2016-02-09 23:12:00 +00:00
|
|
|
// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
|
2017-07-15 23:03:42 +00:00
|
|
|
// This source code is licensed under both the GPLv2 (found in the
|
|
|
|
// COPYING file in the root directory) and Apache 2.0 License
|
|
|
|
// (found in the LICENSE.Apache file in the root directory).
|
2013-10-16 21:59:46 +00:00
|
|
|
//
|
2011-03-18 22:37:00 +00:00
|
|
|
// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
|
|
|
|
// Use of this source code is governed by a BSD-style license that can be
|
|
|
|
// found in the LICENSE file. See the AUTHORS file for names of contributors.
|
|
|
|
|
|
|
|
#include "db/db_iter.h"
|
2014-07-15 23:10:18 +00:00
|
|
|
#include <string>
|
Added support for differential snapshots
Summary:
The motivation for this PR is to add to RocksDB support for differential (incremental) snapshots, as snapshot of the DB changes between two points in time (one can think of it as diff between to sequence numbers, or the diff D which can be thought of as an SST file or just set of KVs that can be applied to sequence number S1 to get the database to the state at sequence number S2).
This feature would be useful for various distributed storages layers built on top of RocksDB, as it should help reduce resources (time and network bandwidth) needed to recover and rebuilt DB instances as replicas in the context of distributed storages.
From the API standpoint that would like client app requesting iterator between (start seqnum) and current DB state, and reading the "diff".
This is a very draft PR for initial review in the discussion on the approach, i'm going to rework some parts and keep updating the PR.
For now, what's done here according to initial discussions:
Preserving deletes:
- We want to be able to optionally preserve recent deletes for some defined period of time, so that if a delete came in recently and might need to be included in the next incremental snapshot it would't get dropped by a compaction. This is done by adding new param to Options (preserve deletes flag) and new variable to DB Impl where we keep track of the sequence number after which we don't want to drop tombstones, even if they are otherwise eligible for deletion.
- I also added a new API call for clients to be able to advance this cutoff seqnum after which we drop deletes; i assume it's more flexible to let clients control this, since otherwise we'd need to keep some kind of timestamp < -- > seqnum mapping inside the DB, which sounds messy and painful to support. Clients could make use of it by periodically calling GetLatestSequenceNumber(), noting the timestamp, doing some calculation and figuring out by how much we need to advance the cutoff seqnum.
- Compaction codepath in compaction_iterator.cc has been modified to avoid dropping tombstones with seqnum > cutoff seqnum.
Iterator changes:
- couple params added to ReadOptions, to optionally allow client to request internal keys instead of user keys (so that client can get the latest value of a key, be it delete marker or a put), as well as min timestamp and min seqnum.
TableCache changes:
- I modified table_cache code to be able to quickly exclude SST files from iterators heep if creation_time on the file is less then iter_start_ts as passed in ReadOptions. That would help a lot in some DB settings (like reading very recent data only or using FIFO compactions), but not so much for universal compaction with more or less long iterator time span.
What's left:
- Still looking at how to best plug that inside DBIter codepath. So far it seems that FindNextUserKeyInternal only parses values as UserKeys, and iter->key() call generally returns user key. Can we add new API to DBIter as internal_key(), and modify this internal method to optionally set saved_key_ to point to the full internal key? I don't need to store actual seqnum there, but I do need to store type.
