedRun* sr = &sorted_runs_[loop];
|
|
|
|
if (sr->being_compacted) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
FileMetaData* f = vstorage_->LevelFiles(0)[loop];
|
|
|
|
start_level_inputs.files.push_back(f);
|
|
|
|
}
|
2018-05-29 22:42:14 +00:00
|
|
|
if (start_level_inputs.size() <= 1) {
|
|
|
|
// If only the last file in L0 is marked for compaction, ignore it
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
inputs.push_back(start_level_inputs);
|
|
|
|
} else {
|
|
|
|
int start_level;
|
|
|
|
|
|
|
|
// For multi-level universal, the strategy is to make this look more like
|
|
|
|
// leveled. We pick one of the files marked for compaction and compact with
|
|
|
|
// overlapping files in the adjacent level.
|
2019-09-16 17:49:20 +00:00
|
|
|
picker_->PickFilesMarkedForCompaction(cf_name_, vstorage_, &start_level,
|
|
|
|
&output_level, &start_level_inputs);
|
2018-05-29 22:42:14 +00:00
|
|
|
if (start_level_inputs.empty()) {
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
|
2023-06-14 18:28:56 +00:00
|
|
|
int max_output_level =
|
|
|
|
vstorage_->MaxOutputLevel(ioptions_.allow_ingest_behind);
|
2018-05-29 22:42:14 +00:00
|
|
|
// Pick the first non-empty level after the start_level
|
2023-06-14 18:28:56 +00:00
|
|
|
for (output_level = start_level + 1; output_level <= max_output_level;
|
2018-05-29 22:42:14 +00:00
|
|
|
output_level++) {
|
2019-09-16 17:49:20 +00:00
|
|
|
if (vstorage_->NumLevelFiles(output_level) != 0) {
|
2018-05-29 22:42:14 +00:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// If all higher levels are empty, pick the highest level as output level
|
2023-06-14 18:28:56 +00:00
|
|
|
if (output_level > max_output_level) {
|
2018-05-29 22:42:14 +00:00
|
|
|
if (start_level == 0) {
|
2023-06-14 18:28:56 +00:00
|
|
|
output_level = max_output_level;
|
2018-05-29 22:42:14 +00:00
|
|
|
} else {
|
|
|
|
// If start level is non-zero and all higher levels are empty, this
|
|
|
|
// compaction will translate into a trivial move. Since the idea is
|
|
|
|
// to reclaim space and trivial move doesn't help with that, we
|
|
|
|
// skip compaction in this case and return nullptr
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
}
|
2023-06-14 18:28:56 +00:00
|
|
|
assert(output_level <= max_output_level);
|
2018-05-29 22:42:14 +00:00
|
|
|
|
|
|
|
if (output_level != 0) {
|
|
|
|
if (start_level == 0) {
|
2019-09-16 17:49:20 +00:00
|
|
|
if (!picker_->GetOverlappingL0Files(vstorage_, &start_level_inputs,
|
|
|
|
output_level, nullptr)) {
|
2018-05-29 22:42:14 +00:00
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
CompactionInputFiles output_level_inputs;
|
|
|
|
int parent_index = -1;
|
|
|
|
|
|
|
|
output_level_inputs.level = output_level;
|
2024-02-03 01:09:42 +00:00
|
|
|
if (!picker_->SetupOtherInputs(cf_name_, vstorage_, &start_level_inputs,
|
|
|
|
&output_level_inputs, &parent_index, -1)) {
|
2018-05-29 22:42:14 +00:00
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
inputs.push_back(start_level_inputs);
|
|
|
|
if (!output_level_inputs.empty()) {
|
|
|
|
inputs.push_back(output_level_inputs);
|
|
|
|
}
|
2022-10-22 15:57:38 +00:00
|
|
|
if (picker_->FilesRangeOverlapWithCompaction(
|
|
|
|
inputs, output_level,
|
|
|
|
Compaction::EvaluatePenultimateLevel(
|
|
|
|
vstorage_, ioptions_, start_level, output_level))) {
|
2018-05-29 22:42:14 +00:00
|
|
|
return nullptr;
|
|
|
|
}
|
Incremental Space Amp Compactions in Universal Style (#8655)
Summary:
This commit introduces incremental compaction in univeral style for space amplification. This follows the first improvement mentioned in https://rocksdb.org/blog/2021/04/12/universal-improvements.html . The implemention simply picks up files about size of max_compaction_bytes to compact and execute if the penalty is not too big. More optimizations can be done in the future, e.g. prioritizing between this compaction and other types. But for now, the feature is supposed to be functional and can often reduce frequency of full compactions, although it can introduce penalty.
