2013-10-16 23:59:46 +02:00
|
|
|
// Copyright (c) 2013, Facebook, Inc. All rights reserved.
|
|
|
|
// This source code is licensed under the BSD-style license found in the
|
|
|
|
// LICENSE file in the root directory of this source tree. An additional grant
|
|
|
|
// of patent rights can be found in the PATENTS file in the same directory.
|
|
|
|
//
|
2011-03-18 23:37:00 +01:00
|
|
|
// 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.
|
|
|
|
|
|
|
|
#include "db/db_impl.h"
|
|
|
|
|
2014-03-18 20:25:08 +01:00
|
|
|
#define __STDC_FORMAT_MACROS
|
|
|
|
#include <inttypes.h>
|
2011-03-18 23:37:00 +01:00
|
|
|
#include <algorithm>
|
2012-11-26 22:56:45 +01:00
|
|
|
#include <climits>
|
|
|
|
#include <cstdio>
|
2011-03-18 23:37:00 +01:00
|
|
|
#include <set>
|
2013-06-21 01:58:59 +02:00
|
|
|
#include <stdexcept>
|
2013-10-16 20:50:50 +02:00
|
|
|
#include <stdint.h>
|
|
|
|
#include <string>
|
[RocksDB] [Performance] Speed up FindObsoleteFiles
Summary:
FindObsoleteFiles was slow, holding the single big lock, resulted in bad p99 behavior.
Didn't profile anything, but several things could be improved:
1. VersionSet::AddLiveFiles works with std::set, which is by itself slow (a tree).
You also don't know how many dynamic allocations occur just for building up this tree.
switched to std::vector, also added logic to pre-calculate total size and do just one allocation
2. Don't see why env_->GetChildren() needs to be mutex proteced, moved to PurgeObsoleteFiles where
mutex could be unlocked.
3. switched std::set to std:unordered_set, the conversion from vector is also inside PurgeObsoleteFiles
I have a feeling this should pretty much fix it.
Test Plan: make check; db_stress
Reviewers: dhruba, heyongqiang, MarkCallaghan
Reviewed By: dhruba
CC: leveldb, zshao
Differential Revision: https://reviews.facebook.net/D10197
2013-04-12 01:49:53 +02:00
|
|
|
#include <unordered_set>
|
2014-01-17 06:56:26 +01:00
|
|
|
#include <utility>
|
2013-10-16 20:50:50 +02:00
|
|
|
#include <vector>
|
2012-11-26 22:56:45 +01:00
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
#include "db/builder.h"
|
2013-10-16 20:50:50 +02:00
|
|
|
#include "db/db_iter.h"
|
2013-12-31 03:33:57 +01:00
|
|
|
#include "db/dbformat.h"
|
2011-03-18 23:37:00 +01:00
|
|
|
#include "db/filename.h"
|
|
|
|
#include "db/log_reader.h"
|
|
|
|
#include "db/log_writer.h"
|
|
|
|
#include "db/memtable.h"
|
2013-12-11 04:03:13 +01:00
|
|
|
#include "db/memtable_list.h"
|
2013-12-03 03:34:05 +01:00
|
|
|
#include "db/merge_context.h"
|
2013-03-21 23:59:47 +01:00
|
|
|
#include "db/merge_helper.h"
|
2013-08-13 23:04:56 +02:00
|
|
|
#include "db/prefix_filter_iterator.h"
|
2011-03-18 23:37:00 +01:00
|
|
|
#include "db/table_cache.h"
|
2013-11-20 01:29:42 +01:00
|
|
|
#include "db/table_properties_collector.h"
|
2014-01-17 06:56:26 +01:00
|
|
|
#include "db/tailing_iter.h"
|
2013-10-16 20:50:50 +02:00
|
|
|
#include "db/transaction_log_impl.h"
|
2011-03-18 23:37:00 +01:00
|
|
|
#include "db/version_set.h"
|
|
|
|
#include "db/write_batch_internal.h"
|
2013-10-16 20:50:50 +02:00
|
|
|
#include "port/port.h"
|
2014-02-27 20:38:55 +01:00
|
|
|
#include "port/likely.h"
|
2013-08-23 17:38:13 +02:00
|
|
|
#include "rocksdb/compaction_filter.h"
|
|
|
|
#include "rocksdb/db.h"
|
|
|
|
#include "rocksdb/env.h"
|
|
|
|
#include "rocksdb/merge_operator.h"
|
|
|
|
#include "rocksdb/statistics.h"
|
|
|
|
#include "rocksdb/status.h"
|
2013-10-29 01:54:09 +01:00
|
|
|
#include "rocksdb/table.h"
|
2011-03-18 23:37:00 +01:00
|
|
|
#include "table/block.h"
|
2013-11-20 07:00:48 +01:00
|
|
|
#include "table/block_based_table_factory.h"
|
2011-03-18 23:37:00 +01:00
|
|
|
#include "table/merger.h"
|
2014-01-28 06:58:46 +01:00
|
|
|
#include "table/table_builder.h"
|
2011-03-18 23:37:00 +01:00
|
|
|
#include "table/two_level_iterator.h"
|
2013-02-15 20:53:17 +01:00
|
|
|
#include "util/auto_roll_logger.h"
|
2014-01-28 06:58:46 +01:00
|
|
|
#include "util/autovector.h"
|
2013-02-15 20:53:17 +01:00
|
|
|
#include "util/build_version.h"
|
2011-03-18 23:37:00 +01:00
|
|
|
#include "util/coding.h"
|
2013-12-03 21:42:15 +01:00
|
|
|
#include "util/hash_skiplist_rep.h"
|
2014-03-10 20:56:46 +01:00
|
|
|
#include "util/hash_linklist_rep.h"
|
2011-03-18 23:37:00 +01:00
|
|
|
#include "util/logging.h"
|
2014-03-10 06:01:13 +01:00
|
|
|
#include "util/log_buffer.h"
|
2011-03-18 23:37:00 +01:00
|
|
|
#include "util/mutexlock.h"
|
2013-10-23 08:26:51 +02:00
|
|
|
#include "util/perf_context_imp.h"
|
2013-02-15 20:53:17 +01:00
|
|
|
#include "util/stop_watch.h"
|
2011-03-18 23:37:00 +01:00
|
|
|
|
2013-10-04 06:49:15 +02:00
|
|
|
namespace rocksdb {
|
2011-03-18 23:37:00 +01:00
|
|
|
|
2014-03-07 23:43:22 +01:00
|
|
|
int DBImpl::SuperVersion::dummy = 0;
|
|
|
|
void* const DBImpl::SuperVersion::kSVInUse = &DBImpl::SuperVersion::dummy;
|
|
|
|
void* const DBImpl::SuperVersion::kSVObsolete = nullptr;
|
|
|
|
|
2013-12-31 03:33:57 +01:00
|
|
|
void DumpLeveldbBuildVersion(Logger * log);
|
2012-09-18 22:18:32 +02:00
|
|
|
|
2012-03-09 01:23:21 +01:00
|
|
|
// Information kept for every waiting writer
|
|
|
|
struct DBImpl::Writer {
|
|
|
|
Status status;
|
|
|
|
WriteBatch* batch;
|
|
|
|
bool sync;
|
2012-07-05 22:39:28 +02:00
|
|
|
bool disableWAL;
|
2012-03-09 01:23:21 +01:00
|
|
|
bool done;
|
|
|
|
port::CondVar cv;
|
|
|
|
|
|
|
|
explicit Writer(port::Mutex* mu) : cv(mu) { }
|
|
|
|
};
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
struct DBImpl::CompactionState {
|
|
|
|
Compaction* const compaction;
|
|
|
|
|
2012-11-27 06:16:21 +01:00
|
|
|
// If there were two snapshots with seq numbers s1 and
|
|
|
|
// s2 and s1 < s2, and if we find two instances of a key k1 then lies
|
|
|
|
// entirely within s1 and s2, then the earlier version of k1 can be safely
|
|
|
|
// deleted because that version is not visible in any snapshot.
|
|
|
|
std::vector<SequenceNumber> existing_snapshots;
|
2011-03-18 23:37:00 +01:00
|
|
|
|
|
|
|
// Files produced by compaction
|
|
|
|
struct Output {
|
|
|
|
uint64_t number;
|
|
|
|
uint64_t file_size;
|
|
|
|
InternalKey smallest, largest;
|
2013-06-14 07:09:08 +02:00
|
|
|
SequenceNumber smallest_seqno, largest_seqno;
|
2011-03-18 23:37:00 +01:00
|
|
|
};
|
|
|
|
std::vector<Output> outputs;
|
2012-10-19 23:00:53 +02:00
|
|
|
std::list<uint64_t> allocated_file_numbers;
|
2011-03-18 23:37:00 +01:00
|
|
|
|
|
|
|
// State kept for output being generated
|
2013-01-20 11:07:13 +01:00
|
|
|
unique_ptr<WritableFile> outfile;
|
|
|
|
unique_ptr<TableBuilder> builder;
|
2011-03-18 23:37:00 +01:00
|
|
|
|
|
|
|
uint64_t total_bytes;
|
|
|
|
|
|
|
|
Output* current_output() { return &outputs[outputs.size()-1]; }
|
|
|
|
|
|
|
|
explicit CompactionState(Compaction* c)
|
|
|
|
: compaction(c),
|
|
|
|
total_bytes(0) {
|
|
|
|
}
|
2013-10-27 07:01:26 +01:00
|
|
|
|
2014-04-02 23:48:53 +02:00
|
|
|
// Create a client visible context of this compaction
|
|
|
|
CompactionFilter::Context GetFilterContextV1() {
|
|
|
|
CompactionFilter::Context context;
|
|
|
|
context.is_full_compaction = compaction->IsFullCompaction();
|
|
|
|
context.is_manual_compaction = compaction->IsManualCompaction();
|
|
|
|
return context;
|
|
|
|
}
|
|
|
|
|
2013-10-27 07:01:26 +01:00
|
|
|
// Create a client visible context of this compaction
|
2014-01-10 02:52:11 +01:00
|
|
|
CompactionFilterContext GetFilterContext() {
|
|
|
|
CompactionFilterContext context;
|
2013-10-27 07:01:26 +01:00
|
|
|
context.is_full_compaction = compaction->IsFullCompaction();
|
2014-02-12 21:24:18 +01:00
|
|
|
context.is_manual_compaction = compaction->IsManualCompaction();
|
2013-10-27 07:01:26 +01:00
|
|
|
return context;
|
|
|
|
}
|
2014-01-10 02:52:11 +01:00
|
|
|
|
|
|
|
std::vector<Slice> key_buf_;
|
|
|
|
std::vector<Slice> existing_value_buf_;
|
|
|
|
std::vector<std::string> key_str_buf_;
|
|
|
|
std::vector<std::string> existing_value_str_buf_;
|
|
|
|
// new_value_buf_ will only be appended if a value changes
|
|
|
|
std::vector<std::string> new_value_buf_;
|
|
|
|
// if values_changed_buf_[i] is true
|
|
|
|
// new_value_buf_ will add a new entry with the changed value
|
|
|
|
std::vector<bool> value_changed_buf_;
|
|
|
|
// to_delete_buf_[i] is true iff key_buf_[i] is deleted
|
|
|
|
std::vector<bool> to_delete_buf_;
|
|
|
|
// buffer for the parsed internal keys, the string buffer is backed
|
|
|
|
// by key_str_buf_
|
|
|
|
std::vector<ParsedInternalKey> ikey_buf_;
|
|
|
|
|
|
|
|
std::vector<Slice> other_key_buf_;
|
|
|
|
std::vector<Slice> other_value_buf_;
|
|
|
|
std::vector<std::string> other_key_str_buf_;
|
|
|
|
std::vector<std::string> other_value_str_buf_;
|
|
|
|
|
|
|
|
std::vector<Slice> combined_key_buf_;
|
|
|
|
std::vector<Slice> combined_value_buf_;
|
|
|
|
|
|
|
|
std::string cur_prefix_;
|
|
|
|
|
|
|
|
// Buffers the kv-pair that will be run through compaction filter V2
|
|
|
|
// in the future.
|
|
|
|
void BufferKeyValueSlices(const Slice& key, const Slice& value) {
|
|
|
|
key_str_buf_.emplace_back(key.ToString());
|
|
|
|
existing_value_str_buf_.emplace_back(value.ToString());
|
|
|
|
key_buf_.emplace_back(Slice(key_str_buf_.back()));
|
|
|
|
existing_value_buf_.emplace_back(Slice(existing_value_str_buf_.back()));
|
|
|
|
|
|
|
|
ParsedInternalKey ikey;
|
|
|
|
ParseInternalKey(key_buf_.back(), &ikey);
|
|
|
|
ikey_buf_.emplace_back(ikey);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Buffers the kv-pair that will not be run through compaction filter V2
|
|
|
|
// in the future.
|
|
|
|
void BufferOtherKeyValueSlices(const Slice& key, const Slice& value) {
|
|
|
|
other_key_str_buf_.emplace_back(key.ToString());
|
|
|
|
other_value_str_buf_.emplace_back(value.ToString());
|
|
|
|
other_key_buf_.emplace_back(Slice(other_key_str_buf_.back()));
|
|
|
|
other_value_buf_.emplace_back(Slice(other_value_str_buf_.back()));
|
|
|
|
}
|
|
|
|
|
|
|
|
// Add a kv-pair to the combined buffer
|
|
|
|
void AddToCombinedKeyValueSlices(const Slice& key, const Slice& value) {
|
|
|
|
// The real strings are stored in the batch buffers
|
|
|
|
combined_key_buf_.emplace_back(key);
|
|
|
|
combined_value_buf_.emplace_back(value);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Merging the two buffers
|
|
|
|
void MergeKeyValueSliceBuffer(const InternalKeyComparator* comparator) {
|
|
|
|
size_t i = 0;
|
|
|
|
size_t j = 0;
|
|
|
|
size_t total_size = key_buf_.size() + other_key_buf_.size();
|
|
|
|
combined_key_buf_.reserve(total_size);
|
|
|
|
combined_value_buf_.reserve(total_size);
|
|
|
|
|
|
|
|
while (i + j < total_size) {
|
|
|
|
int comp_res = 0;
|
|
|
|
if (i < key_buf_.size() && j < other_key_buf_.size()) {
|
|
|
|
comp_res = comparator->Compare(key_buf_[i], other_key_buf_[j]);
|
|
|
|
} else if (i >= key_buf_.size() && j < other_key_buf_.size()) {
|
|
|
|
comp_res = 1;
|
|
|
|
} else if (j >= other_key_buf_.size() && i < key_buf_.size()) {
|
|
|
|
comp_res = -1;
|
|
|
|
}
|
|
|
|
if (comp_res > 0) {
|
|
|
|
AddToCombinedKeyValueSlices(other_key_buf_[j], other_value_buf_[j]);
|
|
|
|
j++;
|
|
|
|
} else if (comp_res < 0) {
|
|
|
|
AddToCombinedKeyValueSlices(key_buf_[i], existing_value_buf_[i]);
|
|
|
|
i++;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void CleanupBatchBuffer() {
|
|
|
|
to_delete_buf_.clear();
|
|
|
|
key_buf_.clear();
|
|
|
|
existing_value_buf_.clear();
|
|
|
|
key_str_buf_.clear();
|
|
|
|
existing_value_str_buf_.clear();
|
|
|
|
new_value_buf_.clear();
|
|
|
|
value_changed_buf_.clear();
|
|
|
|
ikey_buf_.clear();
|
|
|
|
|
|
|
|
to_delete_buf_.shrink_to_fit();
|
|
|
|
key_buf_.shrink_to_fit();
|
|
|
|
existing_value_buf_.shrink_to_fit();
|
|
|
|
key_str_buf_.shrink_to_fit();
|
|
|
|
existing_value_str_buf_.shrink_to_fit();
|
|
|
|
new_value_buf_.shrink_to_fit();
|
|
|
|
value_changed_buf_.shrink_to_fit();
|
|
|
|
ikey_buf_.shrink_to_fit();
|
|
|
|
|
|
|
|
other_key_buf_.clear();
|
|
|
|
other_value_buf_.clear();
|
|
|
|
other_key_str_buf_.clear();
|
|
|
|
other_value_str_buf_.clear();
|
|
|
|
other_key_buf_.shrink_to_fit();
|
|
|
|
other_value_buf_.shrink_to_fit();
|
|
|
|
other_key_str_buf_.shrink_to_fit();
|
|
|
|
other_value_str_buf_.shrink_to_fit();
|
|
|
|
}
|
|
|
|
|
|
|
|
void CleanupMergedBuffer() {
|
|
|
|
combined_key_buf_.clear();
|
|
|
|
combined_value_buf_.clear();
|
|
|
|
combined_key_buf_.shrink_to_fit();
|
|
|
|
combined_value_buf_.shrink_to_fit();
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
};
|
|
|
|
|
|
|
|
// Fix user-supplied options to be reasonable
|
2013-06-05 20:22:38 +02:00
|
|
|
template <class T, class V>
|
2011-03-18 23:37:00 +01:00
|
|
|
static void ClipToRange(T* ptr, V minvalue, V maxvalue) {
|
2011-04-21 00:48:11 +02:00
|
|
|
if (static_cast<V>(*ptr) > maxvalue) *ptr = maxvalue;
|
|
|
|
if (static_cast<V>(*ptr) < minvalue) *ptr = minvalue;
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
Options SanitizeOptions(const std::string& dbname,
|
|
|
|
const InternalKeyComparator* icmp,
|
2012-04-17 17:36:46 +02:00
|
|
|
const InternalFilterPolicy* ipolicy,
|
2011-03-18 23:37:00 +01:00
|
|
|
const Options& src) {
|
|
|
|
Options result = src;
|
2013-02-15 20:53:17 +01:00
|
|
|
result.filter_policy = (src.filter_policy != nullptr) ? ipolicy : nullptr;
|
2014-01-07 05:29:17 +01:00
|
|
|
// result.max_open_files means an "infinite" open files.
|
|
|
|
if (result.max_open_files != -1) {
|
|
|
|
ClipToRange(&result.max_open_files, 20, 1000000);
|
|
|
|
}
|
2013-05-23 20:56:19 +02:00
|
|
|
ClipToRange(&result.write_buffer_size, ((size_t)64)<<10,
|
|
|
|
((size_t)64)<<30);
|
2012-04-17 17:36:46 +02:00
|
|
|
ClipToRange(&result.block_size, 1<<10, 4<<20);
|
2013-06-11 23:23:58 +02:00
|
|
|
|
Make arena block size configurable
Summary:
Add an option for arena block size, default value 4096 bytes. Arena will allocate blocks with such size.
I am not sure about passing parameter to skiplist in the new virtualized framework, though I talked to Jim a bit. So add Jim as reviewer.
Test Plan:
new unit test, I am running db_test.
For passing paramter from configured option to Arena, I tried tests like:
TEST(DBTest, Arena_Option) {
std::string dbname = test::TmpDir() + "/db_arena_option_test";
DestroyDB(dbname, Options());
DB* db = nullptr;
Options opts;
opts.create_if_missing = true;
opts.arena_block_size = 1000000; // tested 99, 999999
Status s = DB::Open(opts, dbname, &db);
db->Put(WriteOptions(), "a", "123");
}
and printed some debug info. The results look good. Any suggestion for such a unit-test?
Reviewers: haobo, dhruba, emayanke, jpaton
Reviewed By: dhruba
CC: leveldb, zshao
Differential Revision: https://reviews.facebook.net/D11799
2013-07-31 21:42:23 +02:00
|
|
|
// if user sets arena_block_size, we trust user to use this value. Otherwise,
|
|
|
|
// calculate a proper value from writer_buffer_size;
|
|
|
|
if (result.arena_block_size <= 0) {
|
|
|
|
result.arena_block_size = result.write_buffer_size / 10;
|
|
|
|
}
|
|
|
|
|
2013-06-11 23:23:58 +02:00
|
|
|
result.min_write_buffer_number_to_merge = std::min(
|
|
|
|
result.min_write_buffer_number_to_merge, result.max_write_buffer_number-1);
|
2013-02-15 20:53:17 +01:00
|
|
|
if (result.info_log == nullptr) {
|
2013-02-05 04:42:40 +01:00
|
|
|
Status s = CreateLoggerFromOptions(dbname, result.db_log_dir, src.env,
|
|
|
|
result, &result.info_log);
|
2011-03-18 23:37:00 +01:00
|
|
|
if (!s.ok()) {
|
|
|
|
// No place suitable for logging
|
2013-02-15 20:53:17 +01:00
|
|
|
result.info_log = nullptr;
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
}
|
2013-02-15 20:53:17 +01:00
|
|
|
if (result.block_cache == nullptr && !result.no_block_cache) {
|
2011-04-12 21:38:58 +02:00
|
|
|
result.block_cache = NewLRUCache(8 << 20);
|
|
|
|
}
|
2013-01-24 19:54:26 +01:00
|
|
|
result.compression_per_level = src.compression_per_level;
|
2013-05-15 19:34:02 +02:00
|
|
|
if (result.block_size_deviation < 0 || result.block_size_deviation > 100) {
|
|
|
|
result.block_size_deviation = 0;
|
|
|
|
}
|
2013-06-09 05:15:10 +02:00
|
|
|
if (result.max_mem_compaction_level >= result.num_levels) {
|
|
|
|
result.max_mem_compaction_level = result.num_levels - 1;
|
|
|
|
}
|
2013-08-06 00:43:49 +02:00
|
|
|
if (result.soft_rate_limit > result.hard_rate_limit) {
|
|
|
|
result.soft_rate_limit = result.hard_rate_limit;
|
|
|
|
}
|
2013-10-27 07:01:26 +01:00
|
|
|
if (result.compaction_filter) {
|
|
|
|
Log(result.info_log, "Compaction filter specified, ignore factory");
|
2013-08-13 19:56:20 +02:00
|
|
|
}
|
2013-08-23 08:10:02 +02:00
|
|
|
if (result.prefix_extractor) {
|
2014-03-10 20:56:46 +01:00
|
|
|
Log(result.info_log, "prefix extractor %s in use.",
|
|
|
|
result.prefix_extractor->Name());
|
|
|
|
} else {
|
|
|
|
assert(result.memtable_factory);
|
|
|
|
Slice name = result.memtable_factory->Name();
|
|
|
|
if (name.compare("HashSkipListRepFactory") == 0 ||
|
|
|
|
name.compare("HashLinkListRepFactory") == 0) {
|
|
|
|
Log(result.info_log, "prefix extractor is not provided while using %s. "
|
|
|
|
"fallback to skiplist", name.ToString().c_str());
|
2013-08-23 08:10:02 +02:00
|
|
|
result.memtable_factory = std::make_shared<SkipListFactory>();
|
|
|
|
}
|
|
|
|
}
|
2013-10-01 23:46:52 +02:00
|
|
|
|
|
|
|
if (result.wal_dir.empty()) {
|
|
|
|
// Use dbname as default
|
|
|
|
result.wal_dir = dbname;
|
|
|
|
}
|
2013-10-16 20:50:50 +02:00
|
|
|
|
2013-11-20 01:29:42 +01:00
|
|
|
// -- Sanitize the table properties collector
|
|
|
|
// All user defined properties collectors will be wrapped by
|
|
|
|
// UserKeyTablePropertiesCollector since for them they only have the
|
|
|
|
// knowledge of the user keys; internal keys are invisible to them.
|
|
|
|
auto& collectors = result.table_properties_collectors;
|
|
|
|
for (size_t i = 0; i < result.table_properties_collectors.size(); ++i) {
|
2013-10-16 20:50:50 +02:00
|
|
|
assert(collectors[i]);
|
|
|
|
collectors[i] =
|
2013-11-20 01:29:42 +01:00
|
|
|
std::make_shared<UserKeyTablePropertiesCollector>(collectors[i]);
|
2013-10-16 20:50:50 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
// Add collector to collect internal key statistics
|
|
|
|
collectors.push_back(
|
2013-11-20 01:29:42 +01:00
|
|
|
std::make_shared<InternalKeyPropertiesCollector>()
|
2013-10-16 20:50:50 +02:00
|
|
|
);
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
2013-10-30 18:52:33 +01:00
|
|
|
CompressionType GetCompressionType(const Options& options, int level,
|
|
|
|
const bool enable_compression) {
|
|
|
|
if (!enable_compression) {
|
|
|
|
// disable compression
|
|
|
|
return kNoCompression;
|
|
|
|
}
|
|
|
|
// If the use has specified a different compression level for each level,
|
|
|
|
// then pick the compresison for that level.
|
|
|
|
if (!options.compression_per_level.empty()) {
|
|
|
|
const int n = options.compression_per_level.size() - 1;
|
|
|
|
// It is possible for level_ to be -1; in that case, we use level
|
|
|
|
// 0's compression. This occurs mostly in backwards compatibility
|
|
|
|
// situations when the builder doesn't know what level the file
|
|
|
|
// belongs to. Likewise, if level_ is beyond the end of the
|
|
|
|
// specified compression levels, use the last value.
|
|
|
|
return options.compression_per_level[std::max(0, std::min(level, n))];
|
|
|
|
} else {
|
|
|
|
return options.compression;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-12-19 19:02:53 +01:00
|
|
|
CompressionType GetCompressionFlush(const Options& options) {
|
|
|
|
// Compressing memtable flushes might not help unless the sequential load
|
|
|
|
// optimization is used for leveled compaction. Otherwise the CPU and
|
|
|
|
// latency overhead is not offset by saving much space.
|
|
|
|
|
|
|
|
bool can_compress;
|
|
|
|
|
|
|
|
if (options.compaction_style == kCompactionStyleUniversal) {
|
|
|
|
can_compress =
|
|
|
|
(options.compaction_options_universal.compression_size_percent < 0);
|
|
|
|
} else {
|
|
|
|
// For leveled compress when min_level_to_compress == 0.
|
|
|
|
can_compress = (GetCompressionType(options, 0, true) != kNoCompression);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (can_compress) {
|
|
|
|
return options.compression;
|
|
|
|
} else {
|
|
|
|
return kNoCompression;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
DBImpl::DBImpl(const Options& options, const std::string& dbname)
|
|
|
|
: env_(options.env),
|
2012-11-06 04:18:49 +01:00
|
|
|
dbname_(dbname),
|
2011-03-18 23:37:00 +01:00
|
|
|
internal_comparator_(options.comparator),
|
2014-01-15 00:27:09 +01:00
|
|
|
options_(SanitizeOptions(dbname, &internal_comparator_,
|
|
|
|
&internal_filter_policy_, options)),
|
2012-11-06 04:18:49 +01:00
|
|
|
internal_filter_policy_(options.filter_policy),
|
2011-03-18 23:37:00 +01:00
|
|
|
owns_info_log_(options_.info_log != options.info_log),
|
2013-02-15 20:53:17 +01:00
|
|
|
db_lock_(nullptr),
|
2013-06-01 01:30:17 +02:00
|
|
|
mutex_(options.use_adaptive_mutex),
|
2013-02-15 20:53:17 +01:00
|
|
|
shutting_down_(nullptr),
|
2011-03-18 23:37:00 +01:00
|
|
|
bg_cv_(&mutex_),
|
2014-01-15 00:27:09 +01:00
|
|
|
mem_(new MemTable(internal_comparator_, options_)),
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
imm_(options_.min_write_buffer_number_to_merge),
|
2011-06-22 04:36:45 +02:00
|
|
|
logfile_number_(0),
|
2013-12-20 18:57:58 +01:00
|
|
|
super_version_(nullptr),
|
2014-01-17 06:56:26 +01:00
|
|
|
super_version_number_(0),
|
2014-02-27 20:38:55 +01:00
|
|
|
local_sv_(new ThreadLocalPtr(&SuperVersionUnrefHandle)),
|
2013-03-28 23:19:28 +01:00
|
|
|
tmp_batch_(),
|
Fix data race against logging data structure because of LogBuffer
Summary:
@igor pointed out that there is a potential data race because of the way we use the newly introduced LogBuffer. After "bg_compaction_scheduled_--" or "bg_flush_scheduled_--", they can both become 0. As soon as the lock is released after that, DBImpl's deconstructor can go ahead and deconstruct all the states inside DB, including the info_log object hold in a shared pointer of the options object it keeps. At that point it is not safe anymore to continue using the info logger to write the delayed logs.
With the patch, lock is released temporarily for log buffer to be flushed before "bg_compaction_scheduled_--" or "bg_flush_scheduled_--". In order to make sure we don't miss any pending flush or compaction, a new flag bg_schedule_needed_ is added, which is set to be true if there is a pending flush or compaction but not scheduled because of the max thread limit. If the flag is set to be true, the scheduling function will be called before compaction or flush thread finishes.
Thanks @igor for this finding!
Test Plan: make all check
Reviewers: haobo, igor
Reviewed By: haobo
CC: dhruba, ljin, yhchiang, igor, leveldb
Differential Revision: https://reviews.facebook.net/D16767
2014-03-11 18:36:30 +01:00
|
|
|
bg_schedule_needed_(false),
|
2012-10-19 23:00:53 +02:00
|
|
|
bg_compaction_scheduled_(0),
|
Fix a deadlock in CompactRange()
Summary:
The way DBImpl::TEST_CompactRange() throttles down the number of bg compactions
can cause it to deadlock when CompactRange() is called concurrently from
multiple threads. Imagine a following scenario with only two threads
(max_background_compactions is 10 and bg_compaction_scheduled_ is initially 0):
1. Thread #1 increments bg_compaction_scheduled_ (to LargeNumber), sets
bg_compaction_scheduled_ to 9 (newvalue), schedules the compaction
(bg_compaction_scheduled_ is now 10) and waits for it to complete.
2. Thread #2 calls TEST_CompactRange(), increments bg_compaction_scheduled_
(now LargeNumber + 10) and waits on a cv for bg_compaction_scheduled_ to
drop to LargeNumber.
3. BG thread completes the first manual compaction, decrements
bg_compaction_scheduled_ and wakes up all threads waiting on bg_cv_.
Thread #1 runs, increments bg_compaction_scheduled_ by LargeNumber again
(now 2*LargeNumber + 9). Since that's more than LargeNumber + newvalue,
thread #2 also goes to sleep (waiting on bg_cv_), without resetting
bg_compaction_scheduled_.
This diff attempts to address the problem by introducing a new counter
bg_manual_only_ (when positive, MaybeScheduleFlushOrCompaction() will only
schedule manual compactions).
Test Plan:
I could pretty much consistently reproduce the deadlock with a program that
calls CompactRange(nullptr, nullptr) immediately after Write() from multiple
threads. This no longer happens with this patch.
Tests (make check) pass.
Reviewers: dhruba, igor, sdong, haobo
Reviewed By: igor
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14799
2013-12-22 00:10:39 +01:00
|
|
|
bg_manual_only_(0),
|
2013-09-13 23:38:37 +02:00
|
|
|
bg_flush_scheduled_(0),
|
2012-08-27 09:50:26 +02:00
|
|
|
bg_logstats_scheduled_(false),
|
2013-02-15 20:53:17 +01:00
|
|
|
manual_compaction_(nullptr),
|
|
|
|
logger_(nullptr),
|
2014-01-02 12:33:42 +01:00
|
|
|
disable_delete_obsolete_files_(0),
|
2013-11-15 03:03:57 +01:00
|
|
|
delete_obsolete_files_last_run_(options.env->NowMicros()),
|
2013-05-06 20:41:01 +02:00
|
|
|
purge_wal_files_last_run_(0),
|
2013-05-11 00:21:04 +02:00
|
|
|
last_stats_dump_time_microsec_(0),
|
2013-11-07 03:46:28 +01:00
|
|
|
default_interval_to_delete_obsolete_WAL_(600),
|
2013-02-19 07:32:07 +01:00
|
|
|
flush_on_destroy_(false),
|
2014-01-30 05:40:41 +01:00
|
|
|
internal_stats_(options.num_levels, options.env,
|
|
|
|
options.statistics.get()),
|
2013-03-04 19:44:04 +01:00
|
|
|
delayed_writes_(0),
|
2013-06-30 08:21:36 +02:00
|
|
|
storage_options_(options),
|
|
|
|
bg_work_gate_closed_(false),
|
2014-02-27 20:38:55 +01:00
|
|
|
refitting_level_(false),
|
|
|
|
opened_successfully_(false) {
|
2011-05-21 04:17:43 +02:00
|
|
|
mem_->Ref();
|
2012-09-06 02:44:13 +02:00
|
|
|
env_->GetAbsolutePath(dbname, &db_absolute_path_);
|
2013-03-02 21:56:04 +01:00
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
// Reserve ten files or so for other uses and give the rest to TableCache.
|
2014-01-07 05:29:17 +01:00
|
|
|
// Give a large number for setting of "infinite" open files.
|
|
|
|
const int table_cache_size =
|
|
|
|
(options_.max_open_files == -1) ?
|
|
|
|
4194304 : options_.max_open_files - 10;
|
2013-03-15 01:00:04 +01:00
|
|
|
table_cache_.reset(new TableCache(dbname_, &options_,
|
|
|
|
storage_options_, table_cache_size));
|
|
|
|
versions_.reset(new VersionSet(dbname_, &options_, storage_options_,
|
|
|
|
table_cache_.get(), &internal_comparator_));
|
2012-08-15 00:20:36 +02:00
|
|
|
|
2013-12-31 03:33:57 +01:00
|
|
|
DumpLeveldbBuildVersion(options_.info_log.get());
|
2013-01-20 11:07:13 +01:00
|
|
|
options_.Dump(options_.info_log.get());
|
2012-08-22 20:43:53 +02:00
|
|
|
|
2012-08-15 00:20:36 +02:00
|
|
|
char name[100];
|
2013-12-31 03:33:57 +01:00
|
|
|
Status s = env_->GetHostName(name, 100L);
|
|
|
|
if (s.ok()) {
|
2012-08-15 00:20:36 +02:00
|
|
|
host_name_ = name;
|
|
|
|
} else {
|
|
|
|
Log(options_.info_log, "Can't get hostname, use localhost as host name.");
|
|
|
|
host_name_ = "localhost";
|
|
|
|
}
|
|
|
|
last_log_ts = 0;
|
2012-11-26 22:56:45 +01:00
|
|
|
|
2013-11-07 20:31:56 +01:00
|
|
|
LogFlush(options_.info_log);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
DBImpl::~DBImpl() {
|
|
|
|
// Wait for background work to finish
|
2013-06-11 23:23:58 +02:00
|
|
|
if (flush_on_destroy_ && mem_->GetFirstSequenceNumber() != 0) {
|
2012-11-06 20:21:57 +01:00
|
|
|
FlushMemTable(FlushOptions());
|
|
|
|
}
|
2014-02-27 20:38:55 +01:00
|
|
|
|
2012-11-09 03:45:19 +01:00
|
|
|
mutex_.Lock();
|
2013-02-15 20:53:17 +01:00
|
|
|
shutting_down_.Release_Store(this); // Any non-nullptr value is ok
|
2013-09-13 23:38:37 +02:00
|
|
|
while (bg_compaction_scheduled_ ||
|
|
|
|
bg_flush_scheduled_ ||
|
|
|
|
bg_logstats_scheduled_) {
|
2011-06-07 16:40:26 +02:00
|
|
|
bg_cv_.Wait();
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
2014-02-27 20:38:55 +01:00
|
|
|
mutex_.Unlock();
|
|
|
|
|
|
|
|
// Release SuperVersion reference kept in ThreadLocalPtr.
|
|
|
|
// This must be done outside of mutex_ since unref handler can lock mutex.
|
|
|
|
// It also needs to be done after FlushMemTable, which can trigger local_sv_
|
|
|
|
// access.
|
|
|
|
delete local_sv_;
|
|
|
|
|
|
|
|
mutex_.Lock();
|
|
|
|
if (options_.allow_thread_local) {
|
|
|
|
// Clean up obsolete files due to SuperVersion release.
|
|
|
|
// (1) Need to delete to obsolete files before closing because RepairDB()
|
|
|
|
// scans all existing files in the file system and builds manifest file.
|
|
|
|
// Keeping obsolete files confuses the repair process.
|
|
|
|
// (2) Need to check if we Open()/Recover() the DB successfully before
|
|
|
|
// deleting because if VersionSet recover fails (may be due to corrupted
|
|
|
|
// manifest file), it is not able to identify live files correctly. As a
|
|
|
|
// result, all "live" files can get deleted by accident. However, corrupted
|
|
|
|
// manifest is recoverable by RepairDB().
|
|
|
|
if (opened_successfully_) {
|
|
|
|
DeletionState deletion_state;
|
|
|
|
FindObsoleteFiles(deletion_state, true);
|
|
|
|
// manifest number starting from 2
|
|
|
|
deletion_state.manifest_file_number = 1;
|
2014-03-10 23:42:14 +01:00
|
|
|
if (deletion_state.HaveSomethingToDelete()) {
|
|
|
|
PurgeObsoleteFiles(deletion_state);
|
|
|
|
}
|
2014-02-27 20:38:55 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-12-20 18:57:58 +01:00
|
|
|
if (super_version_ != nullptr) {
|
|
|
|
bool is_last_reference __attribute__((unused));
|
|
|
|
is_last_reference = super_version_->Unref();
|
|
|
|
assert(is_last_reference);
|
|
|
|
super_version_->Cleanup();
|
|
|
|
delete super_version_;
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
mutex_.Unlock();
|
|
|
|
|
2013-02-15 20:53:17 +01:00
|
|
|
if (db_lock_ != nullptr) {
|
2011-03-18 23:37:00 +01:00
|
|
|
env_->UnlockFile(db_lock_);
|
|
|
|
}
|
|
|
|
|
2013-11-25 20:55:36 +01:00
|
|
|
if (mem_ != nullptr) {
|
|
|
|
delete mem_->Unref();
|
|
|
|
}
|
|
|
|
|
2014-02-27 20:38:55 +01:00
|
|
|
autovector<MemTable*> to_delete;
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
imm_.current()->Unref(&to_delete);
|
2013-11-25 20:55:36 +01:00
|
|
|
for (MemTable* m: to_delete) {
|
|
|
|
delete m;
|
|
|
|
}
|
2014-01-07 05:29:17 +01:00
|
|
|
// versions need to be destroyed before table_cache since it can holds
|
|
|
|
// references to table_cache.
|
|
|
|
versions_.reset();
|
2013-11-07 20:31:56 +01:00
|
|
|
LogFlush(options_.info_log);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
2012-11-29 01:42:36 +01:00
|
|
|
// Do not flush and close database elegantly. Simulate a crash.
|
2012-11-17 00:28:14 +01:00
|
|
|
void DBImpl::TEST_Destroy_DBImpl() {
|
|
|
|
// ensure that no new memtable flushes can occur
|
|
|
|
flush_on_destroy_ = false;
|
|
|
|
|
|
|
|
// wait till all background compactions are done.
|
|
|
|
mutex_.Lock();
|
2013-09-13 23:38:37 +02:00
|
|
|
while (bg_compaction_scheduled_ ||
|
|
|
|
bg_flush_scheduled_ ||
|
|
|
|
bg_logstats_scheduled_) {
|
2012-11-17 00:28:14 +01:00
|
|
|
bg_cv_.Wait();
|
|
|
|
}
|
2014-03-07 03:15:26 +01:00
|
|
|
mutex_.Unlock();
|
|
|
|
|
|
|
|
// Release SuperVersion reference kept in ThreadLocalPtr.
|
|
|
|
// This must be done outside of mutex_ since unref handler can lock mutex.
|
|
|
|
// It also needs to be done after FlushMemTable, which can trigger local_sv_
|
|
|
|
// access.
|
|
|
|
delete local_sv_;
|
|
|
|
|
|
|
|
mutex_.Lock();
|
2013-12-20 18:57:58 +01:00
|
|
|
if (super_version_ != nullptr) {
|
|
|
|
bool is_last_reference __attribute__((unused));
|
|
|
|
is_last_reference = super_version_->Unref();
|
|
|
|
assert(is_last_reference);
|
|
|
|
super_version_->Cleanup();
|
|
|
|
delete super_version_;
|
|
|
|
}
|
2012-11-17 00:28:14 +01:00
|
|
|
|
|
|
|
// Prevent new compactions from occuring.
|
2013-09-13 23:38:37 +02:00
|
|
|
bg_work_gate_closed_ = true;
|
2012-11-17 00:28:14 +01:00
|
|
|
const int LargeNumber = 10000000;
|
|
|
|
bg_compaction_scheduled_ += LargeNumber;
|
2013-09-13 23:38:37 +02:00
|
|
|
|
2012-11-17 00:28:14 +01:00
|
|
|
mutex_.Unlock();
|
2013-11-07 20:31:56 +01:00
|
|
|
LogFlush(options_.info_log);
|
2012-11-17 00:28:14 +01:00
|
|
|
|
|
|
|
// force release the lock file.
|
2013-02-15 20:53:17 +01:00
|
|
|
if (db_lock_ != nullptr) {
|
2012-11-17 00:28:14 +01:00
|
|
|
env_->UnlockFile(db_lock_);
|
|
|
|
}
|
2013-01-20 11:07:13 +01:00
|
|
|
|
|
|
|
log_.reset();
|
|
|
|
versions_.reset();
|
|
|
|
table_cache_.reset();
|
2012-11-17 00:28:14 +01:00
|
|
|
}
|
|
|
|
|
2013-01-11 02:18:50 +01:00
|
|
|
uint64_t DBImpl::TEST_Current_Manifest_FileNo() {
|
|
|
|
return versions_->ManifestFileNumber();
|
|
|
|
}
|
2012-11-17 00:28:14 +01:00
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
Status DBImpl::NewDB() {
|
2014-01-15 00:27:09 +01:00
|
|
|
VersionEdit new_db;
|
2011-03-18 23:37:00 +01:00
|
|
|
new_db.SetComparatorName(user_comparator()->Name());
|
2011-04-12 21:38:58 +02:00
|
|
|
new_db.SetLogNumber(0);
|
2011-03-18 23:37:00 +01:00
|
|
|
new_db.SetNextFile(2);
|
|
|
|
new_db.SetLastSequence(0);
|
|
|
|
|
|
|
|
const std::string manifest = DescriptorFileName(dbname_, 1);
|
2013-01-20 11:07:13 +01:00
|
|
|
unique_ptr<WritableFile> file;
|
2014-03-18 05:52:14 +01:00
|
|
|
Status s = env_->NewWritableFile(
|
|
|
|
manifest, &file, env_->OptimizeForManifestWrite(storage_options_));
|
2011-03-18 23:37:00 +01:00
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
2013-01-15 23:05:42 +01:00
|
|
|
file->SetPreallocationBlockSize(options_.manifest_preallocation_size);
|
2011-03-18 23:37:00 +01:00
|
|
|
{
|
2013-01-20 11:07:13 +01:00
|
|
|
log::Writer log(std::move(file));
|
2011-03-18 23:37:00 +01:00
|
|
|
std::string record;
|
|
|
|
new_db.EncodeTo(&record);
|
|
|
|
s = log.AddRecord(record);
|
|
|
|
}
|
|
|
|
if (s.ok()) {
|
|
|
|
// Make "CURRENT" file that points to the new manifest file.
|
|
|
|
s = SetCurrentFile(env_, dbname_, 1);
|
|
|
|
} else {
|
|
|
|
env_->DeleteFile(manifest);
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
void DBImpl::MaybeIgnoreError(Status* s) const {
|
|
|
|
if (s->ok() || options_.paranoid_checks) {
|
|
|
|
// No change needed
|
|
|
|
} else {
|
2011-07-21 04:40:18 +02:00
|
|
|
Log(options_.info_log, "Ignoring error %s", s->ToString().c_str());
|
2011-03-18 23:37:00 +01:00
|
|
|
*s = Status::OK();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-11-26 22:56:45 +01:00
|
|
|
const Status DBImpl::CreateArchivalDirectory() {
|
2013-11-07 03:46:28 +01:00
|
|
|
if (options_.WAL_ttl_seconds > 0 || options_.WAL_size_limit_MB > 0) {
|
2013-10-01 23:46:52 +02:00
|
|
|
std::string archivalPath = ArchivalDirectory(options_.wal_dir);
|
2012-11-26 22:56:45 +01:00
|
|
|
return env_->CreateDirIfMissing(archivalPath);
|
|
|
|
}
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
2013-05-28 21:35:43 +02:00
|
|
|
void DBImpl::PrintStatistics() {
|
2013-11-22 23:14:05 +01:00
|
|
|
auto dbstats = options_.statistics.get();
|
2013-05-28 21:35:43 +02:00
|
|
|
if (dbstats) {
|
|
|
|
Log(options_.info_log,
|
2013-06-19 05:28:41 +02:00
|
|
|
"STATISTCS:\n %s",
|
|
|
|
dbstats->ToString().c_str());
|
2013-05-28 21:35:43 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-05-11 00:21:04 +02:00
|
|
|
void DBImpl::MaybeDumpStats() {
|
2013-05-24 21:52:45 +02:00
|
|
|
if (options_.stats_dump_period_sec == 0) return;
|
|
|
|
|
|
|
|
const uint64_t now_micros = env_->NowMicros();
|
|
|
|
|
|
|
|
if (last_stats_dump_time_microsec_ +
|
|
|
|
options_.stats_dump_period_sec * 1000000
|
|
|
|
<= now_micros) {
|
|
|
|
// Multiple threads could race in here simultaneously.
|
|
|
|
// However, the last one will update last_stats_dump_time_microsec_
|
|
|
|
// atomically. We could see more than one dump during one dump
|
|
|
|
// period in rare cases.
|
|
|
|
last_stats_dump_time_microsec_ = now_micros;
|
|
|
|
std::string stats;
|
2013-10-05 07:32:05 +02:00
|
|
|
GetProperty("rocksdb.stats", &stats);
|
2013-05-24 21:52:45 +02:00
|
|
|
Log(options_.info_log, "%s", stats.c_str());
|
2013-05-28 21:35:43 +02:00
|
|
|
PrintStatistics();
|
2013-05-11 00:21:04 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-12-20 18:57:58 +01:00
|
|
|
// DBImpl::SuperVersion methods
|
|
|
|
DBImpl::SuperVersion::~SuperVersion() {
|
|
|
|
for (auto td : to_delete) {
|
|
|
|
delete td;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
DBImpl::SuperVersion* DBImpl::SuperVersion::Ref() {
|
|
|
|
refs.fetch_add(1, std::memory_order_relaxed);
|
|
|
|
return this;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool DBImpl::SuperVersion::Unref() {
|
|
|
|
assert(refs > 0);
|
2014-02-11 05:11:19 +01:00
|
|
|
// fetch_sub returns the previous value of yoeref
|
2013-12-20 18:57:58 +01:00
|
|
|
return refs.fetch_sub(1, std::memory_order_relaxed) == 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
void DBImpl::SuperVersion::Cleanup() {
|
2014-03-09 06:12:13 +01:00
|
|
|
db->mutex_.AssertHeld();
|
2013-12-20 18:57:58 +01:00
|
|
|
assert(refs.load(std::memory_order_relaxed) == 0);
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
imm->Unref(&to_delete);
|
2013-12-20 18:57:58 +01:00
|
|
|
MemTable* m = mem->Unref();
|
|
|
|
if (m != nullptr) {
|
|
|
|
to_delete.push_back(m);
|
|
|
|
}
|
|
|
|
current->Unref();
|
|
|
|
}
|
|
|
|
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
void DBImpl::SuperVersion::Init(MemTable* new_mem, MemTableListVersion* new_imm,
|
2013-12-20 18:57:58 +01:00
|
|
|
Version* new_current) {
|
2014-03-09 06:12:13 +01:00
|
|
|
db->mutex_.AssertHeld();
|
2013-12-20 18:57:58 +01:00
|
|
|
mem = new_mem;
|
|
|
|
imm = new_imm;
|
|
|
|
current = new_current;
|
|
|
|
mem->Ref();
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
imm->Ref();
|
2013-12-20 18:57:58 +01:00
|
|
|
current->Ref();
|
|
|
|
refs.store(1, std::memory_order_relaxed);
|
|
|
|
}
|
|
|
|
|
2013-11-12 20:53:26 +01:00
|
|
|
// Returns the list of live files in 'sst_live' and the list
|
2013-12-31 03:33:57 +01:00
|
|
|
// of all files in the filesystem in 'candidate_files'.
|
2013-11-15 03:03:57 +01:00
|
|
|
// no_full_scan = true -- never do the full scan using GetChildren()
|
|
|
|
// force = false -- don't force the full scan, except every
|
|
|
|
// options_.delete_obsolete_files_period_micros
|
|
|
|
// force = true -- force the full scan
|
|
|
|
void DBImpl::FindObsoleteFiles(DeletionState& deletion_state,
|
|
|
|
bool force,
|
|
|
|
bool no_full_scan) {
|
2012-10-21 10:49:48 +02:00
|
|
|
mutex_.AssertHeld();
|
|
|
|
|
2012-09-15 02:11:35 +02:00
|
|
|
// if deletion is disabled, do nothing
|
2014-01-02 12:33:42 +01:00
|
|
|
if (disable_delete_obsolete_files_ > 0) {
|
2012-09-15 02:11:35 +02:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2013-11-15 07:42:02 +01:00
|
|
|
bool doing_the_full_scan = false;
|
|
|
|
|
|
|
|
// logic for figurint out if we're doing the full scan
|
|
|
|
if (no_full_scan) {
|
|
|
|
doing_the_full_scan = false;
|
|
|
|
} else if (force || options_.delete_obsolete_files_period_micros == 0) {
|
|
|
|
doing_the_full_scan = true;
|
|
|
|
} else {
|
|
|
|
const uint64_t now_micros = env_->NowMicros();
|
|
|
|
if (delete_obsolete_files_last_run_ +
|
|
|
|
options_.delete_obsolete_files_period_micros < now_micros) {
|
|
|
|
doing_the_full_scan = true;
|
|
|
|
delete_obsolete_files_last_run_ = now_micros;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-11-15 03:03:57 +01:00
|
|
|
// get obsolete files
|
|
|
|
versions_->GetObsoleteFiles(&deletion_state.sst_delete_files);
|
|
|
|
|
2013-11-09 00:23:46 +01:00
|
|
|
// store the current filenum, lognum, etc
|
|
|
|
deletion_state.manifest_file_number = versions_->ManifestFileNumber();
|
2014-03-18 05:50:15 +01:00
|
|
|
deletion_state.pending_manifest_file_number =
|
|
|
|
versions_->PendingManifestFileNumber();
|
2013-11-09 00:23:46 +01:00
|
|
|
deletion_state.log_number = versions_->LogNumber();
|
|
|
|
deletion_state.prev_log_number = versions_->PrevLogNumber();
|
|
|
|
|
2013-11-15 07:42:02 +01:00
|
|
|
if (!doing_the_full_scan && !deletion_state.HaveSomethingToDelete()) {
|
|
|
|
// avoid filling up sst_live if we're sure that we
|
|
|
|
// are not going to do the full scan and that we don't have
|
|
|
|
// anything to delete at the moment
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
// don't delete live files
|
2013-11-15 03:03:57 +01:00
|
|
|
deletion_state.sst_live.assign(pending_outputs_.begin(),
|
|
|
|
pending_outputs_.end());
|
|
|
|
versions_->AddLiveFiles(&deletion_state.sst_live);
|
|
|
|
|
2013-11-15 07:42:02 +01:00
|
|
|
if (doing_the_full_scan) {
|
2013-12-31 03:33:57 +01:00
|
|
|
// set of all files in the directory. We'll exclude files that are still
|
|
|
|
// alive in the subsequent processings.
|
|
|
|
env_->GetChildren(
|
|
|
|
dbname_, &deletion_state.candidate_files
|
|
|
|
); // Ignore errors
|
2013-11-15 07:42:02 +01:00
|
|
|
|
|
|
|
//Add log files in wal_dir
|
|
|
|
if (options_.wal_dir != dbname_) {
|
|
|
|
std::vector<std::string> log_files;
|
|
|
|
env_->GetChildren(options_.wal_dir, &log_files); // Ignore errors
|
2013-12-31 03:33:57 +01:00
|
|
|
deletion_state.candidate_files.insert(
|
|
|
|
deletion_state.candidate_files.end(),
|
2013-11-15 07:42:02 +01:00
|
|
|
log_files.begin(),
|
|
|
|
log_files.end()
|
|
|
|
);
|
2012-10-16 17:53:46 +02:00
|
|
|
}
|
2013-10-01 23:46:52 +02:00
|
|
|
}
|
2013-07-16 20:56:46 +02:00
|
|
|
}
|
|
|
|
|
2012-10-21 10:49:48 +02:00
|
|
|
// Diffs the files listed in filenames and those that do not
|
2013-11-09 00:23:46 +01:00
|
|
|
// belong to live files are posibly removed. Also, removes all the
|
2013-11-12 20:53:26 +01:00
|
|
|
// files in sst_delete_files and log_delete_files.
|
2013-10-01 23:46:52 +02:00
|
|
|
// It is not necessary to hold the mutex when invoking this method.
|
2012-10-21 10:49:48 +02:00
|
|
|
void DBImpl::PurgeObsoleteFiles(DeletionState& state) {
|
2014-03-09 23:11:34 +01:00
|
|
|
// we'd better have sth to delete
|
|
|
|
assert(state.HaveSomethingToDelete());
|
2013-11-27 23:56:20 +01:00
|
|
|
|
2013-11-15 03:03:57 +01:00
|
|
|
// this checks if FindObsoleteFiles() was run before. If not, don't do
|
|
|
|
// PurgeObsoleteFiles(). If FindObsoleteFiles() was run, we need to also
|
|
|
|
// run PurgeObsoleteFiles(), even if disable_delete_obsolete_files_ is true
|
|
|
|
if (state.manifest_file_number == 0) {
|
2013-11-09 00:23:46 +01:00
|
|
|
return;
|
|
|
|
}
|
2014-03-09 23:11:34 +01:00
|
|
|
|
[RocksDB] [Performance] Speed up FindObsoleteFiles
Summary:
FindObsoleteFiles was slow, holding the single big lock, resulted in bad p99 behavior.