Closes https://github.com/facebook/rocksdb/pull/2999
Differential Revision: D6175602
Pulled By: mikhail-antonov
fbshipit-source-id: c779a6696ee2d574d86c69cec866a3ae095aa900
2017-11-02 01:43:29 +00:00
|
|
|
#include <iostream>
|
2014-08-08 16:44:14 +00:00
|
|
|
#include <limits>
|
2011-03-18 22:37:00 +00:00
|
|
|
|
|
|
|
#include "db/dbformat.h"
|
Eliminate std::deque initialization while iterating over merge operands
Summary:
This patch is similar to D52563, When we iterate over a DB with merge operands we keep creating std::queue to store the operands, optimize this by reusing merge_operands_ data member
Before the patch
```
./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq" --db="/dev/shm/bench_merge_memcpy_on_the_fly/" --merge_operator="put" --merge_keys=10000 --num=10000
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 3.757 micros/op 266141 ops/sec; 29.4 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.413 micros/op 2423538 ops/sec; 268.1 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.451 micros/op 2219071 ops/sec; 245.5 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.420 micros/op 2382039 ops/sec; 263.5 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.408 micros/op 2452017 ops/sec; 271.3 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 3.947 micros/op 253376 ops/sec; 28.0 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.441 micros/op 2266473 ops/sec; 250.7 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.471 micros/op 2122033 ops/sec; 234.8 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.440 micros/op 2271407 ops/sec; 251.3 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.429 micros/op 2331471 ops/sec; 257.9 MB/s
```
with the patch
```
./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq" --db="/dev/shm/bench_merge_memcpy_on_the_fly/" --merge_operator="put" --merge_keys=10000 --num=10000
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 4.080 micros/op 245092 ops/sec; 27.1 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.308 micros/op 3241843 ops/sec; 358.6 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.312 micros/op 3200408 ops/sec; 354.0 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.332 micros/op 3013962 ops/sec; 333.4 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.300 micros/op 3328017 ops/sec; 368.2 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 3.973 micros/op 251705 ops/sec; 27.8 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.320 micros/op 3123752 ops/sec; 345.6 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.335 micros/op 2986641 ops/sec; 330.4 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.339 micros/op 2950047 ops/sec; 326.4 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.319 micros/op 3131565 ops/sec; 346.4 MB/s
```
Test Plan: make check -j64
Reviewers: yhchiang, andrewkr, sdong
Reviewed By: sdong
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D56031
2016-04-01 22:48:55 +00:00
|
|
|
#include "db/merge_context.h"
|
2016-06-13 23:17:26 +00:00
|
|
|
#include "db/merge_helper.h"
|
2016-04-26 19:41:07 +00:00
|
|
|
#include "db/pinned_iterators_manager.h"
|
2017-04-06 02:02:00 +00:00
|
|
|
#include "monitoring/perf_context_imp.h"
|
2013-08-23 15:38:13 +00:00
|
|
|
#include "rocksdb/env.h"
|
|
|
|
#include "rocksdb/iterator.h"
|
|
|
|
#include "rocksdb/merge_operator.h"
|
2016-03-01 02:38:03 +00:00
|
|
|
#include "rocksdb/options.h"
|
2015-10-12 22:06:38 +00:00
|
|
|
#include "table/internal_iterator.h"
|
In DB::NewIterator(), try to allocate the whole iterator tree in an arena
Summary:
In this patch, try to allocate the whole iterator tree starting from DBIter from an arena
1. ArenaWrappedDBIter is created when serves as the entry point of an iterator tree, with an arena in it.
2. Add an option to create iterator from arena for following iterators: DBIter, MergingIterator, MemtableIterator, all mem table's iterators, all table reader's iterators and two level iterator.
3. MergeIteratorBuilder is created to incrementally build the tree of internal iterators. It is passed to mem table list and version set and add iterators to it.
Limitations:
(1) Only DB::NewIterator() without tailing uses the arena. Other cases, including readonly DB and compactions are still from malloc
(2) Two level iterator itself is allocated in arena, but not iterators inside it.
Test Plan: make all check
Reviewers: ljin, haobo
Reviewed By: haobo
Subscribers: leveldb, dhruba, yhchiang, igor
Differential Revision: https://reviews.facebook.net/D18513
2014-06-02 23:38:00 +00:00
|
|
|
#include "util/arena.h"
|
2017-04-04 01:27:24 +00:00
|
|
|
#include "util/filename.h"
|
2011-03-18 22:37:00 +00:00
|
|
|
#include "util/logging.h"
|
|
|
|
#include "util/mutexlock.h"
|
2016-03-01 02:38:03 +00:00
|
|
|
#include "util/string_util.h"
|
2011-03-18 22:37:00 +00:00
|
|
|
|
2013-10-04 04:49:15 +00:00
|
|
|
namespace rocksdb {
|
2011-03-18 22:37:00 +00:00
|
|
|
|
|
|
|
#if 0
|
|
|
|
static void DumpInternalIter(Iterator* iter) {
|
|
|
|
for (iter->SeekToFirst(); iter->Valid(); iter->Next()) {
|
|
|
|
ParsedInternalKey k;
|
|
|
|
if (!ParseInternalKey(iter->key(), &k)) {
|
|
|
|
fprintf(stderr, "Corrupt '%s'\n", EscapeString(iter->key()).c_str());
|
|
|
|
} else {
|
|
|
|
fprintf(stderr, "@ '%s'\n", k.DebugString().c_str());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
|
|
|
// Memtables and sstables that make the DB representation contain
|
|
|
|
// (userkey,seq,type) => uservalue entries. DBIter
|
|
|
|
// combines multiple entries for the same userkey found in the DB
|
|
|
|
// representation into a single entry while accounting for sequence
|
|
|
|
// numbers, deletion markers, overwrites, etc.