In order to add cut files more efficiently so that more files from upper levels can be included, SST file cutting threshold (for current file + overlapping parent level files) is set to 1.5X of target file size. A 2MB target file size will generate files like this: https://gist.github.com/siying/29d2676fba417404f3c95e6c013c7de8 Number of files indeed increases but it is not out of control.
Two set of write benchmarks are run:
1. For ingestion rate limited scenario, we can see full compaction is mostly eliminated: https://gist.github.com/siying/959bc1186066906831cf4c808d6e0a19 . The write amp increased from 7.7 to 9.4, as expected. After applying file cutting, the number is improved to 8.9. In another benchmark, the write amp is even better with the incremental approach: https://gist.github.com/siying/d1c16c286d7c59c4d7bba718ca198163
2. For ingestion rate unlimited scenario, incremental compaction turns out to be too expensive most of the time and is not executed, as expected.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8655
Test Plan: Add unit tests to the functionality.
Reviewed By: ajkr
Differential Revision: D31787034
fbshipit-source-id: ce813e63b15a61d5a56e97bf8902a1b28e011beb
2021-10-20 17:03:03 +00:00
|
|
|
|
|
|
|
picker_->GetGrandparents(vstorage_, start_level_inputs,
|
|
|
|
output_level_inputs, &grandparents);
|
2018-05-29 22:42:14 +00:00
|
|
|
} else {
|
|
|
|
inputs.push_back(start_level_inputs);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t estimated_total_size = 0;
|
|
|
|
// Use size of the output level as estimated file size
|
2019-09-16 17:49:20 +00:00
|
|
|
for (FileMetaData* f : vstorage_->LevelFiles(output_level)) {
|
2018-05-29 22:42:14 +00:00
|
|
|
estimated_total_size += f->fd.GetFileSize();
|
|
|
|
}
|
|
|
|
uint32_t path_id =
|
2019-09-16 17:49:20 +00:00
|
|
|
GetPathId(ioptions_, mutable_cf_options_, estimated_total_size);
|
2018-05-29 22:42:14 +00:00
|
|
|
return new Compaction(
|
2020-07-23 01:31:25 +00:00
|
|
|
vstorage_, ioptions_, mutable_cf_options_, mutable_db_options_,
|
|
|
|
std::move(inputs), output_level,
|
2019-09-16 17:49:20 +00:00
|
|
|
MaxFileSizeForLevel(mutable_cf_options_, output_level,
|
2018-05-29 22:42:14 +00:00
|
|
|
kCompactionStyleUniversal),
|
Incremental Space Amp Compactions in Universal Style (#8655)
Summary:
This commit introduces incremental compaction in univeral style for space amplification. This follows the first improvement mentioned in https://rocksdb.org/blog/2021/04/12/universal-improvements.html . The implemention simply picks up files about size of max_compaction_bytes to compact and execute if the penalty is not too big. More optimizations can be done in the future, e.g. prioritizing between this compaction and other types. But for now, the feature is supposed to be functional and can often reduce frequency of full compactions, although it can introduce penalty.
In order to add cut files more efficiently so that more files from upper levels can be included, SST file cutting threshold (for current file + overlapping parent level files) is set to 1.5X of target file size. A 2MB target file size will generate files like this: https://gist.github.com/siying/29d2676fba417404f3c95e6c013c7de8 Number of files indeed increases but it is not out of control.
Two set of write benchmarks are run:
1. For ingestion rate limited scenario, we can see full compaction is mostly eliminated: https://gist.github.com/siying/959bc1186066906831cf4c808d6e0a19 . The write amp increased from 7.7 to 9.4, as expected. After applying file cutting, the number is improved to 8.9. In another benchmark, the write amp is even better with the incremental approach: https://gist.github.com/siying/d1c16c286d7c59c4d7bba718ca198163
2. For ingestion rate unlimited scenario, incremental compaction turns out to be too expensive most of the time and is not executed, as expected.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8655
Test Plan: Add unit tests to the functionality.