Didn't profile anything, but several things could be improved:
1. VersionSet::AddLiveFiles works with std::set, which is by itself slow (a tree).
You also don't know how many dynamic allocations occur just for building up this tree.
switched to std::vector, also added logic to pre-calculate total size and do just one allocation
2. Don't see why env_->GetChildren() needs to be mutex proteced, moved to PurgeObsoleteFiles where
mutex could be unlocked.
3. switched std::set to std:unordered_set, the conversion from vector is also inside PurgeObsoleteFiles
I have a feeling this should pretty much fix it.
Test Plan: make check; db_stress
Reviewers: dhruba, heyongqiang, MarkCallaghan
Reviewed By: dhruba
CC: leveldb, zshao
Differential Revision: https://reviews.facebook.net/D10197
2013-04-12 01:49:53 +02:00
|
|
|
// Now, convert live list to an unordered set, WITHOUT mutex held;
|
|
|
|
// set is slow.
|
2014-03-18 05:50:15 +01:00
|
|
|
std::unordered_set<uint64_t> sst_live(state.sst_live.begin(),
|
|
|
|
state.sst_live.end());
|
2013-11-09 00:23:46 +01:00
|
|
|
|
2013-12-31 03:33:57 +01:00
|
|
|
auto& candidate_files = state.candidate_files;
|
|
|
|
candidate_files.reserve(
|
|
|
|
candidate_files.size() +
|
|
|
|
state.sst_delete_files.size() +
|
|
|
|
state.log_delete_files.size());
|
|
|
|
// We may ignore the dbname when generating the file names.
|
|
|
|
const char* kDumbDbName = "";
|
2013-11-12 20:53:26 +01:00
|
|
|
for (auto file : state.sst_delete_files) {
|
2013-12-31 03:33:57 +01:00
|
|
|
candidate_files.push_back(
|
|
|
|
TableFileName(kDumbDbName, file->number).substr(1)
|
|
|
|
);
|
2013-11-12 20:53:26 +01:00
|
|
|
delete file;
|
2013-11-09 00:23:46 +01:00
|
|
|
}
|
|
|
|
|
2013-12-31 03:33:57 +01:00
|
|
|
for (auto file_num : state.log_delete_files) {
|
|
|
|
if (file_num > 0) {
|
2014-03-18 05:50:15 +01:00
|
|
|
candidate_files.push_back(LogFileName(kDumbDbName, file_num).substr(1));
|
2013-11-09 00:23:46 +01:00
|
|
|
}
|
|
|
|
}
|
[RocksDB] [Performance] Speed up FindObsoleteFiles
Summary:
FindObsoleteFiles was slow, holding the single big lock, resulted in bad p99 behavior.
Didn't profile anything, but several things could be improved:
1. VersionSet::AddLiveFiles works with std::set, which is by itself slow (a tree).
You also don't know how many dynamic allocations occur just for building up this tree.
switched to std::vector, also added logic to pre-calculate total size and do just one allocation
2. Don't see why env_->GetChildren() needs to be mutex proteced, moved to PurgeObsoleteFiles where
mutex could be unlocked.
3. switched std::set to std:unordered_set, the conversion from vector is also inside PurgeObsoleteFiles
I have a feeling this should pretty much fix it.
Test Plan: make check; db_stress
Reviewers: dhruba, heyongqiang, MarkCallaghan
Reviewed By: dhruba
CC: leveldb, zshao
Differential Revision: https://reviews.facebook.net/D10197
2013-04-12 01:49:53 +02:00
|
|
|
|
2013-12-31 03:33:57 +01:00
|
|
|
// dedup state.candidate_files so we don't try to delete the same
|
2013-11-13 05:32:07 +01:00
|
|
|
// file twice
|
2013-12-31 03:33:57 +01:00
|
|
|
sort(candidate_files.begin(), candidate_files.end());
|
2014-03-18 05:50:15 +01:00
|
|
|
candidate_files.erase(unique(candidate_files.begin(), candidate_files.end()),
|
|
|
|
candidate_files.end());
|
2011-03-18 23:37:00 +01:00
|
|
|
|
2014-03-09 23:11:34 +01:00
|
|
|
std::vector<std::string> old_info_log_files;
|
|
|
|
|
2013-12-31 03:33:57 +01:00
|
|
|
for (const auto& to_delete : candidate_files) {
|
|
|
|
uint64_t number;
|
|
|
|
FileType type;
|
|
|
|
// Ignore file if we cannot recognize it.
|
|
|
|
if (!ParseFileName(to_delete, &number, &type)) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool keep = true;
|
|
|
|
switch (type) {
|
|
|
|
case kLogFile:
|
|
|
|
keep = ((number >= state.log_number) ||
|
|
|
|
(number == state.prev_log_number));
|
|
|
|
break;
|
|
|
|
case kDescriptorFile:
|
|
|
|
// Keep my manifest file, and any newer incarnations'
|
2014-03-18 05:50:15 +01:00
|
|
|
// (can happen during manifest roll)
|
2013-12-31 03:33:57 +01:00
|
|
|
keep = (number >= state.manifest_file_number);
|
|
|
|
break;
|
|
|
|
case kTableFile:
|
|
|
|
keep = (sst_live.find(number) != sst_live.end());
|
|
|
|
break;
|
|
|
|
case kTempFile:
|
|
|
|
// Any temp files that are currently being written to must
|
2014-03-18 05:50:15 +01:00
|
|
|
// be recorded in pending_outputs_, which is inserted into "live".
|
|
|
|
// Also, SetCurrentFile creates a temp file when writing out new
|
|
|
|
// manifest, which is equal to state.pending_manifest_file_number. We
|
|
|
|
// should not delete that file
|
|
|
|
keep = (sst_live.find(number) != sst_live.end()) ||
|
|
|
|
(number == state.pending_manifest_file_number);
|
2013-12-31 03:33:57 +01:00
|
|
|
break;
|
|
|
|
case kInfoLogFile:
|
|
|
|
keep = true;
|
|
|
|
if (number != 0) {
|
2014-03-09 23:11:34 +01:00
|
|
|
old_info_log_files.push_back(to_delete);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
2013-12-31 03:33:57 +01:00
|
|
|
break;
|
|
|
|
case kCurrentFile:
|
|
|
|
case kDBLockFile:
|
|
|
|
case kIdentityFile:
|
|
|
|
case kMetaDatabase:
|
|
|
|
keep = true;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (keep) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (type == kTableFile) {
|
|
|
|
// evict from cache
|
|
|
|
table_cache_->Evict(number);
|
|
|
|
}
|
2014-03-09 23:11:34 +01:00
|
|
|
|
2013-12-31 03:33:57 +01:00
|
|
|
std::string fname = ((type == kLogFile) ? options_.wal_dir : dbname_) +
|
|
|
|
"/" + to_delete;
|
|
|
|
if (type == kLogFile &&
|
|
|
|
(options_.WAL_ttl_seconds > 0 || options_.WAL_size_limit_MB > 0)) {
|
2014-03-09 23:11:34 +01:00
|
|
|
auto archived_log_name = ArchivedLogFileName(options_.wal_dir, number);
|
|
|
|
Status s = env_->RenameFile(fname, archived_log_name);
|
|
|
|
Log(options_.info_log,
|
|
|
|
"Move log file %s to %s -- %s\n",
|
|
|
|
fname.c_str(), archived_log_name.c_str(), s.ToString().c_str());
|
2013-12-31 03:33:57 +01:00
|
|
|
} else {
|
|
|
|
Status s = env_->DeleteFile(fname);
|
2014-03-09 23:11:34 +01:00
|
|
|
Log(options_.info_log, "Delete %s type=%d #%lu -- %s\n",
|
|
|
|
fname.c_str(), type, (unsigned long)number,
|
|
|
|
s.ToString().c_str());
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
}
|
2012-08-18 01:06:05 +02:00
|
|
|
|
2013-07-16 20:56:46 +02:00
|
|
|
// Delete old info log files.
|
2014-03-09 23:11:34 +01:00
|
|
|
size_t old_info_log_file_count = old_info_log_files.size();
|
2013-02-05 04:42:40 +01:00
|
|
|
// NOTE: Currently we only support log purge when options_.db_log_dir is
|
|
|
|
// located in `dbname` directory.
|
2014-03-09 23:11:34 +01:00
|
|
|
if (old_info_log_file_count >= options_.keep_log_file_num &&
|
2013-02-05 04:42:40 +01:00
|
|
|
options_.db_log_dir.empty()) {
|
2014-03-09 23:11:34 +01:00
|
|
|
std::sort(old_info_log_files.begin(), old_info_log_files.end());
|
|
|
|
size_t end = old_info_log_file_count - options_.keep_log_file_num;
|
2013-03-15 02:32:01 +01:00
|
|
|
for (unsigned int i = 0; i <= end; i++) {
|
2014-03-09 23:11:34 +01:00
|
|
|
std::string& to_delete = old_info_log_files.at(i);
|
|
|
|
Log(options_.info_log, "Delete info log file %s\n", to_delete.c_str());
|
|
|
|
Status s = env_->DeleteFile(dbname_ + "/" + to_delete);
|
|
|
|
if (!s.ok()) {
|
|
|
|
Log(options_.info_log, "Delete info log file %s FAILED -- %s\n",
|
|
|
|
to_delete.c_str(), s.ToString().c_str());
|
|
|
|
}
|
2012-08-18 01:06:05 +02:00
|
|
|
}
|
|
|
|
}
|
2013-05-06 20:41:01 +02:00
|
|
|
PurgeObsoleteWALFiles();
|
2013-11-09 00:23:46 +01:00
|
|
|
LogFlush(options_.info_log);
|
2012-10-21 10:49:48 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
void DBImpl::DeleteObsoleteFiles() {
|
|
|
|
mutex_.AssertHeld();
|
|
|
|
DeletionState deletion_state;
|
2013-11-09 00:23:46 +01:00
|
|
|
FindObsoleteFiles(deletion_state, true);
|
2014-03-10 23:42:14 +01:00
|
|
|
if (deletion_state.HaveSomethingToDelete()) {
|
|
|
|
PurgeObsoleteFiles(deletion_state);
|
|
|
|
}
|
2012-11-26 22:56:45 +01:00
|
|
|
}
|
|
|
|
|
2013-11-07 03:46:28 +01:00
|
|
|
// 1. Go through all archived files and
|
|
|
|
// a. if ttl is enabled, delete outdated files
|
|
|
|
// b. if archive size limit is enabled, delete empty files,
|
|
|
|
// compute file number and size.
|
|
|
|
// 2. If size limit is enabled:
|
|
|
|
// a. compute how many files should be deleted
|
|
|
|
// b. get sorted non-empty archived logs
|
|
|
|
// c. delete what should be deleted
|
2012-11-26 22:56:45 +01:00
|
|
|
void DBImpl::PurgeObsoleteWALFiles() {
|
2013-11-07 03:46:28 +01:00
|
|
|
bool const ttl_enabled = options_.WAL_ttl_seconds > 0;
|
|
|
|
bool const size_limit_enabled = options_.WAL_size_limit_MB > 0;
|
|
|
|
if (!ttl_enabled && !size_limit_enabled) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2013-05-06 20:41:01 +02:00
|
|
|
int64_t current_time;
|
|
|
|
Status s = env_->GetCurrentTime(¤t_time);
|
2013-11-07 03:46:28 +01:00
|
|
|
if (!s.ok()) {
|
|
|
|
Log(options_.info_log, "Can't get current time: %s", s.ToString().c_str());
|
|
|
|
assert(false);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
uint64_t const now_seconds = static_cast<uint64_t>(current_time);
|
|
|
|
uint64_t const time_to_check = (ttl_enabled && !size_limit_enabled) ?
|
|
|
|
options_.WAL_ttl_seconds / 2 : default_interval_to_delete_obsolete_WAL_;
|
2013-05-06 20:41:01 +02:00
|
|
|
|
2013-11-07 03:46:28 +01:00
|
|
|
if (purge_wal_files_last_run_ + time_to_check > now_seconds) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
purge_wal_files_last_run_ = now_seconds;
|
|
|
|
|
|
|
|
std::string archival_dir = ArchivalDirectory(options_.wal_dir);
|
|
|
|
std::vector<std::string> files;
|
|
|
|
s = env_->GetChildren(archival_dir, &files);
|
|
|
|
if (!s.ok()) {
|
|
|
|
Log(options_.info_log, "Can't get archive files: %s", s.ToString().c_str());
|
|
|
|
assert(false);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
size_t log_files_num = 0;
|
|
|
|
uint64_t log_file_size = 0;
|
|
|
|
|
|
|
|
for (auto& f : files) {
|
|
|
|
uint64_t number;
|
|
|
|
FileType type;
|
|
|
|
if (ParseFileName(f, &number, &type) && type == kLogFile) {
|
|
|
|
std::string const file_path = archival_dir + "/" + f;
|
|
|
|
if (ttl_enabled) {
|
|
|
|
uint64_t file_m_time;
|
|
|
|
Status const s = env_->GetFileModificationTime(file_path,
|
|
|
|
&file_m_time);
|
|
|
|
if (!s.ok()) {
|
|
|
|
Log(options_.info_log, "Can't get file mod time: %s: %s",
|
|
|
|
file_path.c_str(), s.ToString().c_str());
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
if (now_seconds - file_m_time > options_.WAL_ttl_seconds) {
|
|
|
|
Status const s = env_->DeleteFile(file_path);
|
|
|
|
if (!s.ok()) {
|
|
|
|
Log(options_.info_log, "Can't delete file: %s: %s",
|
|
|
|
file_path.c_str(), s.ToString().c_str());
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
continue;
|
2012-11-26 22:56:45 +01:00
|
|
|
}
|
2013-11-07 03:46:28 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
if (size_limit_enabled) {
|
|
|
|
uint64_t file_size;
|
|
|
|
Status const s = env_->GetFileSize(file_path, &file_size);
|
|
|
|
if (!s.ok()) {
|
|
|
|
Log(options_.info_log, "Can't get file size: %s: %s",
|
|
|
|
file_path.c_str(), s.ToString().c_str());
|
|
|
|
return;
|
|
|
|
} else {
|
|
|
|
if (file_size > 0) {
|
|
|
|
log_file_size = std::max(log_file_size, file_size);
|
|
|
|
++log_files_num;
|
|
|
|
} else {
|
|
|
|
Status s = env_->DeleteFile(file_path);
|
|
|
|
if (!s.ok()) {
|
|
|
|
Log(options_.info_log, "Can't delete file: %s: %s",
|
|
|
|
file_path.c_str(), s.ToString().c_str());
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (0 == log_files_num || !size_limit_enabled) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
size_t const files_keep_num = options_.WAL_size_limit_MB *
|
|
|
|
1024 * 1024 / log_file_size;
|
|
|
|
if (log_files_num <= files_keep_num) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
size_t files_del_num = log_files_num - files_keep_num;
|
|
|
|
VectorLogPtr archived_logs;
|
|
|
|
AppendSortedWalsOfType(archival_dir, archived_logs, kArchivedLogFile);
|
|
|
|
|
|
|
|
if (files_del_num > archived_logs.size()) {
|
|
|
|
Log(options_.info_log, "Trying to delete more archived log files than "
|
|
|
|
"exist. Deleting all");
|
|
|
|
files_del_num = archived_logs.size();
|
|
|
|
}
|
|
|
|
|
|
|
|
for (size_t i = 0; i < files_del_num; ++i) {
|
|
|
|
std::string const file_path = archived_logs[i]->PathName();
|
|
|
|
Status const s = DeleteFile(file_path);
|
|
|
|
if (!s.ok()) {
|
|
|
|
Log(options_.info_log, "Can't delete file: %s: %s",
|
|
|
|
file_path.c_str(), s.ToString().c_str());
|
|
|
|
continue;
|
2012-11-26 22:56:45 +01:00
|
|
|
}
|
|
|
|
}
|
2012-10-21 10:49:48 +02:00
|
|
|
}
|
|
|
|
|
Refactor Recover() code
Summary:
This diff does two things:
* Rethinks how we call Recover() with read_only option. Before, we call it with pointer to memtable where we'd like to apply those changes to. This memtable is set in db_impl_readonly.cc and it's actually DBImpl::mem_. Why don't we just apply updates to mem_ right away? It seems more intuitive.
* Changes when we apply updates to manifest. Before, the process is to recover all the logs, flush it to sst files and then do one giant commit that atomically adds all recovered sst files and sets the next log number. This works good enough, but causes some small troubles for my column family approach, since I can't have one VersionEdit apply to more than single column family[1]. The change here is to commit the files recovered from logs right away. Here is the state of the world before the change:
1. Recover log 5, add new sst files to edit
2. Recover log 7, add new sst files to edit
3. Recover log 8, add new sst files to edit
4. Commit all added sst files to manifest and mark log files 5, 7 and 8 as recoverd (via SetLogNumber(9) function)
After the change, we'll do:
1. Recover log 5, commit the new sst files and set log 5 as recovered
2. Recover log 7, commit the new sst files and set log 7 as recovered
3. Recover log 8, commit the new sst files and set log 8 as recovered
The added (small) benefit is that if we fail after (2), the new recovery will only have to recover log 8. In previous case, we'll have to restart the recovery from the beginning. The bigger benefit will be to enable easier integration of multiple column families in Recovery code path.
[1] I'm happy to dicuss this decison, but I believe this is the cleanest way to go. It also makes backward compatibility much easier. We don't have a requirement of adding multiple column families atomically.
Test Plan: make check
Reviewers: dhruba, haobo, kailiu, sdong
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15237
2014-01-22 19:45:26 +01:00
|
|
|
Status DBImpl::Recover(bool read_only, bool error_if_log_file_exist) {
|
2011-03-18 23:37:00 +01:00
|
|
|
mutex_.AssertHeld();
|
|
|
|
|
2013-02-15 20:53:17 +01:00
|
|
|
assert(db_lock_ == nullptr);
|
Refactor Recover() code
Summary:
This diff does two things:
* Rethinks how we call Recover() with read_only option. Before, we call it with pointer to memtable where we'd like to apply those changes to. This memtable is set in db_impl_readonly.cc and it's actually DBImpl::mem_. Why don't we just apply updates to mem_ right away? It seems more intuitive.
* Changes when we apply updates to manifest. Before, the process is to recover all the logs, flush it to sst files and then do one giant commit that atomically adds all recovered sst files and sets the next log number. This works good enough, but causes some small troubles for my column family approach, since I can't have one VersionEdit apply to more than single column family[1]. The change here is to commit the files recovered from logs right away. Here is the state of the world before the change:
1. Recover log 5, add new sst files to edit
2. Recover log 7, add new sst files to edit
3. Recover log 8, add new sst files to edit
4. Commit all added sst files to manifest and mark log files 5, 7 and 8 as recoverd (via SetLogNumber(9) function)
After the change, we'll do:
1. Recover log 5, commit the new sst files and set log 5 as recovered
2. Recover log 7, commit the new sst files and set log 7 as recovered
3. Recover log 8, commit the new sst files and set log 8 as recovered
The added (small) benefit is that if we fail after (2), the new recovery will only have to recover log 8. In previous case, we'll have to restart the recovery from the beginning. The bigger benefit will be to enable easier integration of multiple column families in Recovery code path.
[1] I'm happy to dicuss this decison, but I believe this is the cleanest way to go. It also makes backward compatibility much easier. We don't have a requirement of adding multiple column families atomically.
Test Plan: make check
Reviewers: dhruba, haobo, kailiu, sdong
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15237
2014-01-22 19:45:26 +01:00
|
|
|
if (!read_only) {
|
2013-01-07 19:11:18 +01:00
|
|
|
// We call CreateDirIfMissing() as the directory may already exist (if we
|
|
|
|
// are reopening a DB), when this happens we don't want creating the
|
|
|
|
// directory to cause an error. However, we need to check if creating the
|
|
|
|
// directory fails or else we may get an obscure message about the lock
|
|
|
|
// file not existing. One real-world example of this occurring is if
|
|
|
|
// env->CreateDirIfMissing() doesn't create intermediate directories, e.g.
|
|
|
|
// when dbname_ is "dir/db" but when "dir" doesn't exist.
|
|
|
|
Status s = env_->CreateDirIfMissing(dbname_);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2014-01-27 20:02:21 +01:00
|
|
|
s = env_->NewDirectory(dbname_, &db_directory_);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2013-01-07 19:11:18 +01:00
|
|
|
s = env_->LockFile(LockFileName(dbname_), &db_lock_);
|
2012-12-17 06:01:02 +01:00
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
|
2012-12-17 06:01:02 +01:00
|
|
|
if (!env_->FileExists(CurrentFileName(dbname_))) {
|
|
|
|
if (options_.create_if_missing) {
|
2013-03-21 23:59:47 +01:00
|
|
|
// TODO: add merge_operator name check
|
2012-12-17 06:01:02 +01:00
|
|
|
s = NewDB();
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
return Status::InvalidArgument(
|
|
|
|
dbname_, "does not exist (create_if_missing is false)");
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
} else {
|
2012-12-17 06:01:02 +01:00
|
|
|
if (options_.error_if_exists) {
|
|
|
|
return Status::InvalidArgument(
|
|
|
|
dbname_, "exists (error_if_exists is true)");
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
2013-10-18 23:50:54 +02:00
|
|
|
// Check for the IDENTITY file and create it if not there
|
|
|
|
if (!env_->FileExists(IdentityFileName(dbname_))) {
|
|
|
|
s = SetIdentityFile(env_, dbname_);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
2012-12-17 06:01:02 +01:00
|
|
|
Status s = versions_->Recover();
|
2014-03-20 22:18:29 +01:00
|
|
|
if (options_.paranoid_checks && s.ok()) {
|
|
|
|
s = CheckConsistency();
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
if (s.ok()) {
|
|
|
|
SequenceNumber max_sequence(0);
|
2011-06-22 04:36:45 +02:00
|
|
|
|
|
|
|
// Recover from all newer log files than the ones named in the
|
|
|
|
// descriptor (new log files may have been added by the previous
|
|
|
|
// incarnation without registering them in the descriptor).
|
|
|
|
//
|
|
|
|
// Note that PrevLogNumber() is no longer used, but we pay
|
|
|
|
// attention to it in case we are recovering a database
|
2013-10-05 07:32:05 +02:00
|
|
|
// produced by an older version of rocksdb.
|
2011-06-22 04:36:45 +02:00
|
|
|
const uint64_t min_log = versions_->LogNumber();
|
|
|
|
const uint64_t prev_log = versions_->PrevLogNumber();
|
|
|
|
std::vector<std::string> filenames;
|
2013-10-01 23:46:52 +02:00
|
|
|
s = env_->GetChildren(options_.wal_dir, &filenames);
|
2011-06-22 04:36:45 +02:00
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
2011-04-12 21:38:58 +02:00
|
|
|
}
|
2013-12-31 03:33:57 +01:00
|
|
|
|
2011-06-22 04:36:45 +02:00
|
|
|
std::vector<uint64_t> logs;
|
|
|
|
for (size_t i = 0; i < filenames.size(); i++) {
|
2013-12-31 03:33:57 +01:00
|
|
|
uint64_t number;
|
|
|
|
FileType type;
|
2011-06-22 04:36:45 +02:00
|
|
|
if (ParseFileName(filenames[i], &number, &type)
|
|
|
|
&& type == kLogFile
|
|
|
|
&& ((number >= min_log) || (number == prev_log))) {
|
|
|
|
logs.push_back(number);
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
2011-06-22 04:36:45 +02:00
|
|
|
|
2012-11-06 04:18:49 +01:00
|
|
|
if (logs.size() > 0 && error_if_log_file_exist) {
|
|
|
|
return Status::Corruption(""
|
|
|
|
"The db was opened in readonly mode with error_if_log_file_exist"
|
|
|
|
"flag but a log file already exists");
|
|
|
|
}
|
|
|
|
|
2011-06-22 04:36:45 +02:00
|
|
|
// Recover in the order in which the logs were generated
|
|
|
|
std::sort(logs.begin(), logs.end());
|
2013-12-31 03:33:57 +01:00
|
|
|
for (const auto& log : logs) {
|
2011-09-01 21:08:02 +02:00
|
|
|
// The previous incarnation may not have written any MANIFEST
|
|
|
|
// records after allocating this log number. So we manually
|
|
|
|
// update the file number allocation counter in VersionSet.
|
2013-12-31 03:33:57 +01:00
|
|
|
versions_->MarkFileNumberUsed(log);
|
2014-01-24 01:32:49 +01:00
|
|
|
s = RecoverLogFile(log, &max_sequence, read_only);
|
2011-06-22 04:36:45 +02:00
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
if (s.ok()) {
|
2011-04-12 21:38:58 +02:00
|
|
|
if (versions_->LastSequence() < max_sequence) {
|
|
|
|
versions_->SetLastSequence(max_sequence);
|
|
|
|
}
|
2013-11-22 23:14:05 +01:00
|
|
|
SetTickerCount(options_.statistics.get(), SEQUENCE_NUMBER,
|
2013-08-16 07:06:26 +02:00
|
|
|
versions_->LastSequence());
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
Refactor Recover() code
Summary:
This diff does two things:
* Rethinks how we call Recover() with read_only option. Before, we call it with pointer to memtable where we'd like to apply those changes to. This memtable is set in db_impl_readonly.cc and it's actually DBImpl::mem_. Why don't we just apply updates to mem_ right away? It seems more intuitive.
* Changes when we apply updates to manifest. Before, the process is to recover all the logs, flush it to sst files and then do one giant commit that atomically adds all recovered sst files and sets the next log number. This works good enough, but causes some small troubles for my column family approach, since I can't have one VersionEdit apply to more than single column family[1]. The change here is to commit the files recovered from logs right away. Here is the state of the world before the change:
1. Recover log 5, add new sst files to edit
2. Recover log 7, add new sst files to edit
3. Recover log 8, add new sst files to edit
4. Commit all added sst files to manifest and mark log files 5, 7 and 8 as recoverd (via SetLogNumber(9) function)
After the change, we'll do:
1. Recover log 5, commit the new sst files and set log 5 as recovered
2. Recover log 7, commit the new sst files and set log 7 as recovered
3. Recover log 8, commit the new sst files and set log 8 as recovered
The added (small) benefit is that if we fail after (2), the new recovery will only have to recover log 8. In previous case, we'll have to restart the recovery from the beginning. The bigger benefit will be to enable easier integration of multiple column families in Recovery code path.
[1] I'm happy to dicuss this decison, but I believe this is the cleanest way to go. It also makes backward compatibility much easier. We don't have a requirement of adding multiple column families atomically.
Test Plan: make check
Reviewers: dhruba, haobo, kailiu, sdong
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15237
2014-01-22 19:45:26 +01:00
|
|
|
Status DBImpl::RecoverLogFile(uint64_t log_number, SequenceNumber* max_sequence,
|
|
|
|
bool read_only) {
|
2011-03-18 23:37:00 +01:00
|
|
|
struct LogReporter : public log::Reader::Reporter {
|
|
|
|
Env* env;
|
2011-07-21 04:40:18 +02:00
|
|
|
Logger* info_log;
|
2011-03-18 23:37:00 +01:00
|
|
|
const char* fname;
|
2013-05-21 20:53:33 +02:00
|
|
|
Status* status; // nullptr if options_.paranoid_checks==false or
|
|
|
|
// options_.skip_log_error_on_recovery==true
|
2011-03-18 23:37:00 +01:00
|
|
|
virtual void Corruption(size_t bytes, const Status& s) {
|
2011-07-21 04:40:18 +02:00
|
|
|
Log(info_log, "%s%s: dropping %d bytes; %s",
|
2013-02-15 20:53:17 +01:00
|
|
|
(this->status == nullptr ? "(ignoring error) " : ""),
|
2011-03-18 23:37:00 +01:00
|
|
|
fname, static_cast<int>(bytes), s.ToString().c_str());
|
2013-02-15 20:53:17 +01:00
|
|
|
if (this->status != nullptr && this->status->ok()) *this->status = s;
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
mutex_.AssertHeld();
|
|
|
|
|
Refactor Recover() code
Summary:
This diff does two things:
* Rethinks how we call Recover() with read_only option. Before, we call it with pointer to memtable where we'd like to apply those changes to. This memtable is set in db_impl_readonly.cc and it's actually DBImpl::mem_. Why don't we just apply updates to mem_ right away? It seems more intuitive.
* Changes when we apply updates to manifest. Before, the process is to recover all the logs, flush it to sst files and then do one giant commit that atomically adds all recovered sst files and sets the next log number. This works good enough, but causes some small troubles for my column family approach, since I can't have one VersionEdit apply to more than single column family[1]. The change here is to commit the files recovered from logs right away. Here is the state of the world before the change:
1. Recover log 5, add new sst files to edit
2. Recover log 7, add new sst files to edit
3. Recover log 8, add new sst files to edit
4. Commit all added sst files to manifest and mark log files 5, 7 and 8 as recoverd (via SetLogNumber(9) function)
After the change, we'll do:
1. Recover log 5, commit the new sst files and set log 5 as recovered
2. Recover log 7, commit the new sst files and set log 7 as recovered
3. Recover log 8, commit the new sst files and set log 8 as recovered
The added (small) benefit is that if we fail after (2), the new recovery will only have to recover log 8. In previous case, we'll have to restart the recovery from the beginning. The bigger benefit will be to enable easier integration of multiple column families in Recovery code path.
[1] I'm happy to dicuss this decison, but I believe this is the cleanest way to go. It also makes backward compatibility much easier. We don't have a requirement of adding multiple column families atomically.
Test Plan: make check
Reviewers: dhruba, haobo, kailiu, sdong
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15237
2014-01-22 19:45:26 +01:00
|
|
|
VersionEdit edit;
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
// Open the log file
|
2013-10-01 23:46:52 +02:00
|
|
|
std::string fname = LogFileName(options_.wal_dir, log_number);
|
2013-01-20 11:07:13 +01:00
|
|
|
unique_ptr<SequentialFile> file;
|
2013-03-15 01:00:04 +01:00
|
|
|
Status status = env_->NewSequentialFile(fname, &file, storage_options_);
|
2011-03-18 23:37:00 +01:00
|
|
|
if (!status.ok()) {
|
|
|
|
MaybeIgnoreError(&status);
|
|
|
|
return status;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Create the log reader.
|
|
|
|
LogReporter reporter;
|
|
|
|
reporter.env = env_;
|
2013-01-20 11:07:13 +01:00
|
|
|
reporter.info_log = options_.info_log.get();
|
2011-03-18 23:37:00 +01:00
|
|
|
reporter.fname = fname.c_str();
|
2013-05-21 20:53:33 +02:00
|
|
|
reporter.status = (options_.paranoid_checks &&
|
|
|
|
!options_.skip_log_error_on_recovery ? &status : nullptr);
|
2011-03-18 23:37:00 +01:00
|
|
|
// We intentially make log::Reader do checksumming even if
|
|
|
|
// paranoid_checks==false so that corruptions cause entire commits
|
|
|
|
// to be skipped instead of propagating bad information (like overly
|
|
|
|
// large sequence numbers).
|
2013-01-20 11:07:13 +01:00
|
|
|
log::Reader reader(std::move(file), &reporter, true/*checksum*/,
|
2011-05-21 04:17:43 +02:00
|
|
|
0/*initial_offset*/);
|
2013-11-13 06:02:03 +01:00
|
|
|
Log(options_.info_log, "Recovering log #%lu",
|
|
|
|
(unsigned long) log_number);
|
2011-03-18 23:37:00 +01:00
|
|
|
|
|
|
|
// Read all the records and add to a memtable
|
|
|
|
std::string scratch;
|
|
|
|
Slice record;
|
|
|
|
WriteBatch batch;
|
Refactor Recover() code
Summary:
This diff does two things:
* Rethinks how we call Recover() with read_only option. Before, we call it with pointer to memtable where we'd like to apply those changes to. This memtable is set in db_impl_readonly.cc and it's actually DBImpl::mem_. Why don't we just apply updates to mem_ right away? It seems more intuitive.
* Changes when we apply updates to manifest. Before, the process is to recover all the logs, flush it to sst files and then do one giant commit that atomically adds all recovered sst files and sets the next log number. This works good enough, but causes some small troubles for my column family approach, since I can't have one VersionEdit apply to more than single column family[1]. The change here is to commit the files recovered from logs right away. Here is the state of the world before the change:
1. Recover log 5, add new sst files to edit
2. Recover log 7, add new sst files to edit
3. Recover log 8, add new sst files to edit
4. Commit all added sst files to manifest and mark log files 5, 7 and 8 as recoverd (via SetLogNumber(9) function)
After the change, we'll do:
1. Recover log 5, commit the new sst files and set log 5 as recovered
2. Recover log 7, commit the new sst files and set log 7 as recovered
3. Recover log 8, commit the new sst files and set log 8 as recovered
The added (small) benefit is that if we fail after (2), the new recovery will only have to recover log 8. In previous case, we'll have to restart the recovery from the beginning. The bigger benefit will be to enable easier integration of multiple column families in Recovery code path.
[1] I'm happy to dicuss this decison, but I believe this is the cleanest way to go. It also makes backward compatibility much easier. We don't have a requirement of adding multiple column families atomically.
Test Plan: make check
Reviewers: dhruba, haobo, kailiu, sdong
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15237
2014-01-22 19:45:26 +01:00
|
|
|
bool memtable_empty = true;
|
|
|
|
while (reader.ReadRecord(&record, &scratch)) {
|
2011-03-18 23:37:00 +01:00
|
|
|
if (record.size() < 12) {
|
|
|
|
reporter.Corruption(
|
|
|
|
record.size(), Status::Corruption("log record too small"));
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
WriteBatchInternal::SetContents(&batch, record);
|
|
|
|
|
Refactor Recover() code
Summary:
This diff does two things:
* Rethinks how we call Recover() with read_only option. Before, we call it with pointer to memtable where we'd like to apply those changes to. This memtable is set in db_impl_readonly.cc and it's actually DBImpl::mem_. Why don't we just apply updates to mem_ right away? It seems more intuitive.
* Changes when we apply updates to manifest. Before, the process is to recover all the logs, flush it to sst files and then do one giant commit that atomically adds all recovered sst files and sets the next log number. This works good enough, but causes some small troubles for my column family approach, since I can't have one VersionEdit apply to more than single column family[1]. The change here is to commit the files recovered from logs right away. Here is the state of the world before the change:
1. Recover log 5, add new sst files to edit
2. Recover log 7, add new sst files to edit
3. Recover log 8, add new sst files to edit
4. Commit all added sst files to manifest and mark log files 5, 7 and 8 as recoverd (via SetLogNumber(9) function)
After the change, we'll do:
1. Recover log 5, commit the new sst files and set log 5 as recovered
2. Recover log 7, commit the new sst files and set log 7 as recovered
3. Recover log 8, commit the new sst files and set log 8 as recovered
The added (small) benefit is that if we fail after (2), the new recovery will only have to recover log 8. In previous case, we'll have to restart the recovery from the beginning. The bigger benefit will be to enable easier integration of multiple column families in Recovery code path.
[1] I'm happy to dicuss this decison, but I believe this is the cleanest way to go. It also makes backward compatibility much easier. We don't have a requirement of adding multiple column families atomically.
Test Plan: make check
Reviewers: dhruba, haobo, kailiu, sdong
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15237
2014-01-22 19:45:26 +01:00
|
|
|
status = WriteBatchInternal::InsertInto(&batch, mem_, &options_);
|
|
|
|
memtable_empty = false;
|
2011-03-18 23:37:00 +01:00
|
|
|
MaybeIgnoreError(&status);
|
|
|
|
if (!status.ok()) {
|
Refactor Recover() code
Summary:
This diff does two things:
* Rethinks how we call Recover() with read_only option. Before, we call it with pointer to memtable where we'd like to apply those changes to. This memtable is set in db_impl_readonly.cc and it's actually DBImpl::mem_. Why don't we just apply updates to mem_ right away? It seems more intuitive.
* Changes when we apply updates to manifest. Before, the process is to recover all the logs, flush it to sst files and then do one giant commit that atomically adds all recovered sst files and sets the next log number. This works good enough, but causes some small troubles for my column family approach, since I can't have one VersionEdit apply to more than single column family[1]. The change here is to commit the files recovered from logs right away. Here is the state of the world before the change:
1. Recover log 5, add new sst files to edit
2. Recover log 7, add new sst files to edit
3. Recover log 8, add new sst files to edit
4. Commit all added sst files to manifest and mark log files 5, 7 and 8 as recoverd (via SetLogNumber(9) function)
After the change, we'll do:
1. Recover log 5, commit the new sst files and set log 5 as recovered
2. Recover log 7, commit the new sst files and set log 7 as recovered
3. Recover log 8, commit the new sst files and set log 8 as recovered
The added (small) benefit is that if we fail after (2), the new recovery will only have to recover log 8. In previous case, we'll have to restart the recovery from the beginning. The bigger benefit will be to enable easier integration of multiple column families in Recovery code path.
[1] I'm happy to dicuss this decison, but I believe this is the cleanest way to go. It also makes backward compatibility much easier. We don't have a requirement of adding multiple column families atomically.
Test Plan: make check
Reviewers: dhruba, haobo, kailiu, sdong
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15237
2014-01-22 19:45:26 +01:00
|
|
|
return status;
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
const SequenceNumber last_seq =
|
|
|
|
WriteBatchInternal::Sequence(&batch) +
|
|
|
|
WriteBatchInternal::Count(&batch) - 1;
|
|
|
|
if (last_seq > *max_sequence) {
|
|
|
|
*max_sequence = last_seq;
|
|
|
|
}
|
|
|
|
|
2014-03-13 00:40:14 +01:00
|
|
|
if (!read_only && mem_->ShouldFlush()) {
|
Refactor Recover() code
Summary:
This diff does two things:
* Rethinks how we call Recover() with read_only option. Before, we call it with pointer to memtable where we'd like to apply those changes to. This memtable is set in db_impl_readonly.cc and it's actually DBImpl::mem_. Why don't we just apply updates to mem_ right away? It seems more intuitive.
* Changes when we apply updates to manifest. Before, the process is to recover all the logs, flush it to sst files and then do one giant commit that atomically adds all recovered sst files and sets the next log number. This works good enough, but causes some small troubles for my column family approach, since I can't have one VersionEdit apply to more than single column family[1]. The change here is to commit the files recovered from logs right away. Here is the state of the world before the change:
1. Recover log 5, add new sst files to edit
2. Recover log 7, add new sst files to edit
3. Recover log 8, add new sst files to edit
4. Commit all added sst files to manifest and mark log files 5, 7 and 8 as recoverd (via SetLogNumber(9) function)
After the change, we'll do:
1. Recover log 5, commit the new sst files and set log 5 as recovered
2. Recover log 7, commit the new sst files and set log 7 as recovered
3. Recover log 8, commit the new sst files and set log 8 as recovered
The added (small) benefit is that if we fail after (2), the new recovery will only have to recover log 8. In previous case, we'll have to restart the recovery from the beginning. The bigger benefit will be to enable easier integration of multiple column families in Recovery code path.
[1] I'm happy to dicuss this decison, but I believe this is the cleanest way to go. It also makes backward compatibility much easier. We don't have a requirement of adding multiple column families atomically.
Test Plan: make check
Reviewers: dhruba, haobo, kailiu, sdong
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15237
2014-01-22 19:45:26 +01:00
|
|
|
status = WriteLevel0TableForRecovery(mem_, &edit);
|
|
|
|
// we still want to clear memtable, even if the recovery failed
|
|
|
|
delete mem_->Unref();
|
|
|
|
mem_ = new MemTable(internal_comparator_, options_);
|
|
|
|
mem_->Ref();
|
|
|
|
memtable_empty = true;
|
2011-03-18 23:37:00 +01:00
|
|
|
if (!status.ok()) {
|
|
|
|
// Reflect errors immediately so that conditions like full
|
|
|
|
// file-systems cause the DB::Open() to fail.
|
Refactor Recover() code
Summary:
This diff does two things:
* Rethinks how we call Recover() with read_only option. Before, we call it with pointer to memtable where we'd like to apply those changes to. This memtable is set in db_impl_readonly.cc and it's actually DBImpl::mem_. Why don't we just apply updates to mem_ right away? It seems more intuitive.
* Changes when we apply updates to manifest. Before, the process is to recover all the logs, flush it to sst files and then do one giant commit that atomically adds all recovered sst files and sets the next log number. This works good enough, but causes some small troubles for my column family approach, since I can't have one VersionEdit apply to more than single column family[1]. The change here is to commit the files recovered from logs right away. Here is the state of the world before the change:
1. Recover log 5, add new sst files to edit
2. Recover log 7, add new sst files to edit
3. Recover log 8, add new sst files to edit
4. Commit all added sst files to manifest and mark log files 5, 7 and 8 as recoverd (via SetLogNumber(9) function)
After the change, we'll do:
1. Recover log 5, commit the new sst files and set log 5 as recovered
2. Recover log 7, commit the new sst files and set log 7 as recovered
3. Recover log 8, commit the new sst files and set log 8 as recovered
The added (small) benefit is that if we fail after (2), the new recovery will only have to recover log 8. In previous case, we'll have to restart the recovery from the beginning. The bigger benefit will be to enable easier integration of multiple column families in Recovery code path.
[1] I'm happy to dicuss this decison, but I believe this is the cleanest way to go. It also makes backward compatibility much easier. We don't have a requirement of adding multiple column families atomically.
Test Plan: make check
Reviewers: dhruba, haobo, kailiu, sdong
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15237
2014-01-22 19:45:26 +01:00
|
|
|
return status;
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Refactor Recover() code
Summary:
This diff does two things:
* Rethinks how we call Recover() with read_only option. Before, we call it with pointer to memtable where we'd like to apply those changes to. This memtable is set in db_impl_readonly.cc and it's actually DBImpl::mem_. Why don't we just apply updates to mem_ right away? It seems more intuitive.
* Changes when we apply updates to manifest. Before, the process is to recover all the logs, flush it to sst files and then do one giant commit that atomically adds all recovered sst files and sets the next log number. This works good enough, but causes some small troubles for my column family approach, since I can't have one VersionEdit apply to more than single column family[1]. The change here is to commit the files recovered from logs right away. Here is the state of the world before the change:
1. Recover log 5, add new sst files to edit
2. Recover log 7, add new sst files to edit
3. Recover log 8, add new sst files to edit
4. Commit all added sst files to manifest and mark log files 5, 7 and 8 as recoverd (via SetLogNumber(9) function)
After the change, we'll do:
1. Recover log 5, commit the new sst files and set log 5 as recovered
2. Recover log 7, commit the new sst files and set log 7 as recovered
3. Recover log 8, commit the new sst files and set log 8 as recovered
The added (small) benefit is that if we fail after (2), the new recovery will only have to recover log 8. In previous case, we'll have to restart the recovery from the beginning. The bigger benefit will be to enable easier integration of multiple column families in Recovery code path.
[1] I'm happy to dicuss this decison, but I believe this is the cleanest way to go. It also makes backward compatibility much easier. We don't have a requirement of adding multiple column families atomically.
Test Plan: make check
Reviewers: dhruba, haobo, kailiu, sdong
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15237
2014-01-22 19:45:26 +01:00
|
|
|
if (!memtable_empty && !read_only) {
|
|
|
|
status = WriteLevel0TableForRecovery(mem_, &edit);
|
|
|
|
delete mem_->Unref();
|
|
|
|
mem_ = new MemTable(internal_comparator_, options_);
|
|
|
|
mem_->Ref();
|
|
|
|
if (!status.ok()) {
|
|
|
|
return status;
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
Refactor Recover() code
Summary:
This diff does two things:
* Rethinks how we call Recover() with read_only option. Before, we call it with pointer to memtable where we'd like to apply those changes to. This memtable is set in db_impl_readonly.cc and it's actually DBImpl::mem_. Why don't we just apply updates to mem_ right away? It seems more intuitive.
* Changes when we apply updates to manifest. Before, the process is to recover all the logs, flush it to sst files and then do one giant commit that atomically adds all recovered sst files and sets the next log number. This works good enough, but causes some small troubles for my column family approach, since I can't have one VersionEdit apply to more than single column family[1]. The change here is to commit the files recovered from logs right away. Here is the state of the world before the change:
1. Recover log 5, add new sst files to edit
2. Recover log 7, add new sst files to edit
3. Recover log 8, add new sst files to edit
4. Commit all added sst files to manifest and mark log files 5, 7 and 8 as recoverd (via SetLogNumber(9) function)
After the change, we'll do:
1. Recover log 5, commit the new sst files and set log 5 as recovered
2. Recover log 7, commit the new sst files and set log 7 as recovered
3. Recover log 8, commit the new sst files and set log 8 as recovered
The added (small) benefit is that if we fail after (2), the new recovery will only have to recover log 8. In previous case, we'll have to restart the recovery from the beginning. The bigger benefit will be to enable easier integration of multiple column families in Recovery code path.
[1] I'm happy to dicuss this decison, but I believe this is the cleanest way to go. It also makes backward compatibility much easier. We don't have a requirement of adding multiple column families atomically.
Test Plan: make check
Reviewers: dhruba, haobo, kailiu, sdong
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15237
2014-01-22 19:45:26 +01:00
|
|
|
if (edit.NumEntries() > 0) {
|
|
|
|
// if read_only, NumEntries() will be 0
|
|
|
|
assert(!read_only);
|
|
|
|
// writing log number in the manifest means that any log file
|
|
|
|
// with number strongly less than (log_number + 1) is already
|
|
|
|
// recovered and should be ignored on next reincarnation.
|
|
|
|
// Since we already recovered log_number, we want all logs
|
|
|
|
// with numbers `<= log_number` (includes this one) to be ignored
|
|
|
|
edit.SetLogNumber(log_number + 1);
|
Mark the log_number file number used
Summary:
VersionSet::next_file_number_ is always assumed to be strictly greater than VersionSet::log_number_. In our new recovery code, we artificially set log_number_ to be (log_number + 1), so that once we flush, we don't recover from the same log file again (this is important because of merge operator non-idempotence)
When we set VersionSet::log_number_ to (log_number + 1), we also have to mark that file number used, such that next_file_number_ is increased to a legal level. Otherwise, VersionSet might assert.
This has not be a problem so far because here's what happens:
1. assume next_file_number is 5, we're recovering log_number 10
2. in DBImpl::Recover() we call MarkFileNumberUsed with 10. This will set VersionSet::next_file_number_ to 11.
3. If there are some updates, we will call WriteTable0ForRecovery(), which will use file number 11 as a new table file and advance VersionSet::next_file_number_ to 12.
4. When we LogAndApply() with log_number 11, assertion is true: assert(11 <= 12);
However, this was a lucky occurrence. Even though this diff doesn't cause a bug, I think the issue is important to fix.
Test Plan: In column families I have different recovery logic and this code path asserted. When adding MarkFileNumberUsed(log_number + 1) assert is gone.
Reviewers: dhruba, kailiu
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15783
2014-01-31 23:43:16 +01:00
|
|
|
// we must mark the next log number as used, even though it's
|
|
|
|
// not actually used. that is because VersionSet assumes
|
|
|
|
// VersionSet::next_file_number_ always to be strictly greater than any log
|
|
|
|
// number
|
|
|
|
versions_->MarkFileNumberUsed(log_number + 1);
|
Refactor Recover() code
Summary:
This diff does two things:
* Rethinks how we call Recover() with read_only option. Before, we call it with pointer to memtable where we'd like to apply those changes to. This memtable is set in db_impl_readonly.cc and it's actually DBImpl::mem_. Why don't we just apply updates to mem_ right away? It seems more intuitive.
* Changes when we apply updates to manifest. Before, the process is to recover all the logs, flush it to sst files and then do one giant commit that atomically adds all recovered sst files and sets the next log number. This works good enough, but causes some small troubles for my column family approach, since I can't have one VersionEdit apply to more than single column family[1]. The change here is to commit the files recovered from logs right away. Here is the state of the world before the change:
1. Recover log 5, add new sst files to edit
2. Recover log 7, add new sst files to edit
3. Recover log 8, add new sst files to edit
4. Commit all added sst files to manifest and mark log files 5, 7 and 8 as recoverd (via SetLogNumber(9) function)
After the change, we'll do:
1. Recover log 5, commit the new sst files and set log 5 as recovered
2. Recover log 7, commit the new sst files and set log 7 as recovered
3. Recover log 8, commit the new sst files and set log 8 as recovered
The added (small) benefit is that if we fail after (2), the new recovery will only have to recover log 8. In previous case, we'll have to restart the recovery from the beginning. The bigger benefit will be to enable easier integration of multiple column families in Recovery code path.
[1] I'm happy to dicuss this decison, but I believe this is the cleanest way to go. It also makes backward compatibility much easier. We don't have a requirement of adding multiple column families atomically.
Test Plan: make check
Reviewers: dhruba, haobo, kailiu, sdong
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15237
2014-01-22 19:45:26 +01:00
|
|
|
status = versions_->LogAndApply(&edit, &mutex_);
|
2013-11-25 20:55:36 +01:00
|
|
|
}
|
Refactor Recover() code
Summary:
This diff does two things:
* Rethinks how we call Recover() with read_only option. Before, we call it with pointer to memtable where we'd like to apply those changes to. This memtable is set in db_impl_readonly.cc and it's actually DBImpl::mem_. Why don't we just apply updates to mem_ right away? It seems more intuitive.
* Changes when we apply updates to manifest. Before, the process is to recover all the logs, flush it to sst files and then do one giant commit that atomically adds all recovered sst files and sets the next log number. This works good enough, but causes some small troubles for my column family approach, since I can't have one VersionEdit apply to more than single column family[1]. The change here is to commit the files recovered from logs right away. Here is the state of the world before the change:
1. Recover log 5, add new sst files to edit
2. Recover log 7, add new sst files to edit
3. Recover log 8, add new sst files to edit
4. Commit all added sst files to manifest and mark log files 5, 7 and 8 as recoverd (via SetLogNumber(9) function)
After the change, we'll do:
1. Recover log 5, commit the new sst files and set log 5 as recovered
2. Recover log 7, commit the new sst files and set log 7 as recovered
3. Recover log 8, commit the new sst files and set log 8 as recovered
The added (small) benefit is that if we fail after (2), the new recovery will only have to recover log 8. In previous case, we'll have to restart the recovery from the beginning. The bigger benefit will be to enable easier integration of multiple column families in Recovery code path.
[1] I'm happy to dicuss this decison, but I believe this is the cleanest way to go. It also makes backward compatibility much easier. We don't have a requirement of adding multiple column families atomically.