|
2017-09-11 18:43:17 +00:00
|
|
|
class DBIter final: public Iterator {
|
2011-03-18 22:37:00 +00:00
|
|
|
public:
|
2013-03-21 22:59:47 +00:00
|
|
|
// The following is grossly complicated. TODO: clean it up
|
2011-03-25 20:27:43 +00:00
|
|
|
// Which direction is the iterator currently moving?
|
|
|
|
// (1) When moving forward, the internal iterator is positioned at
|
|
|
|
// the exact entry that yields this->key(), this->value()
|
|
|
|
// (2) When moving backwards, the internal iterator is positioned
|
|
|
|
// just before all entries whose user key == this->key().
|
|
|
|
enum Direction {
|
|
|
|
kForward,
|
|
|
|
kReverse
|
|
|
|
};
|
|
|
|
|
Aggregate hot Iterator counters in LocalStatistics (DBIter::Next perf regression)
Summary:
This patch bump the counters in the frequent code path DBIter::Next() / DBIter::Prev() in a local data members and send them to Statistics when the iterator is destroyed
A better solution will be to have thread_local implementation for Statistics
New performance
```
readseq : 0.035 micros/op 28597881 ops/sec; 3163.7 MB/s
1,851,568,819 stalled-cycles-frontend # 31.29% frontend cycles idle [49.86%]
884,929,823 stalled-cycles-backend # 14.95% backend cycles idle [50.21%]
readreverse : 0.071 micros/op 14077393 ops/sec; 1557.3 MB/s
3,239,575,993 stalled-cycles-frontend # 27.36% frontend cycles idle [49.96%]
1,558,253,983 stalled-cycles-backend # 13.16% backend cycles idle [50.14%]
```
Existing performance
```
readreverse : 0.174 micros/op 5732342 ops/sec; 634.1 MB/s
20,570,209,389 stalled-cycles-frontend # 70.71% frontend cycles idle [50.01%]
18,422,816,837 stalled-cycles-backend # 63.33% backend cycles idle [50.04%]
readseq : 0.119 micros/op 8400537 ops/sec; 929.3 MB/s
15,634,225,844 stalled-cycles-frontend # 79.07% frontend cycles idle [49.96%]
14,227,427,453 stalled-cycles-backend # 71.95% backend cycles idle [50.09%]
```
Test Plan: unit tests
Reviewers: yhchiang, sdong, igor
Reviewed By: sdong
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D55107
2016-03-12 03:01:12 +00:00
|
|
|
// LocalStatistics contain Statistics counters that will be aggregated per
|
|
|
|
// each iterator instance and then will be sent to the global statistics when
|
|
|
|
// the iterator is destroyed.
|
|
|
|
//
|
|
|
|
// The purpose of this approach is to avoid perf regression happening
|
|
|
|
// when multiple threads bump the atomic counters from a DBIter::Next().