Reviewed By: ajkr
Differential Revision: D31787034
fbshipit-source-id: ce813e63b15a61d5a56e97bf8902a1b28e011beb
2021-10-20 17:03:03 +00:00
|
|
|
/* max_grandparent_overlap_bytes */ GetMaxOverlappingBytes(), path_id,
|
2022-03-08 02:06:19 +00:00
|
|
|
GetCompressionType(vstorage_, mutable_cf_options_, output_level, 1),
|
2020-03-31 19:08:41 +00:00
|
|
|
GetCompressionOptions(mutable_cf_options_, vstorage_, output_level),
|
2021-08-09 19:50:19 +00:00
|
|
|
Temperature::kUnknown,
|
2022-03-12 00:13:23 +00:00
|
|
|
/* max_subcompactions */ 0, grandparents, /* is manual */ false,
|
|
|
|
/* trim_ts */ "", score_, false /* deletion_compaction */,
|
2022-07-01 01:04:23 +00:00
|
|
|
/* l0_files_might_overlap */ true,
|
2018-05-29 22:42:14 +00:00
|
|
|
CompactionReason::kFilesMarkedForCompaction);
|
|
|
|
}
|
2019-10-31 18:16:33 +00:00
|
|
|
|
|
|
|
Compaction* UniversalCompactionBuilder::PickCompactionToOldest(
|
|
|
|
size_t start_index, CompactionReason compaction_reason) {
|
2022-08-08 21:34:36 +00:00
|
|
|
return PickCompactionWithSortedRunRange(start_index, sorted_runs_.size() - 1,
|
|
|
|
compaction_reason);
|
|
|
|
}
|
|
|
|
|
|
|
|
Compaction* UniversalCompactionBuilder::PickCompactionWithSortedRunRange(
|
|
|
|
size_t start_index, size_t end_index, CompactionReason compaction_reason) {
|
2019-11-02 01:31:42 +00:00
|
|
|
assert(start_index < sorted_runs_.size());
|
2019-10-31 18:16:33 +00:00
|
|
|
|
|
|
|
// Estimate total file size
|
|
|
|
uint64_t estimated_total_size = 0;
|
2022-08-08 21:34:36 +00:00
|
|
|
for (size_t loop = start_index; loop <= end_index; loop++) {
|
2019-10-31 18:16:33 +00:00
|
|
|
estimated_total_size += sorted_runs_[loop].size;
|
|
|
|
}
|
|
|
|
uint32_t path_id =
|
|
|
|
GetPathId(ioptions_, mutable_cf_options_, estimated_total_size);
|
|
|
|
int start_level = sorted_runs_[start_index].level;
|
2023-06-14 18:28:56 +00:00
|
|
|
int max_output_level =
|
|
|
|
vstorage_->MaxOutputLevel(ioptions_.allow_ingest_behind);
|
|
|
|
std::vector<CompactionInputFiles> inputs(max_output_level + 1);
|
2019-10-31 18:16:33 +00:00
|
|
|
for (size_t i = 0; i < inputs.size(); ++i) {
|
|
|
|
inputs[i].level = start_level + static_cast<int>(i);
|
|
|
|
}
|
2022-08-08 21:34:36 +00:00
|
|
|
for (size_t loop = start_index; loop <= end_index; loop++) {
|
2019-10-31 18:16:33 +00:00
|
|
|
auto& picking_sr = sorted_runs_[loop];
|
|
|
|
if (picking_sr.level == 0) {
|
|
|
|
FileMetaData* f = picking_sr.file;
|
|
|
|
inputs[0].files.push_back(f);
|
|
|
|
} else {
|
|
|
|
auto& files = inputs[picking_sr.level - start_level].files;
|
|
|
|
for (auto* f : vstorage_->LevelFiles(picking_sr.level)) {
|
|
|
|
files.push_back(f);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
std::string comp_reason_print_string;
|
|
|
|
if (compaction_reason == CompactionReason::kPeriodicCompaction) {
|
|
|
|
comp_reason_print_string = "periodic compaction";
|
|
|
|
} else if (compaction_reason ==
|
|
|
|
CompactionReason::kUniversalSizeAmplification) {
|
|
|
|
comp_reason_print_string = "size amp";
|
|
|
|
} else {
|
|
|
|
assert(false);
|
2020-10-06 20:51:22 +00:00
|
|
|
comp_reason_print_string = "unknown: ";
|
|
|
|
comp_reason_print_string.append(
|
|
|
|
std::to_string(static_cast<int>(compaction_reason)));
|
2019-10-31 18:16:33 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