Test Plan: make check
Reviewers: dhruba, haobo, kailiu, sdong
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15237
2014-01-22 19:45:26 +01:00
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
return status;
|
|
|
|
}
|
|
|
|
|
2012-10-19 23:00:53 +02:00
|
|
|
Status DBImpl::WriteLevel0TableForRecovery(MemTable* mem, VersionEdit* edit) {
|
2011-03-18 23:37:00 +01:00
|
|
|
mutex_.AssertHeld();
|
2011-04-12 21:38:58 +02:00
|
|
|
const uint64_t start_micros = env_->NowMicros();
|
2011-03-18 23:37:00 +01:00
|
|
|
FileMetaData meta;
|
|
|
|
meta.number = versions_->NewFileNumber();
|
|
|
|
pending_outputs_.insert(meta.number);
|
|
|
|
Iterator* iter = mem->NewIterator();
|
2013-02-28 23:09:30 +01:00
|
|
|
const SequenceNumber newest_snapshot = snapshots_.GetNewest();
|
|
|
|
const SequenceNumber earliest_seqno_in_memtable =
|
|
|
|
mem->GetFirstSequenceNumber();
|
2013-11-13 06:02:03 +01:00
|
|
|
Log(options_.info_log, "Level-0 table #%lu: started",
|
|
|
|
(unsigned long) meta.number);
|
2011-04-12 21:38:58 +02:00
|
|
|
|
|
|
|
Status s;
|
|
|
|
{
|
|
|
|
mutex_.Unlock();
|
2013-03-15 01:00:04 +01:00
|
|
|
s = BuildTable(dbname_, env_, options_, storage_options_,
|
2014-01-27 22:53:22 +01:00
|
|
|
table_cache_.get(), iter, &meta, internal_comparator_,
|
|
|
|
newest_snapshot, earliest_seqno_in_memtable,
|
2013-12-19 19:02:53 +01:00
|
|
|
GetCompressionFlush(options_));
|
2013-11-07 20:31:56 +01:00
|
|
|
LogFlush(options_.info_log);
|
2011-04-12 21:38:58 +02:00
|
|
|
mutex_.Lock();
|
|
|
|
}
|
|
|
|
|
2013-11-13 06:02:03 +01:00
|
|
|
Log(options_.info_log, "Level-0 table #%lu: %lu bytes %s",
|
|
|
|
(unsigned long) meta.number,
|
|
|
|
(unsigned long) meta.file_size,
|
2011-03-18 23:37:00 +01:00
|
|
|
s.ToString().c_str());
|
|
|
|
delete iter;
|
2011-04-12 21:38:58 +02:00
|
|
|
|
2012-10-19 23:00:53 +02:00
|
|
|
pending_outputs_.erase(meta.number);
|
2011-06-22 04:36:45 +02:00
|
|
|
|
|
|
|
// Note that if file_size is zero, the file has been deleted and
|
|
|
|
// should not be added to the manifest.
|
|
|
|
int level = 0;
|
|
|
|
if (s.ok() && meta.file_size > 0) {
|
|
|
|
edit->AddFile(level, meta.number, meta.file_size,
|
2013-06-14 07:09:08 +02:00
|
|
|
meta.smallest, meta.largest,
|
|
|
|
meta.smallest_seqno, meta.largest_seqno);
|
2011-06-22 04:36:45 +02:00
|
|
|
}
|
|
|
|
|
2014-01-30 05:40:41 +01:00
|
|
|
InternalStats::CompactionStats stats;
|
2011-04-12 21:38:58 +02:00
|
|
|
stats.micros = env_->NowMicros() - start_micros;
|
|
|
|
stats.bytes_written = meta.file_size;
|
2012-10-23 19:34:09 +02:00
|
|
|
stats.files_out_levelnp1 = 1;
|
2014-01-30 05:40:41 +01:00
|
|
|
internal_stats_.AddCompactionStats(level, stats);
|
Add monitoring for universal compaction and add counters for compaction IO
Summary:
Adds these counters
{ WAL_FILE_SYNCED, "rocksdb.wal.synced" }
number of writes that request a WAL sync
{ WAL_FILE_BYTES, "rocksdb.wal.bytes" },
number of bytes written to the WAL
{ WRITE_DONE_BY_SELF, "rocksdb.write.self" },
number of writes processed by the calling thread
{ WRITE_DONE_BY_OTHER, "rocksdb.write.other" },
number of writes not processed by the calling thread. Instead these were
processed by the current holder of the write lock
{ WRITE_WITH_WAL, "rocksdb.write.wal" },
number of writes that request WAL logging
{ COMPACT_READ_BYTES, "rocksdb.compact.read.bytes" },
number of bytes read during compaction
{ COMPACT_WRITE_BYTES, "rocksdb.compact.write.bytes" },
number of bytes written during compaction
Per-interval stats output was updated with WAL stats and correct stats for universal compaction
including a correct value for write-amplification. It now looks like:
Compactions
Level Files Size(MB) Score Time(sec) Read(MB) Write(MB) Rn(MB) Rnp1(MB) Wnew(MB) RW-Amplify Read(MB/s) Write(MB/s) Rn Rnp1 Wnp1 NewW Count Ln-stall Stall-cnt
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
0 7 464 46.4 281 3411 3875 3411 0 3875 2.1 12.1 13.8 621 0 240 240 628 0.0 0
Uptime(secs): 310.8 total, 2.0 interval
Writes cumulative: 9999999 total, 9999999 batches, 1.0 per batch, 1.22 ingest GB
WAL cumulative: 9999999 WAL writes, 9999999 WAL syncs, 1.00 writes per sync, 1.22 GB written
Compaction IO cumulative (GB): 1.22 new, 3.33 read, 3.78 write, 7.12 read+write
Compaction IO cumulative (MB/sec): 4.0 new, 11.0 read, 12.5 write, 23.4 read+write
Amplification cumulative: 4.1 write, 6.8 compaction
Writes interval: 100000 total, 100000 batches, 1.0 per batch, 12.5 ingest MB
WAL interval: 100000 WAL writes, 100000 WAL syncs, 1.00 writes per sync, 0.01 MB written
Compaction IO interval (MB): 12.49 new, 14.98 read, 21.50 write, 36.48 read+write
Compaction IO interval (MB/sec): 6.4 new, 7.6 read, 11.0 write, 18.6 read+write
Amplification interval: 101.7 write, 102.9 compaction
Stalls(secs): 142.924 level0_slowdown, 0.000 level0_numfiles, 0.805 memtable_compaction, 0.000 leveln_slowdown
Stalls(count): 132461 level0_slowdown, 0 level0_numfiles, 3 memtable_compaction, 0 leveln_slowdown
Task ID: #3329644, #3301695
Blame Rev:
Test Plan:
Revert Plan:
Database Impact:
Memcache Impact:
Other Notes:
EImportant:
- begin *PUBLIC* platform impact section -
Bugzilla: #
- end platform impact -
Reviewers: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14583
2013-12-09 22:43:34 +01:00
|
|
|
RecordTick(options_.statistics.get(), COMPACT_WRITE_BYTES, meta.file_size);
|
2011-03-18 23:37:00 +01:00
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2012-10-19 23:00:53 +02:00
|
|
|
|
2014-01-02 20:26:57 +01:00
|
|
|
Status DBImpl::WriteLevel0Table(autovector<MemTable*>& mems, VersionEdit* edit,
|
2014-03-10 06:01:13 +01:00
|
|
|
uint64_t* filenumber,
|
|
|
|
LogBuffer* log_buffer) {
|
2011-03-18 23:37:00 +01:00
|
|
|
mutex_.AssertHeld();
|
2012-10-19 23:00:53 +02:00
|
|
|
const uint64_t start_micros = env_->NowMicros();
|
|
|
|
FileMetaData meta;
|
|
|
|
meta.number = versions_->NewFileNumber();
|
|
|
|
*filenumber = meta.number;
|
|
|
|
pending_outputs_.insert(meta.number);
|
2013-06-11 23:23:58 +02:00
|
|
|
|
2013-02-28 23:09:30 +01:00
|
|
|
const SequenceNumber newest_snapshot = snapshots_.GetNewest();
|
|
|
|
const SequenceNumber earliest_seqno_in_memtable =
|
2013-06-11 23:23:58 +02:00
|
|
|
mems[0]->GetFirstSequenceNumber();
|
2011-06-22 04:36:45 +02:00
|
|
|
Version* base = versions_->current();
|
2013-06-11 23:23:58 +02:00
|
|
|
base->Ref(); // it is likely that we do not need this reference
|
2012-10-19 23:00:53 +02:00
|
|
|
Status s;
|
|
|
|
{
|
|
|
|
mutex_.Unlock();
|
2014-03-10 06:01:13 +01:00
|
|
|
log_buffer->FlushBufferToLog();
|
2014-01-02 20:26:57 +01:00
|
|
|
std::vector<Iterator*> memtables;
|
2013-12-11 05:03:27 +01:00
|
|
|
for (MemTable* m : mems) {
|
|
|
|
Log(options_.info_log,
|
|
|
|
"Flushing memtable with log file: %lu\n",
|
|
|
|
(unsigned long)m->GetLogNumber());
|
2014-01-02 20:26:57 +01:00
|
|
|
memtables.push_back(m->NewIterator());
|
2013-12-11 05:03:27 +01:00
|
|
|
}
|
2014-01-02 20:26:57 +01:00
|
|
|
Iterator* iter = NewMergingIterator(
|
|
|
|
env_, &internal_comparator_, &memtables[0], memtables.size());
|
2013-12-11 05:03:27 +01:00
|
|
|
Log(options_.info_log,
|
|
|
|
"Level-0 flush table #%lu: started",
|
|
|
|
(unsigned long)meta.number);
|
2013-12-19 19:02:53 +01:00
|
|
|
|
2013-03-15 01:00:04 +01:00
|
|
|
s = BuildTable(dbname_, env_, options_, storage_options_,
|
2014-01-27 22:53:22 +01:00
|
|
|
table_cache_.get(), iter, &meta, internal_comparator_,
|
|
|
|
newest_snapshot, earliest_seqno_in_memtable,
|
|
|
|
GetCompressionFlush(options_));
|
2013-11-07 20:31:56 +01:00
|
|
|
LogFlush(options_.info_log);
|
2013-12-11 05:03:27 +01:00
|
|
|
delete iter;
|
|
|
|
Log(options_.info_log, "Level-0 flush table #%lu: %lu bytes %s",
|
|
|
|
(unsigned long) meta.number,
|
|
|
|
(unsigned long) meta.file_size,
|
|
|
|
s.ToString().c_str());
|
2014-01-27 20:02:21 +01:00
|
|
|
if (!options_.disableDataSync) {
|
|
|
|
db_directory_->Fsync();
|
|
|
|
}
|
2012-10-19 23:00:53 +02:00
|
|
|
mutex_.Lock();
|
|
|
|
}
|
2011-06-22 04:36:45 +02:00
|
|
|
base->Unref();
|
|
|
|
|
2012-10-19 23:00:53 +02:00
|
|
|
// re-acquire the most current version
|
|
|
|
base = versions_->current();
|
|
|
|
|
|
|
|
// There could be multiple threads writing to its own level-0 file.
|
|
|
|
// The pending_outputs cannot be cleared here, otherwise this newly
|
|
|
|
// created file might not be considered as a live-file by another
|
|
|
|
// compaction thread that is concurrently deleting obselete files.
|
|
|
|
// The pending_outputs can be cleared only after the new version is
|
2012-11-29 01:42:36 +01:00
|
|
|
// committed so that other threads can recognize this file as a
|
2012-10-19 23:00:53 +02:00
|
|
|
// valid one.
|
|
|
|
// pending_outputs_.erase(meta.number);
|
|
|
|
|
|
|
|
// Note that if file_size is zero, the file has been deleted and
|
|
|
|
// should not be added to the manifest.
|
|
|
|
int level = 0;
|
|
|
|
if (s.ok() && meta.file_size > 0) {
|
|
|
|
const Slice min_user_key = meta.smallest.user_key();
|
|
|
|
const Slice max_user_key = meta.largest.user_key();
|
|
|
|
// if we have more than 1 background thread, then we cannot
|
|
|
|
// insert files directly into higher levels because some other
|
|
|
|
// threads could be concurrently producing compacted files for
|
|
|
|
// that key range.
|
2013-06-14 07:09:08 +02:00
|
|
|
if (base != nullptr && options_.max_background_compactions <= 1 &&
|
2013-07-04 00:32:49 +02:00
|
|
|
options_.compaction_style == kCompactionStyleLevel) {
|
2012-10-19 23:00:53 +02:00
|
|
|
level = base->PickLevelForMemTableOutput(min_user_key, max_user_key);
|
|
|
|
}
|
|
|
|
edit->AddFile(level, meta.number, meta.file_size,
|
2013-06-14 07:09:08 +02:00
|
|
|
meta.smallest, meta.largest,
|
|
|
|
meta.smallest_seqno, meta.largest_seqno);
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
|
|
|
|
2014-01-30 05:40:41 +01:00
|
|
|
InternalStats::CompactionStats stats;
|
2012-10-19 23:00:53 +02:00
|
|
|
stats.micros = env_->NowMicros() - start_micros;
|
|
|
|
stats.bytes_written = meta.file_size;
|
2014-01-30 05:40:41 +01:00
|
|
|
internal_stats_.AddCompactionStats(level, stats);
|
Add monitoring for universal compaction and add counters for compaction IO
Summary:
Adds these counters
{ WAL_FILE_SYNCED, "rocksdb.wal.synced" }
number of writes that request a WAL sync
{ WAL_FILE_BYTES, "rocksdb.wal.bytes" },
number of bytes written to the WAL
{ WRITE_DONE_BY_SELF, "rocksdb.write.self" },
number of writes processed by the calling thread
{ WRITE_DONE_BY_OTHER, "rocksdb.write.other" },
number of writes not processed by the calling thread. Instead these were
processed by the current holder of the write lock
{ WRITE_WITH_WAL, "rocksdb.write.wal" },
number of writes that request WAL logging
{ COMPACT_READ_BYTES, "rocksdb.compact.read.bytes" },
number of bytes read during compaction
{ COMPACT_WRITE_BYTES, "rocksdb.compact.write.bytes" },
number of bytes written during compaction
Per-interval stats output was updated with WAL stats and correct stats for universal compaction
including a correct value for write-amplification. It now looks like:
Compactions
Level Files Size(MB) Score Time(sec) Read(MB) Write(MB) Rn(MB) Rnp1(MB) Wnew(MB) RW-Amplify Read(MB/s) Write(MB/s) Rn Rnp1 Wnp1 NewW Count Ln-stall Stall-cnt
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
0 7 464 46.4 281 3411 3875 3411 0 3875 2.1 12.1 13.8 621 0 240 240 628 0.0 0
Uptime(secs): 310.8 total, 2.0 interval
Writes cumulative: 9999999 total, 9999999 batches, 1.0 per batch, 1.22 ingest GB
WAL cumulative: 9999999 WAL writes, 9999999 WAL syncs, 1.00 writes per sync, 1.22 GB written
Compaction IO cumulative (GB): 1.22 new, 3.33 read, 3.78 write, 7.12 read+write
Compaction IO cumulative (MB/sec): 4.0 new, 11.0 read, 12.5 write, 23.4 read+write
Amplification cumulative: 4.1 write, 6.8 compaction
Writes interval: 100000 total, 100000 batches, 1.0 per batch, 12.5 ingest MB
WAL interval: 100000 WAL writes, 100000 WAL syncs, 1.00 writes per sync, 0.01 MB written
Compaction IO interval (MB): 12.49 new, 14.98 read, 21.50 write, 36.48 read+write
Compaction IO interval (MB/sec): 6.4 new, 7.6 read, 11.0 write, 18.6 read+write
Amplification interval: 101.7 write, 102.9 compaction
Stalls(secs): 142.924 level0_slowdown, 0.000 level0_numfiles, 0.805 memtable_compaction, 0.000 leveln_slowdown
Stalls(count): 132461 level0_slowdown, 0 level0_numfiles, 3 memtable_compaction, 0 leveln_slowdown
Task ID: #3329644, #3301695
Blame Rev:
Test Plan:
Revert Plan:
Database Impact:
Memcache Impact:
Other Notes:
EImportant:
- begin *PUBLIC* platform impact section -
Bugzilla: #
- end platform impact -
Reviewers: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14583
2013-12-09 22:43:34 +01:00
|
|
|
RecordTick(options_.statistics.get(), COMPACT_WRITE_BYTES, meta.file_size);
|
2012-10-19 23:00:53 +02:00
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2013-11-09 00:23:46 +01:00
|
|
|
Status DBImpl::FlushMemTableToOutputFile(bool* madeProgress,
|
2014-03-10 06:01:13 +01:00
|
|
|
DeletionState& deletion_state,
|
|
|
|
LogBuffer* log_buffer) {
|
2012-10-19 23:00:53 +02:00
|
|
|
mutex_.AssertHeld();
|
|
|
|
assert(imm_.size() != 0);
|
2014-02-12 20:42:54 +01:00
|
|
|
assert(imm_.IsFlushPending());
|
2012-10-19 23:00:53 +02:00
|
|
|
|
|
|
|
// Save the contents of the earliest memtable as a new Table
|
|
|
|
uint64_t file_number;
|
2014-01-02 20:26:57 +01:00
|
|
|
autovector<MemTable*> mems;
|
2013-06-11 23:23:58 +02:00
|
|
|
imm_.PickMemtablesToFlush(&mems);
|
|
|
|
if (mems.empty()) {
|
2014-03-10 06:01:13 +01:00
|
|
|
LogToBuffer(log_buffer, "Nothing in memstore to flush");
|
2014-02-12 20:42:54 +01:00
|
|
|
return Status::OK();
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
// record the logfile_number_ before we release the mutex
|
2013-10-25 00:58:00 +02:00
|
|
|
// entries mems are (implicitly) sorted in ascending order by their created
|
|
|
|
// time. We will use the first memtable's `edit` to keep the meta info for
|
|
|
|
// this flush.
|
2013-06-11 23:23:58 +02:00
|
|
|
MemTable* m = mems[0];
|
2012-10-19 23:00:53 +02:00
|
|
|
VersionEdit* edit = m->GetEdits();
|
|
|
|
edit->SetPrevLogNumber(0);
|
2013-10-25 00:58:00 +02:00
|
|
|
// SetLogNumber(log_num) indicates logs with number smaller than log_num
|
|
|
|
// will no longer be picked up for recovery.
|
2014-03-20 21:10:02 +01:00
|
|
|
edit->SetLogNumber(mems.back()->GetNextLogNumber());
|
2013-10-25 00:58:00 +02:00
|
|
|
|
|
|
|
std::vector<uint64_t> logs_to_delete;
|
|
|
|
for (auto mem : mems) {
|
|
|
|
logs_to_delete.push_back(mem->GetLogNumber());
|
|
|
|
}
|
2012-10-19 23:00:53 +02:00
|
|
|
|
2013-07-16 20:56:46 +02:00
|
|
|
// This will release and re-acquire the mutex.
|
2014-03-10 06:01:13 +01:00
|
|
|
Status s = WriteLevel0Table(mems, edit, &file_number, log_buffer);
|
2012-10-19 23:00:53 +02:00
|
|
|
|
2011-06-22 04:36:45 +02:00
|
|
|
if (s.ok() && shutting_down_.Acquire_Load()) {
|
2014-02-12 20:42:54 +01:00
|
|
|
s = Status::ShutdownInProgress(
|
2014-03-20 21:10:02 +01:00
|
|
|
"Database shutdown started during memtable compaction");
|
2011-06-22 04:36:45 +02:00
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
|
2014-02-12 20:42:54 +01:00
|
|
|
if (!s.ok()) {
|
|
|
|
imm_.RollbackMemtableFlush(mems, file_number, &pending_outputs_);
|
2014-03-20 21:10:02 +01:00
|
|
|
} else {
|
|
|
|
// Replace immutable memtable with the generated Table
|
|
|
|
s = imm_.InstallMemtableFlushResults(
|
|
|
|
mems, versions_.get(), &mutex_, options_.info_log.get(), file_number,
|
|
|
|
pending_outputs_, &deletion_state.memtables_to_free,
|
|
|
|
db_directory_.get());
|
2014-02-12 20:42:54 +01:00
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
if (s.ok()) {
|
2013-12-20 18:57:58 +01:00
|
|
|
InstallSuperVersion(deletion_state);
|
2012-10-19 23:00:53 +02:00
|
|
|
if (madeProgress) {
|
|
|
|
*madeProgress = 1;
|
|
|
|
}
|
2013-07-16 20:56:46 +02:00
|
|
|
|
2012-08-15 00:20:36 +02:00
|
|
|
MaybeScheduleLogDBDeployStats();
|
2013-11-09 00:23:46 +01:00
|
|
|
|
2014-01-02 12:33:42 +01:00
|
|
|
if (disable_delete_obsolete_files_ == 0) {
|
2013-11-09 00:23:46 +01:00
|
|
|
// add to deletion state
|
2013-11-12 20:53:26 +01:00
|
|
|
deletion_state.log_delete_files.insert(
|
|
|
|
deletion_state.log_delete_files.end(),
|
|
|
|
logs_to_delete.begin(),
|
|
|
|
logs_to_delete.end());
|
2013-07-16 20:56:46 +02:00
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
2014-03-20 21:10:02 +01:00
|
|
|
|
|
|
|
if (!s.ok() && !s.IsShutdownInProgress() && options_.paranoid_checks &&
|
|
|
|
bg_error_.ok()) {
|
|
|
|
// if a bad error happened (not ShutdownInProgress) and paranoid_checks is
|
|
|
|
// true, mark DB read-only
|
|
|
|
bg_error_ = s;
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2014-01-22 21:46:24 +01:00
|
|
|
Status DBImpl::CompactRange(const Slice* begin,
|
|
|
|
const Slice* end,
|
|
|
|
bool reduce_level,
|
|
|
|
int target_level) {
|
|
|
|
Status s = FlushMemTable(FlushOptions());
|
|
|
|
if (!s.ok()) {
|
|
|
|
LogFlush(options_.info_log);
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2011-10-06 01:30:28 +02:00
|
|
|
int max_level_with_files = 1;
|
|
|
|
{
|
|
|
|
MutexLock l(&mutex_);
|
|
|
|
Version* base = versions_->current();
|
2012-06-23 04:30:03 +02:00
|
|
|
for (int level = 1; level < NumberLevels(); level++) {
|
2011-10-06 01:30:28 +02:00
|
|
|
if (base->OverlapInLevel(level, begin, end)) {
|
|
|
|
max_level_with_files = level;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2014-01-15 01:19:09 +01:00
|
|
|
for (int level = 0; level <= max_level_with_files; level++) {
|
|
|
|
// in case the compaction is unversal or if we're compacting the
|
|
|
|
// bottom-most level, the output level will be the same as input one
|
|
|
|
if (options_.compaction_style == kCompactionStyleUniversal ||
|
|
|
|
level == max_level_with_files) {
|
2014-01-22 21:46:24 +01:00
|
|
|
s = RunManualCompaction(level, level, begin, end);
|
2014-01-15 01:19:09 +01:00
|
|
|
} else {
|
2014-01-22 21:46:24 +01:00
|
|
|
s = RunManualCompaction(level, level + 1, begin, end);
|
|
|
|
}
|
|
|
|
if (!s.ok()) {
|
|
|
|
LogFlush(options_.info_log);
|
|
|
|
return s;
|
2014-01-15 01:19:09 +01:00
|
|
|
}
|
2011-10-06 01:30:28 +02:00
|
|
|
}
|
2013-06-30 08:21:36 +02:00
|
|
|
|
|
|
|
if (reduce_level) {
|
2014-01-22 21:46:24 +01:00
|
|
|
s = ReFitLevel(max_level_with_files, target_level);
|
2013-06-30 08:21:36 +02:00
|
|
|
}
|
2013-11-07 20:31:56 +01:00
|
|
|
LogFlush(options_.info_log);
|
2014-01-22 21:46:24 +01:00
|
|
|
|
|
|
|
return s;
|
2013-06-30 08:21:36 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
// return the same level if it cannot be moved
|
|
|
|
int DBImpl::FindMinimumEmptyLevelFitting(int level) {
|
|
|
|
mutex_.AssertHeld();
|
2014-01-16 01:18:04 +01:00
|
|
|
Version* current = versions_->current();
|
2013-06-30 08:21:36 +02:00
|
|
|
int minimum_level = level;
|
2013-07-22 21:19:46 +02:00
|
|
|
for (int i = level - 1; i > 0; --i) {
|
2013-06-30 08:21:36 +02:00
|
|
|
// stop if level i is not empty
|
2014-01-16 01:18:04 +01:00
|
|
|
if (current->NumLevelFiles(i) > 0) break;
|
2013-06-30 08:21:36 +02:00
|
|
|
// stop if level i is too small (cannot fit the level files)
|
2014-01-16 01:18:04 +01:00
|
|
|
if (versions_->MaxBytesForLevel(i) < current->NumLevelBytes(level)) break;
|
2013-06-30 08:21:36 +02:00
|
|
|
|
|
|
|
minimum_level = i;
|
|
|
|
}
|
|
|
|
return minimum_level;
|
|
|
|
}
|
|
|
|
|
2014-01-22 21:46:24 +01:00
|
|
|
Status DBImpl::ReFitLevel(int level, int target_level) {
|
2013-06-30 08:21:36 +02:00
|
|
|
assert(level < NumberLevels());
|
|
|
|
|
2013-12-20 18:57:58 +01:00
|
|
|
SuperVersion* superversion_to_free = nullptr;
|
2014-01-02 20:26:57 +01:00
|
|
|
SuperVersion* new_superversion = new SuperVersion();
|
2013-12-20 18:57:58 +01:00
|
|
|
|
|
|
|
mutex_.Lock();
|
2013-06-30 08:21:36 +02:00
|
|
|
|
|
|
|
// only allow one thread refitting
|
|
|
|
if (refitting_level_) {
|
2013-12-20 18:57:58 +01:00
|
|
|
mutex_.Unlock();
|
2013-06-30 08:21:36 +02:00
|
|
|
Log(options_.info_log, "ReFitLevel: another thread is refitting");
|
2013-12-20 18:57:58 +01:00
|
|
|
delete new_superversion;
|
2014-01-22 21:46:24 +01:00
|
|
|
return Status::NotSupported("another thread is refitting");
|
2013-06-30 08:21:36 +02:00
|
|
|
}
|
|
|
|
refitting_level_ = true;
|
|
|
|
|
|
|
|
// wait for all background threads to stop
|
|
|
|
bg_work_gate_closed_ = true;
|
2013-09-13 23:38:37 +02:00
|
|
|
while (bg_compaction_scheduled_ > 0 || bg_flush_scheduled_) {
|
2013-06-30 08:21:36 +02:00
|
|
|
Log(options_.info_log,
|
2013-09-13 23:38:37 +02:00
|
|
|
"RefitLevel: waiting for background threads to stop: %d %d",
|
|
|
|
bg_compaction_scheduled_, bg_flush_scheduled_);
|
2013-06-30 08:21:36 +02:00
|
|
|
bg_cv_.Wait();
|
|
|
|
}
|
|
|
|
|
|
|
|
// move to a smaller level
|
2013-09-04 22:13:08 +02:00
|
|
|
int to_level = target_level;
|
|
|
|
if (target_level < 0) {
|
|
|
|
to_level = FindMinimumEmptyLevelFitting(level);
|
|
|
|
}
|
2013-06-30 08:21:36 +02:00
|
|
|
|
|
|
|
assert(to_level <= level);
|
|
|
|
|
2014-01-22 21:46:24 +01:00
|
|
|
Status status;
|
2013-06-30 08:21:36 +02:00
|
|
|
if (to_level < level) {
|
|
|
|
Log(options_.info_log, "Before refitting:\n%s",
|
|
|
|
versions_->current()->DebugString().data());
|
|
|
|
|
2014-01-15 00:27:09 +01:00
|
|
|
VersionEdit edit;
|
2013-06-30 08:21:36 +02:00
|
|
|
for (const auto& f : versions_->current()->files_[level]) {
|
|
|
|
edit.DeleteFile(level, f->number);
|
2013-08-02 19:22:08 +02:00
|
|
|
edit.AddFile(to_level, f->number, f->file_size, f->smallest, f->largest,
|
|
|
|
f->smallest_seqno, f->largest_seqno);
|
2013-06-30 08:21:36 +02:00
|
|
|
}
|
|
|
|
Log(options_.info_log, "Apply version edit:\n%s",
|
|
|
|
edit.DebugString().data());
|
|
|
|
|
2014-01-27 20:02:21 +01:00
|
|
|
status = versions_->LogAndApply(&edit, &mutex_, db_directory_.get());
|
2013-12-20 18:57:58 +01:00
|
|
|
superversion_to_free = InstallSuperVersion(new_superversion);
|
|
|
|
new_superversion = nullptr;
|
2013-06-30 08:21:36 +02:00
|
|
|
|
|
|
|
Log(options_.info_log, "LogAndApply: %s\n", status.ToString().data());
|
|
|
|
|
|
|
|
if (status.ok()) {
|
|
|
|
Log(options_.info_log, "After refitting:\n%s",
|
|
|
|
versions_->current()->DebugString().data());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
refitting_level_ = false;
|
|
|
|
bg_work_gate_closed_ = false;
|
2013-12-20 18:57:58 +01:00
|
|
|
|
|
|
|
mutex_.Unlock();
|
|
|
|
delete superversion_to_free;
|
|
|
|
delete new_superversion;
|
2014-01-22 21:46:24 +01:00
|
|
|
return status;
|
2011-10-06 01:30:28 +02:00
|
|
|
}
|
|
|
|
|
2012-06-23 04:30:03 +02:00
|
|
|
int DBImpl::NumberLevels() {
|
2012-08-15 00:20:36 +02:00
|
|
|
return options_.num_levels;
|
2012-06-23 04:30:03 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
int DBImpl::MaxMemCompactionLevel() {
|
2012-08-15 00:20:36 +02:00
|
|
|
return options_.max_mem_compaction_level;
|
2012-06-23 04:30:03 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
int DBImpl::Level0StopWriteTrigger() {
|
2012-08-15 00:20:36 +02:00
|
|
|
return options_.level0_stop_writes_trigger;
|
2012-06-23 04:30:03 +02:00
|
|
|
}
|
|
|
|
|
2014-01-17 06:56:26 +01:00
|
|
|
uint64_t DBImpl::CurrentVersionNumber() const {
|
|
|
|
return super_version_number_.load();
|
|
|
|
}
|
|
|
|
|
2012-07-06 20:42:09 +02:00
|
|
|
Status DBImpl::Flush(const FlushOptions& options) {
|
2013-12-27 21:23:17 +01:00
|
|
|
return FlushMemTable(options);
|
2012-07-06 20:42:09 +02:00
|
|
|
}
|
|
|
|
|
2013-10-25 04:09:02 +02:00
|
|
|
SequenceNumber DBImpl::GetLatestSequenceNumber() const {
|
2012-12-11 00:37:00 +01:00
|
|
|
return versions_->LastSequence();
|
|
|
|
}
|
|
|
|
|
2014-02-28 20:50:36 +01:00
|
|
|
Status DBImpl::GetUpdatesSince(
|
|
|
|
SequenceNumber seq, unique_ptr<TransactionLogIterator>* iter,
|
|
|
|
const TransactionLogIterator::ReadOptions& read_options) {
|
2012-11-30 02:28:37 +01:00
|
|
|
|
2013-11-22 23:14:05 +01:00
|
|
|
RecordTick(options_.statistics.get(), GET_UPDATES_SINCE_CALLS);
|
2013-10-25 04:09:02 +02:00
|
|
|
if (seq > versions_->LastSequence()) {
|
2014-02-12 20:42:54 +01:00
|
|
|
return Status::NotFound(
|
|
|
|
"Requested sequence not yet written in the db");
|
2013-08-06 21:54:37 +02:00
|
|
|
}
|
|
|
|
// Get all sorted Wal Files.
|
2012-11-30 02:28:37 +01:00
|
|
|
// Do binary search and open files and find the seq number.
|
|
|
|
|
2013-08-06 21:54:37 +02:00
|
|
|
std::unique_ptr<VectorLogPtr> wal_files(new VectorLogPtr);
|
|
|
|
Status s = GetSortedWalFiles(*wal_files);
|
2012-11-30 02:28:37 +01:00
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2013-08-06 21:54:37 +02:00
|
|
|
s = RetainProbableWalFiles(*wal_files, seq);
|
2012-12-10 20:02:07 +01:00
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
2012-11-30 02:28:37 +01:00
|
|
|
}
|
2014-02-28 20:50:36 +01:00
|
|
|
iter->reset(new TransactionLogIteratorImpl(options_.wal_dir, &options_,
|
|
|
|
read_options, storage_options_,
|
|
|
|
seq, std::move(wal_files), this));
|
2013-10-14 00:28:24 +02:00
|
|
|
return (*iter)->status();
|
2012-11-30 02:28:37 +01:00
|
|
|
}
|
|
|
|
|
2013-08-06 21:54:37 +02:00
|
|
|
Status DBImpl::RetainProbableWalFiles(VectorLogPtr& all_logs,
|
|
|
|
const SequenceNumber target) {
|
2012-12-10 20:02:07 +01:00
|
|
|
long start = 0; // signed to avoid overflow when target is < first file.
|
2013-08-06 21:54:37 +02:00
|
|
|
long end = static_cast<long>(all_logs.size()) - 1;
|
2012-11-30 02:28:37 +01:00
|
|
|
// Binary Search. avoid opening all files.
|
2012-12-10 20:02:07 +01:00
|
|
|
while (end >= start) {
|
|
|
|
long mid = start + (end - start) / 2; // Avoid overflow.
|
2013-08-06 21:54:37 +02:00
|
|
|
SequenceNumber current_seq_num = all_logs.at(mid)->StartSequence();
|
|
|
|
if (current_seq_num == target) {
|
2012-11-30 02:28:37 +01:00
|
|
|
end = mid;
|
2012-12-10 20:02:07 +01:00
|
|
|
break;
|
2013-08-06 21:54:37 +02:00
|
|
|
} else if (current_seq_num < target) {
|
2012-12-10 20:02:07 +01:00
|
|
|
start = mid + 1;
|
2012-11-30 02:28:37 +01:00
|
|
|
} else {
|
2012-12-10 20:02:07 +01:00
|
|
|
end = mid - 1;
|
2012-11-30 02:28:37 +01:00
|
|
|
}
|
|
|
|
}
|
2013-08-06 21:54:37 +02:00
|
|
|
size_t start_index = std::max(0l, end); // end could be -ve.
|
|
|
|
// The last wal file is always included
|
|
|
|
all_logs.erase(all_logs.begin(), all_logs.begin() + start_index);
|
2012-11-30 02:28:37 +01:00
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
2013-08-06 21:54:37 +02:00
|
|
|
bool DBImpl::CheckWalFileExistsAndEmpty(const WalFileType type,
|
|
|
|
const uint64_t number) {
|
|
|
|
const std::string fname = (type == kAliveLogFile) ?
|
2013-10-01 23:46:52 +02:00
|
|
|
LogFileName(options_.wal_dir, number) :
|
|
|
|
ArchivedLogFileName(options_.wal_dir, number);
|
2013-03-18 22:50:59 +01:00
|
|
|
uint64_t file_size;
|
|
|
|
Status s = env_->GetFileSize(fname, &file_size);
|
2013-08-06 21:54:37 +02:00
|
|
|
return (s.ok() && (file_size == 0));
|
2013-03-18 22:50:59 +01:00
|
|
|
}
|
|
|
|
|
2013-08-06 21:54:37 +02:00
|
|
|
Status DBImpl::ReadFirstRecord(const WalFileType type, const uint64_t number,
|
|
|
|
WriteBatch* const result) {
|
2012-11-30 02:28:37 +01:00
|
|
|
|
2013-08-06 21:54:37 +02:00
|
|
|
if (type == kAliveLogFile) {
|
2013-10-01 23:46:52 +02:00
|
|
|
std::string fname = LogFileName(options_.wal_dir, number);
|
2012-11-30 02:28:37 +01:00
|
|
|
Status status = ReadFirstLine(fname, result);
|
2014-02-12 20:42:54 +01:00
|
|
|
if (status.ok() || env_->FileExists(fname)) {
|
|
|
|
// return OK or any error that is not caused non-existing file
|
|
|
|
return status;
|
2012-11-30 02:28:37 +01:00
|
|
|
}
|
2014-02-12 20:42:54 +01:00
|
|
|
|
|
|
|
// check if the file got moved to archive.
|
|
|
|
std::string archived_file =
|
|
|
|
ArchivedLogFileName(options_.wal_dir, number);
|
|
|
|
Status s = ReadFirstLine(archived_file, result);
|
|
|
|
if (s.ok() || env_->FileExists(archived_file)) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
return Status::NotFound("Log File has been deleted: " + archived_file);
|
2013-08-06 21:54:37 +02:00
|
|
|
} else if (type == kArchivedLogFile) {
|
2013-10-01 23:46:52 +02:00
|
|
|
std::string fname = ArchivedLogFileName(options_.wal_dir, number);
|
2012-11-30 02:28:37 +01:00
|
|
|
Status status = ReadFirstLine(fname, result);
|
|
|
|
return status;
|
|
|
|
}
|
2013-11-13 05:05:28 +01:00
|
|
|
return Status::NotSupported("File Type Not Known: " + std::to_string(type));
|
2012-11-30 02:28:37 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
Status DBImpl::ReadFirstLine(const std::string& fname,
|
|
|
|
WriteBatch* const batch) {
|
|
|
|
struct LogReporter : public log::Reader::Reporter {
|
|
|
|
Env* env;
|
|
|
|
Logger* info_log;
|
|
|
|
const char* fname;
|
2014-02-12 20:42:54 +01:00
|
|
|
|
|
|
|
Status* status;
|
|
|
|
bool ignore_error; // true if options_.paranoid_checks==false
|
2012-11-30 02:28:37 +01:00
|
|
|
virtual void Corruption(size_t bytes, const Status& s) {
|
|
|
|
Log(info_log, "%s%s: dropping %d bytes; %s",
|
2014-02-12 20:42:54 +01:00
|
|
|
(this->ignore_error ? "(ignoring error) " : ""),
|
2012-11-30 02:28:37 +01:00
|
|
|
fname, static_cast<int>(bytes), s.ToString().c_str());
|
2014-02-12 20:42:54 +01:00
|
|
|
if (this->status->ok()) {
|
|
|
|
// only keep the first error
|
|
|
|
*this->status = s;
|
|
|
|
}
|
2012-11-30 02:28:37 +01:00
|
|
|
}
|
|
|
|
};
|
|
|
|
|
2013-01-20 11:07:13 +01:00
|
|
|
unique_ptr<SequentialFile> file;
|
2013-03-15 01:00:04 +01:00
|
|
|
Status status = env_->NewSequentialFile(fname, &file, storage_options_);
|
2012-11-30 02:28:37 +01:00
|
|
|
|
|
|
|
if (!status.ok()) {
|
|
|
|
return status;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
LogReporter reporter;
|
|
|
|
reporter.env = env_;
|
2013-01-20 11:07:13 +01:00
|
|
|
reporter.info_log = options_.info_log.get();
|
2012-11-30 02:28:37 +01:00
|
|
|
reporter.fname = fname.c_str();
|
2014-02-12 20:42:54 +01:00
|
|
|
reporter.status = &status;
|
|
|
|
reporter.ignore_error = !options_.paranoid_checks;
|
2013-01-20 11:07:13 +01:00
|
|
|
log::Reader reader(std::move(file), &reporter, true/*checksum*/,
|
2012-11-30 02:28:37 +01:00
|
|
|
0/*initial_offset*/);
|
|
|
|
std::string scratch;
|
|
|
|
Slice record;
|
2013-08-06 21:54:37 +02:00
|
|
|
|
2014-02-12 20:42:54 +01:00
|
|
|
if (reader.ReadRecord(&record, &scratch) &&
|
|
|
|
(status.ok() || !options_.paranoid_checks)) {
|
2012-11-30 02:28:37 +01:00
|
|
|
if (record.size() < 12) {
|
|
|
|
reporter.Corruption(
|
|
|
|
record.size(), Status::Corruption("log record too small"));
|
|
|
|
// TODO read record's till the first no corrupt entry?
|
2014-02-12 20:42:54 +01:00
|
|
|
} else {
|
|
|
|
WriteBatchInternal::SetContents(batch, record);
|
|
|
|
return Status::OK();
|
2012-11-30 02:28:37 +01:00
|
|
|
}
|
|
|
|
}
|
2014-02-12 20:42:54 +01:00
|
|
|
|
|
|
|
// ReadRecord returns false on EOF, which is deemed as OK() by Reader
|
|
|
|
if (status.ok()) {
|
|
|
|
status = Status::Corruption("eof reached");
|
|
|
|
}
|
|
|
|
return status;
|
2012-11-30 02:28:37 +01:00
|
|
|
}
|
|
|
|
|
2013-08-06 21:54:37 +02:00
|
|
|
struct CompareLogByPointer {
|
|
|
|
bool operator() (const unique_ptr<LogFile>& a,
|
|
|
|
const unique_ptr<LogFile>& b) {
|
|
|
|
LogFileImpl* a_impl = dynamic_cast<LogFileImpl*>(a.get());
|
|
|
|
LogFileImpl* b_impl = dynamic_cast<LogFileImpl*>(b.get());
|
|
|
|
return *a_impl < *b_impl;
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
Status DBImpl::AppendSortedWalsOfType(const std::string& path,
|
|
|
|
VectorLogPtr& log_files, WalFileType log_type) {
|
|
|
|
std::vector<std::string> all_files;
|
|
|
|
const Status status = env_->GetChildren(path, &all_files);
|
2012-11-30 02:28:37 +01:00
|
|
|
if (!status.ok()) {
|
|
|
|
return status;
|
|
|
|
}
|
2013-08-06 21:54:37 +02:00
|
|
|
log_files.reserve(log_files.size() + all_files.size());
|
2013-08-29 23:30:52 +02:00
|
|
|
VectorLogPtr::iterator pos_start;
|
|
|
|
if (!log_files.empty()) {
|
|
|
|
pos_start = log_files.end() - 1;
|
|
|
|
} else {
|
|
|
|
pos_start = log_files.begin();
|
|
|
|
}
|
2013-08-06 21:54:37 +02:00
|
|
|
for (const auto& f : all_files) {
|
2012-11-30 02:28:37 +01:00
|
|
|
uint64_t number;
|
|
|
|
FileType type;
|
2013-03-29 02:01:06 +01:00
|
|
|
if (ParseFileName(f, &number, &type) && type == kLogFile){
|
2013-08-06 21:54:37 +02:00
|
|
|
|
|
|
|
WriteBatch batch;
|
|
|
|
Status s = ReadFirstRecord(log_type, number, &batch);
|
|
|
|
if (!s.ok()) {
|
|
|
|
if (CheckWalFileExistsAndEmpty(log_type, number)) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t size_bytes;
|
|
|
|
s = env_->GetFileSize(LogFileName(path, number), &size_bytes);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
log_files.push_back(std::move(unique_ptr<LogFile>(new LogFileImpl(
|
|
|
|
number, log_type, WriteBatchInternal::Sequence(&batch), size_bytes))));
|
2012-11-30 02:28:37 +01:00
|
|
|
}
|
|
|
|
}
|
2013-08-06 21:54:37 +02:00
|
|
|
CompareLogByPointer compare_log_files;
|
2013-08-29 23:30:52 +02:00
|
|
|
std::sort(pos_start, log_files.end(), compare_log_files);
|
2012-11-30 02:28:37 +01:00
|
|
|
return status;
|
|
|
|
}
|
|
|
|
|
2014-01-22 21:46:24 +01:00
|
|
|
Status DBImpl::RunManualCompaction(int input_level,
|
|
|
|
int output_level,
|
|
|
|
const Slice* begin,
|
|
|
|
const Slice* end) {
|
2014-01-15 01:19:09 +01:00
|
|
|
assert(input_level >= 0);
|
2011-06-22 04:36:45 +02:00
|
|
|
|
2011-10-06 01:30:28 +02:00
|
|
|
InternalKey begin_storage, end_storage;
|
|
|
|
|
2011-06-07 16:40:26 +02:00
|
|
|
ManualCompaction manual;
|
2014-01-15 01:19:09 +01:00
|
|
|
manual.input_level = input_level;
|
|
|
|
manual.output_level = output_level;
|
2011-10-06 01:30:28 +02:00
|
|
|
manual.done = false;
|
2012-10-19 23:00:53 +02:00
|
|
|
manual.in_progress = false;
|
2013-08-08 00:20:41 +02:00
|
|
|
// For universal compaction, we enforce every manual compaction to compact
|
|
|
|
// all files.
|
|
|
|
if (begin == nullptr ||
|
|
|
|
options_.compaction_style == kCompactionStyleUniversal) {
|
2013-02-15 20:53:17 +01:00
|
|
|
manual.begin = nullptr;
|
2011-10-06 01:30:28 +02:00
|
|
|
} else {
|
|
|
|
begin_storage = InternalKey(*begin, kMaxSequenceNumber, kValueTypeForSeek);
|
|
|
|
manual.begin = &begin_storage;
|
|
|
|
}
|
2013-08-08 00:20:41 +02:00
|
|
|
if (end == nullptr ||
|
|
|
|
options_.compaction_style == kCompactionStyleUniversal) {
|
2013-02-15 20:53:17 +01:00
|
|
|
manual.end = nullptr;
|
2011-10-06 01:30:28 +02:00
|
|
|
} else {
|
|
|
|
end_storage = InternalKey(*end, 0, static_cast<ValueType>(0));
|
|
|
|
manual.end = &end_storage;
|
|
|
|
}
|
|
|
|
|
|
|
|
MutexLock l(&mutex_);
|
2012-10-19 23:00:53 +02:00
|
|
|
|
Fix a deadlock in CompactRange()
Summary:
The way DBImpl::TEST_CompactRange() throttles down the number of bg compactions
can cause it to deadlock when CompactRange() is called concurrently from
multiple threads. Imagine a following scenario with only two threads
(max_background_compactions is 10 and bg_compaction_scheduled_ is initially 0):
1. Thread #1 increments bg_compaction_scheduled_ (to LargeNumber), sets
bg_compaction_scheduled_ to 9 (newvalue), schedules the compaction
(bg_compaction_scheduled_ is now 10) and waits for it to complete.
2. Thread #2 calls TEST_CompactRange(), increments bg_compaction_scheduled_
(now LargeNumber + 10) and waits on a cv for bg_compaction_scheduled_ to
drop to LargeNumber.
3. BG thread completes the first manual compaction, decrements
bg_compaction_scheduled_ and wakes up all threads waiting on bg_cv_.
Thread #1 runs, increments bg_compaction_scheduled_ by LargeNumber again
(now 2*LargeNumber + 9). Since that's more than LargeNumber + newvalue,
thread #2 also goes to sleep (waiting on bg_cv_), without resetting
bg_compaction_scheduled_.
This diff attempts to address the problem by introducing a new counter
bg_manual_only_ (when positive, MaybeScheduleFlushOrCompaction() will only
schedule manual compactions).
Test Plan:
I could pretty much consistently reproduce the deadlock with a program that
calls CompactRange(nullptr, nullptr) immediately after Write() from multiple
threads. This no longer happens with this patch.
Tests (make check) pass.
Reviewers: dhruba, igor, sdong, haobo
Reviewed By: igor
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14799
2013-12-22 00:10:39 +01:00
|
|
|
// When a manual compaction arrives, temporarily disable scheduling of
|
|
|
|
// non-manual compactions and wait until the number of scheduled compaction
|
|
|
|
// jobs drops to zero. This is needed to ensure that this manual compaction
|
|
|
|
// can compact any range of keys/files.
|
|
|
|
//
|
|
|
|
// bg_manual_only_ is non-zero when at least one thread is inside
|
2014-01-15 01:19:09 +01:00
|
|
|
// RunManualCompaction(), i.e. during that time no other compaction will
|
Fix a deadlock in CompactRange()
Summary:
The way DBImpl::TEST_CompactRange() throttles down the number of bg compactions
can cause it to deadlock when CompactRange() is called concurrently from
multiple threads. Imagine a following scenario with only two threads
(max_background_compactions is 10 and bg_compaction_scheduled_ is initially 0):
1. Thread #1 increments bg_compaction_scheduled_ (to LargeNumber), sets
bg_compaction_scheduled_ to 9 (newvalue), schedules the compaction
(bg_compaction_scheduled_ is now 10) and waits for it to complete.
2. Thread #2 calls TEST_CompactRange(), increments bg_compaction_scheduled_
(now LargeNumber + 10) and waits on a cv for bg_compaction_scheduled_ to
drop to LargeNumber.
3. BG thread completes the first manual compaction, decrements
bg_compaction_scheduled_ and wakes up all threads waiting on bg_cv_.
Thread #1 runs, increments bg_compaction_scheduled_ by LargeNumber again
(now 2*LargeNumber + 9). Since that's more than LargeNumber + newvalue,
thread #2 also goes to sleep (waiting on bg_cv_), without resetting
bg_compaction_scheduled_.
This diff attempts to address the problem by introducing a new counter
bg_manual_only_ (when positive, MaybeScheduleFlushOrCompaction() will only
schedule manual compactions).
Test Plan:
I could pretty much consistently reproduce the deadlock with a program that
calls CompactRange(nullptr, nullptr) immediately after Write() from multiple
threads. This no longer happens with this patch.
Tests (make check) pass.
Reviewers: dhruba, igor, sdong, haobo
Reviewed By: igor
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14799
2013-12-22 00:10:39 +01:00
|
|
|
// get scheduled (see MaybeScheduleFlushOrCompaction).
|
|
|
|
//
|
|
|
|
// Note that the following loop doesn't stop more that one thread calling
|
2014-01-15 01:19:09 +01:00
|
|
|
// RunManualCompaction() from getting to the second while loop below.
|
Fix a deadlock in CompactRange()
Summary:
The way DBImpl::TEST_CompactRange() throttles down the number of bg compactions
can cause it to deadlock when CompactRange() is called concurrently from
multiple threads. Imagine a following scenario with only two threads
(max_background_compactions is 10 and bg_compaction_scheduled_ is initially 0):
1. Thread #1 increments bg_compaction_scheduled_ (to LargeNumber), sets
bg_compaction_scheduled_ to 9 (newvalue), schedules the compaction
(bg_compaction_scheduled_ is now 10) and waits for it to complete.
2. Thread #2 calls TEST_CompactRange(), increments bg_compaction_scheduled_
(now LargeNumber + 10) and waits on a cv for bg_compaction_scheduled_ to
drop to LargeNumber.
3. BG thread completes the first manual compaction, decrements
bg_compaction_scheduled_ and wakes up all threads waiting on bg_cv_.
Thread #1 runs, increments bg_compaction_scheduled_ by LargeNumber again
(now 2*LargeNumber + 9). Since that's more than LargeNumber + newvalue,
thread #2 also goes to sleep (waiting on bg_cv_), without resetting
bg_compaction_scheduled_.
This diff attempts to address the problem by introducing a new counter
bg_manual_only_ (when positive, MaybeScheduleFlushOrCompaction() will only
schedule manual compactions).
Test Plan:
I could pretty much consistently reproduce the deadlock with a program that
calls CompactRange(nullptr, nullptr) immediately after Write() from multiple
threads. This no longer happens with this patch.
Tests (make check) pass.
Reviewers: dhruba, igor, sdong, haobo
Reviewed By: igor
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14799
2013-12-22 00:10:39 +01:00
|
|
|
// However, only one of them will actually schedule compaction, while
|
|
|
|
// others will wait on a condition variable until it completes.
|
|
|
|
|
|
|
|
++bg_manual_only_;
|
|
|
|
while (bg_compaction_scheduled_ > 0) {
|
|
|
|
Log(options_.info_log,
|
|
|
|
"Manual compaction waiting for all other scheduled background "
|
|
|
|
"compactions to finish");
|
2012-10-19 23:00:53 +02:00
|
|
|
bg_cv_.Wait();
|
|
|
|
}
|
Fix a deadlock in CompactRange()
Summary:
The way DBImpl::TEST_CompactRange() throttles down the number of bg compactions
can cause it to deadlock when CompactRange() is called concurrently from
multiple threads. Imagine a following scenario with only two threads
(max_background_compactions is 10 and bg_compaction_scheduled_ is initially 0):
1. Thread #1 increments bg_compaction_scheduled_ (to LargeNumber), sets
bg_compaction_scheduled_ to 9 (newvalue), schedules the compaction
(bg_compaction_scheduled_ is now 10) and waits for it to complete.
2. Thread #2 calls TEST_CompactRange(), increments bg_compaction_scheduled_
(now LargeNumber + 10) and waits on a cv for bg_compaction_scheduled_ to
drop to LargeNumber.
3. BG thread completes the first manual compaction, decrements
bg_compaction_scheduled_ and wakes up all threads waiting on bg_cv_.
Thread #1 runs, increments bg_compaction_scheduled_ by LargeNumber again
(now 2*LargeNumber + 9). Since that's more than LargeNumber + newvalue,
thread #2 also goes to sleep (waiting on bg_cv_), without resetting
bg_compaction_scheduled_.
This diff attempts to address the problem by introducing a new counter
bg_manual_only_ (when positive, MaybeScheduleFlushOrCompaction() will only
schedule manual compactions).
Test Plan:
I could pretty much consistently reproduce the deadlock with a program that
calls CompactRange(nullptr, nullptr) immediately after Write() from multiple
threads. This no longer happens with this patch.
Tests (make check) pass.