char file_num_buf[256];
|
|
|
|
picking_sr.DumpSizeInfo(file_num_buf, sizeof(file_num_buf), loop);
|
|
|
|
ROCKS_LOG_BUFFER(log_buffer_, "[%s] Universal: %s picking %s",
|
|
|
|
cf_name_.c_str(), comp_reason_print_string.c_str(),
|
|
|
|
file_num_buf);
|
|
|
|
}
|
|
|
|
|
2022-08-08 21:34:36 +00:00
|
|
|
int output_level;
|
|
|
|
if (end_index == sorted_runs_.size() - 1) {
|
2023-06-14 18:28:56 +00:00
|
|
|
output_level = max_output_level;
|
2022-08-08 21:34:36 +00:00
|
|
|
} else {
|
|
|
|
// if it's not including all sorted_runs, it can only output to the level
|
|
|
|
// above the `end_index + 1` sorted_run.
|
|
|
|
output_level = sorted_runs_[end_index + 1].level - 1;
|
2019-10-31 18:16:33 +00:00
|
|
|
}
|
|
|
|
|
2022-10-22 15:57:38 +00:00
|
|
|
// intra L0 compactions outputs could have overlap
|
|
|
|
if (output_level != 0 &&
|
|
|
|
picker_->FilesRangeOverlapWithCompaction(
|
|
|
|
inputs, output_level,
|
|
|
|
Compaction::EvaluatePenultimateLevel(vstorage_, ioptions_,
|
|
|
|
start_level, output_level))) {
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
|
2019-10-31 18:16:33 +00:00
|
|
|
// We never check size for
|
|
|
|
// compaction_options_universal.compression_size_percent,
|
|
|
|
// because we always compact all the files, so always compress.
|
|
|
|
return new Compaction(
|
2020-07-23 01:31:25 +00:00
|
|
|
vstorage_, ioptions_, mutable_cf_options_, mutable_db_options_,
|
|
|
|
std::move(inputs), output_level,
|
2019-10-31 18:16:33 +00:00
|
|
|
MaxFileSizeForLevel(mutable_cf_options_, output_level,
|
|
|
|
kCompactionStyleUniversal),
|
Incremental Space Amp Compactions in Universal Style (#8655)
Summary:
This commit introduces incremental compaction in univeral style for space amplification. This follows the first improvement mentioned in https://rocksdb.org/blog/2021/04/12/universal-improvements.html . The implemention simply picks up files about size of max_compaction_bytes to compact and execute if the penalty is not too big. More optimizations can be done in the future, e.g. prioritizing between this compaction and other types. But for now, the feature is supposed to be functional and can often reduce frequency of full compactions, although it can introduce penalty.
In order to add cut files more efficiently so that more files from upper levels can be included, SST file cutting threshold (for current file + overlapping parent level files) is set to 1.5X of target file size. A 2MB target file size will generate files like this: https://gist.github.com/siying/29d2676fba417404f3c95e6c013c7de8 Number of files indeed increases but it is not out of control.
Two set of write benchmarks are run:
1. For ingestion rate limited scenario, we can see full compaction is mostly eliminated: https://gist.github.com/siying/959bc1186066906831cf4c808d6e0a19 . The write amp increased from 7.7 to 9.4, as expected. After applying file cutting, the number is improved to 8.9. In another benchmark, the write amp is even better with the incremental approach: https://gist.github.com/siying/d1c16c286d7c59c4d7bba718ca198163
2. For ingestion rate unlimited scenario, incremental compaction turns out to be too expensive most of the time and is not executed, as expected.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8655
Test Plan: Add unit tests to the functionality.