Reviewers: dhruba, igor, sdong, haobo
Reviewed By: igor
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14799
2013-12-22 00:10:39 +01:00
|
|
|
|
2012-10-19 23:00:53 +02:00
|
|
|
Log(options_.info_log, "Manual compaction starting");
|
|
|
|
|
Fix a deadlock in CompactRange()
Summary:
The way DBImpl::TEST_CompactRange() throttles down the number of bg compactions
can cause it to deadlock when CompactRange() is called concurrently from
multiple threads. Imagine a following scenario with only two threads
(max_background_compactions is 10 and bg_compaction_scheduled_ is initially 0):
1. Thread #1 increments bg_compaction_scheduled_ (to LargeNumber), sets
bg_compaction_scheduled_ to 9 (newvalue), schedules the compaction
(bg_compaction_scheduled_ is now 10) and waits for it to complete.
2. Thread #2 calls TEST_CompactRange(), increments bg_compaction_scheduled_
(now LargeNumber + 10) and waits on a cv for bg_compaction_scheduled_ to
drop to LargeNumber.
3. BG thread completes the first manual compaction, decrements
bg_compaction_scheduled_ and wakes up all threads waiting on bg_cv_.
Thread #1 runs, increments bg_compaction_scheduled_ by LargeNumber again
(now 2*LargeNumber + 9). Since that's more than LargeNumber + newvalue,
thread #2 also goes to sleep (waiting on bg_cv_), without resetting
bg_compaction_scheduled_.
This diff attempts to address the problem by introducing a new counter
bg_manual_only_ (when positive, MaybeScheduleFlushOrCompaction() will only
schedule manual compactions).
Test Plan:
I could pretty much consistently reproduce the deadlock with a program that
calls CompactRange(nullptr, nullptr) immediately after Write() from multiple
threads. This no longer happens with this patch.
Tests (make check) pass.
Reviewers: dhruba, igor, sdong, haobo
Reviewed By: igor
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14799
2013-12-22 00:10:39 +01:00
|
|
|
while (!manual.done && !shutting_down_.Acquire_Load() && bg_error_.ok()) {
|
|
|
|
assert(bg_manual_only_ > 0);
|
|
|
|
if (manual_compaction_ != nullptr) {
|
|
|
|
// Running either this or some other manual compaction
|
2011-10-06 01:30:28 +02:00
|
|
|
bg_cv_.Wait();
|
Fix a deadlock in CompactRange()
Summary:
The way DBImpl::TEST_CompactRange() throttles down the number of bg compactions
can cause it to deadlock when CompactRange() is called concurrently from
multiple threads. Imagine a following scenario with only two threads
(max_background_compactions is 10 and bg_compaction_scheduled_ is initially 0):
1. Thread #1 increments bg_compaction_scheduled_ (to LargeNumber), sets
bg_compaction_scheduled_ to 9 (newvalue), schedules the compaction
(bg_compaction_scheduled_ is now 10) and waits for it to complete.
2. Thread #2 calls TEST_CompactRange(), increments bg_compaction_scheduled_
(now LargeNumber + 10) and waits on a cv for bg_compaction_scheduled_ to
drop to LargeNumber.
3. BG thread completes the first manual compaction, decrements
bg_compaction_scheduled_ and wakes up all threads waiting on bg_cv_.
Thread #1 runs, increments bg_compaction_scheduled_ by LargeNumber again
(now 2*LargeNumber + 9). Since that's more than LargeNumber + newvalue,
thread #2 also goes to sleep (waiting on bg_cv_), without resetting
bg_compaction_scheduled_.
This diff attempts to address the problem by introducing a new counter
bg_manual_only_ (when positive, MaybeScheduleFlushOrCompaction() will only
schedule manual compactions).
Test Plan:
I could pretty much consistently reproduce the deadlock with a program that
calls CompactRange(nullptr, nullptr) immediately after Write() from multiple
threads. This no longer happens with this patch.
Tests (make check) pass.
Reviewers: dhruba, igor, sdong, haobo
Reviewed By: igor
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14799
2013-12-22 00:10:39 +01:00
|
|
|
} else {
|
|
|
|
manual_compaction_ = &manual;
|
|
|
|
MaybeScheduleFlushOrCompaction();
|
2011-10-06 01:30:28 +02:00
|
|
|
}
|
2011-06-07 16:40:26 +02:00
|
|
|
}
|
2012-10-19 23:00:53 +02:00
|
|
|
|
Fix a deadlock in CompactRange()
Summary:
The way DBImpl::TEST_CompactRange() throttles down the number of bg compactions
can cause it to deadlock when CompactRange() is called concurrently from
multiple threads. Imagine a following scenario with only two threads
(max_background_compactions is 10 and bg_compaction_scheduled_ is initially 0):
1. Thread #1 increments bg_compaction_scheduled_ (to LargeNumber), sets
bg_compaction_scheduled_ to 9 (newvalue), schedules the compaction
(bg_compaction_scheduled_ is now 10) and waits for it to complete.
2. Thread #2 calls TEST_CompactRange(), increments bg_compaction_scheduled_
(now LargeNumber + 10) and waits on a cv for bg_compaction_scheduled_ to
drop to LargeNumber.
3. BG thread completes the first manual compaction, decrements
bg_compaction_scheduled_ and wakes up all threads waiting on bg_cv_.
Thread #1 runs, increments bg_compaction_scheduled_ by LargeNumber again
(now 2*LargeNumber + 9). Since that's more than LargeNumber + newvalue,
thread #2 also goes to sleep (waiting on bg_cv_), without resetting
bg_compaction_scheduled_.
This diff attempts to address the problem by introducing a new counter
bg_manual_only_ (when positive, MaybeScheduleFlushOrCompaction() will only
schedule manual compactions).
Test Plan:
I could pretty much consistently reproduce the deadlock with a program that
calls CompactRange(nullptr, nullptr) immediately after Write() from multiple
threads. This no longer happens with this patch.
Tests (make check) pass.
Reviewers: dhruba, igor, sdong, haobo
Reviewed By: igor
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14799
2013-12-22 00:10:39 +01:00
|
|
|
assert(!manual.in_progress);
|
|
|
|
assert(bg_manual_only_ > 0);
|
|
|
|
--bg_manual_only_;
|
2014-01-22 21:46:24 +01:00
|
|
|
return manual.status;
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
2014-01-22 21:46:24 +01:00
|
|
|
Status DBImpl::TEST_CompactRange(int level,
|
|
|
|
const Slice* begin,
|
|
|
|
const Slice* end) {
|
2014-01-15 01:19:09 +01:00
|
|
|
int output_level = (options_.compaction_style == kCompactionStyleUniversal)
|
|
|
|
? level
|
|
|
|
: level + 1;
|
2014-01-22 21:46:24 +01:00
|
|
|
return RunManualCompaction(level, output_level, begin, end);
|
2014-01-15 01:19:09 +01:00
|
|
|
}
|
|
|
|
|
2012-07-06 20:42:09 +02:00
|
|
|
Status DBImpl::FlushMemTable(const FlushOptions& options) {
|
2013-02-15 20:53:17 +01:00
|
|
|
// nullptr batch means just wait for earlier writes to be done
|
|
|
|
Status s = Write(WriteOptions(), nullptr);
|
2012-07-06 20:42:09 +02:00
|
|
|
if (s.ok() && options.wait) {
|
2011-04-12 21:38:58 +02:00
|
|
|
// Wait until the compaction completes
|
2013-10-15 00:12:15 +02:00
|
|
|
s = WaitForFlushMemTable();
|
2011-04-12 21:38:58 +02:00
|
|
|
}
|
|
|
|
return s;
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
2013-10-15 00:12:15 +02:00
|
|
|
Status DBImpl::WaitForFlushMemTable() {
|
2012-08-15 00:20:36 +02:00
|
|
|
Status s;
|
|
|
|
// Wait until the compaction completes
|
|
|
|
MutexLock l(&mutex_);
|
2012-10-19 23:00:53 +02:00
|
|
|
while (imm_.size() > 0 && bg_error_.ok()) {
|
2012-08-15 00:20:36 +02:00
|
|
|
bg_cv_.Wait();
|
|
|
|
}
|
2012-10-19 23:00:53 +02:00
|
|
|
if (imm_.size() != 0) {
|
2012-08-15 00:20:36 +02:00
|
|
|
s = bg_error_;
|
|
|
|
}
|
|
|
|
return s;
|
2012-07-06 20:42:09 +02:00
|
|
|
}
|
|
|
|
|
2014-03-18 20:25:08 +01:00
|
|
|
Status DBImpl::TEST_FlushMemTable(bool wait) {
|
|
|
|
FlushOptions fo;
|
|
|
|
fo.wait = wait;
|
|
|
|
return FlushMemTable(fo);
|
2012-07-06 20:42:09 +02:00
|
|
|
}
|
|
|
|
|
2013-10-15 00:12:15 +02:00
|
|
|
Status DBImpl::TEST_WaitForFlushMemTable() {
|
|
|
|
return WaitForFlushMemTable();
|
2012-06-23 04:30:03 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
Status DBImpl::TEST_WaitForCompact() {
|
2012-08-15 00:20:36 +02:00
|
|
|
// Wait until the compaction completes
|
2013-10-15 00:12:15 +02:00
|
|
|
|
|
|
|
// TODO: a bug here. This function actually does not necessarily
|
|
|
|
// wait for compact. It actually waits for scheduled compaction
|
|
|
|
// OR flush to finish.
|
|
|
|
|
2012-08-15 00:20:36 +02:00
|
|
|
MutexLock l(&mutex_);
|
2013-09-13 23:38:37 +02:00
|
|
|
while ((bg_compaction_scheduled_ || bg_flush_scheduled_) &&
|
|
|
|
bg_error_.ok()) {
|
2012-08-15 00:20:36 +02:00
|
|
|
bg_cv_.Wait();
|
|
|
|
}
|
|
|
|
return bg_error_;
|
2012-06-23 04:30:03 +02:00
|
|
|
}
|
|
|
|
|
2013-10-15 00:12:15 +02:00
|
|
|
void DBImpl::MaybeScheduleFlushOrCompaction() {
|
2011-03-18 23:37:00 +01:00
|
|
|
mutex_.AssertHeld();
|
Fix data race against logging data structure because of LogBuffer
Summary:
@igor pointed out that there is a potential data race because of the way we use the newly introduced LogBuffer. After "bg_compaction_scheduled_--" or "bg_flush_scheduled_--", they can both become 0. As soon as the lock is released after that, DBImpl's deconstructor can go ahead and deconstruct all the states inside DB, including the info_log object hold in a shared pointer of the options object it keeps. At that point it is not safe anymore to continue using the info logger to write the delayed logs.
With the patch, lock is released temporarily for log buffer to be flushed before "bg_compaction_scheduled_--" or "bg_flush_scheduled_--". In order to make sure we don't miss any pending flush or compaction, a new flag bg_schedule_needed_ is added, which is set to be true if there is a pending flush or compaction but not scheduled because of the max thread limit. If the flag is set to be true, the scheduling function will be called before compaction or flush thread finishes.
Thanks @igor for this finding!
Test Plan: make all check
Reviewers: haobo, igor
Reviewed By: haobo
CC: dhruba, ljin, yhchiang, igor, leveldb
Differential Revision: https://reviews.facebook.net/D16767
2014-03-11 18:36:30 +01:00
|
|
|
bg_schedule_needed_ = false;
|
2013-06-30 08:21:36 +02:00
|
|
|
if (bg_work_gate_closed_) {
|
|
|
|
// gate closed for backgrond work
|
2011-03-18 23:37:00 +01:00
|
|
|
} else if (shutting_down_.Acquire_Load()) {
|
|
|
|
// DB is being deleted; no more background compactions
|
|
|
|
} else {
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
bool is_flush_pending = imm_.IsFlushPending();
|
Fix data race against logging data structure because of LogBuffer
Summary:
@igor pointed out that there is a potential data race because of the way we use the newly introduced LogBuffer. After "bg_compaction_scheduled_--" or "bg_flush_scheduled_--", they can both become 0. As soon as the lock is released after that, DBImpl's deconstructor can go ahead and deconstruct all the states inside DB, including the info_log object hold in a shared pointer of the options object it keeps. At that point it is not safe anymore to continue using the info logger to write the delayed logs.
With the patch, lock is released temporarily for log buffer to be flushed before "bg_compaction_scheduled_--" or "bg_flush_scheduled_--". In order to make sure we don't miss any pending flush or compaction, a new flag bg_schedule_needed_ is added, which is set to be true if there is a pending flush or compaction but not scheduled because of the max thread limit. If the flag is set to be true, the scheduling function will be called before compaction or flush thread finishes.
Thanks @igor for this finding!
Test Plan: make all check
Reviewers: haobo, igor
Reviewed By: haobo
CC: dhruba, ljin, yhchiang, igor, leveldb
Differential Revision: https://reviews.facebook.net/D16767
2014-03-11 18:36:30 +01:00
|
|
|
if (is_flush_pending) {
|
|
|
|
if (bg_flush_scheduled_ < options_.max_background_flushes) {
|
|
|
|
// memtable flush needed
|
|
|
|
bg_flush_scheduled_++;
|
|
|
|
env_->Schedule(&DBImpl::BGWorkFlush, this, Env::Priority::HIGH);
|
2014-03-13 19:21:03 +01:00
|
|
|
} else if (options_.max_background_flushes > 0) {
|
Fix data race against logging data structure because of LogBuffer
Summary:
@igor pointed out that there is a potential data race because of the way we use the newly introduced LogBuffer. After "bg_compaction_scheduled_--" or "bg_flush_scheduled_--", they can both become 0. As soon as the lock is released after that, DBImpl's deconstructor can go ahead and deconstruct all the states inside DB, including the info_log object hold in a shared pointer of the options object it keeps. At that point it is not safe anymore to continue using the info logger to write the delayed logs.
With the patch, lock is released temporarily for log buffer to be flushed before "bg_compaction_scheduled_--" or "bg_flush_scheduled_--". In order to make sure we don't miss any pending flush or compaction, a new flag bg_schedule_needed_ is added, which is set to be true if there is a pending flush or compaction but not scheduled because of the max thread limit. If the flag is set to be true, the scheduling function will be called before compaction or flush thread finishes.
Thanks @igor for this finding!
Test Plan: make all check
Reviewers: haobo, igor
Reviewed By: haobo
CC: dhruba, ljin, yhchiang, igor, leveldb
Differential Revision: https://reviews.facebook.net/D16767
2014-03-11 18:36:30 +01:00
|
|
|
bg_schedule_needed_ = true;
|
|
|
|
}
|
2013-09-13 23:38:37 +02:00
|
|
|
}
|
|
|
|
|
Fix a deadlock in CompactRange()
Summary:
The way DBImpl::TEST_CompactRange() throttles down the number of bg compactions
can cause it to deadlock when CompactRange() is called concurrently from
multiple threads. Imagine a following scenario with only two threads
(max_background_compactions is 10 and bg_compaction_scheduled_ is initially 0):
1. Thread #1 increments bg_compaction_scheduled_ (to LargeNumber), sets
bg_compaction_scheduled_ to 9 (newvalue), schedules the compaction
(bg_compaction_scheduled_ is now 10) and waits for it to complete.
2. Thread #2 calls TEST_CompactRange(), increments bg_compaction_scheduled_
(now LargeNumber + 10) and waits on a cv for bg_compaction_scheduled_ to
drop to LargeNumber.
3. BG thread completes the first manual compaction, decrements
bg_compaction_scheduled_ and wakes up all threads waiting on bg_cv_.
Thread #1 runs, increments bg_compaction_scheduled_ by LargeNumber again
(now 2*LargeNumber + 9). Since that's more than LargeNumber + newvalue,
thread #2 also goes to sleep (waiting on bg_cv_), without resetting
bg_compaction_scheduled_.
This diff attempts to address the problem by introducing a new counter
bg_manual_only_ (when positive, MaybeScheduleFlushOrCompaction() will only
schedule manual compactions).
Test Plan:
I could pretty much consistently reproduce the deadlock with a program that
calls CompactRange(nullptr, nullptr) immediately after Write() from multiple
threads. This no longer happens with this patch.
Tests (make check) pass.
Reviewers: dhruba, igor, sdong, haobo
Reviewed By: igor
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14799
2013-12-22 00:10:39 +01:00
|
|
|
// Schedule BGWorkCompaction if there's a compaction pending (or a memtable
|
|
|
|
// flush, but the HIGH pool is not enabled). Do it only if
|
|
|
|
// max_background_compactions hasn't been reached and, in case
|
|
|
|
// bg_manual_only_ > 0, if it's a manual compaction.
|
2013-09-13 23:38:37 +02:00
|
|
|
if ((manual_compaction_ ||
|
2014-01-27 18:59:00 +01:00
|
|
|
versions_->current()->NeedsCompaction() ||
|
2013-09-13 23:38:37 +02:00
|
|
|
(is_flush_pending && (options_.max_background_flushes <= 0))) &&
|
Fix a deadlock in CompactRange()
Summary:
The way DBImpl::TEST_CompactRange() throttles down the number of bg compactions
can cause it to deadlock when CompactRange() is called concurrently from
multiple threads. Imagine a following scenario with only two threads
(max_background_compactions is 10 and bg_compaction_scheduled_ is initially 0):
1. Thread #1 increments bg_compaction_scheduled_ (to LargeNumber), sets
bg_compaction_scheduled_ to 9 (newvalue), schedules the compaction
(bg_compaction_scheduled_ is now 10) and waits for it to complete.
2. Thread #2 calls TEST_CompactRange(), increments bg_compaction_scheduled_
(now LargeNumber + 10) and waits on a cv for bg_compaction_scheduled_ to
drop to LargeNumber.
3. BG thread completes the first manual compaction, decrements
bg_compaction_scheduled_ and wakes up all threads waiting on bg_cv_.
Thread #1 runs, increments bg_compaction_scheduled_ by LargeNumber again
(now 2*LargeNumber + 9). Since that's more than LargeNumber + newvalue,
thread #2 also goes to sleep (waiting on bg_cv_), without resetting
bg_compaction_scheduled_.
This diff attempts to address the problem by introducing a new counter
bg_manual_only_ (when positive, MaybeScheduleFlushOrCompaction() will only
schedule manual compactions).
Test Plan:
I could pretty much consistently reproduce the deadlock with a program that
calls CompactRange(nullptr, nullptr) immediately after Write() from multiple
threads. This no longer happens with this patch.
Tests (make check) pass.
Reviewers: dhruba, igor, sdong, haobo
Reviewed By: igor
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14799
2013-12-22 00:10:39 +01:00
|
|
|
(!bg_manual_only_ || manual_compaction_)) {
|
Fix data race against logging data structure because of LogBuffer
Summary:
@igor pointed out that there is a potential data race because of the way we use the newly introduced LogBuffer. After "bg_compaction_scheduled_--" or "bg_flush_scheduled_--", they can both become 0. As soon as the lock is released after that, DBImpl's deconstructor can go ahead and deconstruct all the states inside DB, including the info_log object hold in a shared pointer of the options object it keeps. At that point it is not safe anymore to continue using the info logger to write the delayed logs.
With the patch, lock is released temporarily for log buffer to be flushed before "bg_compaction_scheduled_--" or "bg_flush_scheduled_--". In order to make sure we don't miss any pending flush or compaction, a new flag bg_schedule_needed_ is added, which is set to be true if there is a pending flush or compaction but not scheduled because of the max thread limit. If the flag is set to be true, the scheduling function will be called before compaction or flush thread finishes.
Thanks @igor for this finding!
Test Plan: make all check
Reviewers: haobo, igor
Reviewed By: haobo
CC: dhruba, ljin, yhchiang, igor, leveldb
Differential Revision: https://reviews.facebook.net/D16767
2014-03-11 18:36:30 +01:00
|
|
|
if (bg_compaction_scheduled_ < options_.max_background_compactions) {
|
|
|
|
bg_compaction_scheduled_++;
|
|
|
|
env_->Schedule(&DBImpl::BGWorkCompaction, this, Env::Priority::LOW);
|
|
|
|
} else {
|
|
|
|
bg_schedule_needed_ = true;
|
|
|
|
}
|
2013-09-13 23:38:37 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void DBImpl::BGWorkFlush(void* db) {
|
|
|
|
reinterpret_cast<DBImpl*>(db)->BackgroundCallFlush();
|
|
|
|
}
|
|
|
|
|
|
|
|
void DBImpl::BGWorkCompaction(void* db) {
|
|
|
|
reinterpret_cast<DBImpl*>(db)->BackgroundCallCompaction();
|
|
|
|
}
|
|
|
|
|
2013-11-09 00:23:46 +01:00
|
|
|
Status DBImpl::BackgroundFlush(bool* madeProgress,
|
2014-03-10 06:01:13 +01:00
|
|
|
DeletionState& deletion_state,
|
|
|
|
LogBuffer* log_buffer) {
|
2013-09-13 23:38:37 +02:00
|
|
|
Status stat;
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
while (stat.ok() && imm_.IsFlushPending()) {
|
2013-09-13 23:38:37 +02:00
|
|
|
Log(options_.info_log,
|
2013-10-15 00:12:15 +02:00
|
|
|
"BackgroundCallFlush doing FlushMemTableToOutputFile, flush slots available %d",
|
2013-09-13 23:38:37 +02:00
|
|
|
options_.max_background_flushes - bg_flush_scheduled_);
|
2014-03-10 06:01:13 +01:00
|
|
|
stat = FlushMemTableToOutputFile(madeProgress, deletion_state, log_buffer);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
2013-09-13 23:38:37 +02:00
|
|
|
return stat;
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
2013-09-13 23:38:37 +02:00
|
|
|
void DBImpl::BackgroundCallFlush() {
|
2013-10-16 22:32:53 +02:00
|
|
|
bool madeProgress = false;
|
2014-01-02 20:26:57 +01:00
|
|
|
DeletionState deletion_state(true);
|
2013-09-13 23:38:37 +02:00
|
|
|
assert(bg_flush_scheduled_);
|
|
|
|
|
2014-03-10 06:01:13 +01:00
|
|
|
LogBuffer log_buffer(INFO, options_.info_log.get());
|
|
|
|
{
|
|
|
|
MutexLock l(&mutex_);
|
|
|
|
|
|
|
|
Status s;
|
|
|
|
if (!shutting_down_.Acquire_Load()) {
|
|
|
|
s = BackgroundFlush(&madeProgress, deletion_state, &log_buffer);
|
|
|
|
if (!s.ok()) {
|
|
|
|
// Wait a little bit before retrying background compaction in
|
|
|
|
// case this is an environmental problem and we do not want to
|
|
|
|
// chew up resources for failed compactions for the duration of
|
|
|
|
// the problem.
|
2014-03-18 20:25:08 +01:00
|
|
|
uint64_t error_cnt = internal_stats_.BumpAndGetBackgroundErrorCount();
|
2014-03-10 06:01:13 +01:00
|
|
|
bg_cv_.SignalAll(); // In case a waiter can proceed despite the error
|
|
|
|
mutex_.Unlock();
|
2014-03-18 20:25:08 +01:00
|
|
|
Log(options_.info_log,
|
|
|
|
"Waiting after background flush error: %s"
|
|
|
|
"Accumulated background error counts: %" PRIu64,
|
|
|
|
s.ToString().c_str(), error_cnt);
|
2014-03-10 06:01:13 +01:00
|
|
|
log_buffer.FlushBufferToLog();
|
|
|
|
LogFlush(options_.info_log);
|
|
|
|
env_->SleepForMicroseconds(1000000);
|
|
|
|
mutex_.Lock();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// If !s.ok(), this means that Flush failed. In that case, we want
|
|
|
|
// to delete all obsolete files and we force FindObsoleteFiles()
|
|
|
|
FindObsoleteFiles(deletion_state, !s.ok());
|
|
|
|
// delete unnecessary files if any, this is done outside the mutex
|
Fix data race against logging data structure because of LogBuffer
Summary:
@igor pointed out that there is a potential data race because of the way we use the newly introduced LogBuffer. After "bg_compaction_scheduled_--" or "bg_flush_scheduled_--", they can both become 0. As soon as the lock is released after that, DBImpl's deconstructor can go ahead and deconstruct all the states inside DB, including the info_log object hold in a shared pointer of the options object it keeps. At that point it is not safe anymore to continue using the info logger to write the delayed logs.
With the patch, lock is released temporarily for log buffer to be flushed before "bg_compaction_scheduled_--" or "bg_flush_scheduled_--". In order to make sure we don't miss any pending flush or compaction, a new flag bg_schedule_needed_ is added, which is set to be true if there is a pending flush or compaction but not scheduled because of the max thread limit. If the flag is set to be true, the scheduling function will be called before compaction or flush thread finishes.
Thanks @igor for this finding!
Test Plan: make all check
Reviewers: haobo, igor
Reviewed By: haobo
CC: dhruba, ljin, yhchiang, igor, leveldb
Differential Revision: https://reviews.facebook.net/D16767
2014-03-11 18:36:30 +01:00
|
|
|
if (deletion_state.HaveSomethingToDelete() || !log_buffer.IsEmpty()) {
|
2013-09-13 23:38:37 +02:00
|
|
|
mutex_.Unlock();
|
Fix data race against logging data structure because of LogBuffer
Summary:
@igor pointed out that there is a potential data race because of the way we use the newly introduced LogBuffer. After "bg_compaction_scheduled_--" or "bg_flush_scheduled_--", they can both become 0. As soon as the lock is released after that, DBImpl's deconstructor can go ahead and deconstruct all the states inside DB, including the info_log object hold in a shared pointer of the options object it keeps. At that point it is not safe anymore to continue using the info logger to write the delayed logs.
With the patch, lock is released temporarily for log buffer to be flushed before "bg_compaction_scheduled_--" or "bg_flush_scheduled_--". In order to make sure we don't miss any pending flush or compaction, a new flag bg_schedule_needed_ is added, which is set to be true if there is a pending flush or compaction but not scheduled because of the max thread limit. If the flag is set to be true, the scheduling function will be called before compaction or flush thread finishes.
Thanks @igor for this finding!
Test Plan: make all check
Reviewers: haobo, igor
Reviewed By: haobo
CC: dhruba, ljin, yhchiang, igor, leveldb
Differential Revision: https://reviews.facebook.net/D16767
2014-03-11 18:36:30 +01:00
|
|
|
// Have to flush the info logs before bg_flush_scheduled_--
|
|
|
|
// because if bg_flush_scheduled_ becomes 0 and the lock is
|
|
|
|
// released, the deconstructor of DB can kick in and destroy all the
|
|
|
|
// states of DB so info_log might not be available after that point.
|
|
|
|
// It also applies to access other states that DB owns.
|
2014-03-10 06:01:13 +01:00
|
|
|
log_buffer.FlushBufferToLog();
|
Fix data race against logging data structure because of LogBuffer
Summary:
@igor pointed out that there is a potential data race because of the way we use the newly introduced LogBuffer. After "bg_compaction_scheduled_--" or "bg_flush_scheduled_--", they can both become 0. As soon as the lock is released after that, DBImpl's deconstructor can go ahead and deconstruct all the states inside DB, including the info_log object hold in a shared pointer of the options object it keeps. At that point it is not safe anymore to continue using the info logger to write the delayed logs.
With the patch, lock is released temporarily for log buffer to be flushed before "bg_compaction_scheduled_--" or "bg_flush_scheduled_--". In order to make sure we don't miss any pending flush or compaction, a new flag bg_schedule_needed_ is added, which is set to be true if there is a pending flush or compaction but not scheduled because of the max thread limit. If the flag is set to be true, the scheduling function will be called before compaction or flush thread finishes.
Thanks @igor for this finding!
Test Plan: make all check
Reviewers: haobo, igor
Reviewed By: haobo
CC: dhruba, ljin, yhchiang, igor, leveldb
Differential Revision: https://reviews.facebook.net/D16767
2014-03-11 18:36:30 +01:00
|
|
|
if (deletion_state.HaveSomethingToDelete()) {
|
|
|
|
PurgeObsoleteFiles(deletion_state);
|
|
|
|
}
|
2013-09-13 23:38:37 +02:00
|
|
|
mutex_.Lock();
|
|
|
|
}
|
|
|
|
|
2014-03-10 06:01:13 +01:00
|
|
|
bg_flush_scheduled_--;
|
Fix data race against logging data structure because of LogBuffer
Summary:
@igor pointed out that there is a potential data race because of the way we use the newly introduced LogBuffer. After "bg_compaction_scheduled_--" or "bg_flush_scheduled_--", they can both become 0. As soon as the lock is released after that, DBImpl's deconstructor can go ahead and deconstruct all the states inside DB, including the info_log object hold in a shared pointer of the options object it keeps. At that point it is not safe anymore to continue using the info logger to write the delayed logs.
With the patch, lock is released temporarily for log buffer to be flushed before "bg_compaction_scheduled_--" or "bg_flush_scheduled_--". In order to make sure we don't miss any pending flush or compaction, a new flag bg_schedule_needed_ is added, which is set to be true if there is a pending flush or compaction but not scheduled because of the max thread limit. If the flag is set to be true, the scheduling function will be called before compaction or flush thread finishes.
Thanks @igor for this finding!
Test Plan: make all check
Reviewers: haobo, igor
Reviewed By: haobo
CC: dhruba, ljin, yhchiang, igor, leveldb
Differential Revision: https://reviews.facebook.net/D16767
2014-03-11 18:36:30 +01:00
|
|
|
// Any time the mutex is released After finding the work to do, another
|
|
|
|
// thread might execute MaybeScheduleFlushOrCompaction(). It is possible
|
|
|
|
// that there is a pending job but it is not scheduled because of the
|
|
|
|
// max thread limit.
|
|
|
|
if (madeProgress || bg_schedule_needed_) {
|
2014-03-10 06:01:13 +01:00
|
|
|
MaybeScheduleFlushOrCompaction();
|
|
|
|
}
|
|
|
|
bg_cv_.SignalAll();
|
2014-03-11 19:27:19 +01:00
|
|
|
// IMPORTANT: there should be no code after calling SignalAll. This call may
|
|
|
|
// signal the DB destructor that it's OK to proceed with destruction. In
|
|
|
|
// that case, all DB variables will be dealloacated and referencing them
|
|
|
|
// will cause trouble.
|
2013-10-16 22:32:53 +02:00
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
2013-09-13 23:38:37 +02:00
|
|
|
|
2013-05-06 20:41:01 +02:00
|
|
|
void DBImpl::TEST_PurgeObsoleteteWAL() {
|
|
|
|
PurgeObsoleteWALFiles();
|
|
|
|
}
|
|
|
|
|
2014-01-16 01:18:04 +01:00
|
|
|
uint64_t DBImpl::TEST_GetLevel0TotalSize() {
|
|
|
|
MutexLock l(&mutex_);
|
|
|
|
return versions_->current()->NumLevelBytes(0);
|
|
|
|
}
|
|
|
|
|
2013-09-13 23:38:37 +02:00
|
|
|
void DBImpl::BackgroundCallCompaction() {
|
2013-01-08 21:16:40 +01:00
|
|
|
bool madeProgress = false;
|
2014-01-02 20:26:57 +01:00
|
|
|
DeletionState deletion_state(true);
|
2013-05-24 21:52:45 +02:00
|
|
|
|
|
|
|
MaybeDumpStats();
|
2014-03-10 06:01:13 +01:00
|
|
|
LogBuffer log_buffer(INFO, options_.info_log.get());
|
Buffer info logs when picking compactions and write them out after releasing the mutex
Summary: Now while the background thread is picking compactions, it writes out multiple info_logs, especially for universal compaction, which introduces a chance of waiting log writing in mutex, which is bad. To remove this risk, write all those info logs to a buffer and flush it after releasing the mutex.
Test Plan:
make all check
check the log lines while running some tests that trigger compactions.
Reviewers: haobo, igor, dhruba
Reviewed By: dhruba
CC: i.am.jin.lei, dhruba, yhchiang, leveldb, nkg-
Differential Revision: https://reviews.facebook.net/D16515
2014-03-04 23:32:55 +01:00
|
|
|
{
|
|
|
|
MutexLock l(&mutex_);
|
|
|
|
// Log(options_.info_log, "XXX BG Thread %llx process new work item",
|
|
|
|
// pthread_self());
|
|
|
|
assert(bg_compaction_scheduled_);
|
|
|
|
Status s;
|
|
|
|
if (!shutting_down_.Acquire_Load()) {
|
|
|
|
s = BackgroundCompaction(&madeProgress, deletion_state, &log_buffer);
|
|
|
|
if (!s.ok()) {
|
|
|
|
// Wait a little bit before retrying background compaction in
|
|
|
|
// case this is an environmental problem and we do not want to
|
|
|
|
// chew up resources for failed compactions for the duration of
|
|
|
|
// the problem.
|
2014-03-18 20:25:08 +01:00
|
|
|
uint64_t error_cnt = internal_stats_.BumpAndGetBackgroundErrorCount();
|
Buffer info logs when picking compactions and write them out after releasing the mutex
Summary: Now while the background thread is picking compactions, it writes out multiple info_logs, especially for universal compaction, which introduces a chance of waiting log writing in mutex, which is bad. To remove this risk, write all those info logs to a buffer and flush it after releasing the mutex.
Test Plan:
make all check
check the log lines while running some tests that trigger compactions.
Reviewers: haobo, igor, dhruba
Reviewed By: dhruba
CC: i.am.jin.lei, dhruba, yhchiang, leveldb, nkg-
Differential Revision: https://reviews.facebook.net/D16515
2014-03-04 23:32:55 +01:00
|
|
|
bg_cv_.SignalAll(); // In case a waiter can proceed despite the error
|
|
|
|
mutex_.Unlock();
|
2014-03-10 06:01:13 +01:00
|
|
|
log_buffer.FlushBufferToLog();
|
2014-03-18 20:25:08 +01:00
|
|
|
Log(options_.info_log,
|
|
|
|
"Waiting after background compaction error: %s, "
|
|
|
|
"Accumulated background error counts: %" PRIu64,
|
|
|
|
s.ToString().c_str(), error_cnt);
|
Buffer info logs when picking compactions and write them out after releasing the mutex
Summary: Now while the background thread is picking compactions, it writes out multiple info_logs, especially for universal compaction, which introduces a chance of waiting log writing in mutex, which is bad. To remove this risk, write all those info logs to a buffer and flush it after releasing the mutex.
Test Plan:
make all check
check the log lines while running some tests that trigger compactions.
Reviewers: haobo, igor, dhruba
Reviewed By: dhruba
CC: i.am.jin.lei, dhruba, yhchiang, leveldb, nkg-
Differential Revision: https://reviews.facebook.net/D16515
2014-03-04 23:32:55 +01:00
|
|
|
LogFlush(options_.info_log);
|
|
|
|
env_->SleepForMicroseconds(1000000);
|
|
|
|
mutex_.Lock();
|
|
|
|
}
|
|
|
|
}
|
2013-05-24 21:52:45 +02:00
|
|
|
|
Buffer info logs when picking compactions and write them out after releasing the mutex
Summary: Now while the background thread is picking compactions, it writes out multiple info_logs, especially for universal compaction, which introduces a chance of waiting log writing in mutex, which is bad. To remove this risk, write all those info logs to a buffer and flush it after releasing the mutex.
Test Plan:
make all check
check the log lines while running some tests that trigger compactions.
Reviewers: haobo, igor, dhruba
Reviewed By: dhruba
CC: i.am.jin.lei, dhruba, yhchiang, leveldb, nkg-
Differential Revision: https://reviews.facebook.net/D16515
2014-03-04 23:32:55 +01:00
|
|
|
// If !s.ok(), this means that Compaction failed. In that case, we want
|
|
|
|
// to delete all obsolete files we might have created and we force
|
|
|
|
// FindObsoleteFiles(). This is because deletion_state does not catch
|
|
|
|
// all created files if compaction failed.
|
|
|
|
FindObsoleteFiles(deletion_state, !s.ok());
|
|
|
|
|
|
|
|
// delete unnecessary files if any, this is done outside the mutex
|
Fix data race against logging data structure because of LogBuffer
Summary:
@igor pointed out that there is a potential data race because of the way we use the newly introduced LogBuffer. After "bg_compaction_scheduled_--" or "bg_flush_scheduled_--", they can both become 0. As soon as the lock is released after that, DBImpl's deconstructor can go ahead and deconstruct all the states inside DB, including the info_log object hold in a shared pointer of the options object it keeps. At that point it is not safe anymore to continue using the info logger to write the delayed logs.
With the patch, lock is released temporarily for log buffer to be flushed before "bg_compaction_scheduled_--" or "bg_flush_scheduled_--". In order to make sure we don't miss any pending flush or compaction, a new flag bg_schedule_needed_ is added, which is set to be true if there is a pending flush or compaction but not scheduled because of the max thread limit. If the flag is set to be true, the scheduling function will be called before compaction or flush thread finishes.
Thanks @igor for this finding!
Test Plan: make all check
Reviewers: haobo, igor
Reviewed By: haobo
CC: dhruba, ljin, yhchiang, igor, leveldb
Differential Revision: https://reviews.facebook.net/D16767
2014-03-11 18:36:30 +01:00
|
|
|
if (deletion_state.HaveSomethingToDelete() || !log_buffer.IsEmpty()) {
|
2012-08-23 01:57:51 +02:00
|
|
|
mutex_.Unlock();
|
Fix data race against logging data structure because of LogBuffer
Summary:
@igor pointed out that there is a potential data race because of the way we use the newly introduced LogBuffer. After "bg_compaction_scheduled_--" or "bg_flush_scheduled_--", they can both become 0. As soon as the lock is released after that, DBImpl's deconstructor can go ahead and deconstruct all the states inside DB, including the info_log object hold in a shared pointer of the options object it keeps. At that point it is not safe anymore to continue using the info logger to write the delayed logs.
With the patch, lock is released temporarily for log buffer to be flushed before "bg_compaction_scheduled_--" or "bg_flush_scheduled_--". In order to make sure we don't miss any pending flush or compaction, a new flag bg_schedule_needed_ is added, which is set to be true if there is a pending flush or compaction but not scheduled because of the max thread limit. If the flag is set to be true, the scheduling function will be called before compaction or flush thread finishes.
Thanks @igor for this finding!
Test Plan: make all check
Reviewers: haobo, igor
Reviewed By: haobo
CC: dhruba, ljin, yhchiang, igor, leveldb
Differential Revision: https://reviews.facebook.net/D16767
2014-03-11 18:36:30 +01:00
|
|
|
// Have to flush the info logs before bg_compaction_scheduled_--
|
|
|
|
// because if bg_flush_scheduled_ becomes 0 and the lock is
|
|
|
|
// released, the deconstructor of DB can kick in and destroy all the
|
|
|
|
// states of DB so info_log might not be available after that point.
|
|
|
|
// It also applies to access other states that DB owns.
|
2014-03-10 06:01:13 +01:00
|
|
|
log_buffer.FlushBufferToLog();
|
Fix data race against logging data structure because of LogBuffer
Summary:
@igor pointed out that there is a potential data race because of the way we use the newly introduced LogBuffer. After "bg_compaction_scheduled_--" or "bg_flush_scheduled_--", they can both become 0. As soon as the lock is released after that, DBImpl's deconstructor can go ahead and deconstruct all the states inside DB, including the info_log object hold in a shared pointer of the options object it keeps. At that point it is not safe anymore to continue using the info logger to write the delayed logs.
With the patch, lock is released temporarily for log buffer to be flushed before "bg_compaction_scheduled_--" or "bg_flush_scheduled_--". In order to make sure we don't miss any pending flush or compaction, a new flag bg_schedule_needed_ is added, which is set to be true if there is a pending flush or compaction but not scheduled because of the max thread limit. If the flag is set to be true, the scheduling function will be called before compaction or flush thread finishes.
Thanks @igor for this finding!
Test Plan: make all check
Reviewers: haobo, igor
Reviewed By: haobo
CC: dhruba, ljin, yhchiang, igor, leveldb
Differential Revision: https://reviews.facebook.net/D16767
2014-03-11 18:36:30 +01:00
|
|
|
if (deletion_state.HaveSomethingToDelete()) {
|
|
|
|
PurgeObsoleteFiles(deletion_state);
|
|
|
|
}
|
2012-08-23 01:57:51 +02:00
|
|
|
mutex_.Lock();
|
|
|
|
}
|
|
|
|
|
Buffer info logs when picking compactions and write them out after releasing the mutex
Summary: Now while the background thread is picking compactions, it writes out multiple info_logs, especially for universal compaction, which introduces a chance of waiting log writing in mutex, which is bad. To remove this risk, write all those info logs to a buffer and flush it after releasing the mutex.
Test Plan:
make all check
check the log lines while running some tests that trigger compactions.
Reviewers: haobo, igor, dhruba
Reviewed By: dhruba
CC: i.am.jin.lei, dhruba, yhchiang, leveldb, nkg-
Differential Revision: https://reviews.facebook.net/D16515
2014-03-04 23:32:55 +01:00
|
|
|
bg_compaction_scheduled_--;
|
2012-10-21 10:49:48 +02:00
|
|
|
|
Buffer info logs when picking compactions and write them out after releasing the mutex
Summary: Now while the background thread is picking compactions, it writes out multiple info_logs, especially for universal compaction, which introduces a chance of waiting log writing in mutex, which is bad. To remove this risk, write all those info logs to a buffer and flush it after releasing the mutex.
Test Plan:
make all check
check the log lines while running some tests that trigger compactions.
Reviewers: haobo, igor, dhruba
Reviewed By: dhruba
CC: i.am.jin.lei, dhruba, yhchiang, leveldb, nkg-
Differential Revision: https://reviews.facebook.net/D16515
2014-03-04 23:32:55 +01:00
|
|
|
MaybeScheduleLogDBDeployStats();
|
2012-08-15 00:20:36 +02:00
|
|
|
|
Buffer info logs when picking compactions and write them out after releasing the mutex
Summary: Now while the background thread is picking compactions, it writes out multiple info_logs, especially for universal compaction, which introduces a chance of waiting log writing in mutex, which is bad. To remove this risk, write all those info logs to a buffer and flush it after releasing the mutex.
Test Plan:
make all check
check the log lines while running some tests that trigger compactions.
Reviewers: haobo, igor, dhruba
Reviewed By: dhruba
CC: i.am.jin.lei, dhruba, yhchiang, leveldb, nkg-
Differential Revision: https://reviews.facebook.net/D16515
2014-03-04 23:32:55 +01:00
|
|
|
// Previous compaction may have produced too many files in a level,
|
|
|
|
// So reschedule another compaction if we made progress in the
|
|
|
|
// last compaction.
|
Fix data race against logging data structure because of LogBuffer
Summary:
@igor pointed out that there is a potential data race because of the way we use the newly introduced LogBuffer. After "bg_compaction_scheduled_--" or "bg_flush_scheduled_--", they can both become 0. As soon as the lock is released after that, DBImpl's deconstructor can go ahead and deconstruct all the states inside DB, including the info_log object hold in a shared pointer of the options object it keeps. At that point it is not safe anymore to continue using the info logger to write the delayed logs.
With the patch, lock is released temporarily for log buffer to be flushed before "bg_compaction_scheduled_--" or "bg_flush_scheduled_--". In order to make sure we don't miss any pending flush or compaction, a new flag bg_schedule_needed_ is added, which is set to be true if there is a pending flush or compaction but not scheduled because of the max thread limit. If the flag is set to be true, the scheduling function will be called before compaction or flush thread finishes.
Thanks @igor for this finding!
Test Plan: make all check
Reviewers: haobo, igor
Reviewed By: haobo
CC: dhruba, ljin, yhchiang, igor, leveldb
Differential Revision: https://reviews.facebook.net/D16767
2014-03-11 18:36:30 +01:00
|
|
|
//
|
|
|
|
// Also, any time the mutex is released After finding the work to do,
|
|
|
|
// another thread might execute MaybeScheduleFlushOrCompaction(). It is
|
|
|
|
// possible that there is a pending job but it is not scheduled because of
|
|
|
|
// the max thread limit.
|
|
|
|
if (madeProgress || bg_schedule_needed_) {
|
Buffer info logs when picking compactions and write them out after releasing the mutex
Summary: Now while the background thread is picking compactions, it writes out multiple info_logs, especially for universal compaction, which introduces a chance of waiting log writing in mutex, which is bad. To remove this risk, write all those info logs to a buffer and flush it after releasing the mutex.
Test Plan:
make all check
check the log lines while running some tests that trigger compactions.
Reviewers: haobo, igor, dhruba
Reviewed By: dhruba
CC: i.am.jin.lei, dhruba, yhchiang, leveldb, nkg-
Differential Revision: https://reviews.facebook.net/D16515
2014-03-04 23:32:55 +01:00
|
|
|
MaybeScheduleFlushOrCompaction();
|
|
|
|
}
|
|
|
|
bg_cv_.SignalAll();
|
2014-03-11 19:27:19 +01:00
|
|
|
// IMPORTANT: there should be no code after calling SignalAll. This call may
|
|
|
|
// signal the DB destructor that it's OK to proceed with destruction. In
|
|
|
|
// that case, all DB variables will be dealloacated and referencing them
|
|
|
|
// will cause trouble.
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
2012-11-29 01:42:36 +01:00
|
|
|
Status DBImpl::BackgroundCompaction(bool* madeProgress,
|
Buffer info logs when picking compactions and write them out after releasing the mutex
Summary: Now while the background thread is picking compactions, it writes out multiple info_logs, especially for universal compaction, which introduces a chance of waiting log writing in mutex, which is bad. To remove this risk, write all those info logs to a buffer and flush it after releasing the mutex.
Test Plan:
make all check
check the log lines while running some tests that trigger compactions.
Reviewers: haobo, igor, dhruba
Reviewed By: dhruba
CC: i.am.jin.lei, dhruba, yhchiang, leveldb, nkg-
Differential Revision: https://reviews.facebook.net/D16515
2014-03-04 23:32:55 +01:00
|
|
|
DeletionState& deletion_state,
|
|
|
|
LogBuffer* log_buffer) {
|
2012-10-19 23:00:53 +02:00
|
|
|
*madeProgress = false;
|
2011-03-18 23:37:00 +01:00
|
|
|
mutex_.AssertHeld();
|
2011-04-12 21:38:58 +02:00
|
|
|
|
2014-01-30 00:55:58 +01:00
|
|
|
bool is_manual = (manual_compaction_ != nullptr) &&
|
|
|
|
(manual_compaction_->in_progress == false);
|
|
|
|
if (is_manual) {
|
|
|
|
// another thread cannot pick up the same work
|
|
|
|
manual_compaction_->in_progress = true;
|
|
|
|
}
|
|
|
|
|
2013-09-13 23:38:37 +02:00
|
|
|
// TODO: remove memtable flush from formal compaction
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
while (imm_.IsFlushPending()) {
|
Buffer info logs when picking compactions and write them out after releasing the mutex
Summary: Now while the background thread is picking compactions, it writes out multiple info_logs, especially for universal compaction, which introduces a chance of waiting log writing in mutex, which is bad. To remove this risk, write all those info logs to a buffer and flush it after releasing the mutex.
Test Plan:
make all check
check the log lines while running some tests that trigger compactions.
Reviewers: haobo, igor, dhruba
Reviewed By: dhruba
CC: i.am.jin.lei, dhruba, yhchiang, leveldb, nkg-
Differential Revision: https://reviews.facebook.net/D16515
2014-03-04 23:32:55 +01:00
|
|
|
LogToBuffer(log_buffer,
|
|
|
|
"BackgroundCompaction doing FlushMemTableToOutputFile, "
|
|
|
|
"compaction slots "
|
|
|
|
"available %d",
|
|
|
|
options_.max_background_compactions - bg_compaction_scheduled_);
|
2014-03-10 06:01:13 +01:00
|
|
|
Status stat = FlushMemTableToOutputFile(madeProgress, deletion_state,
|
|
|
|
log_buffer);
|
2012-10-19 23:00:53 +02:00
|
|
|
if (!stat.ok()) {
|
2014-01-30 00:55:58 +01:00
|
|
|
if (is_manual) {
|
|
|
|
manual_compaction_->status = stat;
|
|
|
|
manual_compaction_->done = true;
|
|
|
|
manual_compaction_->in_progress = false;
|
|
|
|
manual_compaction_ = nullptr;
|
|
|
|
}
|
2012-10-19 23:00:53 +02:00
|
|
|
return stat;
|
|
|
|
}
|
2011-04-12 21:38:58 +02:00
|
|
|
}
|
|
|
|
|
2013-01-20 11:07:13 +01:00
|
|
|
unique_ptr<Compaction> c;
|
2014-01-15 01:19:09 +01:00
|
|
|
InternalKey manual_end_storage;
|
|
|
|
InternalKey* manual_end = &manual_end_storage;
|
2011-06-07 16:40:26 +02:00
|
|
|
if (is_manual) {
|
2011-10-06 01:30:28 +02:00
|
|
|
ManualCompaction* m = manual_compaction_;
|
2014-01-30 00:55:58 +01:00
|
|
|
assert(m->in_progress);
|
2014-01-15 01:19:09 +01:00
|
|
|
c.reset(versions_->CompactRange(
|
|
|
|
m->input_level, m->output_level, m->begin, m->end, &manual_end));
|
|
|
|
if (!c) {
|
2013-01-20 11:07:13 +01:00
|
|
|
m->done = true;
|
2011-10-06 01:30:28 +02:00
|
|
|
}
|
Buffer info logs when picking compactions and write them out after releasing the mutex
Summary: Now while the background thread is picking compactions, it writes out multiple info_logs, especially for universal compaction, which introduces a chance of waiting log writing in mutex, which is bad. To remove this risk, write all those info logs to a buffer and flush it after releasing the mutex.
Test Plan:
make all check
check the log lines while running some tests that trigger compactions.
Reviewers: haobo, igor, dhruba
Reviewed By: dhruba
CC: i.am.jin.lei, dhruba, yhchiang, leveldb, nkg-
Differential Revision: https://reviews.facebook.net/D16515
2014-03-04 23:32:55 +01:00
|
|
|
LogToBuffer(
|
|
|
|
log_buffer,
|
2014-01-15 01:19:09 +01:00
|
|
|
"Manual compaction from level-%d to level-%d from %s .. %s; will stop "
|
|
|
|
"at %s\n",
|
Buffer info logs when picking compactions and write them out after releasing the mutex
Summary: Now while the background thread is picking compactions, it writes out multiple info_logs, especially for universal compaction, which introduces a chance of waiting log writing in mutex, which is bad. To remove this risk, write all those info logs to a buffer and flush it after releasing the mutex.
Test Plan:
make all check
check the log lines while running some tests that trigger compactions.
Reviewers: haobo, igor, dhruba
Reviewed By: dhruba
CC: i.am.jin.lei, dhruba, yhchiang, leveldb, nkg-
Differential Revision: https://reviews.facebook.net/D16515
2014-03-04 23:32:55 +01:00
|
|
|
m->input_level, m->output_level,
|
2011-10-06 01:30:28 +02:00
|
|
|
(m->begin ? m->begin->DebugString().c_str() : "(begin)"),
|
|
|
|
(m->end ? m->end->DebugString().c_str() : "(end)"),
|
2014-01-15 01:19:09 +01:00
|
|
|
((m->done || manual_end == nullptr)
|
|
|
|
? "(end)"
|
|
|
|
: manual_end->DebugString().c_str()));
|
2012-11-21 00:45:41 +01:00
|
|
|
} else if (!options_.disable_auto_compactions) {
|
Buffer info logs when picking compactions and write them out after releasing the mutex
Summary: Now while the background thread is picking compactions, it writes out multiple info_logs, especially for universal compaction, which introduces a chance of waiting log writing in mutex, which is bad. To remove this risk, write all those info logs to a buffer and flush it after releasing the mutex.
Test Plan:
make all check
check the log lines while running some tests that trigger compactions.
Reviewers: haobo, igor, dhruba
Reviewed By: dhruba
CC: i.am.jin.lei, dhruba, yhchiang, leveldb, nkg-
Differential Revision: https://reviews.facebook.net/D16515
2014-03-04 23:32:55 +01:00
|
|
|
c.reset(versions_->PickCompaction(log_buffer));
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
Status status;
|
2013-01-20 11:07:13 +01:00
|
|
|
if (!c) {
|
2011-06-07 16:40:26 +02:00
|
|
|
// Nothing to do
|
Buffer info logs when picking compactions and write them out after releasing the mutex
Summary: Now while the background thread is picking compactions, it writes out multiple info_logs, especially for universal compaction, which introduces a chance of waiting log writing in mutex, which is bad. To remove this risk, write all those info logs to a buffer and flush it after releasing the mutex.
Test Plan:
make all check
check the log lines while running some tests that trigger compactions.
Reviewers: haobo, igor, dhruba
Reviewed By: dhruba
CC: i.am.jin.lei, dhruba, yhchiang, leveldb, nkg-
Differential Revision: https://reviews.facebook.net/D16515
2014-03-04 23:32:55 +01:00
|
|
|
LogToBuffer(log_buffer, "Compaction nothing to do");
|
2011-06-07 16:40:26 +02:00
|
|
|
} else if (!is_manual && c->IsTrivialMove()) {
|
2011-03-18 23:37:00 +01:00
|
|
|
// Move file to next level
|
2011-03-22 19:32:49 +01:00
|
|
|
assert(c->num_input_files(0) == 1);
|
2011-03-18 23:37:00 +01:00
|
|
|
FileMetaData* f = c->input(0, 0);
|
|
|
|
c->edit()->DeleteFile(c->level(), f->number);
|
|
|
|
c->edit()->AddFile(c->level() + 1, f->number, f->file_size,
|
2013-06-14 07:09:08 +02:00
|
|
|
f->smallest, f->largest,
|
|
|
|
f->smallest_seqno, f->largest_seqno);
|
2014-01-27 20:02:21 +01:00
|
|
|
status = versions_->LogAndApply(c->edit(), &mutex_, db_directory_.get());
|
2013-12-20 18:57:58 +01:00
|
|
|
InstallSuperVersion(deletion_state);
|
2014-01-16 01:18:04 +01:00
|
|
|
Version::LevelSummaryStorage tmp;
|
Buffer info logs when picking compactions and write them out after releasing the mutex
Summary: Now while the background thread is picking compactions, it writes out multiple info_logs, especially for universal compaction, which introduces a chance of waiting log writing in mutex, which is bad. To remove this risk, write all those info logs to a buffer and flush it after releasing the mutex.
Test Plan:
make all check
check the log lines while running some tests that trigger compactions.
Reviewers: haobo, igor, dhruba
Reviewed By: dhruba
CC: i.am.jin.lei, dhruba, yhchiang, leveldb, nkg-
Differential Revision: https://reviews.facebook.net/D16515
2014-03-04 23:32:55 +01:00
|
|
|
LogToBuffer(log_buffer, "Moved #%lld to level-%d %lld bytes %s: %s\n",
|
|
|
|
static_cast<unsigned long long>(f->number), c->level() + 1,
|
|
|
|
static_cast<unsigned long long>(f->file_size),
|
|
|
|
status.ToString().c_str(),
|
|
|
|
versions_->current()->LevelSummary(&tmp));
|
2013-01-20 11:07:13 +01:00
|
|
|
versions_->ReleaseCompactionFiles(c.get(), status);
|
2012-10-19 23:00:53 +02:00
|
|
|
*madeProgress = true;
|
2011-03-18 23:37:00 +01:00
|
|
|
} else {
|
2013-10-15 00:12:15 +02:00
|
|
|
MaybeScheduleFlushOrCompaction(); // do more compaction work in parallel.
|
2013-01-20 11:07:13 +01:00
|
|
|
CompactionState* compact = new CompactionState(c.get());
|
2014-03-10 06:01:13 +01:00
|
|
|
status = DoCompactionWork(compact, deletion_state, log_buffer);
|
2013-09-02 08:23:40 +02:00
|
|
|
CleanupCompaction(compact, status);
|
2013-01-20 11:07:13 +01:00
|
|
|
versions_->ReleaseCompactionFiles(c.get(), status);
|
2012-01-25 23:56:52 +01:00
|
|
|
c->ReleaseInputs();
|
2012-10-19 23:00:53 +02:00
|
|
|
*madeProgress = true;
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
2013-01-20 11:07:13 +01:00
|
|
|
c.reset();
|
2011-03-18 23:37:00 +01:00
|
|
|
|
|
|
|
if (status.ok()) {
|
|
|
|
// Done
|
|
|
|
} else if (shutting_down_.Acquire_Load()) {
|
|
|
|
// Ignore compaction errors found during shutting down
|
|
|
|
} else {
|
Buffer info logs when picking compactions and write them out after releasing the mutex
Summary: Now while the background thread is picking compactions, it writes out multiple info_logs, especially for universal compaction, which introduces a chance of waiting log writing in mutex, which is bad. To remove this risk, write all those info logs to a buffer and flush it after releasing the mutex.
Test Plan:
make all check
check the log lines while running some tests that trigger compactions.
Reviewers: haobo, igor, dhruba
Reviewed By: dhruba
CC: i.am.jin.lei, dhruba, yhchiang, leveldb, nkg-
Differential Revision: https://reviews.facebook.net/D16515
2014-03-04 23:32:55 +01:00
|
|
|
Log(WARN, options_.info_log, "Compaction error: %s",
|
|
|
|
status.ToString().c_str());
|
2011-03-18 23:37:00 +01:00
|
|
|
if (options_.paranoid_checks && bg_error_.ok()) {
|
|
|
|
bg_error_ = status;
|
|
|
|
}
|
|
|
|
}
|
2011-06-07 16:40:26 +02:00
|
|
|
|
|
|
|
if (is_manual) {
|
2011-10-06 01:30:28 +02:00
|
|
|
ManualCompaction* m = manual_compaction_;
|
2012-01-25 23:56:52 +01:00
|
|
|
if (!status.ok()) {
|
2014-01-22 21:46:24 +01:00
|
|
|
m->status = status;
|
2012-01-25 23:56:52 +01:00
|
|
|
m->done = true;
|
|
|
|
}
|
2013-08-02 20:46:47 +02:00
|
|
|
// For universal compaction:
|
|
|
|
// Because universal compaction always happens at level 0, so one
|
|
|
|
// compaction will pick up all overlapped files. No files will be
|
|
|
|
// filtered out due to size limit and left for a successive compaction.
|
|
|
|
// So we can safely conclude the current compaction.
|
|
|
|
//
|
|
|
|
// Also note that, if we don't stop here, then the current compaction
|
|
|
|
// writes a new file back to level 0, which will be used in successive
|
|
|
|
// compaction. Hence the manual compaction will never finish.
|
2014-01-15 01:19:09 +01:00
|
|
|
//
|
|
|
|
// Stop the compaction if manual_end points to nullptr -- this means
|
|
|
|
// that we compacted the whole range. manual_end should always point
|
|
|
|
// to nullptr in case of universal compaction
|
|
|
|
if (manual_end == nullptr) {
|
2013-08-02 20:46:47 +02:00
|
|
|
m->done = true;
|
|
|
|
}
|
2011-10-06 01:30:28 +02:00
|
|
|
if (!m->done) {
|
|
|
|
// We only compacted part of the requested range. Update *m
|
|
|
|
// to the range that is left to be compacted.
|
2014-01-15 01:19:09 +01:00
|
|
|
// Universal compaction should always compact the whole range
|
|
|
|
assert(options_.compaction_style != kCompactionStyleUniversal);
|
|
|
|
m->tmp_storage = *manual_end;
|
2011-10-06 01:30:28 +02:00
|
|
|
m->begin = &m->tmp_storage;
|
|
|
|
}
|
2012-10-19 23:00:53 +02:00
|
|
|
m->in_progress = false; // not being processed anymore
|
2013-02-15 20:53:17 +01:00
|
|
|
manual_compaction_ = nullptr;
|
2011-06-07 16:40:26 +02:00
|
|
|
}
|
2012-08-23 01:57:51 +02:00
|
|
|
return status;
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
2013-09-02 08:23:40 +02:00
|
|
|
void DBImpl::CleanupCompaction(CompactionState* compact, Status status) {
|
2011-03-18 23:37:00 +01:00
|
|
|
mutex_.AssertHeld();
|
2013-02-15 20:53:17 +01:00
|
|
|
if (compact->builder != nullptr) {
|
2011-03-18 23:37:00 +01:00
|
|
|
// May happen if we get a shutdown call in the middle of compaction
|
|
|
|
compact->builder->Abandon();
|
2013-01-20 11:07:13 +01:00
|
|
|
compact->builder.reset();
|
2011-03-18 23:37:00 +01:00
|
|
|
} else {
|
2013-02-15 20:53:17 +01:00
|
|
|
assert(compact->outfile == nullptr);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
2011-04-21 00:48:11 +02:00
|
|
|
for (size_t i = 0; i < compact->outputs.size(); i++) {
|
2011-03-18 23:37:00 +01:00
|
|
|
const CompactionState::Output& out = compact->outputs[i];
|
|
|
|
pending_outputs_.erase(out.number);
|
2013-09-02 08:23:40 +02:00
|
|
|
|
|
|
|
// If this file was inserted into the table cache then remove
|
|
|
|
// them here because this compaction was not committed.
|
|
|
|
if (!status.ok()) {
|
|
|
|
table_cache_->Evict(out.number);
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
delete compact;
|
|
|
|
}
|
|
|
|
|
2012-10-19 23:00:53 +02:00
|
|
|
// Allocate the file numbers for the output file. We allocate as
|
2014-01-15 01:19:09 +01:00
|
|
|
// many output file numbers as there are files in level+1 (at least one)
|
2012-10-19 23:00:53 +02:00
|
|
|
// Insert them into pending_outputs so that they do not get deleted.
|
|
|
|
void DBImpl::AllocateCompactionOutputFileNumbers(CompactionState* compact) {
|
|
|
|
mutex_.AssertHeld();
|
2013-02-15 20:53:17 +01:00
|
|
|
assert(compact != nullptr);
|
|
|
|
assert(compact->builder == nullptr);
|
2012-10-19 23:00:53 +02:00
|
|
|
int filesNeeded = compact->compaction->num_input_files(1);
|
2014-01-15 01:19:09 +01:00
|
|
|
for (int i = 0; i < std::max(filesNeeded, 1); i++) {
|
2012-10-19 23:00:53 +02:00
|
|
|
uint64_t file_number = versions_->NewFileNumber();
|
|
|
|
pending_outputs_.insert(file_number);
|
|
|
|
compact->allocated_file_numbers.push_back(file_number);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Frees up unused file number.
|
|
|
|
void DBImpl::ReleaseCompactionUnusedFileNumbers(CompactionState* compact) {
|
|
|
|
mutex_.AssertHeld();
|
2013-03-29 02:01:06 +01:00
|
|
|
for (const auto file_number : compact->allocated_file_numbers) {
|
2012-10-19 23:00:53 +02:00
|
|
|
pending_outputs_.erase(file_number);
|
|
|
|
// Log(options_.info_log, "XXX releasing unused file num %d", file_number);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
Status DBImpl::OpenCompactionOutputFile(CompactionState* compact) {
|
2013-02-15 20:53:17 +01:00
|
|
|
assert(compact != nullptr);
|
|
|
|
assert(compact->builder == nullptr);
|
2011-03-18 23:37:00 +01:00
|
|
|
uint64_t file_number;
|
2012-10-19 23:00:53 +02:00
|
|
|
// If we have not yet exhausted the pre-allocated file numbers,
|
|
|
|
// then use the one from the front. Otherwise, we have to acquire
|
|
|
|
// the heavyweight lock and allocate a new file number.
|
|
|
|
if (!compact->allocated_file_numbers.empty()) {
|
|
|
|
file_number = compact->allocated_file_numbers.front();
|
|
|
|
compact->allocated_file_numbers.pop_front();
|
|
|
|
} else {
|
2011-03-18 23:37:00 +01:00
|
|
|
mutex_.Lock();
|
|
|
|
file_number = versions_->NewFileNumber();
|
|
|
|
pending_outputs_.insert(file_number);
|
|
|
|
mutex_.Unlock();
|
|
|
|
}
|
2012-10-19 23:00:53 +02:00
|
|
|
CompactionState::Output out;
|
|
|
|
out.number = file_number;
|
|
|
|
out.smallest.Clear();
|
|
|
|
out.largest.Clear();
|
2013-06-14 07:09:08 +02:00
|
|
|
out.smallest_seqno = out.largest_seqno = 0;
|
2012-10-19 23:00:53 +02:00
|
|
|
compact->outputs.push_back(out);
|
2011-03-18 23:37:00 +01:00
|
|
|
|
|
|
|
// Make the output file
|
|
|
|
std::string fname = TableFileName(dbname_, file_number);
|
2013-03-15 01:00:04 +01:00
|
|
|
Status s = env_->NewWritableFile(fname, &compact->outfile, storage_options_);
|
2013-01-15 23:05:42 +01:00
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
if (s.ok()) {
|
2013-04-10 17:37:03 +02:00
|
|
|
// Over-estimate slightly so we don't end up just barely crossing
|
|
|
|
// the threshold.
|
|
|
|
compact->outfile->SetPreallocationBlockSize(
|
2013-06-14 07:09:08 +02:00
|
|
|
1.1 * versions_->MaxFileSizeForLevel(compact->compaction->output_level()));
|
2013-04-10 17:37:03 +02:00
|
|
|
|
2013-10-30 18:52:33 +01:00
|
|
|
CompressionType compression_type = GetCompressionType(
|
|
|
|
options_, compact->compaction->output_level(),
|
|
|
|
compact->compaction->enable_compression());
|
|
|
|
|
2014-01-27 22:53:22 +01:00
|
|
|
compact->builder.reset(NewTableBuilder(options_, internal_comparator_,
|
|
|
|
compact->outfile.get(),
|
|
|
|
compression_type));
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
2013-11-07 20:31:56 +01:00
|
|
|
LogFlush(options_.info_log);
|
2011-03-18 23:37:00 +01:00
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status DBImpl::FinishCompactionOutputFile(CompactionState* compact,
|
|
|
|
Iterator* input) {
|
2013-02-15 20:53:17 +01:00
|
|
|
assert(compact != nullptr);
|
2013-01-20 11:07:13 +01:00
|
|
|
assert(compact->outfile);
|
2013-02-15 20:53:17 +01:00
|
|
|
assert(compact->builder != nullptr);
|
2011-03-18 23:37:00 +01:00
|
|
|
|
|
|
|
const uint64_t output_number = compact->current_output()->number;
|
|
|
|
assert(output_number != 0);
|
|
|
|
|
|
|
|
// Check for iterator errors
|
|
|
|
Status s = input->status();
|
|
|
|
const uint64_t current_entries = compact->builder->NumEntries();
|
|
|
|
if (s.ok()) {
|
|
|
|
s = compact->builder->Finish();
|
|
|
|
} else {
|
|
|
|
compact->builder->Abandon();
|
|
|
|
}
|
|
|
|
const uint64_t current_bytes = compact->builder->FileSize();
|
|
|
|
compact->current_output()->file_size = current_bytes;
|
|
|
|
compact->total_bytes += current_bytes;
|
2013-01-20 11:07:13 +01:00
|
|
|
compact->builder.reset();
|
2011-03-18 23:37:00 +01:00
|
|
|
|
|
|
|
// Finish and check for file errors
|
2012-08-04 00:20:58 +02:00
|
|
|
if (s.ok() && !options_.disableDataSync) {
|
2012-08-27 21:10:26 +02:00
|
|
|
if (options_.use_fsync) {
|
2013-11-22 23:14:05 +01:00
|
|
|
StopWatch sw(env_, options_.statistics.get(),
|
2014-01-07 23:33:15 +01:00
|
|
|
COMPACTION_OUTFILE_SYNC_MICROS, false);
|
2012-08-27 21:10:26 +02:00
|
|
|
s = compact->outfile->Fsync();
|
|
|
|
} else {
|
2013-11-22 23:14:05 +01:00
|
|
|
StopWatch sw(env_, options_.statistics.get(),
|
2014-01-07 23:33:15 +01:00
|
|
|
COMPACTION_OUTFILE_SYNC_MICROS, false);
|
2012-08-27 21:10:26 +02:00
|
|
|
s = compact->outfile->Sync();
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
if (s.ok()) {
|
|
|
|
s = compact->outfile->Close();
|
|
|
|
}
|
2013-01-20 11:07:13 +01:00
|
|
|
compact->outfile.reset();
|
2011-03-18 23:37:00 +01:00
|
|
|
|
|
|
|
if (s.ok() && current_entries > 0) {
|
|
|
|
// Verify that the table is usable
|
2014-01-07 05:29:17 +01:00
|
|
|
FileMetaData meta(output_number, current_bytes);
|
2014-01-27 22:53:22 +01:00
|
|
|
Iterator* iter = table_cache_->NewIterator(ReadOptions(), storage_options_,
|
|
|
|
internal_comparator_, meta);
|
2011-03-18 23:37:00 +01:00
|
|
|
s = iter->status();
|
|
|
|
delete iter;
|
|
|
|
if (s.ok()) {
|
2011-07-21 04:40:18 +02:00
|
|
|
Log(options_.info_log,
|
2013-11-13 06:02:03 +01:00
|
|
|
"Generated table #%lu: %lu keys, %lu bytes",
|
|
|
|
(unsigned long) output_number,
|
|
|
|
(unsigned long) current_entries,
|
|
|
|
(unsigned long) current_bytes);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
Status DBImpl::InstallCompactionResults(CompactionState* compact) {
|
|
|
|
mutex_.AssertHeld();
|
2012-10-19 23:00:53 +02:00
|
|
|
|
|
|
|
// paranoia: verify that the files that we started with
|
|
|
|
// still exist in the current version and in the same original level.
|
|
|
|
// This ensures that a concurrent compaction did not erroneously
|
|
|
|
// pick the same files to compact.
|
2013-03-06 22:28:54 +01:00
|
|
|
if (!versions_->VerifyCompactionFileConsistency(compact->compaction)) {
|
2012-10-19 23:00:53 +02:00
|
|
|
Log(options_.info_log, "Compaction %d@%d + %d@%d files aborted",
|
|
|
|
compact->compaction->num_input_files(0),
|
|
|
|
compact->compaction->level(),
|
|
|
|
compact->compaction->num_input_files(1),
|
|
|
|
compact->compaction->level() + 1);
|
2014-02-12 20:42:54 +01:00
|
|
|
return Status::Corruption("Compaction input files inconsistent");
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
|
|
|
|
2011-07-21 04:40:18 +02:00
|
|
|
Log(options_.info_log, "Compacted %d@%d + %d@%d files => %lld bytes",
|
2011-03-18 23:37:00 +01:00
|
|
|
compact->compaction->num_input_files(0),
|
|
|
|
compact->compaction->level(),
|
|
|
|
compact->compaction->num_input_files(1),
|
|
|
|
compact->compaction->level() + 1,
|
|
|
|
static_cast<long long>(compact->total_bytes));
|
|
|
|
|
|
|
|
// Add compaction outputs
|
|
|
|
compact->compaction->AddInputDeletions(compact->compaction->edit());
|
2011-04-21 00:48:11 +02:00
|
|
|
for (size_t i = 0; i < compact->outputs.size(); i++) {
|
2011-03-18 23:37:00 +01:00
|
|
|
const CompactionState::Output& out = compact->outputs[i];
|
|
|
|
compact->compaction->edit()->AddFile(
|
2014-01-15 01:19:09 +01:00
|
|
|
compact->compaction->output_level(), out.number, out.file_size,
|
|
|
|
out.smallest, out.largest, out.smallest_seqno, out.largest_seqno);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
2014-01-27 20:02:21 +01:00
|
|
|
return versions_->LogAndApply(compact->compaction->edit(), &mutex_,
|
|
|
|
db_directory_.get());
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
2012-11-27 06:16:21 +01:00
|
|
|
//
|
|
|
|
// Given a sequence number, return the sequence number of the
|
|
|
|
// earliest snapshot that this sequence number is visible in.
|
|
|
|
// The snapshots themselves are arranged in ascending order of
|
|
|
|
// sequence numbers.
|
|
|
|
// Employ a sequential search because the total number of
|
|
|
|
// snapshots are typically small.
|
|
|
|
inline SequenceNumber DBImpl::findEarliestVisibleSnapshot(
|
2013-03-21 23:59:47 +01:00
|
|
|
SequenceNumber in, std::vector<SequenceNumber>& snapshots,
|
|
|
|
SequenceNumber* prev_snapshot) {
|
2013-01-14 21:39:24 +01:00
|
|
|
SequenceNumber prev __attribute__((unused)) = 0;
|
2013-03-29 02:01:06 +01:00
|
|
|
for (const auto cur : snapshots) {
|
|
|
|
assert(prev <= cur);
|
|
|
|
if (cur >= in) {
|
2013-03-21 23:59:47 +01:00
|
|
|
*prev_snapshot = prev;
|
2013-03-29 02:01:06 +01:00
|
|
|
return cur;
|
2012-11-27 06:16:21 +01:00
|
|
|
}
|
2013-03-29 02:01:06 +01:00
|
|
|
prev = cur; // assignment
|
|
|
|
assert(prev);
|
2012-11-27 06:16:21 +01:00
|
|
|
}
|
|
|
|
Log(options_.info_log,
|
2013-11-13 06:02:03 +01:00
|
|
|
"Looking for seqid %lu but maxseqid is %lu",
|
|
|
|
(unsigned long)in,
|
|
|
|
(unsigned long)snapshots[snapshots.size()-1]);
|
2012-11-27 06:16:21 +01:00
|
|
|
assert(0);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2014-01-10 02:52:11 +01:00
|
|
|
Status DBImpl::ProcessKeyValueCompaction(
|
|
|
|
SequenceNumber visible_at_tip,
|
|
|
|
SequenceNumber earliest_snapshot,
|
|
|
|
SequenceNumber latest_snapshot,
|
|
|
|
DeletionState& deletion_state,
|
|
|
|
bool bottommost_level,
|
|
|
|
int64_t& imm_micros,
|
|
|
|
Iterator* input,
|
|
|
|
CompactionState* compact,
|
|
|
|
bool is_compaction_v2,
|
|
|
|
LogBuffer* log_buffer) {
|
|
|
|
size_t combined_idx = 0;
|
2011-03-18 23:37:00 +01:00
|
|
|
Status status;
|
2014-01-10 02:52:11 +01:00
|
|
|
std::string compaction_filter_value;
|
2011-03-18 23:37:00 +01:00
|
|
|
ParsedInternalKey ikey;
|
|
|
|
std::string current_user_key;
|
|
|
|
bool has_current_user_key = false;
|
2014-01-10 02:52:11 +01:00
|
|
|
std::vector<char> delete_key; // for compaction filter
|
2013-01-14 21:39:24 +01:00
|
|
|
SequenceNumber last_sequence_for_key __attribute__((unused)) =
|
|
|
|
kMaxSequenceNumber;
|
2012-11-27 06:16:21 +01:00
|
|
|
SequenceNumber visible_in_snapshot = kMaxSequenceNumber;
|
2013-08-20 22:35:28 +02:00
|
|
|
MergeHelper merge(user_comparator(), options_.merge_operator.get(),
|
2013-03-21 23:59:47 +01:00
|
|
|
options_.info_log.get(),
|
2014-03-25 01:57:13 +01:00
|
|
|
options_.min_partial_merge_operands,
|
2013-03-21 23:59:47 +01:00
|
|
|
false /* internal key corruption is expected */);
|
2013-08-13 19:56:20 +02:00
|
|
|
auto compaction_filter = options_.compaction_filter;
|
|
|
|
std::unique_ptr<CompactionFilter> compaction_filter_from_factory = nullptr;
|
|
|
|
if (!compaction_filter) {
|
2014-04-02 23:48:53 +02:00
|
|
|
auto context = compact->GetFilterContextV1();
|
2013-10-27 07:01:26 +01:00
|
|
|
compaction_filter_from_factory =
|
|
|
|
options_.compaction_filter_factory->CreateCompactionFilter(context);
|
2013-08-13 19:56:20 +02:00
|
|
|
compaction_filter = compaction_filter_from_factory.get();
|
|
|
|
}
|
2013-10-27 07:01:26 +01:00
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
for (; input->Valid() && !shutting_down_.Acquire_Load(); ) {
|
2011-04-12 21:38:58 +02:00
|
|
|
// Prioritize immutable compaction work
|
2013-09-13 23:38:37 +02:00
|
|
|
// TODO: remove memtable flush from normal compaction work
|
2013-02-15 20:53:17 +01:00
|
|
|
if (imm_.imm_flush_needed.NoBarrier_Load() != nullptr) {
|
2011-04-12 21:38:58 +02:00
|
|
|
const uint64_t imm_start = env_->NowMicros();
|
2013-11-07 20:31:56 +01:00
|
|
|
LogFlush(options_.info_log);
|
2011-04-12 21:38:58 +02:00
|
|
|
mutex_.Lock();
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
if (imm_.IsFlushPending()) {
|
2014-03-10 06:01:13 +01:00
|
|
|
FlushMemTableToOutputFile(nullptr, deletion_state, log_buffer);
|
2011-06-07 16:40:26 +02:00
|
|
|
bg_cv_.SignalAll(); // Wakeup MakeRoomForWrite() if necessary
|
2011-04-12 21:38:58 +02:00
|
|
|
}
|
|
|
|
mutex_.Unlock();
|
2014-03-10 06:01:13 +01:00
|
|
|
log_buffer->FlushBufferToLog();
|
2011-04-12 21:38:58 +02:00
|
|
|
imm_micros += (env_->NowMicros() - imm_start);
|
|
|
|
}
|
|
|
|
|
2014-01-10 02:52:11 +01:00
|
|
|
Slice key;
|
|
|
|
Slice value;
|
|
|
|
// If is_compaction_v2 is on, kv-pairs are reset to the prefix batch.
|
|
|
|
// This prefix batch should contain results after calling
|
|
|
|
// compaction_filter_v2.
|
|
|
|
//
|
|
|
|
// If is_compaction_v2 is off, this function will go through all the
|
|
|
|
// kv-pairs in input.
|
|
|
|
if (!is_compaction_v2) {
|
|
|
|
key = input->key();
|
|
|
|
value = input->value();
|
|
|
|
} else {
|
|
|
|
if (combined_idx >= compact->combined_key_buf_.size()) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
assert(combined_idx < compact->combined_key_buf_.size());
|
|
|
|
key = compact->combined_key_buf_[combined_idx];
|
|
|
|
value = compact->combined_value_buf_[combined_idx];
|
|
|
|
|
|
|
|
++combined_idx;
|
|
|
|
}
|
2013-04-18 06:30:21 +02:00
|
|
|
|
2011-05-21 04:17:43 +02:00
|
|
|
if (compact->compaction->ShouldStopBefore(key) &&
|
2013-02-15 20:53:17 +01:00
|
|
|
compact->builder != nullptr) {
|
2014-01-10 02:52:11 +01:00
|
|
|
status = FinishCompactionOutputFile(compact, input);
|
2011-03-22 19:32:49 +01:00
|
|
|
if (!status.ok()) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Handle key/value, add to state, etc.
|
2011-03-18 23:37:00 +01:00
|
|
|
bool drop = false;
|
[RocksDB] [MergeOperator] The new Merge Interface! Uses merge sequences.
Summary:
Here are the major changes to the Merge Interface. It has been expanded
to handle cases where the MergeOperator is not associative. It does so by stacking
up merge operations while scanning through the key history (i.e.: during Get() or
Compaction), until a valid Put/Delete/end-of-history is encountered; it then
applies all of the merge operations in the correct sequence starting with the
base/sentinel value.
I have also introduced an "AssociativeMerge" function which allows the user to
take advantage of associative merge operations (such as in the case of counters).
The implementation will always attempt to merge the operations/operands themselves
together when they are encountered, and will resort to the "stacking" method if
and only if the "associative-merge" fails.
This implementation is conjectured to allow MergeOperator to handle the general
case, while still providing the user with the ability to take advantage of certain
efficiencies in their own merge-operator / data-structure.
NOTE: This is a preliminary diff. This must still go through a lot of review,
revision, and testing. Feedback welcome!
Test Plan:
-This is a preliminary diff. I have only just begun testing/debugging it.
-I will be testing this with the existing MergeOperator use-cases and unit-tests
(counters, string-append, and redis-lists)
-I will be "desk-checking" and walking through the code with the help gdb.
-I will find a way of stress-testing the new interface / implementation using
db_bench, db_test, merge_test, and/or db_stress.
-I will ensure that my tests cover all cases: Get-Memtable,
Get-Immutable-Memtable, Get-from-Disk, Iterator-Range-Scan, Flush-Memtable-to-L0,
Compaction-L0-L1, Compaction-Ln-L(n+1), Put/Delete found, Put/Delete not-found,
end-of-history, end-of-file, etc.
-A lot of feedback from the reviewers.
Reviewers: haobo, dhruba, zshao, emayanke
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11499
2013-08-06 05:14:32 +02:00
|
|
|
bool current_entry_is_merging = false;
|
2011-03-18 23:37:00 +01:00
|
|
|
if (!ParseInternalKey(key, &ikey)) {
|
|
|
|
// Do not hide error keys
|
2013-03-21 23:59:47 +01:00
|
|
|
// TODO: error key stays in db forever? Figure out the intention/rationale
|
|
|
|
// v10 error v8 : we cannot hide v8 even though it's pretty obvious.
|
2011-03-18 23:37:00 +01:00
|
|
|
current_user_key.clear();
|
|
|
|
has_current_user_key = false;
|
|
|
|
last_sequence_for_key = kMaxSequenceNumber;
|
2012-11-27 06:16:21 +01:00
|
|
|
visible_in_snapshot = kMaxSequenceNumber;
|
2011-03-18 23:37:00 +01:00
|
|
|
} else {
|
|
|
|
if (!has_current_user_key ||
|
|
|
|
user_comparator()->Compare(ikey.user_key,
|
2014-01-10 02:52:11 +01:00
|
|
|
Slice(current_user_key)) != 0) {
|
2011-03-18 23:37:00 +01:00
|
|
|
// First occurrence of this user key
|
|
|
|
current_user_key.assign(ikey.user_key.data(), ikey.user_key.size());
|
|
|
|
has_current_user_key = true;
|
|
|
|
last_sequence_for_key = kMaxSequenceNumber;
|
2012-11-27 06:16:21 +01:00
|
|
|
visible_in_snapshot = kMaxSequenceNumber;
|
2013-06-06 19:31:37 +02:00
|
|
|
// apply the compaction filter to the first occurrence of the user key
|
2014-01-10 02:52:11 +01:00
|
|
|
if (compaction_filter && !is_compaction_v2 &&
|
2013-06-06 19:31:37 +02:00
|
|
|
ikey.type == kTypeValue &&
|
|
|
|
(visible_at_tip || ikey.sequence > latest_snapshot)) {
|
|
|
|
// If the user has specified a compaction filter and the sequence
|
|
|
|
// number is greater than any external snapshot, then invoke the
|
|
|
|
// filter.
|
|
|
|
// If the return value of the compaction filter is true, replace
|
|
|
|
// the entry with a delete marker.
|
|
|
|
bool value_changed = false;
|
|
|
|
compaction_filter_value.clear();
|
|
|
|
bool to_delete =
|
2013-08-13 19:56:20 +02:00
|
|
|
compaction_filter->Filter(compact->compaction->level(),
|
2014-01-10 02:52:11 +01:00
|
|
|
ikey.user_key, value,
|
|
|
|
&compaction_filter_value,
|
|
|
|
&value_changed);
|
2013-06-06 19:31:37 +02:00
|
|
|
if (to_delete) {
|
|
|
|
// make a copy of the original key
|
|
|
|
delete_key.assign(key.data(), key.data() + key.size());
|
|
|
|
// convert it to a delete
|
|
|
|
UpdateInternalKey(&delete_key[0], delete_key.size(),
|
2014-01-10 02:52:11 +01:00
|
|
|
ikey.sequence, kTypeDeletion);
|
2013-06-06 19:31:37 +02:00
|
|
|
// anchor the key again
|
|
|
|
key = Slice(&delete_key[0], delete_key.size());
|
|
|
|
// needed because ikey is backed by key
|
|
|
|
ParseInternalKey(key, &ikey);
|
|
|
|
// no value associated with delete
|
|
|
|
value.clear();
|
2013-11-22 23:14:05 +01:00
|
|
|
RecordTick(options_.statistics.get(), COMPACTION_KEY_DROP_USER);
|
2013-06-06 19:31:37 +02:00
|
|
|
} else if (value_changed) {
|
|
|
|
value = compaction_filter_value;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
2012-11-27 06:16:21 +01:00
|
|
|
// If there are no snapshots, then this kv affect visibility at tip.
|
|
|
|
// Otherwise, search though all existing snapshots to find
|
|
|
|
// the earlist snapshot that is affected by this kv.
|
2013-03-21 23:59:47 +01:00
|
|
|
SequenceNumber prev_snapshot = 0; // 0 means no previous snapshot
|
|
|
|
SequenceNumber visible = visible_at_tip ?
|
|
|
|
visible_at_tip :
|
|
|
|
findEarliestVisibleSnapshot(ikey.sequence,
|
2014-01-10 02:52:11 +01:00
|
|
|
compact->existing_snapshots,
|
|
|
|
&prev_snapshot);
|
2012-11-27 06:16:21 +01:00
|
|
|
|
|
|
|
if (visible_in_snapshot == visible) {
|
|
|
|
// If the earliest snapshot is which this key is visible in
|
|
|
|
// is the same as the visibily of a previous instance of the
|
|
|
|
// same key, then this kv is not visible in any snapshot.
|
2011-03-18 23:37:00 +01:00
|
|
|
// Hidden by an newer entry for same user key
|
2013-03-21 23:59:47 +01:00
|
|
|
// TODO: why not > ?
|
2012-11-27 06:16:21 +01:00
|
|
|
assert(last_sequence_for_key >= ikey.sequence);
|
2011-03-18 23:37:00 +01:00
|
|
|
drop = true; // (A)
|
2013-11-22 23:14:05 +01:00
|
|
|
RecordTick(options_.statistics.get(), COMPACTION_KEY_DROP_NEWER_ENTRY);
|
2011-03-18 23:37:00 +01:00
|
|
|
} else if (ikey.type == kTypeDeletion &&
|
2014-01-10 02:52:11 +01:00
|
|
|
ikey.sequence <= earliest_snapshot &&
|
|
|
|
compact->compaction->IsBaseLevelForKey(ikey.user_key)) {
|
2011-03-18 23:37:00 +01:00
|
|
|
// For this user key:
|
|
|
|
// (1) there is no data in higher levels
|
|
|
|
// (2) data in lower levels will have larger sequence numbers
|
|
|
|
// (3) data in layers that are being compacted here and have
|
|
|
|
// smaller sequence numbers will be dropped in the next
|
|
|
|
// few iterations of this loop (by rule (A) above).
|
|
|
|
// Therefore this deletion marker is obsolete and can be dropped.
|
|
|
|
drop = true;
|
2013-11-22 23:14:05 +01:00
|
|
|
RecordTick(options_.statistics.get(), COMPACTION_KEY_DROP_OBSOLETE);
|
2013-03-21 23:59:47 +01:00
|
|
|
} else if (ikey.type == kTypeMerge) {
|
|
|
|
// We know the merge type entry is not hidden, otherwise we would
|
|
|
|
// have hit (A)
|
|
|
|
// We encapsulate the merge related state machine in a different
|
|
|
|
// object to minimize change to the existing flow. Turn out this
|
|
|
|
// logic could also be nicely re-used for memtable flush purge
|
|
|
|
// optimization in BuildTable.
|
2014-01-10 02:52:11 +01:00
|
|
|
int steps = 0;
|
|
|
|
merge.MergeUntil(input, prev_snapshot, bottommost_level,
|
|
|
|
options_.statistics.get(), &steps);
|
|
|
|
// Skip the Merge ops
|
|
|
|
combined_idx = combined_idx - 1 + steps;
|
|
|
|
|
[RocksDB] [MergeOperator] The new Merge Interface! Uses merge sequences.
Summary:
Here are the major changes to the Merge Interface. It has been expanded
to handle cases where the MergeOperator is not associative. It does so by stacking
up merge operations while scanning through the key history (i.e.: during Get() or
Compaction), until a valid Put/Delete/end-of-history is encountered; it then
applies all of the merge operations in the correct sequence starting with the
base/sentinel value.
I have also introduced an "AssociativeMerge" function which allows the user to
take advantage of associative merge operations (such as in the case of counters).
The implementation will always attempt to merge the operations/operands themselves
together when they are encountered, and will resort to the "stacking" method if
and only if the "associative-merge" fails.
This implementation is conjectured to allow MergeOperator to handle the general
case, while still providing the user with the ability to take advantage of certain
efficiencies in their own merge-operator / data-structure.
NOTE: This is a preliminary diff. This must still go through a lot of review,
revision, and testing. Feedback welcome!
Test Plan:
-This is a preliminary diff. I have only just begun testing/debugging it.
-I will be testing this with the existing MergeOperator use-cases and unit-tests
(counters, string-append, and redis-lists)
-I will be "desk-checking" and walking through the code with the help gdb.
-I will find a way of stress-testing the new interface / implementation using
db_bench, db_test, merge_test, and/or db_stress.
-I will ensure that my tests cover all cases: Get-Memtable,
Get-Immutable-Memtable, Get-from-Disk, Iterator-Range-Scan, Flush-Memtable-to-L0,
Compaction-L0-L1, Compaction-Ln-L(n+1), Put/Delete found, Put/Delete not-found,
end-of-history, end-of-file, etc.
-A lot of feedback from the reviewers.
Reviewers: haobo, dhruba, zshao, emayanke
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11499
2013-08-06 05:14:32 +02:00
|
|
|
current_entry_is_merging = true;
|
|
|
|
if (merge.IsSuccess()) {
|
|
|
|
// Successfully found Put/Delete/(end-of-key-range) while merging
|
|
|
|
// Get the merge result
|
|
|
|
key = merge.key();
|
|
|
|
ParseInternalKey(key, &ikey);
|
|
|
|
value = merge.value();
|
|
|
|
} else {
|
|
|
|
// Did not find a Put/Delete/(end-of-key-range) while merging
|
|
|
|
// We now have some stack of merge operands to write out.
|
|
|
|
// NOTE: key,value, and ikey are now referring to old entries.
|
|
|
|
// These will be correctly set below.
|
|
|
|
assert(!merge.keys().empty());
|
|
|
|
assert(merge.keys().size() == merge.values().size());
|
|
|
|
|
|
|
|
// Hack to make sure last_sequence_for_key is correct
|
|
|
|
ParseInternalKey(merge.keys().front(), &ikey);
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
last_sequence_for_key = ikey.sequence;
|
2012-11-27 06:16:21 +01:00
|
|
|
visible_in_snapshot = visible;
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
#if 0
|
2011-07-21 04:40:18 +02:00
|
|
|
Log(options_.info_log,
|
2011-03-18 23:37:00 +01:00
|
|
|
" Compact: %s, seq %d, type: %d %d, drop: %d, is_base: %d, "
|
2013-02-15 23:31:24 +01:00
|
|
|
"%d smallest_snapshot: %d level: %d bottommost %d",
|
2011-03-18 23:37:00 +01:00
|
|
|
ikey.user_key.ToString().c_str(),
|
2011-04-21 00:48:11 +02:00
|
|
|
(int)ikey.sequence, ikey.type, kTypeValue, drop,
|
2011-03-18 23:37:00 +01:00
|
|
|
compact->compaction->IsBaseLevelForKey(ikey.user_key),
|
2013-02-15 23:31:24 +01:00
|
|
|
(int)last_sequence_for_key, (int)earliest_snapshot,
|
|
|
|
compact->compaction->level(), bottommost_level);
|
2011-03-18 23:37:00 +01:00
|
|
|
#endif
|
|
|
|
|
|
|
|
if (!drop) {
|
[RocksDB] [MergeOperator] The new Merge Interface! Uses merge sequences.
Summary:
Here are the major changes to the Merge Interface. It has been expanded
to handle cases where the MergeOperator is not associative. It does so by stacking
up merge operations while scanning through the key history (i.e.: during Get() or
Compaction), until a valid Put/Delete/end-of-history is encountered; it then
applies all of the merge operations in the correct sequence starting with the
base/sentinel value.
I have also introduced an "AssociativeMerge" function which allows the user to
take advantage of associative merge operations (such as in the case of counters).
The implementation will always attempt to merge the operations/operands themselves
together when they are encountered, and will resort to the "stacking" method if
and only if the "associative-merge" fails.
This implementation is conjectured to allow MergeOperator to handle the general
case, while still providing the user with the ability to take advantage of certain
efficiencies in their own merge-operator / data-structure.
NOTE: This is a preliminary diff. This must still go through a lot of review,
revision, and testing. Feedback welcome!
Test Plan:
-This is a preliminary diff. I have only just begun testing/debugging it.
-I will be testing this with the existing MergeOperator use-cases and unit-tests
(counters, string-append, and redis-lists)
-I will be "desk-checking" and walking through the code with the help gdb.
-I will find a way of stress-testing the new interface / implementation using
db_bench, db_test, merge_test, and/or db_stress.
-I will ensure that my tests cover all cases: Get-Memtable,
Get-Immutable-Memtable, Get-from-Disk, Iterator-Range-Scan, Flush-Memtable-to-L0,
Compaction-L0-L1, Compaction-Ln-L(n+1), Put/Delete found, Put/Delete not-found,
end-of-history, end-of-file, etc.
-A lot of feedback from the reviewers.
Reviewers: haobo, dhruba, zshao, emayanke
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11499
2013-08-06 05:14:32 +02:00
|
|
|
// We may write a single key (e.g.: for Put/Delete or successful merge).
|
|
|
|
// Or we may instead have to write a sequence/list of keys.
|
|
|
|
// We have to write a sequence iff we have an unsuccessful merge
|
|
|
|
bool has_merge_list = current_entry_is_merging && !merge.IsSuccess();
|
|
|
|
const std::deque<std::string>* keys = nullptr;
|
|
|
|
const std::deque<std::string>* values = nullptr;
|
|
|
|
std::deque<std::string>::const_reverse_iterator key_iter;
|
|
|
|
std::deque<std::string>::const_reverse_iterator value_iter;
|
|
|
|
if (has_merge_list) {
|
|
|
|
keys = &merge.keys();
|
|
|
|
values = &merge.values();
|
|
|
|
key_iter = keys->rbegin(); // The back (*rbegin()) is the first key
|
|
|
|
value_iter = values->rbegin();
|
|
|
|
|
|
|
|
key = Slice(*key_iter);
|
|
|
|
value = Slice(*value_iter);
|
2013-02-15 23:31:24 +01:00
|
|
|
}
|
2013-02-21 04:22:13 +01:00
|
|
|
|
[RocksDB] [MergeOperator] The new Merge Interface! Uses merge sequences.
Summary:
Here are the major changes to the Merge Interface. It has been expanded
to handle cases where the MergeOperator is not associative. It does so by stacking
up merge operations while scanning through the key history (i.e.: during Get() or
Compaction), until a valid Put/Delete/end-of-history is encountered; it then
applies all of the merge operations in the correct sequence starting with the
base/sentinel value.
I have also introduced an "AssociativeMerge" function which allows the user to
take advantage of associative merge operations (such as in the case of counters).
The implementation will always attempt to merge the operations/operands themselves
together when they are encountered, and will resort to the "stacking" method if
and only if the "associative-merge" fails.
This implementation is conjectured to allow MergeOperator to handle the general
case, while still providing the user with the ability to take advantage of certain
efficiencies in their own merge-operator / data-structure.
NOTE: This is a preliminary diff. This must still go through a lot of review,
revision, and testing. Feedback welcome!
Test Plan:
-This is a preliminary diff. I have only just begun testing/debugging it.
-I will be testing this with the existing MergeOperator use-cases and unit-tests
(counters, string-append, and redis-lists)
-I will be "desk-checking" and walking through the code with the help gdb.
-I will find a way of stress-testing the new interface / implementation using
db_bench, db_test, merge_test, and/or db_stress.
-I will ensure that my tests cover all cases: Get-Memtable,
Get-Immutable-Memtable, Get-from-Disk, Iterator-Range-Scan, Flush-Memtable-to-L0,
Compaction-L0-L1, Compaction-Ln-L(n+1), Put/Delete found, Put/Delete not-found,
end-of-history, end-of-file, etc.
-A lot of feedback from the reviewers.
Reviewers: haobo, dhruba, zshao, emayanke
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11499
2013-08-06 05:14:32 +02:00
|
|
|
// If we have a list of keys to write, traverse the list.
|
|
|
|
// If we have a single key to write, simply write that key.
|
|
|
|
while (true) {
|
|
|
|
// Invariant: key,value,ikey will always be the next entry to write
|
|
|
|
char* kptr = (char*)key.data();
|
|
|
|
std::string kstr;
|
|
|
|
|
|
|
|
// Zeroing out the sequence number leads to better compression.
|
|
|
|
// If this is the bottommost level (no files in lower levels)
|
|
|
|
// and the earliest snapshot is larger than this seqno
|
|
|
|
// then we can squash the seqno to zero.
|
2014-01-27 22:53:22 +01:00
|
|
|
if (bottommost_level && ikey.sequence < earliest_snapshot &&
|
[RocksDB] [MergeOperator] The new Merge Interface! Uses merge sequences.
Summary:
Here are the major changes to the Merge Interface. It has been expanded
to handle cases where the MergeOperator is not associative. It does so by stacking
up merge operations while scanning through the key history (i.e.: during Get() or
Compaction), until a valid Put/Delete/end-of-history is encountered; it then
applies all of the merge operations in the correct sequence starting with the
base/sentinel value.
I have also introduced an "AssociativeMerge" function which allows the user to
take advantage of associative merge operations (such as in the case of counters).
The implementation will always attempt to merge the operations/operands themselves
together when they are encountered, and will resort to the "stacking" method if
and only if the "associative-merge" fails.
This implementation is conjectured to allow MergeOperator to handle the general
case, while still providing the user with the ability to take advantage of certain
efficiencies in their own merge-operator / data-structure.
NOTE: This is a preliminary diff. This must still go through a lot of review,
revision, and testing. Feedback welcome!
Test Plan:
-This is a preliminary diff. I have only just begun testing/debugging it.
-I will be testing this with the existing MergeOperator use-cases and unit-tests
(counters, string-append, and redis-lists)
-I will be "desk-checking" and walking through the code with the help gdb.
-I will find a way of stress-testing the new interface / implementation using
db_bench, db_test, merge_test, and/or db_stress.
-I will ensure that my tests cover all cases: Get-Memtable,
Get-Immutable-Memtable, Get-from-Disk, Iterator-Range-Scan, Flush-Memtable-to-L0,
Compaction-L0-L1, Compaction-Ln-L(n+1), Put/Delete found, Put/Delete not-found,
end-of-history, end-of-file, etc.
-A lot of feedback from the reviewers.
Reviewers: haobo, dhruba, zshao, emayanke
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11499
2013-08-06 05:14:32 +02:00
|
|
|
ikey.type != kTypeMerge) {
|
|
|
|
assert(ikey.type != kTypeDeletion);
|
|
|
|
// make a copy because updating in place would cause problems
|
|
|
|
// with the priority queue that is managing the input key iterator
|
|
|
|
kstr.assign(key.data(), key.size());
|
|
|
|
kptr = (char *)kstr.c_str();
|
|
|
|
UpdateInternalKey(kptr, key.size(), (uint64_t)0, ikey.type);
|
|
|
|
}
|
|
|
|
|
|
|
|
Slice newkey(kptr, key.size());
|
|
|
|
assert((key.clear(), 1)); // we do not need 'key' anymore
|
2013-03-06 06:53:28 +01:00
|
|
|
|
[RocksDB] [MergeOperator] The new Merge Interface! Uses merge sequences.
Summary:
Here are the major changes to the Merge Interface. It has been expanded
to handle cases where the MergeOperator is not associative. It does so by stacking
up merge operations while scanning through the key history (i.e.: during Get() or
Compaction), until a valid Put/Delete/end-of-history is encountered; it then
applies all of the merge operations in the correct sequence starting with the
base/sentinel value.
I have also introduced an "AssociativeMerge" function which allows the user to
take advantage of associative merge operations (such as in the case of counters).
The implementation will always attempt to merge the operations/operands themselves
together when they are encountered, and will resort to the "stacking" method if
and only if the "associative-merge" fails.
This implementation is conjectured to allow MergeOperator to handle the general
case, while still providing the user with the ability to take advantage of certain
efficiencies in their own merge-operator / data-structure.
NOTE: This is a preliminary diff. This must still go through a lot of review,
revision, and testing. Feedback welcome!
Test Plan:
-This is a preliminary diff. I have only just begun testing/debugging it.
-I will be testing this with the existing MergeOperator use-cases and unit-tests
(counters, string-append, and redis-lists)
-I will be "desk-checking" and walking through the code with the help gdb.
-I will find a way of stress-testing the new interface / implementation using
db_bench, db_test, merge_test, and/or db_stress.
-I will ensure that my tests cover all cases: Get-Memtable,
Get-Immutable-Memtable, Get-from-Disk, Iterator-Range-Scan, Flush-Memtable-to-L0,
Compaction-L0-L1, Compaction-Ln-L(n+1), Put/Delete found, Put/Delete not-found,
end-of-history, end-of-file, etc.
-A lot of feedback from the reviewers.
Reviewers: haobo, dhruba, zshao, emayanke
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11499
2013-08-06 05:14:32 +02:00
|
|
|
// Open output file if necessary
|
|
|
|
if (compact->builder == nullptr) {
|
|
|
|
status = OpenCompactionOutputFile(compact);
|
|
|
|
if (!status.ok()) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
2013-08-07 20:58:06 +02:00
|
|
|
|
|
|
|
SequenceNumber seqno = GetInternalKeySeqno(newkey);
|
[RocksDB] [MergeOperator] The new Merge Interface! Uses merge sequences.
Summary:
Here are the major changes to the Merge Interface. It has been expanded
to handle cases where the MergeOperator is not associative. It does so by stacking
up merge operations while scanning through the key history (i.e.: during Get() or
Compaction), until a valid Put/Delete/end-of-history is encountered; it then
applies all of the merge operations in the correct sequence starting with the
base/sentinel value.
I have also introduced an "AssociativeMerge" function which allows the user to
take advantage of associative merge operations (such as in the case of counters).
The implementation will always attempt to merge the operations/operands themselves
together when they are encountered, and will resort to the "stacking" method if
and only if the "associative-merge" fails.
This implementation is conjectured to allow MergeOperator to handle the general
case, while still providing the user with the ability to take advantage of certain
efficiencies in their own merge-operator / data-structure.
NOTE: This is a preliminary diff. This must still go through a lot of review,
revision, and testing. Feedback welcome!
Test Plan:
-This is a preliminary diff. I have only just begun testing/debugging it.
-I will be testing this with the existing MergeOperator use-cases and unit-tests
(counters, string-append, and redis-lists)
-I will be "desk-checking" and walking through the code with the help gdb.
-I will find a way of stress-testing the new interface / implementation using
db_bench, db_test, merge_test, and/or db_stress.
-I will ensure that my tests cover all cases: Get-Memtable,
Get-Immutable-Memtable, Get-from-Disk, Iterator-Range-Scan, Flush-Memtable-to-L0,
Compaction-L0-L1, Compaction-Ln-L(n+1), Put/Delete found, Put/Delete not-found,
end-of-history, end-of-file, etc.
-A lot of feedback from the reviewers.
Reviewers: haobo, dhruba, zshao, emayanke
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11499
2013-08-06 05:14:32 +02:00
|
|
|
if (compact->builder->NumEntries() == 0) {
|
|
|
|
compact->current_output()->smallest.DecodeFrom(newkey);
|
2013-08-07 20:58:06 +02:00
|
|
|
compact->current_output()->smallest_seqno = seqno;
|
|
|
|
} else {
|
|
|
|
compact->current_output()->smallest_seqno =
|
|
|
|
std::min(compact->current_output()->smallest_seqno, seqno);
|
[RocksDB] [MergeOperator] The new Merge Interface! Uses merge sequences.
Summary:
Here are the major changes to the Merge Interface. It has been expanded
to handle cases where the MergeOperator is not associative. It does so by stacking
up merge operations while scanning through the key history (i.e.: during Get() or
Compaction), until a valid Put/Delete/end-of-history is encountered; it then
applies all of the merge operations in the correct sequence starting with the
base/sentinel value.
I have also introduced an "AssociativeMerge" function which allows the user to
take advantage of associative merge operations (such as in the case of counters).
The implementation will always attempt to merge the operations/operands themselves
together when they are encountered, and will resort to the "stacking" method if
and only if the "associative-merge" fails.
This implementation is conjectured to allow MergeOperator to handle the general
case, while still providing the user with the ability to take advantage of certain
efficiencies in their own merge-operator / data-structure.
NOTE: This is a preliminary diff. This must still go through a lot of review,
revision, and testing. Feedback welcome!
Test Plan:
-This is a preliminary diff. I have only just begun testing/debugging it.
-I will be testing this with the existing MergeOperator use-cases and unit-tests
(counters, string-append, and redis-lists)
-I will be "desk-checking" and walking through the code with the help gdb.
-I will find a way of stress-testing the new interface / implementation using
db_bench, db_test, merge_test, and/or db_stress.
-I will ensure that my tests cover all cases: Get-Memtable,
Get-Immutable-Memtable, Get-from-Disk, Iterator-Range-Scan, Flush-Memtable-to-L0,
Compaction-L0-L1, Compaction-Ln-L(n+1), Put/Delete found, Put/Delete not-found,
end-of-history, end-of-file, etc.