Reviewed By: ajkr
Differential Revision: D31787034
fbshipit-source-id: ce813e63b15a61d5a56e97bf8902a1b28e011beb
2021-10-20 17:03:03 +00:00
|
|
|
GetMaxOverlappingBytes(), path_id,
|
2022-03-08 02:06:19 +00:00
|
|
|
GetCompressionType(vstorage_, mutable_cf_options_, output_level, 1,
|
|
|
|
true /* enable_compression */),
|
2020-09-04 06:16:06 +00:00
|
|
|
GetCompressionOptions(mutable_cf_options_, vstorage_, output_level,
|
2019-10-31 18:16:33 +00:00
|
|
|
true /* enable_compression */),
|
2021-08-09 19:50:19 +00:00
|
|
|
Temperature::kUnknown,
|
2019-10-31 18:16:33 +00:00
|
|
|
/* max_subcompactions */ 0, /* grandparents */ {}, /* is manual */ false,
|
2022-03-12 00:13:23 +00:00
|
|
|
/* trim_ts */ "", score_, false /* deletion_compaction */,
|
2022-07-01 01:04:23 +00:00
|
|
|
/* l0_files_might_overlap */ true, compaction_reason);
|
2019-10-31 18:16:33 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
Compaction* UniversalCompactionBuilder::PickPeriodicCompaction() {
|
|
|
|
ROCKS_LOG_BUFFER(log_buffer_, "[%s] Universal: Periodic Compaction",
|
|
|
|
cf_name_.c_str());
|
|
|
|
|
|
|
|
// In universal compaction, sorted runs contain older data are almost always
|
|
|
|
// generated earlier too. To simplify the problem, we just try to trigger
|
|
|
|
// a full compaction. We start from the oldest sorted run and include
|
|
|
|
// all sorted runs, until we hit a sorted already being compacted.
|
|
|
|
// Since usually the largest (which is usually the oldest) sorted run is
|
|
|
|
// included anyway, doing a full compaction won't increase write
|
|
|
|
// amplification much.
|
|
|
|
|
|
|
|
// Get some information from marked files to check whether a file is
|
|
|
|
// included in the compaction.
|
|
|
|
|
|
|
|
size_t start_index = sorted_runs_.size();
|
|
|
|
while (start_index > 0 && !sorted_runs_[start_index - 1].being_compacted) {
|
|
|
|
start_index--;
|
|
|
|
}
|
|
|
|
if (start_index == sorted_runs_.size()) {
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
// There is a rare corner case where we can't pick up all the files
|
|
|
|
// because some files are being compacted and we end up with picking files
|
|
|
|
// but none of them need periodic compaction. Unless we simply recompact
|
|
|
|
// the last sorted run (either the last level or last L0 file), we would just
|
|
|
|
// execute the compaction, in order to simplify the logic.
|
|
|
|
if (start_index == sorted_runs_.size() - 1) {
|
|
|
|
bool included_file_marked = false;
|
|
|
|
int start_level = sorted_runs_[start_index].level;
|
|
|
|
FileMetaData* start_file = sorted_runs_[start_index].file;
|
|
|
|
for (const std::pair<int, FileMetaData*>& level_file_pair :
|
|
|
|
vstorage_->FilesMarkedForPeriodicCompaction()) {
|
|
|
|
if (start_level != 0) {
|
|
|
|
// Last sorted run is a level
|
|
|
|
if (start_level == level_file_pair.first) {
|
|
|
|
included_file_marked = true;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
// Last sorted run is a L0 file.
|
|
|
|
if (start_file == level_file_pair.second) {
|
|
|
|
included_file_marked = true;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (!included_file_marked) {
|
|
|
|
ROCKS_LOG_BUFFER(log_buffer_,
|
|
|
|
"[%s] Universal: Cannot form a compaction covering file "
|
|
|
|
"marked for periodic compaction",
|
|
|
|
cf_name_.c_str());
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
Compaction* c = PickCompactionToOldest(start_index,
|
|
|
|
CompactionReason::kPeriodicCompaction);
|
|
|
|
|
|
|
|
TEST_SYNC_POINT_CALLBACK(
|
|
|
|
"UniversalCompactionPicker::PickPeriodicCompaction:Return", c);
|
|
|
|
|
|
|
|
return c;
|
|
|
|
}
|
Incremental Space Amp Compactions in Universal Style (#8655)
Summary:
This commit introduces incremental compaction in univeral style for space amplification. This follows the first improvement mentioned in https://rocksdb.org/blog/2021/04/12/universal-improvements.html . The implemention simply picks up files about size of max_compaction_bytes to compact and execute if the penalty is not too big. More optimizations can be done in the future, e.g. prioritizing between this compaction and other types. But for now, the feature is supposed to be functional and can often reduce frequency of full compactions, although it can introduce penalty.