-A lot of feedback from the reviewers.
Reviewers: haobo, dhruba, zshao, emayanke
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11499
2013-08-06 05:14:32 +02:00
|
|
|
}
|
|
|
|
compact->current_output()->largest.DecodeFrom(newkey);
|
|
|
|
compact->builder->Add(newkey, value);
|
2013-08-07 20:58:06 +02:00
|
|
|
compact->current_output()->largest_seqno =
|
|
|
|
std::max(compact->current_output()->largest_seqno, seqno);
|
[RocksDB] [MergeOperator] The new Merge Interface! Uses merge sequences.
Summary:
Here are the major changes to the Merge Interface. It has been expanded
to handle cases where the MergeOperator is not associative. It does so by stacking
up merge operations while scanning through the key history (i.e.: during Get() or
Compaction), until a valid Put/Delete/end-of-history is encountered; it then
applies all of the merge operations in the correct sequence starting with the
base/sentinel value.
I have also introduced an "AssociativeMerge" function which allows the user to
take advantage of associative merge operations (such as in the case of counters).
The implementation will always attempt to merge the operations/operands themselves
together when they are encountered, and will resort to the "stacking" method if
and only if the "associative-merge" fails.
This implementation is conjectured to allow MergeOperator to handle the general
case, while still providing the user with the ability to take advantage of certain
efficiencies in their own merge-operator / data-structure.
NOTE: This is a preliminary diff. This must still go through a lot of review,
revision, and testing. Feedback welcome!
Test Plan:
-This is a preliminary diff. I have only just begun testing/debugging it.
-I will be testing this with the existing MergeOperator use-cases and unit-tests
(counters, string-append, and redis-lists)
-I will be "desk-checking" and walking through the code with the help gdb.
-I will find a way of stress-testing the new interface / implementation using
db_bench, db_test, merge_test, and/or db_stress.
-I will ensure that my tests cover all cases: Get-Memtable,
Get-Immutable-Memtable, Get-from-Disk, Iterator-Range-Scan, Flush-Memtable-to-L0,
Compaction-L0-L1, Compaction-Ln-L(n+1), Put/Delete found, Put/Delete not-found,
end-of-history, end-of-file, etc.
-A lot of feedback from the reviewers.
Reviewers: haobo, dhruba, zshao, emayanke
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11499
2013-08-06 05:14:32 +02:00
|
|
|
|
|
|
|
// Close output file if it is big enough
|
|
|
|
if (compact->builder->FileSize() >=
|
|
|
|
compact->compaction->MaxOutputFileSize()) {
|
2014-01-10 02:52:11 +01:00
|
|
|
status = FinishCompactionOutputFile(compact, input);
|
[RocksDB] [MergeOperator] The new Merge Interface! Uses merge sequences.
Summary:
Here are the major changes to the Merge Interface. It has been expanded
to handle cases where the MergeOperator is not associative. It does so by stacking
up merge operations while scanning through the key history (i.e.: during Get() or
Compaction), until a valid Put/Delete/end-of-history is encountered; it then
applies all of the merge operations in the correct sequence starting with the
base/sentinel value.
I have also introduced an "AssociativeMerge" function which allows the user to
take advantage of associative merge operations (such as in the case of counters).
The implementation will always attempt to merge the operations/operands themselves
together when they are encountered, and will resort to the "stacking" method if
and only if the "associative-merge" fails.
This implementation is conjectured to allow MergeOperator to handle the general
case, while still providing the user with the ability to take advantage of certain
efficiencies in their own merge-operator / data-structure.
NOTE: This is a preliminary diff. This must still go through a lot of review,
revision, and testing. Feedback welcome!
Test Plan:
-This is a preliminary diff. I have only just begun testing/debugging it.
-I will be testing this with the existing MergeOperator use-cases and unit-tests
(counters, string-append, and redis-lists)
-I will be "desk-checking" and walking through the code with the help gdb.
-I will find a way of stress-testing the new interface / implementation using
db_bench, db_test, merge_test, and/or db_stress.
-I will ensure that my tests cover all cases: Get-Memtable,
Get-Immutable-Memtable, Get-from-Disk, Iterator-Range-Scan, Flush-Memtable-to-L0,
Compaction-L0-L1, Compaction-Ln-L(n+1), Put/Delete found, Put/Delete not-found,
end-of-history, end-of-file, etc.
-A lot of feedback from the reviewers.
Reviewers: haobo, dhruba, zshao, emayanke
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11499
2013-08-06 05:14:32 +02:00
|
|
|
if (!status.ok()) {
|
|
|
|
break;
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
[RocksDB] [MergeOperator] The new Merge Interface! Uses merge sequences.
Summary:
Here are the major changes to the Merge Interface. It has been expanded
to handle cases where the MergeOperator is not associative. It does so by stacking
up merge operations while scanning through the key history (i.e.: during Get() or
Compaction), until a valid Put/Delete/end-of-history is encountered; it then
applies all of the merge operations in the correct sequence starting with the
base/sentinel value.
I have also introduced an "AssociativeMerge" function which allows the user to
take advantage of associative merge operations (such as in the case of counters).
The implementation will always attempt to merge the operations/operands themselves
together when they are encountered, and will resort to the "stacking" method if
and only if the "associative-merge" fails.
This implementation is conjectured to allow MergeOperator to handle the general
case, while still providing the user with the ability to take advantage of certain
efficiencies in their own merge-operator / data-structure.
NOTE: This is a preliminary diff. This must still go through a lot of review,
revision, and testing. Feedback welcome!
Test Plan:
-This is a preliminary diff. I have only just begun testing/debugging it.
-I will be testing this with the existing MergeOperator use-cases and unit-tests
(counters, string-append, and redis-lists)
-I will be "desk-checking" and walking through the code with the help gdb.
-I will find a way of stress-testing the new interface / implementation using
db_bench, db_test, merge_test, and/or db_stress.
-I will ensure that my tests cover all cases: Get-Memtable,
Get-Immutable-Memtable, Get-from-Disk, Iterator-Range-Scan, Flush-Memtable-to-L0,
Compaction-L0-L1, Compaction-Ln-L(n+1), Put/Delete found, Put/Delete not-found,
end-of-history, end-of-file, etc.
-A lot of feedback from the reviewers.
Reviewers: haobo, dhruba, zshao, emayanke
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11499
2013-08-06 05:14:32 +02:00
|
|
|
// If we have a list of entries, move to next element
|
|
|
|
// If we only had one entry, then break the loop.
|
|
|
|
if (has_merge_list) {
|
|
|
|
++key_iter;
|
|
|
|
++value_iter;
|
|
|
|
|
|
|
|
// If at end of list
|
|
|
|
if (key_iter == keys->rend() || value_iter == values->rend()) {
|
|
|
|
// Sanity Check: if one ends, then both end
|
|
|
|
assert(key_iter == keys->rend() && value_iter == values->rend());
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Otherwise not at end of list. Update key, value, and ikey.
|
|
|
|
key = Slice(*key_iter);
|
|
|
|
value = Slice(*value_iter);
|
|
|
|
ParseInternalKey(key, &ikey);
|
|
|
|
|
|
|
|
} else{
|
|
|
|
// Only had one item to begin with (Put/Delete)
|
2011-03-18 23:37:00 +01:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-03-21 23:59:47 +01:00
|
|
|
// MergeUntil has moved input to the next entry
|
[RocksDB] [MergeOperator] The new Merge Interface! Uses merge sequences.
Summary:
Here are the major changes to the Merge Interface. It has been expanded
to handle cases where the MergeOperator is not associative. It does so by stacking
up merge operations while scanning through the key history (i.e.: during Get() or
Compaction), until a valid Put/Delete/end-of-history is encountered; it then
applies all of the merge operations in the correct sequence starting with the
base/sentinel value.
I have also introduced an "AssociativeMerge" function which allows the user to
take advantage of associative merge operations (such as in the case of counters).
The implementation will always attempt to merge the operations/operands themselves
together when they are encountered, and will resort to the "stacking" method if
and only if the "associative-merge" fails.
This implementation is conjectured to allow MergeOperator to handle the general
case, while still providing the user with the ability to take advantage of certain
efficiencies in their own merge-operator / data-structure.
NOTE: This is a preliminary diff. This must still go through a lot of review,
revision, and testing. Feedback welcome!
Test Plan:
-This is a preliminary diff. I have only just begun testing/debugging it.
-I will be testing this with the existing MergeOperator use-cases and unit-tests
(counters, string-append, and redis-lists)
-I will be "desk-checking" and walking through the code with the help gdb.
-I will find a way of stress-testing the new interface / implementation using
db_bench, db_test, merge_test, and/or db_stress.
-I will ensure that my tests cover all cases: Get-Memtable,
Get-Immutable-Memtable, Get-from-Disk, Iterator-Range-Scan, Flush-Memtable-to-L0,
Compaction-L0-L1, Compaction-Ln-L(n+1), Put/Delete found, Put/Delete not-found,
end-of-history, end-of-file, etc.
-A lot of feedback from the reviewers.
Reviewers: haobo, dhruba, zshao, emayanke
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11499
2013-08-06 05:14:32 +02:00
|
|
|
if (!current_entry_is_merging) {
|
2013-03-21 23:59:47 +01:00
|
|
|
input->Next();
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
2014-01-10 02:52:11 +01:00
|
|
|
return status;
|
|
|
|
}
|
|
|
|
|
|
|
|
void DBImpl::CallCompactionFilterV2(CompactionState* compact,
|
|
|
|
CompactionFilterV2* compaction_filter_v2) {
|
|
|
|
if (compact == nullptr || compaction_filter_v2 == nullptr) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
std::vector<Slice> user_key_buf;
|
|
|
|
for (const auto& key : compact->ikey_buf_) {
|
|
|
|
user_key_buf.emplace_back(key.user_key);
|
|
|
|
}
|
|
|
|
|
|
|
|
// If the user has specified a compaction filter and the sequence
|
|
|
|
// number is greater than any external snapshot, then invoke the
|
|
|
|
// filter.
|
|
|
|
// If the return value of the compaction filter is true, replace
|
|
|
|
// the entry with a delete marker.
|
|
|
|
compact->to_delete_buf_ = compaction_filter_v2->Filter(
|
|
|
|
compact->compaction->level(),
|
|
|
|
user_key_buf, compact->existing_value_buf_,
|
|
|
|
&compact->new_value_buf_,
|
|
|
|
&compact->value_changed_buf_);
|
|
|
|
|
|
|
|
// new_value_buf_.size() <= to_delete__buf_.size(). "=" iff all
|
|
|
|
// kv-pairs in this compaction run needs to be deleted.
|
|
|
|
assert(compact->to_delete_buf_.size() ==
|
|
|
|
compact->key_buf_.size());
|
|
|
|
assert(compact->to_delete_buf_.size() ==
|
|
|
|
compact->existing_value_buf_.size());
|
|
|
|
assert(compact->to_delete_buf_.size() ==
|
|
|
|
compact->value_changed_buf_.size());
|
|
|
|
|
|
|
|
int new_value_idx = 0;
|
|
|
|
for (unsigned int i = 0; i < compact->to_delete_buf_.size(); ++i) {
|
|
|
|
if (compact->to_delete_buf_[i]) {
|
|
|
|
// update the string buffer directly
|
|
|
|
// the Slice buffer points to the updated buffer
|
|
|
|
UpdateInternalKey(&compact->key_str_buf_[i][0],
|
|
|
|
compact->key_str_buf_[i].size(),
|
|
|
|
compact->ikey_buf_[i].sequence,
|
|
|
|
kTypeDeletion);
|
|
|
|
|
|
|
|
// no value associated with delete
|
|
|
|
compact->existing_value_buf_[i].clear();
|
|
|
|
RecordTick(options_.statistics.get(), COMPACTION_KEY_DROP_USER);
|
|
|
|
} else if (compact->value_changed_buf_[i]) {
|
|
|
|
compact->existing_value_buf_[i] =
|
|
|
|
Slice(compact->new_value_buf_[new_value_idx++]);
|
|
|
|
}
|
|
|
|
} // for
|
|
|
|
}
|
|
|
|
|
|
|
|
Status DBImpl::DoCompactionWork(CompactionState* compact,
|
|
|
|
DeletionState& deletion_state,
|
|
|
|
LogBuffer* log_buffer) {
|
|
|
|
assert(compact);
|
|
|
|
compact->CleanupBatchBuffer();
|
|
|
|
compact->CleanupMergedBuffer();
|
2014-03-25 19:34:31 +01:00
|
|
|
bool prefix_initialized = false;
|
2014-01-10 02:52:11 +01:00
|
|
|
|
|
|
|
int64_t imm_micros = 0; // Micros spent doing imm_ compactions
|
|
|
|
Log(options_.info_log,
|
|
|
|
"Compacting %d@%d + %d@%d files, score %.2f slots available %d",
|
|
|
|
compact->compaction->num_input_files(0),
|
|
|
|
compact->compaction->level(),
|
|
|
|
compact->compaction->num_input_files(1),
|
|
|
|
compact->compaction->output_level(),
|
|
|
|
compact->compaction->score(),
|
|
|
|
options_.max_background_compactions - bg_compaction_scheduled_);
|
|
|
|
char scratch[2345];
|
|
|
|
compact->compaction->Summary(scratch, sizeof(scratch));
|
|
|
|
Log(options_.info_log, "Compaction start summary: %s\n", scratch);
|
|
|
|
|
|
|
|
assert(versions_->current()->NumLevelFiles(compact->compaction->level()) > 0);
|
|
|
|
assert(compact->builder == nullptr);
|
|
|
|
assert(!compact->outfile);
|
|
|
|
|
|
|
|
SequenceNumber visible_at_tip = 0;
|
|
|
|
SequenceNumber earliest_snapshot;
|
|
|
|
SequenceNumber latest_snapshot = 0;
|
|
|
|
snapshots_.getAll(compact->existing_snapshots);
|
|
|
|
if (compact->existing_snapshots.size() == 0) {
|
|
|
|
// optimize for fast path if there are no snapshots
|
|
|
|
visible_at_tip = versions_->LastSequence();
|
|
|
|
earliest_snapshot = visible_at_tip;
|
|
|
|
} else {
|
|
|
|
latest_snapshot = compact->existing_snapshots.back();
|
|
|
|
// Add the current seqno as the 'latest' virtual
|
|
|
|
// snapshot to the end of this list.
|
|
|
|
compact->existing_snapshots.push_back(versions_->LastSequence());
|
|
|
|
earliest_snapshot = compact->existing_snapshots[0];
|
|
|
|
}
|
|
|
|
|
|
|
|
// Is this compaction producing files at the bottommost level?
|
|
|
|
bool bottommost_level = compact->compaction->BottomMostLevel();
|
|
|
|
|
|
|
|
// Allocate the output file numbers before we release the lock
|
|
|
|
AllocateCompactionOutputFileNumbers(compact);
|
|
|
|
|
|
|
|
// Release mutex while we're actually doing the compaction work
|
|
|
|
mutex_.Unlock();
|
|
|
|
|
|
|
|
const uint64_t start_micros = env_->NowMicros();
|
|
|
|
unique_ptr<Iterator> input(versions_->MakeInputIterator(compact->compaction));
|
|
|
|
input->SeekToFirst();
|
|
|
|
shared_ptr<Iterator> backup_input(
|
|
|
|
versions_->MakeInputIterator(compact->compaction));
|
|
|
|
backup_input->SeekToFirst();
|
|
|
|
|
|
|
|
Status status;
|
|
|
|
ParsedInternalKey ikey;
|
|
|
|
std::unique_ptr<CompactionFilterV2> compaction_filter_from_factory_v2
|
|
|
|
= nullptr;
|
|
|
|
auto context = compact->GetFilterContext();
|
|
|
|
compaction_filter_from_factory_v2 =
|
|
|
|
options_.compaction_filter_factory_v2->CreateCompactionFilterV2(context);
|
|
|
|
auto compaction_filter_v2 =
|
|
|
|
compaction_filter_from_factory_v2.get();
|
|
|
|
|
|
|
|
// temp_backup_input always point to the start of the current buffer
|
|
|
|
// temp_backup_input = backup_input;
|
|
|
|
// iterate through input,
|
|
|
|
// 1) buffer ineligible keys and value keys into 2 separate buffers;
|
|
|
|
// 2) send value_buffer to compaction filter and alternate the values;
|
|
|
|
// 3) merge value_buffer with ineligible_value_buffer;
|
|
|
|
// 4) run the modified "compaction" using the old for loop.
|
|
|
|
if (compaction_filter_v2) {
|
|
|
|
for (; backup_input->Valid() && !shutting_down_.Acquire_Load(); ) {
|
|
|
|
// Prioritize immutable compaction work
|
|
|
|
if (imm_.imm_flush_needed.NoBarrier_Load() != nullptr) {
|
|
|
|
const uint64_t imm_start = env_->NowMicros();
|
|
|
|
LogFlush(options_.info_log);
|
|
|
|
mutex_.Lock();
|
|
|
|
if (imm_.IsFlushPending()) {
|
|
|
|
FlushMemTableToOutputFile(nullptr, deletion_state, log_buffer);
|
|
|
|
bg_cv_.SignalAll(); // Wakeup MakeRoomForWrite() if necessary
|
|
|
|
}
|
|
|
|
mutex_.Unlock();
|
|
|
|
imm_micros += (env_->NowMicros() - imm_start);
|
|
|
|
}
|
|
|
|
|
|
|
|
Slice key = backup_input->key();
|
|
|
|
Slice value = backup_input->value();
|
|
|
|
|
|
|
|
const SliceTransform* transformer =
|
|
|
|
options_.compaction_filter_factory_v2->GetPrefixExtractor();
|
|
|
|
std::string key_prefix = transformer->Transform(key).ToString();
|
2014-03-25 19:34:31 +01:00
|
|
|
if (!prefix_initialized) {
|
2014-01-10 02:52:11 +01:00
|
|
|
compact->cur_prefix_ = key_prefix;
|
2014-03-25 19:34:31 +01:00
|
|
|
prefix_initialized = true;
|
2014-01-10 02:52:11 +01:00
|
|
|
}
|
|
|
|
if (!ParseInternalKey(key, &ikey)) {
|
|
|
|
// log error
|
|
|
|
Log(options_.info_log, "Failed to parse key: %s",
|
|
|
|
key.ToString().c_str());
|
|
|
|
continue;
|
|
|
|
} else {
|
|
|
|
// If the prefix remains the same, keep buffering
|
|
|
|
if (key_prefix == compact->cur_prefix_) {
|
|
|
|
// Apply the compaction filter V2 to all the kv pairs sharing
|
|
|
|
// the same prefix
|
|
|
|
if (ikey.type == kTypeValue &&
|
|
|
|
(visible_at_tip || ikey.sequence > latest_snapshot)) {
|
|
|
|
// Buffer all keys sharing the same prefix for CompactionFilterV2
|
|
|
|
// Iterate through keys to check prefix
|
|
|
|
compact->BufferKeyValueSlices(key, value);
|
|
|
|
} else {
|
|
|
|
// buffer ineligible keys
|
|
|
|
compact->BufferOtherKeyValueSlices(key, value);
|
|
|
|
}
|
|
|
|
backup_input->Next();
|
|
|
|
continue;
|
|
|
|
// finish changing values for eligible keys
|
|
|
|
} else {
|
|
|
|
// Now prefix changes, this batch is done.
|
|
|
|
// Call compaction filter on the buffered values to change the value
|
|
|
|
if (compact->key_buf_.size() > 0) {
|
|
|
|
CallCompactionFilterV2(compact, compaction_filter_v2);
|
|
|
|
}
|
|
|
|
compact->cur_prefix_ = key_prefix;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Merge this batch of data (values + ineligible keys)
|
|
|
|
compact->MergeKeyValueSliceBuffer(&internal_comparator_);
|
|
|
|
|
|
|
|
// Done buffering for the current prefix. Spit it out to disk
|
|
|
|
// Now just iterate through all the kv-pairs
|
|
|
|
status = ProcessKeyValueCompaction(
|
|
|
|
visible_at_tip,
|
|
|
|
earliest_snapshot,
|
|
|
|
latest_snapshot,
|
|
|
|
deletion_state,
|
|
|
|
bottommost_level,
|
|
|
|
imm_micros,
|
|
|
|
input.get(),
|
|
|
|
compact,
|
|
|
|
true,
|
|
|
|
log_buffer);
|
|
|
|
|
|
|
|
if (!status.ok()) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
// After writing the kv-pairs, we can safely remove the reference
|
|
|
|
// to the string buffer and clean them up
|
|
|
|
compact->CleanupBatchBuffer();
|
|
|
|
compact->CleanupMergedBuffer();
|
|
|
|
// Buffer the key that triggers the mismatch in prefix
|
|
|
|
if (ikey.type == kTypeValue &&
|
|
|
|
(visible_at_tip || ikey.sequence > latest_snapshot)) {
|
|
|
|
compact->BufferKeyValueSlices(key, value);
|
|
|
|
} else {
|
|
|
|
compact->BufferOtherKeyValueSlices(key, value);
|
|
|
|
}
|
|
|
|
backup_input->Next();
|
|
|
|
if (!backup_input->Valid()) {
|
|
|
|
// If this is the single last value, we need to merge it.
|
|
|
|
if (compact->key_buf_.size() > 0) {
|
|
|
|
CallCompactionFilterV2(compact, compaction_filter_v2);
|
|
|
|
}
|
|
|
|
compact->MergeKeyValueSliceBuffer(&internal_comparator_);
|
|
|
|
|
|
|
|
status = ProcessKeyValueCompaction(
|
|
|
|
visible_at_tip,
|
|
|
|
earliest_snapshot,
|
|
|
|
latest_snapshot,
|
|
|
|
deletion_state,
|
|
|
|
bottommost_level,
|
|
|
|
imm_micros,
|
|
|
|
input.get(),
|
|
|
|
compact,
|
|
|
|
true,
|
|
|
|
log_buffer);
|
|
|
|
|
|
|
|
compact->CleanupBatchBuffer();
|
|
|
|
compact->CleanupMergedBuffer();
|
|
|
|
}
|
|
|
|
} // done processing all prefix batches
|
|
|
|
// finish the last batch
|
|
|
|
if (compact->key_buf_.size() > 0) {
|
|
|
|
CallCompactionFilterV2(compact, compaction_filter_v2);
|
|
|
|
}
|
|
|
|
compact->MergeKeyValueSliceBuffer(&internal_comparator_);
|
|
|
|
status = ProcessKeyValueCompaction(
|
|
|
|
visible_at_tip,
|
|
|
|
earliest_snapshot,
|
|
|
|
latest_snapshot,
|
|
|
|
deletion_state,
|
|
|
|
bottommost_level,
|
|
|
|
imm_micros,
|
|
|
|
input.get(),
|
|
|
|
compact,
|
|
|
|
true,
|
|
|
|
log_buffer);
|
|
|
|
} // checking for compaction filter v2
|
|
|
|
|
|
|
|
if (!compaction_filter_v2) {
|
|
|
|
status = ProcessKeyValueCompaction(
|
|
|
|
visible_at_tip,
|
|
|
|
earliest_snapshot,
|
|
|
|
latest_snapshot,
|
|
|
|
deletion_state,
|
|
|
|
bottommost_level,
|
|
|
|
imm_micros,
|
|
|
|
input.get(),
|
|
|
|
compact,
|
|
|
|
false,
|
|
|
|
log_buffer);
|
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
if (status.ok() && shutting_down_.Acquire_Load()) {
|
2014-02-12 20:42:54 +01:00
|
|
|
status = Status::ShutdownInProgress(
|
|
|
|
"Database shutdown started during compaction");
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
2013-02-15 20:53:17 +01:00
|
|
|
if (status.ok() && compact->builder != nullptr) {
|
2013-01-20 11:07:13 +01:00
|
|
|
status = FinishCompactionOutputFile(compact, input.get());
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
if (status.ok()) {
|
|
|
|
status = input->status();
|
|
|
|
}
|
2013-01-20 11:07:13 +01:00
|
|
|
input.reset();
|
2011-03-18 23:37:00 +01:00
|
|
|
|
2014-01-27 20:02:21 +01:00
|
|
|
if (!options_.disableDataSync) {
|
|
|
|
db_directory_->Fsync();
|
|
|
|
}
|
2014-01-30 05:40:41 +01:00
|
|
|
|
|
|
|
InternalStats::CompactionStats stats;
|
2011-04-12 21:38:58 +02:00
|
|
|
stats.micros = env_->NowMicros() - start_micros - imm_micros;
|
2014-01-17 21:46:06 +01:00
|
|
|
MeasureTime(options_.statistics.get(), COMPACTION_TIME, stats.micros);
|
2012-10-23 19:34:09 +02:00
|
|
|
stats.files_in_leveln = compact->compaction->num_input_files(0);
|
|
|
|
stats.files_in_levelnp1 = compact->compaction->num_input_files(1);
|
2013-01-17 19:04:45 +01:00
|
|
|
|
|
|
|
int num_output_files = compact->outputs.size();
|
2013-02-15 20:53:17 +01:00
|
|
|
if (compact->builder != nullptr) {
|
2013-11-16 12:21:34 +01:00
|
|
|
// An error occurred so ignore the last output.
|
2013-01-17 19:04:45 +01:00
|
|
|
assert(num_output_files > 0);
|
|
|
|
--num_output_files;
|
|
|
|
}
|
|
|
|
stats.files_out_levelnp1 = num_output_files;
|
2012-10-23 19:34:09 +02:00
|
|
|
|
Add monitoring for universal compaction and add counters for compaction IO
Summary:
Adds these counters
{ WAL_FILE_SYNCED, "rocksdb.wal.synced" }
number of writes that request a WAL sync
{ WAL_FILE_BYTES, "rocksdb.wal.bytes" },
number of bytes written to the WAL
{ WRITE_DONE_BY_SELF, "rocksdb.write.self" },
number of writes processed by the calling thread
{ WRITE_DONE_BY_OTHER, "rocksdb.write.other" },
number of writes not processed by the calling thread. Instead these were
processed by the current holder of the write lock
{ WRITE_WITH_WAL, "rocksdb.write.wal" },
number of writes that request WAL logging
{ COMPACT_READ_BYTES, "rocksdb.compact.read.bytes" },
number of bytes read during compaction
{ COMPACT_WRITE_BYTES, "rocksdb.compact.write.bytes" },
number of bytes written during compaction
Per-interval stats output was updated with WAL stats and correct stats for universal compaction
including a correct value for write-amplification. It now looks like:
Compactions
Level Files Size(MB) Score Time(sec) Read(MB) Write(MB) Rn(MB) Rnp1(MB) Wnew(MB) RW-Amplify Read(MB/s) Write(MB/s) Rn Rnp1 Wnp1 NewW Count Ln-stall Stall-cnt
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
0 7 464 46.4 281 3411 3875 3411 0 3875 2.1 12.1 13.8 621 0 240 240 628 0.0 0
Uptime(secs): 310.8 total, 2.0 interval
Writes cumulative: 9999999 total, 9999999 batches, 1.0 per batch, 1.22 ingest GB
WAL cumulative: 9999999 WAL writes, 9999999 WAL syncs, 1.00 writes per sync, 1.22 GB written
Compaction IO cumulative (GB): 1.22 new, 3.33 read, 3.78 write, 7.12 read+write
Compaction IO cumulative (MB/sec): 4.0 new, 11.0 read, 12.5 write, 23.4 read+write
Amplification cumulative: 4.1 write, 6.8 compaction
Writes interval: 100000 total, 100000 batches, 1.0 per batch, 12.5 ingest MB
WAL interval: 100000 WAL writes, 100000 WAL syncs, 1.00 writes per sync, 0.01 MB written
Compaction IO interval (MB): 12.49 new, 14.98 read, 21.50 write, 36.48 read+write
Compaction IO interval (MB/sec): 6.4 new, 7.6 read, 11.0 write, 18.6 read+write
Amplification interval: 101.7 write, 102.9 compaction
Stalls(secs): 142.924 level0_slowdown, 0.000 level0_numfiles, 0.805 memtable_compaction, 0.000 leveln_slowdown
Stalls(count): 132461 level0_slowdown, 0 level0_numfiles, 3 memtable_compaction, 0 leveln_slowdown
Task ID: #3329644, #3301695
Blame Rev:
Test Plan:
Revert Plan:
Database Impact:
Memcache Impact:
Other Notes:
EImportant:
- begin *PUBLIC* platform impact section -
Bugzilla: #
- end platform impact -
Reviewers: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14583
2013-12-09 22:43:34 +01:00
|
|
|
for (int i = 0; i < compact->compaction->num_input_files(0); i++) {
|
2012-10-23 19:34:09 +02:00
|
|
|
stats.bytes_readn += compact->compaction->input(0, i)->file_size;
|
Add monitoring for universal compaction and add counters for compaction IO
Summary:
Adds these counters
{ WAL_FILE_SYNCED, "rocksdb.wal.synced" }
number of writes that request a WAL sync
{ WAL_FILE_BYTES, "rocksdb.wal.bytes" },
number of bytes written to the WAL
{ WRITE_DONE_BY_SELF, "rocksdb.write.self" },
number of writes processed by the calling thread
{ WRITE_DONE_BY_OTHER, "rocksdb.write.other" },
number of writes not processed by the calling thread. Instead these were
processed by the current holder of the write lock
{ WRITE_WITH_WAL, "rocksdb.write.wal" },
number of writes that request WAL logging
{ COMPACT_READ_BYTES, "rocksdb.compact.read.bytes" },
number of bytes read during compaction
{ COMPACT_WRITE_BYTES, "rocksdb.compact.write.bytes" },
number of bytes written during compaction
Per-interval stats output was updated with WAL stats and correct stats for universal compaction
including a correct value for write-amplification. It now looks like:
Compactions
Level Files Size(MB) Score Time(sec) Read(MB) Write(MB) Rn(MB) Rnp1(MB) Wnew(MB) RW-Amplify Read(MB/s) Write(MB/s) Rn Rnp1 Wnp1 NewW Count Ln-stall Stall-cnt
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
0 7 464 46.4 281 3411 3875 3411 0 3875 2.1 12.1 13.8 621 0 240 240 628 0.0 0
Uptime(secs): 310.8 total, 2.0 interval
Writes cumulative: 9999999 total, 9999999 batches, 1.0 per batch, 1.22 ingest GB
WAL cumulative: 9999999 WAL writes, 9999999 WAL syncs, 1.00 writes per sync, 1.22 GB written
Compaction IO cumulative (GB): 1.22 new, 3.33 read, 3.78 write, 7.12 read+write
Compaction IO cumulative (MB/sec): 4.0 new, 11.0 read, 12.5 write, 23.4 read+write
Amplification cumulative: 4.1 write, 6.8 compaction
Writes interval: 100000 total, 100000 batches, 1.0 per batch, 12.5 ingest MB
WAL interval: 100000 WAL writes, 100000 WAL syncs, 1.00 writes per sync, 0.01 MB written
Compaction IO interval (MB): 12.49 new, 14.98 read, 21.50 write, 36.48 read+write
Compaction IO interval (MB/sec): 6.4 new, 7.6 read, 11.0 write, 18.6 read+write
Amplification interval: 101.7 write, 102.9 compaction
Stalls(secs): 142.924 level0_slowdown, 0.000 level0_numfiles, 0.805 memtable_compaction, 0.000 leveln_slowdown
Stalls(count): 132461 level0_slowdown, 0 level0_numfiles, 3 memtable_compaction, 0 leveln_slowdown
Task ID: #3329644, #3301695
Blame Rev:
Test Plan:
Revert Plan:
Database Impact:
Memcache Impact:
Other Notes:
EImportant:
- begin *PUBLIC* platform impact section -
Bugzilla: #
- end platform impact -
Reviewers: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14583
2013-12-09 22:43:34 +01:00
|
|
|
RecordTick(options_.statistics.get(), COMPACT_READ_BYTES,
|
|
|
|
compact->compaction->input(0, i)->file_size);
|
|
|
|
}
|
2012-10-23 19:34:09 +02:00
|
|
|
|
Add monitoring for universal compaction and add counters for compaction IO
Summary:
Adds these counters
{ WAL_FILE_SYNCED, "rocksdb.wal.synced" }
number of writes that request a WAL sync
{ WAL_FILE_BYTES, "rocksdb.wal.bytes" },
number of bytes written to the WAL
{ WRITE_DONE_BY_SELF, "rocksdb.write.self" },
number of writes processed by the calling thread
{ WRITE_DONE_BY_OTHER, "rocksdb.write.other" },
number of writes not processed by the calling thread. Instead these were
processed by the current holder of the write lock
{ WRITE_WITH_WAL, "rocksdb.write.wal" },
number of writes that request WAL logging
{ COMPACT_READ_BYTES, "rocksdb.compact.read.bytes" },
number of bytes read during compaction
{ COMPACT_WRITE_BYTES, "rocksdb.compact.write.bytes" },
number of bytes written during compaction
Per-interval stats output was updated with WAL stats and correct stats for universal compaction
including a correct value for write-amplification. It now looks like:
Compactions
Level Files Size(MB) Score Time(sec) Read(MB) Write(MB) Rn(MB) Rnp1(MB) Wnew(MB) RW-Amplify Read(MB/s) Write(MB/s) Rn Rnp1 Wnp1 NewW Count Ln-stall Stall-cnt
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
0 7 464 46.4 281 3411 3875 3411 0 3875 2.1 12.1 13.8 621 0 240 240 628 0.0 0
Uptime(secs): 310.8 total, 2.0 interval
Writes cumulative: 9999999 total, 9999999 batches, 1.0 per batch, 1.22 ingest GB
WAL cumulative: 9999999 WAL writes, 9999999 WAL syncs, 1.00 writes per sync, 1.22 GB written
Compaction IO cumulative (GB): 1.22 new, 3.33 read, 3.78 write, 7.12 read+write
Compaction IO cumulative (MB/sec): 4.0 new, 11.0 read, 12.5 write, 23.4 read+write
Amplification cumulative: 4.1 write, 6.8 compaction
Writes interval: 100000 total, 100000 batches, 1.0 per batch, 12.5 ingest MB
WAL interval: 100000 WAL writes, 100000 WAL syncs, 1.00 writes per sync, 0.01 MB written
Compaction IO interval (MB): 12.49 new, 14.98 read, 21.50 write, 36.48 read+write
Compaction IO interval (MB/sec): 6.4 new, 7.6 read, 11.0 write, 18.6 read+write
Amplification interval: 101.7 write, 102.9 compaction
Stalls(secs): 142.924 level0_slowdown, 0.000 level0_numfiles, 0.805 memtable_compaction, 0.000 leveln_slowdown
Stalls(count): 132461 level0_slowdown, 0 level0_numfiles, 3 memtable_compaction, 0 leveln_slowdown
Task ID: #3329644, #3301695
Blame Rev:
Test Plan:
Revert Plan:
Database Impact:
Memcache Impact:
Other Notes:
EImportant:
- begin *PUBLIC* platform impact section -
Bugzilla: #
- end platform impact -
Reviewers: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14583
2013-12-09 22:43:34 +01:00
|
|
|
for (int i = 0; i < compact->compaction->num_input_files(1); i++) {
|
2012-10-23 19:34:09 +02:00
|
|
|
stats.bytes_readnp1 += compact->compaction->input(1, i)->file_size;
|
Add monitoring for universal compaction and add counters for compaction IO
Summary:
Adds these counters
{ WAL_FILE_SYNCED, "rocksdb.wal.synced" }
number of writes that request a WAL sync
{ WAL_FILE_BYTES, "rocksdb.wal.bytes" },
number of bytes written to the WAL
{ WRITE_DONE_BY_SELF, "rocksdb.write.self" },
number of writes processed by the calling thread
{ WRITE_DONE_BY_OTHER, "rocksdb.write.other" },
number of writes not processed by the calling thread. Instead these were
processed by the current holder of the write lock
{ WRITE_WITH_WAL, "rocksdb.write.wal" },
number of writes that request WAL logging
{ COMPACT_READ_BYTES, "rocksdb.compact.read.bytes" },
number of bytes read during compaction
{ COMPACT_WRITE_BYTES, "rocksdb.compact.write.bytes" },
number of bytes written during compaction
Per-interval stats output was updated with WAL stats and correct stats for universal compaction
including a correct value for write-amplification. It now looks like:
Compactions
Level Files Size(MB) Score Time(sec) Read(MB) Write(MB) Rn(MB) Rnp1(MB) Wnew(MB) RW-Amplify Read(MB/s) Write(MB/s) Rn Rnp1 Wnp1 NewW Count Ln-stall Stall-cnt
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
0 7 464 46.4 281 3411 3875 3411 0 3875 2.1 12.1 13.8 621 0 240 240 628 0.0 0
Uptime(secs): 310.8 total, 2.0 interval
Writes cumulative: 9999999 total, 9999999 batches, 1.0 per batch, 1.22 ingest GB
WAL cumulative: 9999999 WAL writes, 9999999 WAL syncs, 1.00 writes per sync, 1.22 GB written
Compaction IO cumulative (GB): 1.22 new, 3.33 read, 3.78 write, 7.12 read+write
Compaction IO cumulative (MB/sec): 4.0 new, 11.0 read, 12.5 write, 23.4 read+write
Amplification cumulative: 4.1 write, 6.8 compaction
Writes interval: 100000 total, 100000 batches, 1.0 per batch, 12.5 ingest MB
WAL interval: 100000 WAL writes, 100000 WAL syncs, 1.00 writes per sync, 0.01 MB written
Compaction IO interval (MB): 12.49 new, 14.98 read, 21.50 write, 36.48 read+write
Compaction IO interval (MB/sec): 6.4 new, 7.6 read, 11.0 write, 18.6 read+write
Amplification interval: 101.7 write, 102.9 compaction
Stalls(secs): 142.924 level0_slowdown, 0.000 level0_numfiles, 0.805 memtable_compaction, 0.000 leveln_slowdown
Stalls(count): 132461 level0_slowdown, 0 level0_numfiles, 3 memtable_compaction, 0 leveln_slowdown
Task ID: #3329644, #3301695
Blame Rev:
Test Plan:
Revert Plan:
Database Impact:
Memcache Impact:
Other Notes:
EImportant:
- begin *PUBLIC* platform impact section -
Bugzilla: #
- end platform impact -
Reviewers: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14583
2013-12-09 22:43:34 +01:00
|
|
|
RecordTick(options_.statistics.get(), COMPACT_READ_BYTES,
|
|
|
|
compact->compaction->input(1, i)->file_size);
|
|
|
|
}
|
2012-10-23 19:34:09 +02:00
|
|
|
|
2013-01-17 19:04:45 +01:00
|
|
|
for (int i = 0; i < num_output_files; i++) {
|
2011-04-12 21:38:58 +02:00
|
|
|
stats.bytes_written += compact->outputs[i].file_size;
|
Add monitoring for universal compaction and add counters for compaction IO
Summary:
Adds these counters
{ WAL_FILE_SYNCED, "rocksdb.wal.synced" }
number of writes that request a WAL sync
{ WAL_FILE_BYTES, "rocksdb.wal.bytes" },
number of bytes written to the WAL
{ WRITE_DONE_BY_SELF, "rocksdb.write.self" },
number of writes processed by the calling thread
{ WRITE_DONE_BY_OTHER, "rocksdb.write.other" },
number of writes not processed by the calling thread. Instead these were
processed by the current holder of the write lock
{ WRITE_WITH_WAL, "rocksdb.write.wal" },
number of writes that request WAL logging
{ COMPACT_READ_BYTES, "rocksdb.compact.read.bytes" },
number of bytes read during compaction
{ COMPACT_WRITE_BYTES, "rocksdb.compact.write.bytes" },
number of bytes written during compaction
Per-interval stats output was updated with WAL stats and correct stats for universal compaction
including a correct value for write-amplification. It now looks like:
Compactions
Level Files Size(MB) Score Time(sec) Read(MB) Write(MB) Rn(MB) Rnp1(MB) Wnew(MB) RW-Amplify Read(MB/s) Write(MB/s) Rn Rnp1 Wnp1 NewW Count Ln-stall Stall-cnt
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
0 7 464 46.4 281 3411 3875 3411 0 3875 2.1 12.1 13.8 621 0 240 240 628 0.0 0
Uptime(secs): 310.8 total, 2.0 interval
Writes cumulative: 9999999 total, 9999999 batches, 1.0 per batch, 1.22 ingest GB
WAL cumulative: 9999999 WAL writes, 9999999 WAL syncs, 1.00 writes per sync, 1.22 GB written
Compaction IO cumulative (GB): 1.22 new, 3.33 read, 3.78 write, 7.12 read+write
Compaction IO cumulative (MB/sec): 4.0 new, 11.0 read, 12.5 write, 23.4 read+write
Amplification cumulative: 4.1 write, 6.8 compaction
Writes interval: 100000 total, 100000 batches, 1.0 per batch, 12.5 ingest MB
WAL interval: 100000 WAL writes, 100000 WAL syncs, 1.00 writes per sync, 0.01 MB written
Compaction IO interval (MB): 12.49 new, 14.98 read, 21.50 write, 36.48 read+write
Compaction IO interval (MB/sec): 6.4 new, 7.6 read, 11.0 write, 18.6 read+write
Amplification interval: 101.7 write, 102.9 compaction
Stalls(secs): 142.924 level0_slowdown, 0.000 level0_numfiles, 0.805 memtable_compaction, 0.000 leveln_slowdown
Stalls(count): 132461 level0_slowdown, 0 level0_numfiles, 3 memtable_compaction, 0 leveln_slowdown
Task ID: #3329644, #3301695
Blame Rev:
Test Plan:
Revert Plan:
Database Impact:
Memcache Impact:
Other Notes:
EImportant:
- begin *PUBLIC* platform impact section -
Bugzilla: #
- end platform impact -
Reviewers: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14583
2013-12-09 22:43:34 +01:00
|
|
|
RecordTick(options_.statistics.get(), COMPACT_WRITE_BYTES,
|
|
|
|
compact->outputs[i].file_size);
|
2011-04-12 21:38:58 +02:00
|
|
|
}
|
|
|
|
|
2013-11-07 20:31:56 +01:00
|
|
|
LogFlush(options_.info_log);
|
2011-03-18 23:37:00 +01:00
|
|
|
mutex_.Lock();
|
2014-01-30 05:40:41 +01:00
|
|
|
internal_stats_.AddCompactionStats(compact->compaction->output_level(),
|
|
|
|
stats);
|
2011-03-18 23:37:00 +01:00
|
|
|
|
2012-10-19 23:00:53 +02:00
|
|
|
// if there were any unused file number (mostly in case of
|
|
|
|
// compaction error), free up the entry from pending_putputs
|
|
|
|
ReleaseCompactionUnusedFileNumbers(compact);
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
if (status.ok()) {
|
|
|
|
status = InstallCompactionResults(compact);
|
2013-12-20 18:57:58 +01:00
|
|
|
InstallSuperVersion(deletion_state);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
2014-01-16 01:18:04 +01:00
|
|
|
Version::LevelSummaryStorage tmp;
|
2011-07-21 04:40:18 +02:00
|
|
|
Log(options_.info_log,
|
2012-10-23 19:34:09 +02:00
|
|
|
"compacted to: %s, %.1f MB/sec, level %d, files in(%d, %d) out(%d) "
|
2013-08-23 07:37:13 +02:00
|
|
|
"MB in(%.1f, %.1f) out(%.1f), read-write-amplify(%.1f) "
|
|
|
|
"write-amplify(%.1f) %s\n",
|
2014-01-16 01:18:04 +01:00
|
|
|
versions_->current()->LevelSummary(&tmp),
|
2012-10-23 19:34:09 +02:00
|
|
|
(stats.bytes_readn + stats.bytes_readnp1 + stats.bytes_written) /
|
2014-01-16 01:18:04 +01:00
|
|
|
(double)stats.micros,
|
|
|
|
compact->compaction->output_level(), stats.files_in_leveln,
|
|
|
|
stats.files_in_levelnp1, stats.files_out_levelnp1,
|
|
|
|
stats.bytes_readn / 1048576.0, stats.bytes_readnp1 / 1048576.0,
|
2012-10-23 19:34:09 +02:00
|
|
|
stats.bytes_written / 1048576.0,
|
2013-08-23 07:37:13 +02:00
|
|
|
(stats.bytes_written + stats.bytes_readnp1 + stats.bytes_readn) /
|
2014-01-16 01:18:04 +01:00
|
|
|
(double)stats.bytes_readn,
|
|
|
|
stats.bytes_written / (double)stats.bytes_readn,
|
2012-11-28 07:10:20 +01:00
|
|
|
status.ToString().c_str());
|
2012-10-23 19:34:09 +02:00
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
return status;
|
|
|
|
}
|
|
|
|
|
2011-05-28 02:53:58 +02:00
|
|
|
namespace {
|
|
|
|
struct IterState {
|
2014-02-03 22:13:36 +01:00
|
|
|
IterState(DBImpl* db, port::Mutex* mu, DBImpl::SuperVersion* super_version)
|
|
|
|
: db(db), mu(mu), super_version(super_version) {}
|
|
|
|
|
|
|
|
DBImpl* db;
|
2011-05-28 02:53:58 +02:00
|
|
|
port::Mutex* mu;
|
2014-02-03 22:13:36 +01:00
|
|
|
DBImpl::SuperVersion* super_version;
|
2011-05-28 02:53:58 +02:00
|
|
|
};
|
|
|
|
|
|
|
|
static void CleanupIteratorState(void* arg1, void* arg2) {
|
|
|
|
IterState* state = reinterpret_cast<IterState*>(arg1);
|
2014-02-03 22:13:36 +01:00
|
|
|
|
|
|
|
bool need_cleanup = state->super_version->Unref();
|
|
|
|
if (need_cleanup) {
|
2014-03-05 03:17:27 +01:00
|
|
|
DBImpl::DeletionState deletion_state;
|
|
|
|
|
2014-02-03 22:13:36 +01:00
|
|
|
state->mu->Lock();
|
|
|
|
state->super_version->Cleanup();
|
|
|
|
state->db->FindObsoleteFiles(deletion_state, false, true);
|
|
|
|
state->mu->Unlock();
|
|
|
|
|
|
|
|
delete state->super_version;
|
2014-03-10 23:42:14 +01:00
|
|
|
if (deletion_state.HaveSomethingToDelete()) {
|
|
|
|
state->db->PurgeObsoleteFiles(deletion_state);
|
|
|
|
}
|
2013-12-11 20:56:36 +01:00
|
|
|
}
|
2014-01-17 06:56:26 +01:00
|
|
|
|
2011-05-28 02:53:58 +02:00
|
|
|
delete state;
|
|
|
|
}
|
2011-10-31 18:22:06 +01:00
|
|
|
} // namespace
|
2011-05-28 02:53:58 +02:00
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
Iterator* DBImpl::NewInternalIterator(const ReadOptions& options,
|
|
|
|
SequenceNumber* latest_snapshot) {
|
2013-12-12 19:54:03 +01:00
|
|
|
mutex_.Lock();
|
|
|
|
*latest_snapshot = versions_->LastSequence();
|
2014-02-03 22:13:36 +01:00
|
|
|
SuperVersion* super_version = super_version_->Ref();
|
2013-12-12 19:54:03 +01:00
|
|
|
mutex_.Unlock();
|
|
|
|
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
std::vector<Iterator*> iterator_list;
|
2014-02-03 22:13:36 +01:00
|
|
|
// Collect iterator for mutable mem
|
|
|
|
iterator_list.push_back(super_version->mem->NewIterator(options));
|
2014-01-17 06:56:26 +01:00
|
|
|
// Collect all needed child iterators for immutable memtables
|
2014-02-03 22:13:36 +01:00
|
|
|
super_version->imm->AddIterators(options, &iterator_list);
|
2014-01-17 06:56:26 +01:00
|
|
|
// Collect iterators for files in L0 - Ln
|
2014-02-03 22:13:36 +01:00
|
|
|
super_version->current->AddIterators(options, storage_options_,
|
|
|
|
&iterator_list);
|
2014-01-02 20:26:57 +01:00
|
|
|
Iterator* internal_iter = NewMergingIterator(
|
2014-01-28 19:35:48 +01:00
|
|
|
env_, &internal_comparator_, &iterator_list[0], iterator_list.size());
|
2014-02-03 22:13:36 +01:00
|
|
|
|
|
|
|
IterState* cleanup = new IterState(this, &mutex_, super_version);
|
2013-02-15 20:53:17 +01:00
|
|
|
internal_iter->RegisterCleanup(CleanupIteratorState, cleanup, nullptr);
|
2011-03-18 23:37:00 +01:00
|
|
|
|
|
|
|
return internal_iter;
|
|
|
|
}
|
|
|
|
|
|
|
|
Iterator* DBImpl::TEST_NewInternalIterator() {
|
|
|
|
SequenceNumber ignored;
|
2014-02-08 01:25:38 +01:00
|
|
|
ReadOptions read_options;
|
|
|
|
// Use prefix_seek to make the test function more useful.
|
|
|
|
read_options.prefix_seek = true;
|
|
|
|
return NewInternalIterator(read_options, &ignored);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
2014-01-17 06:56:26 +01:00
|
|
|
std::pair<Iterator*, Iterator*> DBImpl::GetTailingIteratorPair(
|
|
|
|
const ReadOptions& options,
|
|
|
|
uint64_t* superversion_number) {
|
|
|
|
|
|
|
|
mutex_.Lock();
|
2014-02-03 22:13:36 +01:00
|
|
|
SuperVersion* super_version = super_version_->Ref();
|
2014-01-17 06:56:26 +01:00
|
|
|
if (superversion_number != nullptr) {
|
|
|
|
*superversion_number = CurrentVersionNumber();
|
|
|
|
}
|
|
|
|
mutex_.Unlock();
|
|
|
|
|
2014-02-03 22:13:36 +01:00
|
|
|
Iterator* mutable_iter = super_version->mem->NewIterator(options);
|
2014-01-17 06:56:26 +01:00
|
|
|
// create a DBIter that only uses memtable content; see NewIterator()
|
|
|
|
mutable_iter = NewDBIterator(&dbname_, env_, options_, user_comparator(),
|
|
|
|
mutable_iter, kMaxSequenceNumber);
|
|
|
|
|
|
|
|
std::vector<Iterator*> list;
|
2014-02-03 22:13:36 +01:00
|
|
|
super_version->imm->AddIterators(options, &list);
|
|
|
|
super_version->current->AddIterators(options, storage_options_, &list);
|
|
|
|
Iterator* immutable_iter =
|
2014-02-06 05:56:14 +01:00
|
|
|
NewMergingIterator(env_, &internal_comparator_, &list[0], list.size());
|
2014-01-17 06:56:26 +01:00
|
|
|
|
|
|
|
// create a DBIter that only uses memtable content; see NewIterator()
|
|
|
|
immutable_iter = NewDBIterator(&dbname_, env_, options_, user_comparator(),
|
|
|
|
immutable_iter, kMaxSequenceNumber);
|
|
|
|
|
2014-02-03 22:13:36 +01:00
|
|
|
// register cleanups
|
|
|
|
mutable_iter->RegisterCleanup(CleanupIteratorState,
|
|
|
|
new IterState(this, &mutex_, super_version), nullptr);
|
|
|
|
|
|
|
|
// bump the ref one more time since it will be Unref'ed twice
|
|
|
|
immutable_iter->RegisterCleanup(CleanupIteratorState,
|
|
|
|
new IterState(this, &mutex_, super_version->Ref()), nullptr);
|
|
|
|
|
2014-01-17 06:56:26 +01:00
|
|
|
return std::make_pair(mutable_iter, immutable_iter);
|
|
|
|
}
|
|
|
|
|
2011-03-23 00:24:02 +01:00
|
|
|
int64_t DBImpl::TEST_MaxNextLevelOverlappingBytes() {
|
2011-03-22 19:32:49 +01:00
|
|
|
MutexLock l(&mutex_);
|
2014-01-16 01:18:04 +01:00
|
|
|
return versions_->current()->MaxNextLevelOverlappingBytes();
|
2011-03-22 19:32:49 +01:00
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
Status DBImpl::Get(const ReadOptions& options,
|
|
|
|
const Slice& key,
|
|
|
|
std::string* value) {
|
2013-07-06 03:49:18 +02:00
|
|
|
return GetImpl(options, key, value);
|
|
|
|
}
|
|
|
|
|
2013-12-20 18:57:58 +01:00
|
|
|
// DeletionState gets created and destructed outside of the lock -- we
|
|
|
|
// use this convinently to:
|
|
|
|
// * malloc one SuperVersion() outside of the lock -- new_superversion
|
2014-02-27 20:38:55 +01:00
|
|
|
// * delete SuperVersion()s outside of the lock -- superversions_to_free
|
2013-12-20 18:57:58 +01:00
|
|
|
//
|
|
|
|
// However, if InstallSuperVersion() gets called twice with the same,
|
|
|
|
// deletion_state, we can't reuse the SuperVersion() that got malloced because
|
|
|
|
// first call already used it. In that rare case, we take a hit and create a
|
2014-02-27 20:38:55 +01:00
|
|
|
// new SuperVersion() inside of the mutex.
|
2013-12-20 18:57:58 +01:00
|
|
|
void DBImpl::InstallSuperVersion(DeletionState& deletion_state) {
|
2014-02-27 20:38:55 +01:00
|
|
|
mutex_.AssertHeld();
|
2013-12-20 18:57:58 +01:00
|
|
|
// if new_superversion == nullptr, it means somebody already used it
|
|
|
|
SuperVersion* new_superversion =
|
|
|
|
(deletion_state.new_superversion != nullptr) ?
|
|
|
|
deletion_state.new_superversion : new SuperVersion();
|
|
|
|
SuperVersion* old_superversion = InstallSuperVersion(new_superversion);
|
|
|
|
deletion_state.new_superversion = nullptr;
|
2014-02-27 20:38:55 +01:00
|
|
|
deletion_state.superversions_to_free.push_back(old_superversion);
|
2014-03-07 23:43:22 +01:00
|
|
|
// Reset SuperVersions cached in thread local storage
|
|
|
|
if (options_.allow_thread_local) {
|
|
|
|
ResetThreadLocalSuperVersions(&deletion_state);
|
|
|
|
}
|
2013-12-20 18:57:58 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
DBImpl::SuperVersion* DBImpl::InstallSuperVersion(
|
|
|
|
SuperVersion* new_superversion) {
|
|
|
|
mutex_.AssertHeld();
|
2014-03-26 19:24:52 +01:00
|
|
|
new_superversion->db = this;
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
new_superversion->Init(mem_, imm_.current(), versions_->current());
|
2013-12-20 18:57:58 +01:00
|
|
|
SuperVersion* old_superversion = super_version_;
|
|
|
|
super_version_ = new_superversion;
|
2014-01-17 06:56:26 +01:00
|
|
|
++super_version_number_;
|
2014-02-27 20:38:55 +01:00
|
|
|
super_version_->version_number = super_version_number_;
|
|
|
|
|
2013-12-20 18:57:58 +01:00
|
|
|
if (old_superversion != nullptr && old_superversion->Unref()) {
|
|
|
|
old_superversion->Cleanup();
|
|
|
|
return old_superversion; // will let caller delete outside of mutex
|
|
|
|
}
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
|
2014-02-27 20:38:55 +01:00
|
|
|
void DBImpl::ResetThreadLocalSuperVersions(DeletionState* deletion_state) {
|
|
|
|
mutex_.AssertHeld();
|
|
|
|
autovector<void*> sv_ptrs;
|
2014-03-07 23:43:22 +01:00
|
|
|
local_sv_->Scrape(&sv_ptrs, SuperVersion::kSVObsolete);
|
2014-02-27 20:38:55 +01:00
|
|
|
for (auto ptr : sv_ptrs) {
|
|
|
|
assert(ptr);
|
2014-03-07 23:43:22 +01:00
|
|
|
if (ptr == SuperVersion::kSVInUse) {
|
|
|
|
continue;
|
|
|
|
}
|
2014-02-27 20:38:55 +01:00
|
|
|
auto sv = static_cast<SuperVersion*>(ptr);
|
|
|
|
if (static_cast<SuperVersion*>(ptr)->Unref()) {
|
|
|
|
sv->Cleanup();
|
|
|
|
deletion_state->superversions_to_free.push_back(sv);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-07-06 03:49:18 +02:00
|
|
|
Status DBImpl::GetImpl(const ReadOptions& options,
|
|
|
|
const Slice& key,
|
|
|
|
std::string* value,
|
2013-07-26 21:57:01 +02:00
|
|
|
bool* value_found) {
|
2011-06-22 04:36:45 +02:00
|
|
|
Status s;
|
2013-03-26 02:01:47 +01:00
|
|
|
|
2014-01-07 23:33:15 +01:00
|
|
|
StopWatch sw(env_, options_.statistics.get(), DB_GET, false);
|
2013-11-18 20:32:54 +01:00
|
|
|
StopWatchNano snapshot_timer(env_, false);
|
|
|
|
StartPerfTimer(&snapshot_timer);
|
2011-06-22 04:36:45 +02:00
|
|
|
SequenceNumber snapshot;
|
2013-12-27 21:23:17 +01:00
|
|
|
|
2013-02-15 20:53:17 +01:00
|
|
|
if (options.snapshot != nullptr) {
|
2011-06-22 04:36:45 +02:00
|
|
|
snapshot = reinterpret_cast<const SnapshotImpl*>(options.snapshot)->number_;
|
|
|
|
} else {
|
|
|
|
snapshot = versions_->LastSequence();
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
2011-06-22 04:36:45 +02:00
|
|
|
|
2014-02-27 20:38:55 +01:00
|
|
|
// Acquire SuperVersion
|
|
|
|
SuperVersion* sv = nullptr;
|
|
|
|
if (LIKELY(options_.allow_thread_local)) {
|
|
|
|
// The SuperVersion is cached in thread local storage to avoid acquiring
|
|
|
|
// mutex when SuperVersion does not change since the last use. When a new
|
|
|
|
// SuperVersion is installed, the compaction or flush thread cleans up
|
|
|
|
// cached SuperVersion in all existing thread local storage. To avoid
|
|
|
|
// acquiring mutex for this operation, we use atomic Swap() on the thread
|
|
|
|
// local pointer to guarantee exclusive access. If the thread local pointer
|
|
|
|
// is being used while a new SuperVersion is installed, the cached
|
2014-03-09 06:12:13 +01:00
|
|
|
// SuperVersion can become stale. In that case, the background thread would
|
|
|
|
// have swapped in kSVObsolete. We re-check the value at the end of
|
|
|
|
// Get, with an atomic compare and swap. The superversion will be released
|
|
|
|
// if detected to be stale.
|
2014-03-07 23:43:22 +01:00
|
|
|
void* ptr = local_sv_->Swap(SuperVersion::kSVInUse);
|
|
|
|
// Invariant:
|
|
|
|
// (1) Scrape (always) installs kSVObsolete in ThreadLocal storage
|
|
|
|
// (2) the Swap above (always) installs kSVInUse, ThreadLocal storage
|
|
|
|
// should only keep kSVInUse during a GetImpl.
|
|
|
|
assert(ptr != SuperVersion::kSVInUse);
|
|
|
|
sv = static_cast<SuperVersion*>(ptr);
|
|
|
|
if (sv == SuperVersion::kSVObsolete ||
|
|
|
|
sv->version_number != super_version_number_.load(
|
|
|
|
std::memory_order_relaxed)) {
|
|
|
|
RecordTick(options_.statistics.get(), NUMBER_SUPERVERSION_ACQUIRES);
|
2014-02-27 20:38:55 +01:00
|
|
|
SuperVersion* sv_to_delete = nullptr;
|
|
|
|
|
|
|
|
if (sv && sv->Unref()) {
|
2014-03-09 06:12:13 +01:00
|
|
|
RecordTick(options_.statistics.get(), NUMBER_SUPERVERSION_CLEANUPS);
|
2014-02-27 20:38:55 +01:00
|
|
|
mutex_.Lock();
|
2014-03-09 06:12:13 +01:00
|
|
|
// TODO underlying resources held by superversion (sst files) might
|
|
|
|
// not be released until the next background job.
|
2014-02-27 20:38:55 +01:00
|
|
|
sv->Cleanup();
|
|
|
|
sv_to_delete = sv;
|
|
|
|
} else {
|
|
|
|
mutex_.Lock();
|
|
|
|
}
|
|
|
|
sv = super_version_->Ref();
|
|
|
|
mutex_.Unlock();
|
|
|
|
|
|
|
|
delete sv_to_delete;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
mutex_.Lock();
|
|
|
|
sv = super_version_->Ref();
|
|
|
|
mutex_.Unlock();
|
|
|
|
}
|
2013-12-20 18:57:58 +01:00
|
|
|
|
2011-08-22 23:08:51 +02:00
|
|
|
bool have_stat_update = false;
|
2011-06-22 04:36:45 +02:00
|
|
|
Version::GetStats stats;
|
2013-03-21 23:59:47 +01:00
|
|
|
|
[RocksDB] [MergeOperator] The new Merge Interface! Uses merge sequences.
Summary:
Here are the major changes to the Merge Interface. It has been expanded
to handle cases where the MergeOperator is not associative. It does so by stacking
up merge operations while scanning through the key history (i.e.: during Get() or
Compaction), until a valid Put/Delete/end-of-history is encountered; it then
applies all of the merge operations in the correct sequence starting with the
base/sentinel value.
I have also introduced an "AssociativeMerge" function which allows the user to
take advantage of associative merge operations (such as in the case of counters).
The implementation will always attempt to merge the operations/operands themselves
together when they are encountered, and will resort to the "stacking" method if
and only if the "associative-merge" fails.
This implementation is conjectured to allow MergeOperator to handle the general
case, while still providing the user with the ability to take advantage of certain
efficiencies in their own merge-operator / data-structure.
NOTE: This is a preliminary diff. This must still go through a lot of review,
revision, and testing. Feedback welcome!
Test Plan:
-This is a preliminary diff. I have only just begun testing/debugging it.
-I will be testing this with the existing MergeOperator use-cases and unit-tests
(counters, string-append, and redis-lists)
-I will be "desk-checking" and walking through the code with the help gdb.
-I will find a way of stress-testing the new interface / implementation using
db_bench, db_test, merge_test, and/or db_stress.
-I will ensure that my tests cover all cases: Get-Memtable,
Get-Immutable-Memtable, Get-from-Disk, Iterator-Range-Scan, Flush-Memtable-to-L0,
Compaction-L0-L1, Compaction-Ln-L(n+1), Put/Delete found, Put/Delete not-found,
end-of-history, end-of-file, etc.
-A lot of feedback from the reviewers.
Reviewers: haobo, dhruba, zshao, emayanke
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11499
2013-08-06 05:14:32 +02:00
|
|
|
// Prepare to store a list of merge operations if merge occurs.
|
2013-12-03 03:34:05 +01:00
|
|
|
MergeContext merge_context;
|
[RocksDB] [MergeOperator] The new Merge Interface! Uses merge sequences.
Summary:
Here are the major changes to the Merge Interface. It has been expanded
to handle cases where the MergeOperator is not associative. It does so by stacking
up merge operations while scanning through the key history (i.e.: during Get() or
Compaction), until a valid Put/Delete/end-of-history is encountered; it then
applies all of the merge operations in the correct sequence starting with the
base/sentinel value.
I have also introduced an "AssociativeMerge" function which allows the user to
take advantage of associative merge operations (such as in the case of counters).
The implementation will always attempt to merge the operations/operands themselves
together when they are encountered, and will resort to the "stacking" method if
and only if the "associative-merge" fails.
This implementation is conjectured to allow MergeOperator to handle the general
case, while still providing the user with the ability to take advantage of certain
efficiencies in their own merge-operator / data-structure.
NOTE: This is a preliminary diff. This must still go through a lot of review,
revision, and testing. Feedback welcome!
Test Plan:
-This is a preliminary diff. I have only just begun testing/debugging it.
-I will be testing this with the existing MergeOperator use-cases and unit-tests
(counters, string-append, and redis-lists)
-I will be "desk-checking" and walking through the code with the help gdb.
-I will find a way of stress-testing the new interface / implementation using
db_bench, db_test, merge_test, and/or db_stress.
-I will ensure that my tests cover all cases: Get-Memtable,
Get-Immutable-Memtable, Get-from-Disk, Iterator-Range-Scan, Flush-Memtable-to-L0,
Compaction-L0-L1, Compaction-Ln-L(n+1), Put/Delete found, Put/Delete not-found,
end-of-history, end-of-file, etc.
-A lot of feedback from the reviewers.
Reviewers: haobo, dhruba, zshao, emayanke
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11499
2013-08-06 05:14:32 +02:00
|
|
|
|
2013-03-26 02:01:47 +01:00
|
|
|
// First look in the memtable, then in the immutable memtable (if any).
|
2013-03-21 23:59:47 +01:00
|
|
|
// s is both in/out. When in, s could either be OK or MergeInProgress.
|
[RocksDB] [MergeOperator] The new Merge Interface! Uses merge sequences.
Summary:
Here are the major changes to the Merge Interface. It has been expanded
to handle cases where the MergeOperator is not associative. It does so by stacking
up merge operations while scanning through the key history (i.e.: during Get() or
Compaction), until a valid Put/Delete/end-of-history is encountered; it then
applies all of the merge operations in the correct sequence starting with the
base/sentinel value.
I have also introduced an "AssociativeMerge" function which allows the user to
take advantage of associative merge operations (such as in the case of counters).
The implementation will always attempt to merge the operations/operands themselves
together when they are encountered, and will resort to the "stacking" method if
and only if the "associative-merge" fails.
This implementation is conjectured to allow MergeOperator to handle the general
case, while still providing the user with the ability to take advantage of certain
efficiencies in their own merge-operator / data-structure.
NOTE: This is a preliminary diff. This must still go through a lot of review,
revision, and testing. Feedback welcome!
Test Plan:
-This is a preliminary diff. I have only just begun testing/debugging it.
-I will be testing this with the existing MergeOperator use-cases and unit-tests
(counters, string-append, and redis-lists)
-I will be "desk-checking" and walking through the code with the help gdb.
-I will find a way of stress-testing the new interface / implementation using
db_bench, db_test, merge_test, and/or db_stress.
-I will ensure that my tests cover all cases: Get-Memtable,
Get-Immutable-Memtable, Get-from-Disk, Iterator-Range-Scan, Flush-Memtable-to-L0,
Compaction-L0-L1, Compaction-Ln-L(n+1), Put/Delete found, Put/Delete not-found,
end-of-history, end-of-file, etc.
-A lot of feedback from the reviewers.
Reviewers: haobo, dhruba, zshao, emayanke
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11499
2013-08-06 05:14:32 +02:00
|
|
|
// merge_operands will contain the sequence of merges in the latter case.
|
2013-03-26 02:01:47 +01:00
|
|
|
LookupKey lkey(key, snapshot);
|
2013-11-18 20:32:54 +01:00
|
|
|
BumpPerfTime(&perf_context.get_snapshot_time, &snapshot_timer);
|
2014-02-27 20:38:55 +01:00
|
|
|
if (sv->mem->Get(lkey, value, &s, merge_context, options_)) {
|
2013-03-26 02:01:47 +01:00
|
|
|
// Done
|
2013-12-03 21:59:53 +01:00
|
|
|
RecordTick(options_.statistics.get(), MEMTABLE_HIT);
|
2014-02-27 20:38:55 +01:00
|
|
|
} else if (sv->imm->Get(lkey, value, &s, merge_context, options_)) {
|
2013-03-26 02:01:47 +01:00
|
|
|
// Done
|
2013-12-03 21:59:53 +01:00
|
|
|
RecordTick(options_.statistics.get(), MEMTABLE_HIT);
|
2013-03-26 02:01:47 +01:00
|
|
|
} else {
|
2013-11-18 20:32:54 +01:00
|
|
|
StopWatchNano from_files_timer(env_, false);
|
|
|
|
StartPerfTimer(&from_files_timer);
|
|
|
|
|
2014-02-27 20:38:55 +01:00
|
|
|
sv->current->Get(options, lkey, value, &s, &merge_context, &stats,
|
|
|
|
options_, value_found);
|
2013-03-26 02:01:47 +01:00
|
|
|
have_stat_update = true;
|
2013-11-18 20:32:54 +01:00
|
|
|
BumpPerfTime(&perf_context.get_from_output_files_time, &from_files_timer);
|
2013-12-03 21:59:53 +01:00
|
|
|
RecordTick(options_.statistics.get(), MEMTABLE_MISS);
|
2011-06-22 04:36:45 +02:00
|
|
|
}
|
2013-11-18 20:32:54 +01:00
|
|
|
|
|
|
|
StopWatchNano post_process_timer(env_, false);
|
|
|
|
StartPerfTimer(&post_process_timer);
|
2011-08-22 23:08:51 +02:00
|
|
|
|
2013-12-20 18:57:58 +01:00
|
|
|
if (!options_.disable_seek_compaction && have_stat_update) {
|
|
|
|
mutex_.Lock();
|
2014-02-27 20:38:55 +01:00
|
|
|
if (sv->current->UpdateStats(stats)) {
|
2013-12-20 18:57:58 +01:00
|
|
|
MaybeScheduleFlushOrCompaction();
|
|
|
|
}
|
|
|
|
mutex_.Unlock();
|
2014-02-27 20:38:55 +01:00
|
|
|
}
|
|
|
|
|
2014-03-07 23:43:22 +01:00
|
|
|
bool unref_sv = true;
|
2014-02-27 20:38:55 +01:00
|
|
|
if (LIKELY(options_.allow_thread_local)) {
|
|
|
|
// Put the SuperVersion back
|
2014-03-07 23:43:22 +01:00
|
|
|
void* expected = SuperVersion::kSVInUse;
|
|
|
|
if (local_sv_->CompareAndSwap(static_cast<void*>(sv), expected)) {
|
|
|
|
// When we see kSVInUse in the ThreadLocal, we are sure ThreadLocal
|
|
|
|
// storage has not been altered and no Scrape has happend. The
|
|
|
|
// SuperVersion is still current.
|
|
|
|
unref_sv = false;
|
|
|
|
} else {
|
|
|
|
// ThreadLocal scrape happened in the process of this GetImpl call (after
|
|
|
|
// thread local Swap() at the beginning and before CompareAndSwap()).
|
|
|
|
// This means the SuperVersion it holds is obsolete.
|
|
|
|
assert(expected == SuperVersion::kSVObsolete);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (unref_sv) {
|
|
|
|
// Release SuperVersion
|
2014-02-27 20:38:55 +01:00
|
|
|
if (sv->Unref()) {
|
2013-12-20 18:57:58 +01:00
|
|
|
mutex_.Lock();
|
2014-02-27 20:38:55 +01:00
|
|
|
sv->Cleanup();
|
2013-12-20 18:57:58 +01:00
|
|
|
mutex_.Unlock();
|
2014-02-27 20:38:55 +01:00
|
|
|
delete sv;
|
2014-03-09 06:12:13 +01:00
|
|
|
RecordTick(options_.statistics.get(), NUMBER_SUPERVERSION_CLEANUPS);
|
2013-12-20 18:57:58 +01:00
|
|
|
}
|
2014-03-07 23:43:22 +01:00
|
|
|
RecordTick(options_.statistics.get(), NUMBER_SUPERVERSION_RELEASES);
|
2011-06-22 04:36:45 +02:00
|
|
|
}
|
2013-11-25 20:55:36 +01:00
|
|
|
|
2013-08-27 07:23:47 +02:00
|
|
|
// Note, tickers are atomic now - no lock protection needed any more.
|
2013-11-22 23:14:05 +01:00
|
|
|
RecordTick(options_.statistics.get(), NUMBER_KEYS_READ);
|
|
|
|
RecordTick(options_.statistics.get(), BYTES_READ, value->size());
|
2013-11-27 20:47:40 +01:00
|
|
|
BumpPerfTime(&perf_context.get_post_process_time, &post_process_timer);
|
2011-06-22 04:36:45 +02:00
|
|
|
return s;
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
2013-06-05 20:22:38 +02:00
|
|
|
std::vector<Status> DBImpl::MultiGet(const ReadOptions& options,
|
|
|
|
const std::vector<Slice>& keys,
|
|
|
|
std::vector<std::string>* values) {
|
2014-01-07 23:33:15 +01:00
|
|
|
StopWatch sw(env_, options_.statistics.get(), DB_MULTIGET, false);
|
2013-11-18 20:32:54 +01:00
|
|
|
StopWatchNano snapshot_timer(env_, false);
|
|
|
|
StartPerfTimer(&snapshot_timer);
|
2013-11-27 20:47:40 +01:00
|
|
|
|
2013-06-05 20:22:38 +02:00
|
|
|
SequenceNumber snapshot;
|
2013-11-25 20:55:36 +01:00
|
|
|
|
2013-08-27 07:23:47 +02:00
|
|
|
mutex_.Lock();
|
2013-06-05 20:22:38 +02:00
|
|
|
if (options.snapshot != nullptr) {
|
|
|
|
snapshot = reinterpret_cast<const SnapshotImpl*>(options.snapshot)->number_;
|
|
|
|
} else {
|
|
|
|
snapshot = versions_->LastSequence();
|
|
|
|
}
|
|
|
|
|
2014-02-03 22:13:36 +01:00
|
|
|
SuperVersion* get_version = super_version_->Ref();
|
2013-06-05 20:22:38 +02:00
|
|
|
mutex_.Unlock();
|
2014-02-03 22:13:36 +01:00
|
|
|
|
2013-06-05 20:22:38 +02:00
|
|
|
bool have_stat_update = false;
|
|
|
|
Version::GetStats stats;
|
|
|
|
|
2013-12-03 03:34:05 +01:00
|
|
|
// Contain a list of merge operations if merge occurs.
|
|
|
|
MergeContext merge_context;
|
[RocksDB] [MergeOperator] The new Merge Interface! Uses merge sequences.
Summary:
Here are the major changes to the Merge Interface. It has been expanded
to handle cases where the MergeOperator is not associative. It does so by stacking
up merge operations while scanning through the key history (i.e.: during Get() or
Compaction), until a valid Put/Delete/end-of-history is encountered; it then
applies all of the merge operations in the correct sequence starting with the
base/sentinel value.
I have also introduced an "AssociativeMerge" function which allows the user to
take advantage of associative merge operations (such as in the case of counters).
The implementation will always attempt to merge the operations/operands themselves
together when they are encountered, and will resort to the "stacking" method if
and only if the "associative-merge" fails.
This implementation is conjectured to allow MergeOperator to handle the general
case, while still providing the user with the ability to take advantage of certain
efficiencies in their own merge-operator / data-structure.
NOTE: This is a preliminary diff. This must still go through a lot of review,
revision, and testing. Feedback welcome!
Test Plan:
-This is a preliminary diff. I have only just begun testing/debugging it.
-I will be testing this with the existing MergeOperator use-cases and unit-tests
(counters, string-append, and redis-lists)
-I will be "desk-checking" and walking through the code with the help gdb.
-I will find a way of stress-testing the new interface / implementation using
db_bench, db_test, merge_test, and/or db_stress.
-I will ensure that my tests cover all cases: Get-Memtable,
Get-Immutable-Memtable, Get-from-Disk, Iterator-Range-Scan, Flush-Memtable-to-L0,
Compaction-L0-L1, Compaction-Ln-L(n+1), Put/Delete found, Put/Delete not-found,
end-of-history, end-of-file, etc.
-A lot of feedback from the reviewers.
Reviewers: haobo, dhruba, zshao, emayanke
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11499
2013-08-06 05:14:32 +02:00
|
|
|
|
2013-06-05 20:22:38 +02:00
|
|
|
// Note: this always resizes the values array
|
|
|
|
int numKeys = keys.size();
|
|
|
|
std::vector<Status> statList(numKeys);
|
|
|
|
values->resize(numKeys);
|
|
|
|
|
|
|
|
// Keep track of bytes that we read for statistics-recording later
|
|
|
|
uint64_t bytesRead = 0;
|
2013-11-18 20:32:54 +01:00
|
|
|
BumpPerfTime(&perf_context.get_snapshot_time, &snapshot_timer);
|
2013-06-05 20:22:38 +02:00
|
|
|
|
|
|
|
// For each of the given keys, apply the entire "get" process as follows:
|
|
|
|
// First look in the memtable, then in the immutable memtable (if any).
|
|
|
|
// s is both in/out. When in, s could either be OK or MergeInProgress.
|
[RocksDB] [MergeOperator] The new Merge Interface! Uses merge sequences.
Summary:
Here are the major changes to the Merge Interface. It has been expanded
to handle cases where the MergeOperator is not associative. It does so by stacking
up merge operations while scanning through the key history (i.e.: during Get() or
Compaction), until a valid Put/Delete/end-of-history is encountered; it then
applies all of the merge operations in the correct sequence starting with the
base/sentinel value.
I have also introduced an "AssociativeMerge" function which allows the user to
take advantage of associative merge operations (such as in the case of counters).
The implementation will always attempt to merge the operations/operands themselves
together when they are encountered, and will resort to the "stacking" method if
and only if the "associative-merge" fails.
This implementation is conjectured to allow MergeOperator to handle the general
case, while still providing the user with the ability to take advantage of certain
efficiencies in their own merge-operator / data-structure.
NOTE: This is a preliminary diff. This must still go through a lot of review,
revision, and testing. Feedback welcome!
Test Plan:
-This is a preliminary diff. I have only just begun testing/debugging it.
-I will be testing this with the existing MergeOperator use-cases and unit-tests
(counters, string-append, and redis-lists)
-I will be "desk-checking" and walking through the code with the help gdb.
-I will find a way of stress-testing the new interface / implementation using
db_bench, db_test, merge_test, and/or db_stress.
-I will ensure that my tests cover all cases: Get-Memtable,
Get-Immutable-Memtable, Get-from-Disk, Iterator-Range-Scan, Flush-Memtable-to-L0,
Compaction-L0-L1, Compaction-Ln-L(n+1), Put/Delete found, Put/Delete not-found,
end-of-history, end-of-file, etc.
-A lot of feedback from the reviewers.
Reviewers: haobo, dhruba, zshao, emayanke
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11499
2013-08-06 05:14:32 +02:00
|
|
|
// merge_operands will contain the sequence of merges in the latter case.
|
|
|
|
for (int i=0; i<numKeys; ++i) {
|
2013-12-03 03:34:05 +01:00
|
|
|
merge_context.Clear();
|
2013-06-05 20:22:38 +02:00
|
|
|
Status& s = statList[i];
|
|
|
|
std::string* value = &(*values)[i];
|
|
|
|
|
|
|
|
LookupKey lkey(keys[i], snapshot);
|
2014-02-03 22:13:36 +01:00
|
|
|
if (get_version->mem->Get(lkey, value, &s, merge_context, options_)) {
|
2013-06-05 20:22:38 +02:00
|
|
|
// Done
|
2014-02-03 22:13:36 +01:00
|
|
|
} else if (get_version->imm->Get(lkey, value, &s, merge_context,
|
|
|
|
options_)) {
|
2013-06-05 20:22:38 +02:00
|
|
|
// Done
|
|
|
|
} else {
|
2014-02-03 22:13:36 +01:00
|
|
|
get_version->current->Get(options, lkey, value, &s, &merge_context,
|
|
|
|
&stats, options_);
|
2013-06-05 20:22:38 +02:00
|
|
|
have_stat_update = true;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (s.ok()) {
|
|
|
|
bytesRead += value->size();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Post processing (decrement reference counts and record statistics)
|
2013-11-18 20:32:54 +01:00
|
|
|
StopWatchNano post_process_timer(env_, false);
|
|
|
|
StartPerfTimer(&post_process_timer);
|
2014-02-03 22:13:36 +01:00
|
|
|
bool delete_get_version = false;
|
|
|
|
if (!options_.disable_seek_compaction && have_stat_update) {
|
|
|
|
mutex_.Lock();
|
|
|
|
if (get_version->current->UpdateStats(stats)) {
|
|
|
|
MaybeScheduleFlushOrCompaction();
|
|
|
|
}
|
|
|
|
if (get_version->Unref()) {
|
|
|
|
get_version->Cleanup();
|
|
|
|
delete_get_version = true;
|
|
|
|
}
|
|
|
|
mutex_.Unlock();
|
|
|
|
} else {
|
|
|
|
if (get_version->Unref()) {
|
|
|
|
mutex_.Lock();
|
|
|
|
get_version->Cleanup();
|
|
|
|
mutex_.Unlock();
|
|
|
|
delete_get_version = true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (delete_get_version) {
|
|
|
|
delete get_version;
|
2013-06-05 20:22:38 +02:00
|
|
|
}
|
2013-11-25 20:55:36 +01:00
|
|
|
|
2013-11-22 23:14:05 +01:00
|
|
|
RecordTick(options_.statistics.get(), NUMBER_MULTIGET_CALLS);
|
|
|
|
RecordTick(options_.statistics.get(), NUMBER_MULTIGET_KEYS_READ, numKeys);
|
|
|
|
RecordTick(options_.statistics.get(), NUMBER_MULTIGET_BYTES_READ, bytesRead);
|
2013-11-18 20:32:54 +01:00
|
|
|
BumpPerfTime(&perf_context.get_post_process_time, &post_process_timer);
|
2013-06-05 20:22:38 +02:00
|
|
|
|
|
|
|
return statList;
|
|
|
|
}
|
|
|
|
|
2013-07-26 21:57:01 +02:00
|
|
|
bool DBImpl::KeyMayExist(const ReadOptions& options,
|
|
|
|
const Slice& key,
|
|
|
|
std::string* value,
|
|
|
|
bool* value_found) {
|
|
|
|
if (value_found != nullptr) {
|
2013-11-17 07:59:22 +01:00
|
|
|
// falsify later if key-may-exist but can't fetch value
|
|
|
|
*value_found = true;
|
2013-07-26 21:57:01 +02:00
|
|
|
}
|
2013-08-25 07:48:51 +02:00
|
|
|
ReadOptions roptions = options;
|
|
|
|
roptions.read_tier = kBlockCacheTier; // read from block cache only
|
2013-11-17 07:59:22 +01:00
|
|
|
auto s = GetImpl(roptions, key, value, value_found);
|
|
|
|
|
|
|
|
// If options.block_cache != nullptr and the index block of the table didn't
|
|
|
|
// not present in block_cache, the return value will be Status::Incomplete.
|
|
|
|
// In this case, key may still exist in the table.
|
|
|
|
return s.ok() || s.IsIncomplete();
|
2013-07-06 03:49:18 +02:00
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
Iterator* DBImpl::NewIterator(const ReadOptions& options) {
|
2014-01-17 06:56:26 +01:00
|
|
|
Iterator* iter;
|
|
|
|
|
|
|
|
if (options.tailing) {
|
|
|
|
iter = new TailingIterator(this, options, user_comparator());
|
|
|
|
} else {
|
|
|
|
SequenceNumber latest_snapshot;
|
|
|
|
iter = NewInternalIterator(options, &latest_snapshot);
|
|
|
|
|
|
|
|
iter = NewDBIterator(
|
|
|
|
&dbname_, env_, options_, user_comparator(), iter,
|
|
|
|
(options.snapshot != nullptr
|
|
|
|
? reinterpret_cast<const SnapshotImpl*>(options.snapshot)->number_
|
|
|
|
: latest_snapshot));
|
|
|
|
}
|
|
|
|
|
2013-08-13 23:04:56 +02:00
|
|
|
if (options.prefix) {
|
|
|
|
// use extra wrapper to exclude any keys from the results which
|
|
|
|
// don't begin with the prefix
|
|
|
|
iter = new PrefixFilterIterator(iter, *options.prefix,
|
2014-03-10 20:56:46 +01:00
|
|
|
options_.prefix_extractor.get());
|
2013-08-13 23:04:56 +02:00
|
|
|
}
|
|
|
|
return iter;
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
const Snapshot* DBImpl::GetSnapshot() {
|
|
|
|
MutexLock l(&mutex_);
|
2011-04-12 21:38:58 +02:00
|
|
|
return snapshots_.New(versions_->LastSequence());
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
void DBImpl::ReleaseSnapshot(const Snapshot* s) {
|
|
|
|
MutexLock l(&mutex_);
|
2011-05-21 04:17:43 +02:00
|
|
|
snapshots_.Delete(reinterpret_cast<const SnapshotImpl*>(s));
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
// Convenience methods
|
|
|
|
Status DBImpl::Put(const WriteOptions& o, const Slice& key, const Slice& val) {
|
|
|
|
return DB::Put(o, key, val);
|
|
|
|
}
|
|
|
|
|
2013-03-21 23:59:47 +01:00
|
|
|
Status DBImpl::Merge(const WriteOptions& o, const Slice& key,
|
|
|
|
const Slice& val) {
|
|
|
|
if (!options_.merge_operator) {
|
|
|
|
return Status::NotSupported("Provide a merge_operator when opening DB");
|
|
|
|
} else {
|
|
|
|
return DB::Merge(o, key, val);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
Status DBImpl::Delete(const WriteOptions& options, const Slice& key) {
|
|
|
|
return DB::Delete(options, key);
|
|
|
|
}
|
|
|
|
|
2012-03-09 01:23:21 +01:00
|
|
|
Status DBImpl::Write(const WriteOptions& options, WriteBatch* my_batch) {
|
2013-11-18 20:32:54 +01:00
|
|
|
StopWatchNano pre_post_process_timer(env_, false);
|
|
|
|
StartPerfTimer(&pre_post_process_timer);
|
2012-03-09 01:23:21 +01:00
|
|
|
Writer w(&mutex_);
|
|
|
|
w.batch = my_batch;
|
|
|
|
w.sync = options.sync;
|
2012-07-05 22:39:28 +02:00
|
|
|
w.disableWAL = options.disableWAL;
|
2012-03-09 01:23:21 +01:00
|
|
|
w.done = false;
|
2011-09-01 21:08:02 +02:00
|
|
|
|
2014-01-07 23:33:15 +01:00
|
|
|
StopWatch sw(env_, options_.statistics.get(), DB_WRITE, false);
|
2013-12-20 18:57:58 +01:00
|
|
|
mutex_.Lock();
|
2013-07-24 19:01:13 +02:00
|
|
|
writers_.push_back(&w);
|
|
|
|
while (!w.done && &w != writers_.front()) {
|
|
|
|
w.cv.Wait();
|
|
|
|
}
|
Add monitoring for universal compaction and add counters for compaction IO
Summary:
Adds these counters
{ WAL_FILE_SYNCED, "rocksdb.wal.synced" }
number of writes that request a WAL sync
{ WAL_FILE_BYTES, "rocksdb.wal.bytes" },
number of bytes written to the WAL
{ WRITE_DONE_BY_SELF, "rocksdb.write.self" },
number of writes processed by the calling thread
{ WRITE_DONE_BY_OTHER, "rocksdb.write.other" },
number of writes not processed by the calling thread. Instead these were
processed by the current holder of the write lock
{ WRITE_WITH_WAL, "rocksdb.write.wal" },
number of writes that request WAL logging
{ COMPACT_READ_BYTES, "rocksdb.compact.read.bytes" },
number of bytes read during compaction
{ COMPACT_WRITE_BYTES, "rocksdb.compact.write.bytes" },
number of bytes written during compaction
Per-interval stats output was updated with WAL stats and correct stats for universal compaction
including a correct value for write-amplification. It now looks like:
Compactions
Level Files Size(MB) Score Time(sec) Read(MB) Write(MB) Rn(MB) Rnp1(MB) Wnew(MB) RW-Amplify Read(MB/s) Write(MB/s) Rn Rnp1 Wnp1 NewW Count Ln-stall Stall-cnt
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
0 7 464 46.4 281 3411 3875 3411 0 3875 2.1 12.1 13.8 621 0 240 240 628 0.0 0
Uptime(secs): 310.8 total, 2.0 interval
Writes cumulative: 9999999 total, 9999999 batches, 1.0 per batch, 1.22 ingest GB
WAL cumulative: 9999999 WAL writes, 9999999 WAL syncs, 1.00 writes per sync, 1.22 GB written
Compaction IO cumulative (GB): 1.22 new, 3.33 read, 3.78 write, 7.12 read+write
Compaction IO cumulative (MB/sec): 4.0 new, 11.0 read, 12.5 write, 23.4 read+write
Amplification cumulative: 4.1 write, 6.8 compaction
Writes interval: 100000 total, 100000 batches, 1.0 per batch, 12.5 ingest MB
WAL interval: 100000 WAL writes, 100000 WAL syncs, 1.00 writes per sync, 0.01 MB written
Compaction IO interval (MB): 12.49 new, 14.98 read, 21.50 write, 36.48 read+write
Compaction IO interval (MB/sec): 6.4 new, 7.6 read, 11.0 write, 18.6 read+write
Amplification interval: 101.7 write, 102.9 compaction
Stalls(secs): 142.924 level0_slowdown, 0.000 level0_numfiles, 0.805 memtable_compaction, 0.000 leveln_slowdown
Stalls(count): 132461 level0_slowdown, 0 level0_numfiles, 3 memtable_compaction, 0 leveln_slowdown
Task ID: #3329644, #3301695
Blame Rev:
Test Plan:
Revert Plan:
Database Impact:
Memcache Impact:
Other Notes:
EImportant:
- begin *PUBLIC* platform impact section -
Bugzilla: #
- end platform impact -
Reviewers: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14583
2013-12-09 22:43:34 +01:00
|
|
|
|
|
|
|
if (!options.disableWAL) {
|
|
|
|
RecordTick(options_.statistics.get(), WRITE_WITH_WAL, 1);
|
|
|
|
}
|
|
|
|
|
2013-07-24 19:01:13 +02:00
|
|
|
if (w.done) {
|
2013-12-20 18:57:58 +01:00
|
|
|
mutex_.Unlock();
|
Add monitoring for universal compaction and add counters for compaction IO
Summary:
Adds these counters
{ WAL_FILE_SYNCED, "rocksdb.wal.synced" }
number of writes that request a WAL sync
{ WAL_FILE_BYTES, "rocksdb.wal.bytes" },
number of bytes written to the WAL
{ WRITE_DONE_BY_SELF, "rocksdb.write.self" },
number of writes processed by the calling thread
{ WRITE_DONE_BY_OTHER, "rocksdb.write.other" },
number of writes not processed by the calling thread. Instead these were
processed by the current holder of the write lock
{ WRITE_WITH_WAL, "rocksdb.write.wal" },
number of writes that request WAL logging
{ COMPACT_READ_BYTES, "rocksdb.compact.read.bytes" },
number of bytes read during compaction
{ COMPACT_WRITE_BYTES, "rocksdb.compact.write.bytes" },
number of bytes written during compaction
Per-interval stats output was updated with WAL stats and correct stats for universal compaction
including a correct value for write-amplification. It now looks like:
Compactions
Level Files Size(MB) Score Time(sec) Read(MB) Write(MB) Rn(MB) Rnp1(MB) Wnew(MB) RW-Amplify Read(MB/s) Write(MB/s) Rn Rnp1 Wnp1 NewW Count Ln-stall Stall-cnt
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
0 7 464 46.4 281 3411 3875 3411 0 3875 2.1 12.1 13.8 621 0 240 240 628 0.0 0
Uptime(secs): 310.8 total, 2.0 interval
Writes cumulative: 9999999 total, 9999999 batches, 1.0 per batch, 1.22 ingest GB
WAL cumulative: 9999999 WAL writes, 9999999 WAL syncs, 1.00 writes per sync, 1.22 GB written
Compaction IO cumulative (GB): 1.22 new, 3.33 read, 3.78 write, 7.12 read+write
Compaction IO cumulative (MB/sec): 4.0 new, 11.0 read, 12.5 write, 23.4 read+write
Amplification cumulative: 4.1 write, 6.8 compaction
Writes interval: 100000 total, 100000 batches, 1.0 per batch, 12.5 ingest MB
WAL interval: 100000 WAL writes, 100000 WAL syncs, 1.00 writes per sync, 0.01 MB written
Compaction IO interval (MB): 12.49 new, 14.98 read, 21.50 write, 36.48 read+write
Compaction IO interval (MB/sec): 6.4 new, 7.6 read, 11.0 write, 18.6 read+write
Amplification interval: 101.7 write, 102.9 compaction
Stalls(secs): 142.924 level0_slowdown, 0.000 level0_numfiles, 0.805 memtable_compaction, 0.000 leveln_slowdown
Stalls(count): 132461 level0_slowdown, 0 level0_numfiles, 3 memtable_compaction, 0 leveln_slowdown
Task ID: #3329644, #3301695
Blame Rev:
Test Plan:
Revert Plan:
Database Impact:
Memcache Impact:
Other Notes:
EImportant:
- begin *PUBLIC* platform impact section -
Bugzilla: #
- end platform impact -
Reviewers: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14583
2013-12-09 22:43:34 +01:00
|
|
|
RecordTick(options_.statistics.get(), WRITE_DONE_BY_OTHER, 1);
|
2013-07-24 19:01:13 +02:00
|
|
|
return w.status;
|
Add monitoring for universal compaction and add counters for compaction IO
Summary:
Adds these counters
{ WAL_FILE_SYNCED, "rocksdb.wal.synced" }
number of writes that request a WAL sync
{ WAL_FILE_BYTES, "rocksdb.wal.bytes" },
number of bytes written to the WAL
{ WRITE_DONE_BY_SELF, "rocksdb.write.self" },
number of writes processed by the calling thread
{ WRITE_DONE_BY_OTHER, "rocksdb.write.other" },
number of writes not processed by the calling thread. Instead these were
processed by the current holder of the write lock
{ WRITE_WITH_WAL, "rocksdb.write.wal" },
number of writes that request WAL logging
{ COMPACT_READ_BYTES, "rocksdb.compact.read.bytes" },
number of bytes read during compaction
{ COMPACT_WRITE_BYTES, "rocksdb.compact.write.bytes" },
number of bytes written during compaction
Per-interval stats output was updated with WAL stats and correct stats for universal compaction
including a correct value for write-amplification. It now looks like:
Compactions
Level Files Size(MB) Score Time(sec) Read(MB) Write(MB) Rn(MB) Rnp1(MB) Wnew(MB) RW-Amplify Read(MB/s) Write(MB/s) Rn Rnp1 Wnp1 NewW Count Ln-stall Stall-cnt
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
0 7 464 46.4 281 3411 3875 3411 0 3875 2.1 12.1 13.8 621 0 240 240 628 0.0 0
Uptime(secs): 310.8 total, 2.0 interval
Writes cumulative: 9999999 total, 9999999 batches, 1.0 per batch, 1.22 ingest GB
WAL cumulative: 9999999 WAL writes, 9999999 WAL syncs, 1.00 writes per sync, 1.22 GB written
Compaction IO cumulative (GB): 1.22 new, 3.33 read, 3.78 write, 7.12 read+write
Compaction IO cumulative (MB/sec): 4.0 new, 11.0 read, 12.5 write, 23.4 read+write
Amplification cumulative: 4.1 write, 6.8 compaction
Writes interval: 100000 total, 100000 batches, 1.0 per batch, 12.5 ingest MB
WAL interval: 100000 WAL writes, 100000 WAL syncs, 1.00 writes per sync, 0.01 MB written
Compaction IO interval (MB): 12.49 new, 14.98 read, 21.50 write, 36.48 read+write
Compaction IO interval (MB/sec): 6.4 new, 7.6 read, 11.0 write, 18.6 read+write
Amplification interval: 101.7 write, 102.9 compaction
Stalls(secs): 142.924 level0_slowdown, 0.000 level0_numfiles, 0.805 memtable_compaction, 0.000 leveln_slowdown
Stalls(count): 132461 level0_slowdown, 0 level0_numfiles, 3 memtable_compaction, 0 leveln_slowdown
Task ID: #3329644, #3301695
Blame Rev:
Test Plan:
Revert Plan:
Database Impact:
Memcache Impact:
Other Notes:
EImportant:
- begin *PUBLIC* platform impact section -
Bugzilla: #
- end platform impact -
Reviewers: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14583
2013-12-09 22:43:34 +01:00
|
|
|
} else {
|
|
|
|
RecordTick(options_.statistics.get(), WRITE_DONE_BY_SELF, 1);
|
2012-03-09 01:23:21 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
// May temporarily unlock and wait.
|
2013-12-20 18:57:58 +01:00
|
|
|
SuperVersion* superversion_to_free = nullptr;
|
|
|
|
Status status = MakeRoomForWrite(my_batch == nullptr, &superversion_to_free);
|
2011-04-21 00:48:11 +02:00
|
|
|
uint64_t last_sequence = versions_->LastSequence();
|
2012-03-09 01:23:21 +01:00
|
|
|
Writer* last_writer = &w;
|
2013-02-15 20:53:17 +01:00
|
|
|
if (status.ok() && my_batch != nullptr) { // nullptr batch is for compactions
|
2014-01-14 23:49:31 +01:00
|
|
|
autovector<WriteBatch*> write_batch_group;
|
|
|
|
BuildBatchGroup(&last_writer, &write_batch_group);
|
2013-08-27 07:23:47 +02:00
|
|
|
|
2012-03-09 01:23:21 +01:00
|
|
|
// Add to log and apply to memtable. We can release the lock
|
|
|
|
// during this phase since &w is currently responsible for logging
|
|
|
|
// and protects against concurrent loggers and concurrent writes
|
|
|
|
// into mem_.
|
2011-09-01 21:08:02 +02:00
|
|
|
{
|
2013-07-24 19:01:13 +02:00
|
|
|
mutex_.Unlock();
|
2014-01-14 23:49:31 +01:00
|
|
|
WriteBatch* updates = nullptr;
|
|
|
|
if (write_batch_group.size() == 1) {
|
|
|
|
updates = write_batch_group[0];
|
|
|
|
} else {
|
|
|
|
updates = &tmp_batch_;
|
|
|
|
for (size_t i = 0; i < write_batch_group.size(); ++i) {
|
|
|
|
WriteBatchInternal::Append(updates, write_batch_group[i]);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-08-27 07:23:47 +02:00
|
|
|
const SequenceNumber current_sequence = last_sequence + 1;
|
|
|
|
WriteBatchInternal::SetSequence(updates, current_sequence);
|
|
|
|
int my_batch_count = WriteBatchInternal::Count(updates);
|
|
|
|
last_sequence += my_batch_count;
|
|
|
|
// Record statistics
|
2013-11-22 23:14:05 +01:00
|
|
|
RecordTick(options_.statistics.get(),
|
|
|
|
NUMBER_KEYS_WRITTEN, my_batch_count);
|
|
|
|
RecordTick(options_.statistics.get(),
|
2013-08-27 07:23:47 +02:00
|
|
|
BYTES_WRITTEN,
|
|
|
|
WriteBatchInternal::ByteSize(updates));
|
2012-11-06 20:21:57 +01:00
|
|
|
if (options.disableWAL) {
|
|
|
|
flush_on_destroy_ = true;
|
2013-07-24 19:01:13 +02:00
|
|
|
}
|
2013-11-18 20:32:54 +01:00
|
|
|
BumpPerfTime(&perf_context.write_pre_and_post_process_time,
|
|
|
|
&pre_post_process_timer);
|
2013-07-24 19:01:13 +02:00
|
|
|
|
|
|
|
if (!options.disableWAL) {
|
2013-10-23 08:26:51 +02:00
|
|
|
StopWatchNano timer(env_);
|
|
|
|
StartPerfTimer(&timer);
|
Add monitoring for universal compaction and add counters for compaction IO
Summary:
Adds these counters
{ WAL_FILE_SYNCED, "rocksdb.wal.synced" }
number of writes that request a WAL sync
{ WAL_FILE_BYTES, "rocksdb.wal.bytes" },
number of bytes written to the WAL
{ WRITE_DONE_BY_SELF, "rocksdb.write.self" },
number of writes processed by the calling thread
{ WRITE_DONE_BY_OTHER, "rocksdb.write.other" },
number of writes not processed by the calling thread. Instead these were
processed by the current holder of the write lock
{ WRITE_WITH_WAL, "rocksdb.write.wal" },
number of writes that request WAL logging
{ COMPACT_READ_BYTES, "rocksdb.compact.read.bytes" },
number of bytes read during compaction
{ COMPACT_WRITE_BYTES, "rocksdb.compact.write.bytes" },
number of bytes written during compaction
Per-interval stats output was updated with WAL stats and correct stats for universal compaction
including a correct value for write-amplification. It now looks like:
Compactions
Level Files Size(MB) Score Time(sec) Read(MB) Write(MB) Rn(MB) Rnp1(MB) Wnew(MB) RW-Amplify Read(MB/s) Write(MB/s) Rn Rnp1 Wnp1 NewW Count Ln-stall Stall-cnt
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
0 7 464 46.4 281 3411 3875 3411 0 3875 2.1 12.1 13.8 621 0 240 240 628 0.0 0
Uptime(secs): 310.8 total, 2.0 interval
Writes cumulative: 9999999 total, 9999999 batches, 1.0 per batch, 1.22 ingest GB
WAL cumulative: 9999999 WAL writes, 9999999 WAL syncs, 1.00 writes per sync, 1.22 GB written
Compaction IO cumulative (GB): 1.22 new, 3.33 read, 3.78 write, 7.12 read+write
Compaction IO cumulative (MB/sec): 4.0 new, 11.0 read, 12.5 write, 23.4 read+write
Amplification cumulative: 4.1 write, 6.8 compaction
Writes interval: 100000 total, 100000 batches, 1.0 per batch, 12.5 ingest MB
WAL interval: 100000 WAL writes, 100000 WAL syncs, 1.00 writes per sync, 0.01 MB written
Compaction IO interval (MB): 12.49 new, 14.98 read, 21.50 write, 36.48 read+write
Compaction IO interval (MB/sec): 6.4 new, 7.6 read, 11.0 write, 18.6 read+write
Amplification interval: 101.7 write, 102.9 compaction
Stalls(secs): 142.924 level0_slowdown, 0.000 level0_numfiles, 0.805 memtable_compaction, 0.000 leveln_slowdown
Stalls(count): 132461 level0_slowdown, 0 level0_numfiles, 3 memtable_compaction, 0 leveln_slowdown
Task ID: #3329644, #3301695
Blame Rev:
Test Plan:
Revert Plan:
Database Impact:
Memcache Impact:
Other Notes:
EImportant:
- begin *PUBLIC* platform impact section -
Bugzilla: #
- end platform impact -
Reviewers: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14583
2013-12-09 22:43:34 +01:00
|
|
|
Slice log_entry = WriteBatchInternal::Contents(updates);
|
|
|
|
status = log_->AddRecord(log_entry);
|
|
|
|
RecordTick(options_.statistics.get(), WAL_FILE_SYNCED, 1);
|
|
|
|
RecordTick(options_.statistics.get(), WAL_FILE_BYTES, log_entry.size());
|
2012-07-05 22:39:28 +02:00
|
|
|
if (status.ok() && options.sync) {
|
2012-08-27 21:10:26 +02:00
|
|
|
if (options_.use_fsync) {
|
2013-11-22 23:14:05 +01:00
|
|
|
StopWatch(env_, options_.statistics.get(), WAL_FILE_SYNC_MICROS);
|
2013-01-20 11:07:13 +01:00
|
|
|
status = log_->file()->Fsync();
|
2012-08-27 21:10:26 +02:00
|
|
|
} else {
|
2013-11-22 23:14:05 +01:00
|
|
|
StopWatch(env_, options_.statistics.get(), WAL_FILE_SYNC_MICROS);
|
2013-01-20 11:07:13 +01:00
|
|
|
status = log_->file()->Sync();
|
2012-08-27 21:10:26 +02:00
|
|
|
}
|
2012-07-05 22:39:28 +02:00
|
|
|
}
|
2013-11-18 20:32:54 +01:00
|
|
|
BumpPerfTime(&perf_context.write_wal_time, &timer);
|
2011-09-01 21:08:02 +02:00
|
|
|
}
|
|
|
|
if (status.ok()) {
|
2013-11-18 20:32:54 +01:00
|
|
|
StopWatchNano write_memtable_timer(env_, false);
|
|
|
|
StartPerfTimer(&write_memtable_timer);
|
2013-07-13 01:56:52 +02:00
|
|
|
status = WriteBatchInternal::InsertInto(updates, mem_, &options_, this,
|
|
|
|
options_.filter_deletes);
|
2013-11-18 20:32:54 +01:00
|
|
|
BumpPerfTime(&perf_context.write_memtable_time, &write_memtable_timer);
|
2013-06-21 01:58:59 +02:00
|
|
|
if (!status.ok()) {
|
|
|
|
// Panic for in-memory corruptions
|
|
|
|
// Note that existing logic was not sound. Any partial failure writing
|
|
|
|
// into the memtable would result in a state that some write ops might
|
|
|
|
// have succeeded in memtable but Status reports error for all writes.
|
|
|
|
throw std::runtime_error("In memory WriteBatch corruption!");
|
|
|
|
}
|
2014-01-17 21:46:06 +01:00
|
|
|
SetTickerCount(options_.statistics.get(), SEQUENCE_NUMBER,
|
|
|
|
last_sequence);
|
2013-09-27 01:04:12 +02:00
|
|
|
}
|
2013-11-18 20:32:54 +01:00
|
|
|
StartPerfTimer(&pre_post_process_timer);
|
2014-01-14 23:49:31 +01:00
|
|
|
if (updates == &tmp_batch_) tmp_batch_.Clear();
|
2013-09-27 01:04:12 +02:00
|
|
|
mutex_.Lock();
|
|
|
|
if (status.ok()) {
|
2013-06-21 01:58:59 +02:00
|
|
|
versions_->SetLastSequence(last_sequence);
|
2011-09-01 21:08:02 +02:00
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
}
|
2013-10-28 20:36:02 +01:00
|
|
|
if (options_.paranoid_checks && !status.ok() && bg_error_.ok()) {
|
|
|
|
bg_error_ = status; // stop compaction & fail any further writes
|
|
|
|
}
|
2012-03-09 01:23:21 +01:00
|
|
|
|
2013-07-24 19:01:13 +02:00
|
|
|
while (true) {
|
|
|
|
Writer* ready = writers_.front();
|
|
|
|
writers_.pop_front();
|
|
|
|
if (ready != &w) {
|
|
|
|
ready->status = status;
|
|
|
|
ready->done = true;
|
|
|
|
ready->cv.Signal();
|
2012-03-09 01:23:21 +01:00
|
|
|
}
|
2013-07-24 19:01:13 +02:00
|
|
|
if (ready == last_writer) break;
|
|
|
|
}
|
2012-03-09 01:23:21 +01:00
|
|
|
|
2013-07-24 19:01:13 +02:00
|
|
|
// Notify new head of write queue
|
|
|
|
if (!writers_.empty()) {
|
|
|
|
writers_.front()->cv.Signal();
|
2012-03-09 01:23:21 +01:00
|
|
|
}
|
2013-12-20 18:57:58 +01:00
|
|
|
mutex_.Unlock();
|
|
|
|
delete superversion_to_free;
|
2013-11-18 20:32:54 +01:00
|
|
|
BumpPerfTime(&perf_context.write_pre_and_post_process_time,
|
|
|
|
&pre_post_process_timer);
|
2011-03-18 23:37:00 +01:00
|
|
|
return status;
|
|
|
|
}
|
|
|
|
|
2012-03-09 01:23:21 +01:00
|
|
|
// REQUIRES: Writer list must be non-empty
|
2013-02-15 20:53:17 +01:00
|
|
|
// REQUIRES: First writer must have a non-nullptr batch
|
2014-01-14 23:49:31 +01:00
|
|
|
void DBImpl::BuildBatchGroup(Writer** last_writer,
|
|
|
|
autovector<WriteBatch*>* write_batch_group) {
|
2012-03-09 01:23:21 +01:00
|
|
|
assert(!writers_.empty());
|
|
|
|
Writer* first = writers_.front();
|
2014-01-14 23:49:31 +01:00
|
|
|
assert(first->batch != nullptr);
|
2012-03-09 01:23:21 +01:00
|
|
|
|
|
|
|
size_t size = WriteBatchInternal::ByteSize(first->batch);
|
2014-01-14 23:49:31 +01:00
|
|
|
write_batch_group->push_back(first->batch);
|
2012-03-09 01:23:21 +01:00
|
|
|
|
|
|
|
// Allow the group to grow up to a maximum size, but if the
|
|
|
|
// original write is small, limit the growth so we do not slow
|
|
|
|
// down the small write too much.
|
|
|
|
size_t max_size = 1 << 20;
|
|
|
|
if (size <= (128<<10)) {
|
|
|
|
max_size = size + (128<<10);
|
|
|
|
}
|
|
|
|
|
|
|
|
*last_writer = first;
|
|
|
|
std::deque<Writer*>::iterator iter = writers_.begin();
|
|
|
|
++iter; // Advance past "first"
|
|
|
|
for (; iter != writers_.end(); ++iter) {
|
|
|
|
Writer* w = *iter;
|
|
|
|
if (w->sync && !first->sync) {
|
|
|
|
// Do not include a sync write into a batch handled by a non-sync write.