In order to add cut files more efficiently so that more files from upper levels can be included, SST file cutting threshold (for current file + overlapping parent level files) is set to 1.5X of target file size. A 2MB target file size will generate files like this: https://gist.github.com/siying/29d2676fba417404f3c95e6c013c7de8 Number of files indeed increases but it is not out of control.
Two set of write benchmarks are run:
1. For ingestion rate limited scenario, we can see full compaction is mostly eliminated: https://gist.github.com/siying/959bc1186066906831cf4c808d6e0a19 . The write amp increased from 7.7 to 9.4, as expected. After applying file cutting, the number is improved to 8.9. In another benchmark, the write amp is even better with the incremental approach: https://gist.github.com/siying/d1c16c286d7c59c4d7bba718ca198163
2. For ingestion rate unlimited scenario, incremental compaction turns out to be too expensive most of the time and is not executed, as expected.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8655
Test Plan: Add unit tests to the functionality.
Reviewed By: ajkr
Differential Revision: D31787034
fbshipit-source-id: ce813e63b15a61d5a56e97bf8902a1b28e011beb
2021-10-20 17:03:03 +00:00
|
|
|
|
|
|
|
uint64_t UniversalCompactionBuilder::GetMaxOverlappingBytes() const {
|
|
|
|
if (!mutable_cf_options_.compaction_options_universal.incremental) {
|
2022-05-05 20:08:21 +00:00
|
|
|
return std::numeric_limits<uint64_t>::max();
|
Incremental Space Amp Compactions in Universal Style (#8655)
Summary:
This commit introduces incremental compaction in univeral style for space amplification. This follows the first improvement mentioned in https://rocksdb.org/blog/2021/04/12/universal-improvements.html . The implemention simply picks up files about size of max_compaction_bytes to compact and execute if the penalty is not too big. More optimizations can be done in the future, e.g. prioritizing between this compaction and other types. But for now, the feature is supposed to be functional and can often reduce frequency of full compactions, although it can introduce penalty.
In order to add cut files more efficiently so that more files from upper levels can be included, SST file cutting threshold (for current file + overlapping parent level files) is set to 1.5X of target file size. A 2MB target file size will generate files like this: https://gist.github.com/siying/29d2676fba417404f3c95e6c013c7de8 Number of files indeed increases but it is not out of control.
Two set of write benchmarks are run:
1. For ingestion rate limited scenario, we can see full compaction is mostly eliminated: https://gist.github.com/siying/959bc1186066906831cf4c808d6e0a19 . The write amp increased from 7.7 to 9.4, as expected. After applying file cutting, the number is improved to 8.9. In another benchmark, the write amp is even better with the incremental approach: https://gist.github.com/siying/d1c16c286d7c59c4d7bba718ca198163
2. For ingestion rate unlimited scenario, incremental compaction turns out to be too expensive most of the time and is not executed, as expected.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8655
Test Plan: Add unit tests to the functionality.
Reviewed By: ajkr
Differential Revision: D31787034
fbshipit-source-id: ce813e63b15a61d5a56e97bf8902a1b28e011beb
2021-10-20 17:03:03 +00:00
|
|
|
} else {
|
|
|
|
// Try to align cutting boundary with files at the next level if the
|
|
|
|
// file isn't end up with 1/2 of target size, or it would overlap
|
|
|
|
// with two full size files at the next level.
|
|
|
|
return mutable_cf_options_.target_file_size_base / 2 * 3;
|
|
|
|
}
|
|
|
|
}
|
2020-02-20 20:07:53 +00:00
|
|
|
} // namespace ROCKSDB_NAMESPACE
|