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2012-07-05 22:39:28 +02:00
|
|
|
if (!w->disableWAL && first->disableWAL) {
|
|
|
|
// Do not include a write that needs WAL into a batch that has
|
|
|
|
// WAL disabled.
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2013-02-15 20:53:17 +01:00
|
|
|
if (w->batch != nullptr) {
|
2012-03-09 01:23:21 +01:00
|
|
|
size += WriteBatchInternal::ByteSize(w->batch);
|
|
|
|
if (size > max_size) {
|
|
|
|
// Do not make batch too big
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2014-01-14 23:49:31 +01:00
|
|
|
write_batch_group->push_back(w->batch);
|
2012-03-09 01:23:21 +01:00
|
|
|
}
|
|
|
|
*last_writer = w;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-08-01 01:20:48 +02:00
|
|
|
// This function computes the amount of time in microseconds by which a write
|
|
|
|
// should be delayed based on the number of level-0 files according to the
|
|
|
|
// following formula:
|
2013-08-06 00:43:49 +02:00
|
|
|
// if n < bottom, return 0;
|
|
|
|
// if n >= top, return 1000;
|
|
|
|
// otherwise, let r = (n - bottom) /
|
|
|
|
// (top - bottom)
|
2013-08-01 01:20:48 +02:00
|
|
|
// and return r^2 * 1000.
|
|
|
|
// The goal of this formula is to gradually increase the rate at which writes
|
|
|
|
// are slowed. We also tried linear delay (r * 1000), but it seemed to do
|
|
|
|
// slightly worse. There is no other particular reason for choosing quadratic.
|
Fix SlowdownAmount
Summary:
This had a few bugs.
1) bottom and top were reversed. top is for the max value but the callers were passing the max
value to bottom. The result is that the max sleep is used when n >= bottom.
2) one of the callers passed values with type double and these values are frequently between
1.0 and 2.0 so rounding will do some bad things
3) sometimes the function returned 0 when there should be a stall
With this change and one other diff (out for review soon) there are slightly fewer stalls on one workload.
With the fix.
Stalls(secs): 160.166 level0_slowdown, 0.000 level0_numfiles, 0.000 memtable_compaction, 58.495 leveln_slowdown
Stalls(count): 910261 level0_slowdown, 0 level0_numfiles, 0 memtable_compaction, 54526 leveln_slowdown
Without the fix.
Stalls(secs): 172.227 level0_slowdown, 0.000 level0_numfiles, 0.000 memtable_compaction, 56.538 leveln_slowdown
Stalls(count): 160831 level0_slowdown, 0 level0_numfiles, 0 memtable_compaction, 52845 leveln_slowdown
Task ID: #
Blame Rev:
Test Plan:
run db_bench for --benchmarks=overwrite with IO-bound database
Revert Plan:
Database Impact:
Memcache Impact:
Other Notes:
EImportant:
- begin *PUBLIC* platform impact section -
Bugzilla: #
- end platform impact -
Reviewers: haobo
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15243
2014-01-17 03:44:23 +01:00
|
|
|
uint64_t DBImpl::SlowdownAmount(int n, double bottom, double top) {
|
2013-08-01 01:20:48 +02:00
|
|
|
uint64_t delay;
|
2013-08-06 00:43:49 +02:00
|
|
|
if (n >= top) {
|
2013-08-01 01:20:48 +02:00
|
|
|
delay = 1000;
|
|
|
|
}
|
2013-08-06 00:43:49 +02:00
|
|
|
else if (n < bottom) {
|
2013-08-01 01:20:48 +02:00
|
|
|
delay = 0;
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
// If we are here, we know that:
|
2013-08-06 00:43:49 +02:00
|
|
|
// level0_start_slowdown <= n < level0_slowdown
|
2013-08-01 01:20:48 +02:00
|
|
|
// since the previous two conditions are false.
|
Fix SlowdownAmount
Summary:
This had a few bugs.
1) bottom and top were reversed. top is for the max value but the callers were passing the max
value to bottom. The result is that the max sleep is used when n >= bottom.
2) one of the callers passed values with type double and these values are frequently between
1.0 and 2.0 so rounding will do some bad things
3) sometimes the function returned 0 when there should be a stall
With this change and one other diff (out for review soon) there are slightly fewer stalls on one workload.
With the fix.
Stalls(secs): 160.166 level0_slowdown, 0.000 level0_numfiles, 0.000 memtable_compaction, 58.495 leveln_slowdown
Stalls(count): 910261 level0_slowdown, 0 level0_numfiles, 0 memtable_compaction, 54526 leveln_slowdown
Without the fix.
Stalls(secs): 172.227 level0_slowdown, 0.000 level0_numfiles, 0.000 memtable_compaction, 56.538 leveln_slowdown
Stalls(count): 160831 level0_slowdown, 0 level0_numfiles, 0 memtable_compaction, 52845 leveln_slowdown
Task ID: #
Blame Rev:
Test Plan:
run db_bench for --benchmarks=overwrite with IO-bound database
Revert Plan:
Database Impact:
Memcache Impact:
Other Notes:
EImportant:
- begin *PUBLIC* platform impact section -
Bugzilla: #
- end platform impact -
Reviewers: haobo
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15243
2014-01-17 03:44:23 +01:00
|
|
|
double how_much =
|
|
|
|
(double) (n - bottom) /
|
2013-08-06 00:43:49 +02:00
|
|
|
(top - bottom);
|
Fix SlowdownAmount
Summary:
This had a few bugs.
1) bottom and top were reversed. top is for the max value but the callers were passing the max
value to bottom. The result is that the max sleep is used when n >= bottom.
2) one of the callers passed values with type double and these values are frequently between
1.0 and 2.0 so rounding will do some bad things
3) sometimes the function returned 0 when there should be a stall
With this change and one other diff (out for review soon) there are slightly fewer stalls on one workload.
With the fix.
Stalls(secs): 160.166 level0_slowdown, 0.000 level0_numfiles, 0.000 memtable_compaction, 58.495 leveln_slowdown
Stalls(count): 910261 level0_slowdown, 0 level0_numfiles, 0 memtable_compaction, 54526 leveln_slowdown
Without the fix.
Stalls(secs): 172.227 level0_slowdown, 0.000 level0_numfiles, 0.000 memtable_compaction, 56.538 leveln_slowdown
Stalls(count): 160831 level0_slowdown, 0 level0_numfiles, 0 memtable_compaction, 52845 leveln_slowdown
Task ID: #
Blame Rev:
Test Plan:
run db_bench for --benchmarks=overwrite with IO-bound database
Revert Plan:
Database Impact:
Memcache Impact:
Other Notes:
EImportant:
- begin *PUBLIC* platform impact section -
Bugzilla: #
- end platform impact -
Reviewers: haobo
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15243
2014-01-17 03:44:23 +01:00
|
|
|
delay = std::max(how_much * how_much * 1000, 100.0);
|
2013-08-01 01:20:48 +02:00
|
|
|
}
|
|
|
|
assert(delay <= 1000);
|
|
|
|
return delay;
|
|
|
|
}
|
|
|
|
|
2011-09-01 21:08:02 +02:00
|
|
|
// REQUIRES: mutex_ is held
|
2012-03-09 01:23:21 +01:00
|
|
|
// REQUIRES: this thread is currently at the front of the writer queue
|
2013-12-20 18:57:58 +01:00
|
|
|
Status DBImpl::MakeRoomForWrite(bool force,
|
|
|
|
SuperVersion** superversion_to_free) {
|
2011-04-12 21:38:58 +02:00
|
|
|
mutex_.AssertHeld();
|
2013-07-24 19:01:13 +02:00
|
|
|
assert(!writers_.empty());
|
2011-05-21 04:17:43 +02:00
|
|
|
bool allow_delay = !force;
|
2013-08-06 00:43:49 +02:00
|
|
|
bool allow_hard_rate_limit_delay = !force;
|
|
|
|
bool allow_soft_rate_limit_delay = !force;
|
2013-03-02 21:56:04 +01:00
|
|
|
uint64_t rate_limit_delay_millis = 0;
|
2011-04-12 21:38:58 +02:00
|
|
|
Status s;
|
2012-10-26 22:37:21 +02:00
|
|
|
double score;
|
2013-12-20 18:57:58 +01:00
|
|
|
*superversion_to_free = nullptr;
|
2012-06-23 04:30:03 +02:00
|
|
|
|
2011-04-12 21:38:58 +02:00
|
|
|
while (true) {
|
|
|
|
if (!bg_error_.ok()) {
|
|
|
|
// Yield previous error
|
|
|
|
s = bg_error_;
|
|
|
|
break;
|
2014-01-16 01:18:04 +01:00
|
|
|
} else if (allow_delay && versions_->NeedSlowdownForNumLevel0Files()) {
|
2011-05-21 04:17:43 +02:00
|
|
|
// We are getting close to hitting a hard limit on the number of
|
|
|
|
// L0 files. Rather than delaying a single write by several
|
|
|
|
// seconds when we hit the hard limit, start delaying each
|
2013-08-01 01:20:48 +02:00
|
|
|
// individual write by 0-1ms to reduce latency variance. Also,
|
2011-05-21 04:17:43 +02:00
|
|
|
// this delay hands over some CPU to the compaction thread in
|
|
|
|
// case it is sharing the same core as the writer.
|
2014-01-18 01:38:54 +01:00
|
|
|
uint64_t slowdown =
|
|
|
|
SlowdownAmount(versions_->current()->NumLevelFiles(0),
|
|
|
|
options_.level0_slowdown_writes_trigger,
|
|
|
|
options_.level0_stop_writes_trigger);
|
2011-05-21 04:17:43 +02:00
|
|
|
mutex_.Unlock();
|
2013-07-30 00:46:36 +02:00
|
|
|
uint64_t delayed;
|
2013-07-29 19:34:23 +02:00
|
|
|
{
|
2013-11-22 23:14:05 +01:00
|
|
|
StopWatch sw(env_, options_.statistics.get(), STALL_L0_SLOWDOWN_COUNT);
|
2014-01-18 01:38:54 +01:00
|
|
|
env_->SleepForMicroseconds(slowdown);
|
2013-07-30 00:46:36 +02:00
|
|
|
delayed = sw.ElapsedMicros();
|
2013-07-29 19:34:23 +02:00
|
|
|
}
|
2013-11-22 23:14:05 +01:00
|
|
|
RecordTick(options_.statistics.get(), STALL_L0_SLOWDOWN_MICROS, delayed);
|
2014-01-30 05:40:41 +01:00
|
|
|
internal_stats_.RecordWriteStall(InternalStats::LEVEL0_SLOWDOWN, delayed);
|
2011-05-21 04:17:43 +02:00
|
|
|
allow_delay = false; // Do not delay a single write more than once
|
|
|
|
mutex_.Lock();
|
2012-10-19 23:00:53 +02:00
|
|
|
delayed_writes_++;
|
2014-03-13 00:40:14 +01:00
|
|
|
} else if (!force && !mem_->ShouldFlush()) {
|
2011-04-12 21:38:58 +02:00
|
|
|
// There is room in current memtable
|
2012-10-19 23:00:53 +02:00
|
|
|
if (allow_delay) {
|
|
|
|
DelayLoggingAndReset();
|
|
|
|
}
|
2011-04-12 21:38:58 +02:00
|
|
|
break;
|
2012-10-19 23:00:53 +02:00
|
|
|
} else if (imm_.size() == options_.max_write_buffer_number - 1) {
|
2011-04-12 21:38:58 +02:00
|
|
|
// We have filled up the current memtable, but the previous
|
2014-02-25 21:04:14 +01:00
|
|
|
// ones are still being flushed, so we wait.
|
2012-10-19 23:00:53 +02:00
|
|
|
DelayLoggingAndReset();
|
2014-02-25 21:04:14 +01:00
|
|
|
Log(options_.info_log, "wait for memtable flush...\n");
|
|
|
|
MaybeScheduleFlushOrCompaction();
|
2013-07-30 00:46:36 +02:00
|
|
|
uint64_t stall;
|
2013-07-29 19:34:23 +02:00
|
|
|
{
|
2013-11-22 23:14:05 +01:00
|
|
|
StopWatch sw(env_, options_.statistics.get(),
|
2013-07-29 19:34:23 +02:00
|
|
|
STALL_MEMTABLE_COMPACTION_COUNT);
|
|
|
|
bg_cv_.Wait();
|
2013-07-30 00:46:36 +02:00
|
|
|
stall = sw.ElapsedMicros();
|
2013-07-29 19:34:23 +02:00
|
|
|
}
|
2013-11-22 23:14:05 +01:00
|
|
|
RecordTick(options_.statistics.get(),
|
|
|
|
STALL_MEMTABLE_COMPACTION_MICROS, stall);
|
2014-01-30 05:40:41 +01:00
|
|
|
internal_stats_.RecordWriteStall(InternalStats::MEMTABLE_COMPACTION,
|
|
|
|
stall);
|
2014-01-16 01:15:43 +01:00
|
|
|
} else if (versions_->current()->NumLevelFiles(0) >=
|
2012-11-27 06:16:21 +01:00
|
|
|
options_.level0_stop_writes_trigger) {
|
2011-05-21 04:17:43 +02:00
|
|
|
// There are too many level-0 files.
|
2012-10-19 23:00:53 +02:00
|
|
|
DelayLoggingAndReset();
|
|
|
|
Log(options_.info_log, "wait for fewer level0 files...\n");
|
2013-07-30 00:46:36 +02:00
|
|
|
uint64_t stall;
|
2013-07-29 19:34:23 +02:00
|
|
|
{
|
2013-11-22 23:14:05 +01:00
|
|
|
StopWatch sw(env_, options_.statistics.get(),
|
|
|
|
STALL_L0_NUM_FILES_COUNT);
|
2013-07-29 19:34:23 +02:00
|
|
|
bg_cv_.Wait();
|
2013-07-30 00:46:36 +02:00
|
|
|
stall = sw.ElapsedMicros();
|
2013-07-29 19:34:23 +02:00
|
|
|
}
|
2013-11-22 23:14:05 +01:00
|
|
|
RecordTick(options_.statistics.get(), STALL_L0_NUM_FILES_MICROS, stall);
|
2014-01-30 05:40:41 +01:00
|
|
|
internal_stats_.RecordWriteStall(InternalStats::LEVEL0_NUM_FILES, stall);
|
2014-01-27 18:59:00 +01:00
|
|
|
} else if (allow_hard_rate_limit_delay && options_.hard_rate_limit > 1.0 &&
|
|
|
|
(score = versions_->current()->MaxCompactionScore()) >
|
|
|
|
options_.hard_rate_limit) {
|
2012-10-26 22:37:21 +02:00
|
|
|
// Delay a write when the compaction score for any level is too large.
|
2014-01-27 18:59:00 +01:00
|
|
|
int max_level = versions_->current()->MaxCompactionScoreLevel();
|
2012-10-26 22:37:21 +02:00
|
|
|
mutex_.Unlock();
|
2013-07-30 00:46:36 +02:00
|
|
|
uint64_t delayed;
|
2013-07-29 19:34:23 +02:00
|
|
|
{
|
2013-11-22 23:14:05 +01:00
|
|
|
StopWatch sw(env_, options_.statistics.get(),
|
|
|
|
HARD_RATE_LIMIT_DELAY_COUNT);
|
2013-07-29 19:34:23 +02:00
|
|
|
env_->SleepForMicroseconds(1000);
|
2013-07-30 00:46:36 +02:00
|
|
|
delayed = sw.ElapsedMicros();
|
2013-07-29 19:34:23 +02:00
|
|
|
}
|
2014-01-30 05:40:41 +01:00
|
|
|
internal_stats_.RecordLevelNSlowdown(max_level, delayed);
|
2013-03-02 21:56:04 +01:00
|
|
|
// Make sure the following value doesn't round to zero.
|
2013-05-10 23:41:45 +02:00
|
|
|
uint64_t rate_limit = std::max((delayed / 1000), (uint64_t) 1);
|
|
|
|
rate_limit_delay_millis += rate_limit;
|
2013-11-22 23:14:05 +01:00
|
|
|
RecordTick(options_.statistics.get(),
|
|
|
|
RATE_LIMIT_DELAY_MILLIS, rate_limit);
|
2013-08-06 00:43:49 +02:00
|
|
|
if (options_.rate_limit_delay_max_milliseconds > 0 &&
|
|
|
|
rate_limit_delay_millis >=
|
|
|
|
(unsigned)options_.rate_limit_delay_max_milliseconds) {
|
|
|
|
allow_hard_rate_limit_delay = false;
|
2013-03-02 21:56:04 +01:00
|
|
|
}
|
2012-10-26 22:37:21 +02:00
|
|
|
mutex_.Lock();
|
2014-01-27 18:59:00 +01:00
|
|
|
} else if (allow_soft_rate_limit_delay && options_.soft_rate_limit > 0.0 &&
|
|
|
|
(score = versions_->current()->MaxCompactionScore()) >
|
|
|
|
options_.soft_rate_limit) {
|
2013-08-06 00:43:49 +02:00
|
|
|
// Delay a write when the compaction score for any level is too large.
|
|
|
|
// TODO: add statistics
|
|
|
|
mutex_.Unlock();
|
2013-08-06 03:45:23 +02:00
|
|
|
{
|
2013-11-22 23:14:05 +01:00
|
|
|
StopWatch sw(env_, options_.statistics.get(),
|
|
|
|
SOFT_RATE_LIMIT_DELAY_COUNT);
|
2013-08-06 03:45:23 +02:00
|
|
|
env_->SleepForMicroseconds(SlowdownAmount(
|
|
|
|
score,
|
|
|
|
options_.soft_rate_limit,
|
|
|
|
options_.hard_rate_limit)
|
|
|
|
);
|
|
|
|
rate_limit_delay_millis += sw.ElapsedMicros();
|
|
|
|
}
|
2013-08-06 00:43:49 +02:00
|
|
|
allow_soft_rate_limit_delay = false;
|
|
|
|
mutex_.Lock();
|
2013-11-22 21:52:33 +01:00
|
|
|
|
2011-04-12 21:38:58 +02:00
|
|
|
} else {
|
2013-11-22 21:52:33 +01:00
|
|
|
unique_ptr<WritableFile> lfile;
|
2013-12-31 03:33:57 +01:00
|
|
|
MemTable* new_mem = nullptr;
|
2013-11-22 21:52:33 +01:00
|
|
|
|
2014-02-25 21:04:14 +01:00
|
|
|
// Attempt to switch to a new memtable and trigger flush of old.
|
2013-11-22 21:52:33 +01:00
|
|
|
// Do this without holding the dbmutex lock.
|
2011-04-12 21:38:58 +02:00
|
|
|
assert(versions_->PrevLogNumber() == 0);
|
|
|
|
uint64_t new_log_number = versions_->NewFileNumber();
|
2013-12-20 18:57:58 +01:00
|
|
|
SuperVersion* new_superversion = nullptr;
|
2013-11-22 21:52:33 +01:00
|
|
|
mutex_.Unlock();
|
|
|
|
{
|
|
|
|
DelayLoggingAndReset();
|
2014-01-15 00:27:09 +01:00
|
|
|
s = env_->NewWritableFile(LogFileName(options_.wal_dir, new_log_number),
|
2014-03-18 05:52:14 +01:00
|
|
|
&lfile,
|
|
|
|
env_->OptimizeForLogWrite(storage_options_));
|
2013-11-22 21:52:33 +01:00
|
|
|
if (s.ok()) {
|
|
|
|
// Our final size should be less than write_buffer_size
|
|
|
|
// (compression, etc) but err on the side of caution.
|
|
|
|
lfile->SetPreallocationBlockSize(1.1 * options_.write_buffer_size);
|
2014-01-16 08:12:31 +01:00
|
|
|
new_mem = new MemTable(internal_comparator_, options_);
|
2014-01-02 20:26:57 +01:00
|
|
|
new_superversion = new SuperVersion();
|
2013-11-22 21:52:33 +01:00
|
|
|
}
|
2014-04-03 01:34:19 +02:00
|
|
|
Log(options_.info_log,
|
|
|
|
"New memtable created with log file: #%lu\n",
|
|
|
|
(unsigned long)new_log_number);
|
2013-11-22 21:52:33 +01:00
|
|
|
}
|
|
|
|
mutex_.Lock();
|
2011-04-12 21:38:58 +02:00
|
|
|
if (!s.ok()) {
|
2012-08-27 08:45:35 +02:00
|
|
|
// Avoid chewing through file number space in a tight loop.
|
2013-01-20 11:07:13 +01:00
|
|
|
versions_->ReuseFileNumber(new_log_number);
|
2013-12-31 03:33:57 +01:00
|
|
|
assert (!new_mem);
|
2011-04-12 21:38:58 +02:00
|
|
|
break;
|
|
|
|
}
|
2011-06-22 04:36:45 +02:00
|
|
|
logfile_number_ = new_log_number;
|
2013-01-20 11:07:13 +01:00
|
|
|
log_.reset(new log::Writer(std::move(lfile)));
|
2013-07-16 20:56:46 +02:00
|
|
|
mem_->SetNextLogNumber(logfile_number_);
|
2012-10-19 23:00:53 +02:00
|
|
|
imm_.Add(mem_);
|
2013-09-05 02:24:35 +02:00
|
|
|
if (force) {
|
|
|
|
imm_.FlushRequested();
|
|
|
|
}
|
2013-12-31 03:33:57 +01:00
|
|
|
mem_ = new_mem;
|
2011-05-21 04:17:43 +02:00
|
|
|
mem_->Ref();
|
2013-07-16 20:56:46 +02:00
|
|
|
mem_->SetLogNumber(logfile_number_);
|
2011-04-12 21:38:58 +02:00
|
|
|
force = false; // Do not force another compaction if have room
|
2013-10-15 00:12:15 +02:00
|
|
|
MaybeScheduleFlushOrCompaction();
|
2013-12-20 18:57:58 +01:00
|
|
|
*superversion_to_free = InstallSuperVersion(new_superversion);
|
2011-04-12 21:38:58 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2014-02-14 01:28:21 +01:00
|
|
|
Status DBImpl::GetPropertiesOfAllTables(TablePropertiesCollection* props) {
|
|
|
|
// Increment the ref count
|
|
|
|
mutex_.Lock();
|
|
|
|
auto version = versions_->current();
|
|
|
|
version->Ref();
|
|
|
|
mutex_.Unlock();
|
|
|
|
|
|
|
|
auto s = version->GetPropertiesOfAllTables(props);
|
|
|
|
|
|
|
|
// Decrement the ref count
|
|
|
|
mutex_.Lock();
|
|
|
|
version->Unref();
|
|
|
|
mutex_.Unlock();
|
|
|
|
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
[RocksDB] BackupableDB
Summary:
In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you.
Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes.
There are multiple things you can configure:
1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like.
2. sync - if true, it *guarantees* backup consistency on machine reboot
3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files.
4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup.
5. More things you can find in BackupableDBOptions
Here is the directory structure I use:
backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot
0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files
files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file
files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files
All the files are ref counted and deleted immediatelly when they get out of scope.
Some other stuff in this diff:
1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do.
2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB.
Test Plan:
I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff.
Also, `make asan_check`
Reviewers: dhruba, haobo, emayanke
Reviewed By: dhruba
CC: leveldb, haobo
Differential Revision: https://reviews.facebook.net/D14295
2013-12-09 23:06:52 +01:00
|
|
|
const std::string& DBImpl::GetName() const {
|
|
|
|
return dbname_;
|
|
|
|
}
|
|
|
|
|
2013-11-25 21:39:23 +01:00
|
|
|
Env* DBImpl::GetEnv() const {
|
|
|
|
return env_;
|
|
|
|
}
|
|
|
|
|
2013-11-26 00:51:50 +01:00
|
|
|
const Options& DBImpl::GetOptions() const {
|
|
|
|
return options_;
|
|
|
|
}
|
|
|
|
|
2011-04-12 21:38:58 +02:00
|
|
|
bool DBImpl::GetProperty(const Slice& property, std::string* value) {
|
|
|
|
value->clear();
|
2014-03-18 20:37:42 +01:00
|
|
|
DBPropertyType property_type = GetPropertyType(property);
|
2011-03-18 23:37:00 +01:00
|
|
|
MutexLock l(&mutex_);
|
2014-03-27 19:59:37 +01:00
|
|
|
return internal_stats_.GetProperty(property_type, property, value, this);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
void DBImpl::GetApproximateSizes(
|
|
|
|
const Range* range, int n,
|
|
|
|
uint64_t* sizes) {
|
|
|
|
// TODO(opt): better implementation
|
|
|
|
Version* v;
|
|
|
|
{
|
|
|
|
MutexLock l(&mutex_);
|
|
|
|
versions_->current()->Ref();
|
|
|
|
v = versions_->current();
|
|
|
|
}
|
|
|
|
|
|
|
|
for (int i = 0; i < n; i++) {
|
|
|
|
// Convert user_key into a corresponding internal key.
|
|
|
|
InternalKey k1(range[i].start, kMaxSequenceNumber, kValueTypeForSeek);
|
|
|
|
InternalKey k2(range[i].limit, kMaxSequenceNumber, kValueTypeForSeek);
|
|
|
|
uint64_t start = versions_->ApproximateOffsetOf(v, k1);
|
|
|
|
uint64_t limit = versions_->ApproximateOffsetOf(v, k2);
|
|
|
|
sizes[i] = (limit >= start ? limit - start : 0);
|
|
|
|
}
|
|
|
|
|
|
|
|
{
|
|
|
|
MutexLock l(&mutex_);
|
|
|
|
v->Unref();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-10-19 23:00:53 +02:00
|
|
|
inline void DBImpl::DelayLoggingAndReset() {
|
|
|
|
if (delayed_writes_ > 0) {
|
|
|
|
Log(options_.info_log, "delayed %d write...\n", delayed_writes_ );
|
|
|
|
delayed_writes_ = 0;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-08-22 23:32:53 +02:00
|
|
|
Status DBImpl::DeleteFile(std::string name) {
|
|
|
|
uint64_t number;
|
|
|
|
FileType type;
|
2013-10-24 08:39:23 +02:00
|
|
|
WalFileType log_type;
|
|
|
|
if (!ParseFileName(name, &number, &type, &log_type) ||
|
|
|
|
(type != kTableFile && type != kLogFile)) {
|
|
|
|
Log(options_.info_log, "DeleteFile %s failed.\n", name.c_str());
|
2013-08-22 23:32:53 +02:00
|
|
|
return Status::InvalidArgument("Invalid file name");
|
|
|
|
}
|
|
|
|
|
2013-10-24 08:39:23 +02:00
|
|
|
Status status;
|
|
|
|
if (type == kLogFile) {
|
|
|
|
// Only allow deleting archived log files
|
|
|
|
if (log_type != kArchivedLogFile) {
|
2014-03-09 23:11:34 +01:00
|
|
|
Log(options_.info_log, "DeleteFile %s failed - not archived log.\n",
|
|
|
|
name.c_str());
|
2013-10-24 08:39:23 +02:00
|
|
|
return Status::NotSupported("Delete only supported for archived logs");
|
|
|
|
}
|
|
|
|
status = env_->DeleteFile(options_.wal_dir + "/" + name.c_str());
|
|
|
|
if (!status.ok()) {
|
2014-03-09 23:11:34 +01:00
|
|
|
Log(options_.info_log, "DeleteFile %s failed -- %s.\n",
|
|
|
|
name.c_str(), status.ToString().c_str());
|
2013-10-24 08:39:23 +02:00
|
|
|
}
|
|
|
|
return status;
|
|
|
|
}
|
|
|
|
|
2013-08-22 23:32:53 +02:00
|
|
|
int level;
|
2014-01-07 05:29:17 +01:00
|
|
|
FileMetaData* metadata;
|
2013-08-22 23:32:53 +02:00
|
|
|
int maxlevel = NumberLevels();
|
2014-01-15 00:27:09 +01:00
|
|
|
VersionEdit edit;
|
2014-01-02 20:26:57 +01:00
|
|
|
DeletionState deletion_state(true);
|
2013-08-27 23:54:06 +02:00
|
|
|
{
|
|
|
|
MutexLock l(&mutex_);
|
|
|
|
status = versions_->GetMetadataForFile(number, &level, &metadata);
|
|
|
|
if (!status.ok()) {
|
2013-10-24 08:39:23 +02:00
|
|
|
Log(options_.info_log, "DeleteFile %s failed. File not found\n",
|
|
|
|
name.c_str());
|
2013-08-27 23:54:06 +02:00
|
|
|
return Status::InvalidArgument("File not found");
|
|
|
|
}
|
|
|
|
assert((level > 0) && (level < maxlevel));
|
2013-08-22 23:32:53 +02:00
|
|
|
|
2013-08-27 23:54:06 +02:00
|
|
|
// If the file is being compacted no need to delete.
|
2014-01-07 05:29:17 +01:00
|
|
|
if (metadata->being_compacted) {
|
2013-08-22 23:32:53 +02:00
|
|
|
Log(options_.info_log,
|
2013-10-24 08:39:23 +02:00
|
|
|
"DeleteFile %s Skipped. File about to be compacted\n", name.c_str());
|
2013-08-27 23:54:06 +02:00
|
|
|
return Status::OK();
|
2013-08-22 23:32:53 +02:00
|
|
|
}
|
|
|
|
|
2013-08-27 23:54:06 +02:00
|
|
|
// Only the files in the last level can be deleted externally.
|
|
|
|
// This is to make sure that any deletion tombstones are not
|
|
|
|
// lost. Check that the level passed is the last level.
|
|
|
|
for (int i = level + 1; i < maxlevel; i++) {
|
2014-01-16 01:15:43 +01:00
|
|
|
if (versions_->current()->NumLevelFiles(i) != 0) {
|
2013-08-27 23:54:06 +02:00
|
|
|
Log(options_.info_log,
|
2013-10-24 08:39:23 +02:00
|
|
|
"DeleteFile %s FAILED. File not in last level\n", name.c_str());
|
2013-08-27 23:54:06 +02:00
|
|
|
return Status::InvalidArgument("File not in last level");
|
|
|
|
}
|
|
|
|
}
|
|
|
|
edit.DeleteFile(level, number);
|
2014-01-27 20:02:21 +01:00
|
|
|
status = versions_->LogAndApply(&edit, &mutex_, db_directory_.get());
|
2013-12-20 18:57:58 +01:00
|
|
|
if (status.ok()) {
|
|
|
|
InstallSuperVersion(deletion_state);
|
|
|
|
}
|
2013-11-12 06:03:41 +01:00
|
|
|
FindObsoleteFiles(deletion_state, false);
|
2013-08-27 23:54:06 +02:00
|
|
|
} // lock released here
|
2013-11-07 20:31:56 +01:00
|
|
|
LogFlush(options_.info_log);
|
2013-12-20 18:57:58 +01:00
|
|
|
// remove files outside the db-lock
|
2014-03-10 23:42:14 +01:00
|
|
|
if (deletion_state.HaveSomethingToDelete()) {
|
|
|
|
PurgeObsoleteFiles(deletion_state);
|
|
|
|
}
|
2014-02-25 01:00:13 +01:00
|
|
|
{
|
|
|
|
MutexLock l(&mutex_);
|
|
|
|
// schedule flush if file deletion means we freed the space for flushes to
|
|
|
|
// continue
|
|
|
|
MaybeScheduleFlushOrCompaction();
|
|
|
|
}
|
2013-08-22 23:32:53 +02:00
|
|
|
return status;
|
|
|
|
}
|
|
|
|
|
2014-03-20 22:18:29 +01:00
|
|
|
void DBImpl::GetLiveFilesMetaData(std::vector<LiveFileMetaData>* metadata) {
|
2013-08-22 23:32:53 +02:00
|
|
|
MutexLock l(&mutex_);
|
|
|
|
return versions_->GetLiveFilesMetaData(metadata);
|
|
|
|
}
|
|
|
|
|
2014-03-20 22:18:29 +01:00
|
|
|
Status DBImpl::CheckConsistency() {
|
|
|
|
mutex_.AssertHeld();
|
|
|
|
std::vector<LiveFileMetaData> metadata;
|
|
|
|
versions_->GetLiveFilesMetaData(&metadata);
|
|
|
|
|
|
|
|
std::string corruption_messages;
|
|
|
|
for (const auto& md : metadata) {
|
|
|
|
std::string file_path = dbname_ + md.name;
|
|
|
|
uint64_t fsize = 0;
|
|
|
|
Status s = env_->GetFileSize(file_path, &fsize);
|
|
|
|
if (!s.ok()) {
|
|
|
|
corruption_messages +=
|
|
|
|
"Can't access " + md.name + ": " + s.ToString() + "\n";
|
|
|
|
} else if (fsize != md.size) {
|
|
|
|
corruption_messages += "Sst file size mismatch: " + md.name +
|
|
|
|
". Size recorded in manifest " +
|
|
|
|
std::to_string(md.size) + ", actual size " +
|
|
|
|
std::to_string(fsize) + "\n";
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (corruption_messages.size() == 0) {
|
|
|
|
return Status::OK();
|
|
|
|
} else {
|
|
|
|
return Status::Corruption(corruption_messages);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-02-12 19:43:27 +01:00
|
|
|
void DBImpl::TEST_GetFilesMetaData(
|
|
|
|
std::vector<std::vector<FileMetaData>>* metadata) {
|
|
|
|
MutexLock l(&mutex_);
|
|
|
|
metadata->resize(NumberLevels());
|
|
|
|
for (int level = 0; level < NumberLevels(); level++) {
|
|
|
|
const std::vector<FileMetaData*>& files =
|
|
|
|
versions_->current()->files_[level];
|
|
|
|
|
|
|
|
(*metadata)[level].clear();
|
|
|
|
for (const auto& f : files) {
|
|
|
|
(*metadata)[level].push_back(*f);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-12-03 15:39:07 +01:00
|
|
|
Status DBImpl::GetDbIdentity(std::string& identity) {
|
|
|
|
std::string idfilename = IdentityFileName(dbname_);
|
|
|
|
unique_ptr<SequentialFile> idfile;
|
|
|
|
const EnvOptions soptions;
|
|
|
|
Status s = env_->NewSequentialFile(idfilename, &idfile, soptions);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
uint64_t file_size;
|
|
|
|
s = env_->GetFileSize(idfilename, &file_size);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
char buffer[file_size];
|
|
|
|
Slice id;
|
|
|
|
s = idfile->Read(file_size, &id, buffer);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
identity.assign(id.ToString());
|
|
|
|
// If last character is '\n' remove it from identity
|
|
|
|
if (identity.size() > 0 && identity.back() == '\n') {
|
|
|
|
identity.pop_back();
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
// Default implementations of convenience methods that subclasses of DB
|
|
|
|
// can call if they wish
|
|
|
|
Status DB::Put(const WriteOptions& opt, const Slice& key, const Slice& value) {
|
2014-01-14 19:42:36 +01:00
|
|
|
// Pre-allocate size of write batch conservatively.
|
|
|
|
// 8 bytes are taken by header, 4 bytes for count, 1 byte for type,
|
|
|
|
// and we allocate 11 extra bytes for key length, as well as value length.
|
|
|
|
WriteBatch batch(key.size() + value.size() + 24);
|
2011-03-18 23:37:00 +01:00
|
|
|
batch.Put(key, value);
|
|
|
|
return Write(opt, &batch);
|
|
|
|
}
|
|
|
|
|
|
|
|
Status DB::Delete(const WriteOptions& opt, const Slice& key) {
|
|
|
|
WriteBatch batch;
|
|
|
|
batch.Delete(key);
|
|
|
|
return Write(opt, &batch);
|
|
|
|
}
|
|
|
|
|
2013-03-21 23:59:47 +01:00
|
|
|
Status DB::Merge(const WriteOptions& opt, const Slice& key,
|
|
|
|
const Slice& value) {
|
|
|
|
WriteBatch batch;
|
|
|
|
batch.Merge(key, value);
|
|
|
|
return Write(opt, &batch);
|
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
DB::~DB() { }
|
|
|
|
|
2013-07-23 23:42:27 +02:00
|
|
|
Status DB::Open(const Options& options, const std::string& dbname, DB** dbptr) {
|
2013-02-15 20:53:17 +01:00
|
|
|
*dbptr = nullptr;
|
2011-03-18 23:37:00 +01:00
|
|
|
|
2013-02-15 20:53:17 +01:00
|
|
|
if (options.block_cache != nullptr && options.no_block_cache) {
|
2012-11-01 01:02:24 +01:00
|
|
|
return Status::InvalidArgument(
|
2013-02-15 20:53:17 +01:00
|
|
|
"no_block_cache is true while block_cache is not nullptr");
|
2012-11-01 01:02:24 +01:00
|
|
|
}
|
2013-10-01 23:46:52 +02:00
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
DBImpl* impl = new DBImpl(options, dbname);
|
2013-10-01 23:46:52 +02:00
|
|
|
Status s = impl->env_->CreateDirIfMissing(impl->options_.wal_dir);
|
|
|
|
if (!s.ok()) {
|
|
|
|
delete impl;
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
s = impl->CreateArchivalDirectory();
|
2012-11-26 22:56:45 +01:00
|
|
|
if (!s.ok()) {
|
|
|
|
delete impl;
|
|
|
|
return s;
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
impl->mutex_.Lock();
|
Refactor Recover() code
Summary:
This diff does two things:
* Rethinks how we call Recover() with read_only option. Before, we call it with pointer to memtable where we'd like to apply those changes to. This memtable is set in db_impl_readonly.cc and it's actually DBImpl::mem_. Why don't we just apply updates to mem_ right away? It seems more intuitive.
* Changes when we apply updates to manifest. Before, the process is to recover all the logs, flush it to sst files and then do one giant commit that atomically adds all recovered sst files and sets the next log number. This works good enough, but causes some small troubles for my column family approach, since I can't have one VersionEdit apply to more than single column family[1]. The change here is to commit the files recovered from logs right away. Here is the state of the world before the change:
1. Recover log 5, add new sst files to edit
2. Recover log 7, add new sst files to edit
3. Recover log 8, add new sst files to edit
4. Commit all added sst files to manifest and mark log files 5, 7 and 8 as recoverd (via SetLogNumber(9) function)
After the change, we'll do:
1. Recover log 5, commit the new sst files and set log 5 as recovered
2. Recover log 7, commit the new sst files and set log 7 as recovered
3. Recover log 8, commit the new sst files and set log 8 as recovered
The added (small) benefit is that if we fail after (2), the new recovery will only have to recover log 8. In previous case, we'll have to restart the recovery from the beginning. The bigger benefit will be to enable easier integration of multiple column families in Recovery code path.
[1] I'm happy to dicuss this decison, but I believe this is the cleanest way to go. It also makes backward compatibility much easier. We don't have a requirement of adding multiple column families atomically.
Test Plan: make check
Reviewers: dhruba, haobo, kailiu, sdong
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15237
2014-01-22 19:45:26 +01:00
|
|
|
s = impl->Recover(); // Handles create_if_missing, error_if_exists
|
2011-03-18 23:37:00 +01:00
|
|
|
if (s.ok()) {
|
2011-04-12 21:38:58 +02:00
|
|
|
uint64_t new_log_number = impl->versions_->NewFileNumber();
|
2013-01-20 11:07:13 +01:00
|
|
|
unique_ptr<WritableFile> lfile;
|
2014-02-28 22:19:47 +01:00
|
|
|
EnvOptions soptions(options);
|
2013-11-16 02:17:13 +01:00
|
|
|
s = impl->options_.env->NewWritableFile(
|
2014-02-28 22:19:47 +01:00
|
|
|
LogFileName(impl->options_.wal_dir, new_log_number), &lfile,
|
2014-03-18 05:52:14 +01:00
|
|
|
impl->options_.env->OptimizeForLogWrite(soptions));
|
2011-03-18 23:37:00 +01:00
|
|
|
if (s.ok()) {
|
2013-11-16 02:17:13 +01:00
|
|
|
lfile->SetPreallocationBlockSize(1.1 * impl->options_.write_buffer_size);
|
Refactor Recover() code
Summary:
This diff does two things:
* Rethinks how we call Recover() with read_only option. Before, we call it with pointer to memtable where we'd like to apply those changes to. This memtable is set in db_impl_readonly.cc and it's actually DBImpl::mem_. Why don't we just apply updates to mem_ right away? It seems more intuitive.
* Changes when we apply updates to manifest. Before, the process is to recover all the logs, flush it to sst files and then do one giant commit that atomically adds all recovered sst files and sets the next log number. This works good enough, but causes some small troubles for my column family approach, since I can't have one VersionEdit apply to more than single column family[1]. The change here is to commit the files recovered from logs right away. Here is the state of the world before the change:
1. Recover log 5, add new sst files to edit
2. Recover log 7, add new sst files to edit
3. Recover log 8, add new sst files to edit
4. Commit all added sst files to manifest and mark log files 5, 7 and 8 as recoverd (via SetLogNumber(9) function)
After the change, we'll do:
1. Recover log 5, commit the new sst files and set log 5 as recovered
2. Recover log 7, commit the new sst files and set log 7 as recovered
3. Recover log 8, commit the new sst files and set log 8 as recovered
The added (small) benefit is that if we fail after (2), the new recovery will only have to recover log 8. In previous case, we'll have to restart the recovery from the beginning. The bigger benefit will be to enable easier integration of multiple column families in Recovery code path.
[1] I'm happy to dicuss this decison, but I believe this is the cleanest way to go. It also makes backward compatibility much easier. We don't have a requirement of adding multiple column families atomically.
Test Plan: make check
Reviewers: dhruba, haobo, kailiu, sdong
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15237
2014-01-22 19:45:26 +01:00
|
|
|
VersionEdit edit;
|
2011-04-12 21:38:58 +02:00
|
|
|
edit.SetLogNumber(new_log_number);
|
2011-06-22 04:36:45 +02:00
|
|
|
impl->logfile_number_ = new_log_number;
|
2013-01-20 11:07:13 +01:00
|
|
|
impl->log_.reset(new log::Writer(std::move(lfile)));
|
2014-01-27 20:02:21 +01:00
|
|
|
s = impl->versions_->LogAndApply(&edit, &impl->mutex_,
|
|
|
|
impl->db_directory_.get());
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
if (s.ok()) {
|
2013-12-20 18:57:58 +01:00
|
|
|
delete impl->InstallSuperVersion(new DBImpl::SuperVersion());
|
2013-07-16 20:56:46 +02:00
|
|
|
impl->mem_->SetLogNumber(impl->logfile_number_);
|
2011-03-18 23:37:00 +01:00
|
|
|
impl->DeleteObsoleteFiles();
|
2013-10-15 00:12:15 +02:00
|
|
|
impl->MaybeScheduleFlushOrCompaction();
|
2012-08-15 00:20:36 +02:00
|
|
|
impl->MaybeScheduleLogDBDeployStats();
|
2014-01-27 20:02:21 +01:00
|
|
|
s = impl->db_directory_->Fsync();
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
}
|
2013-09-04 22:13:08 +02:00
|
|
|
|
2014-01-16 01:15:43 +01:00
|
|
|
if (s.ok() && impl->options_.compaction_style == kCompactionStyleUniversal) {
|
|
|
|
Version* current = impl->versions_->current();
|
2013-09-04 22:13:08 +02:00
|
|
|
for (int i = 1; i < impl->NumberLevels(); i++) {
|
2014-01-16 01:15:43 +01:00
|
|
|
int num_files = current->NumLevelFiles(i);
|
2013-09-04 22:13:08 +02:00
|
|
|
if (num_files > 0) {
|
|
|
|
s = Status::InvalidArgument("Not all files are at level 0. Cannot "
|
|
|
|
"open with universal compaction style.");
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-01-16 01:15:43 +01:00
|
|
|
impl->mutex_.Unlock();
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
if (s.ok()) {
|
2014-02-27 20:38:55 +01:00
|
|
|
impl->opened_successfully_ = true;
|
2011-03-18 23:37:00 +01:00
|
|
|
*dbptr = impl;
|
|
|
|
} else {
|
|
|
|
delete impl;
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2011-05-21 04:17:43 +02:00
|
|
|
Snapshot::~Snapshot() {
|
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
Status DestroyDB(const std::string& dbname, const Options& options) {
|
2013-10-01 23:46:52 +02:00
|
|
|
const InternalKeyComparator comparator(options.comparator);
|
|
|
|
const InternalFilterPolicy filter_policy(options.filter_policy);
|
|
|
|
const Options& soptions(SanitizeOptions(
|
|
|
|
dbname, &comparator, &filter_policy, options));
|
|
|
|
Env* env = soptions.env;
|
2011-03-18 23:37:00 +01:00
|
|
|
std::vector<std::string> filenames;
|
2012-12-10 20:02:07 +01:00
|
|
|
std::vector<std::string> archiveFiles;
|
|
|
|
|
2013-10-01 23:46:52 +02:00
|
|
|
std::string archivedir = ArchivalDirectory(dbname);
|
2011-03-18 23:37:00 +01:00
|
|
|
// Ignore error in case directory does not exist
|
|
|
|
env->GetChildren(dbname, &filenames);
|
2013-10-01 23:46:52 +02:00
|
|
|
|
|
|
|
if (dbname != soptions.wal_dir) {
|
|
|
|
std::vector<std::string> logfilenames;
|
|
|
|
env->GetChildren(soptions.wal_dir, &logfilenames);
|
|
|
|
filenames.insert(filenames.end(), logfilenames.begin(), logfilenames.end());
|
|
|
|
archivedir = ArchivalDirectory(soptions.wal_dir);
|
|
|
|
}
|
2012-12-10 20:02:07 +01:00
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
if (filenames.empty()) {
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
FileLock* lock;
|
2011-07-15 02:20:57 +02:00
|
|
|
const std::string lockname = LockFileName(dbname);
|
|
|
|
Status result = env->LockFile(lockname, &lock);
|
2011-03-18 23:37:00 +01:00
|
|
|
if (result.ok()) {
|
|
|
|
uint64_t number;
|
|
|
|
FileType type;
|
2011-04-21 00:48:11 +02:00
|
|
|
for (size_t i = 0; i < filenames.size(); i++) {
|
2011-07-15 02:20:57 +02:00
|
|
|
if (ParseFileName(filenames[i], &number, &type) &&
|
2012-03-09 16:51:04 +01:00
|
|
|
type != kDBLockFile) { // Lock file will be deleted at end
|
2012-12-17 20:26:59 +01:00
|
|
|
Status del;
|
|
|
|
if (type == kMetaDatabase) {
|
|
|
|
del = DestroyDB(dbname + "/" + filenames[i], options);
|
2013-10-01 23:46:52 +02:00
|
|
|
} else if (type == kLogFile) {
|
|
|
|
del = env->DeleteFile(soptions.wal_dir + "/" + filenames[i]);
|
2012-12-17 20:26:59 +01:00
|
|
|
} else {
|
|
|
|
del = env->DeleteFile(dbname + "/" + filenames[i]);
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
if (result.ok() && !del.ok()) {
|
|
|
|
result = del;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2012-12-10 20:02:07 +01:00
|
|
|
|
2013-10-01 23:46:52 +02:00
|
|
|
env->GetChildren(archivedir, &archiveFiles);
|
2012-12-10 20:02:07 +01:00
|
|
|
// Delete archival files.
|
|
|
|
for (size_t i = 0; i < archiveFiles.size(); ++i) {
|
2013-10-11 03:02:10 +02:00
|
|
|
if (ParseFileName(archiveFiles[i], &number, &type) &&
|
|
|
|
type == kLogFile) {
|
2013-10-01 23:46:52 +02:00
|
|
|
Status del = env->DeleteFile(archivedir + "/" + archiveFiles[i]);
|
2012-12-10 20:02:07 +01:00
|
|
|
if (result.ok() && !del.ok()) {
|
|
|
|
result = del;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2012-12-14 22:21:53 +01:00
|
|
|
// ignore case where no archival directory is present.
|
2013-10-01 23:46:52 +02:00
|
|
|
env->DeleteDir(archivedir);
|
2012-12-10 20:02:07 +01:00
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
env->UnlockFile(lock); // Ignore error since state is already gone
|
2011-07-15 02:20:57 +02:00
|
|
|
env->DeleteFile(lockname);
|
2011-03-18 23:37:00 +01:00
|
|
|
env->DeleteDir(dbname); // Ignore error in case dir contains other files
|
2013-10-01 23:46:52 +02:00
|
|
|
env->DeleteDir(soptions.wal_dir);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
2012-08-21 09:33:21 +02:00
|
|
|
//
|
|
|
|
// A global method that can dump out the build version
|
2013-12-31 03:33:57 +01:00
|
|
|
void DumpLeveldbBuildVersion(Logger * log) {
|
2013-10-05 07:32:05 +02:00
|
|
|
Log(log, "Git sha %s", rocksdb_build_git_sha);
|
2013-01-24 20:45:11 +01:00
|
|
|
Log(log, "Compile time %s %s",
|
2013-10-05 07:32:05 +02:00
|
|
|
rocksdb_build_compile_time, rocksdb_build_compile_date);
|
2012-08-21 09:33:21 +02:00
|
|
|
}
|
|
|
|
|
2013-10-04 06:49:15 +02:00
|
|
|
} // namespace rocksdb
|