|
|
|
|
struct LocalStatistics {
|
|
|
|
explicit LocalStatistics() { ResetCounters(); }
|
|
|
|
|
|
|
|
void ResetCounters() {
|
|
|
|
next_count_ = 0;
|
|
|
|
next_found_count_ = 0;
|
|
|
|
prev_count_ = 0;
|
|
|
|
prev_found_count_ = 0;
|
|
|
|
bytes_read_ = 0;
|
2017-11-21 05:12:55 +00:00
|
|
|
skip_count_ = 0;
|
Aggregate hot Iterator counters in LocalStatistics (DBIter::Next perf regression)
Summary:
This patch bump the counters in the frequent code path DBIter::Next() / DBIter::Prev() in a local data members and send them to Statistics when the iterator is destroyed
A better solution will be to have thread_local implementation for Statistics
New performance
```
readseq : 0.035 micros/op 28597881 ops/sec; 3163.7 MB/s
1,851,568,819 stalled-cycles-frontend # 31.29% frontend cycles idle [49.86%]
884,929,823 stalled-cycles-backend # 14.95% backend cycles idle [50.21%]
readreverse : 0.071 micros/op 14077393 ops/sec; 1557.3 MB/s
3,239,575,993 stalled-cycles-frontend # 27.36% frontend cycles idle [49.96%]
1,558,253,983 stalled-cycles-backend # 13.16% backend cycles idle [50.14%]
```
Existing performance
```
readreverse : 0.174 micros/op 5732342 ops/sec; 634.1 MB/s
20,570,209,389 stalled-cycles-frontend # 70.71% frontend cycles idle [50.01%]
18,422,816,837 stalled-cycles-backend # 63.33% backend cycles idle [50.04%]
readseq : 0.119 micros/op 8400537 ops/sec; 929.3 MB/s
15,634,225,844 stalled-cycles-frontend # 79.07% frontend cycles idle [49.96%]
14,227,427,453 stalled-cycles-backend # 71.95% backend cycles idle [50.09%]
```
Test Plan: unit tests
Reviewers: yhchiang, sdong, igor
Reviewed By: sdong
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D55107
2016-03-12 03:01:12 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
void BumpGlobalStatistics(Statistics* global_statistics) {
|
|
|
|
RecordTick(global_statistics, NUMBER_DB_NEXT, next_count_);
|
|
|
|
RecordTick(global_statistics, NUMBER_DB_NEXT_FOUND, next_found_count_);
|
|
|
|
RecordTick(global_statistics, NUMBER_DB_PREV, prev_count_);
|
|
|
|
RecordTick(global_statistics, NUMBER_DB_PREV_FOUND, prev_found_count_);
|
|
|
|
RecordTick(global_statistics, ITER_BYTES_READ, bytes_read_);
|
2017-11-21 05:12:55 +00:00
|
|
|
RecordTick(global_statistics, NUMBER_ITER_SKIP, skip_count_);
|
2017-08-18 18:40:36 +00:00
|
|
|
PERF_COUNTER_ADD(iter_read_bytes, bytes_read_);
|
Aggregate hot Iterator counters in LocalStatistics (DBIter::Next perf regression)
Summary:
This patch bump the counters in the frequent code path DBIter::Next() / DBIter::Prev() in a local data members and send them to Statistics when the iterator is destroyed
A better solution will be to have thread_local implementation for Statistics
New performance
```
readseq : 0.035 micros/op 28597881 ops/sec; 3163.7 MB/s
1,851,568,819 stalled-cycles-frontend # 31.29% frontend cycles idle [49.86%]
884,929,823 stalled-cycles-backend # 14.95% backend cycles idle [50.21%]
readreverse : 0.071 micros/op 14077393 ops/sec; 1557.3 MB/s
3,239,575,993 stalled-cycles-frontend # 27.36% frontend cycles idle [49.96%]
1,558,253,983 stalled-cycles-backend # 13.16% backend cycles idle [50.14%]
```
Existing performance
```
readreverse : 0.174 micros/op 5732342 ops/sec; 634.1 MB/s
20,570,209,389 stalled-cycles-frontend # 70.71% frontend cycles idle [50.01%]
18,422,816,837 stalled-cycles-backend # 63.33% backend cycles idle [50.04%]
readseq : 0.119 micros/op 8400537 ops/sec; 929.3 MB/s
15,634,225,844 stalled-cycles-frontend # 79.07% frontend cycles idle [49.96%]
14,227,427,453 stalled-cycles-backend # 71.95% backend cycles idle [50.09%]
```
Test Plan: unit tests
Reviewers: yhchiang, sdong, igor
Reviewed By: sdong
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D55107
2016-03-12 03:01:12 +00:00
|
|
|
ResetCounters();
|
|
|
|
}
|
|
|
|
|
|
|
|
// Map to Tickers::NUMBER_DB_NEXT
|
|
|
|
uint64_t next_count_;
|
|
|
|
// Map to Tickers::NUMBER_DB_NEXT_FOUND
|
|
|
|
uint64_t next_found_count_;
|
|
|
|
// Map to Tickers::NUMBER_DB_PREV
|
|
|
|
uint64_t prev_count_;
|
|
|
|
// Map to Tickers::NUMBER_DB_PREV_FOUND
|
|
|
|
uint64_t prev_found_count_;
|
|
|
|
// Map to Tickers::ITER_BYTES_READ
|
|
|
|
uint64_t bytes_read_;
|
2017-11-21 05:12:55 +00:00
|
|
|
// Map to Tickers::NUMBER_ITER_SKIP
|
|
|
|
uint64_t skip_count_;
|
Aggregate hot Iterator counters in LocalStatistics (DBIter::Next perf regression)
Summary:
This patch bump the counters in the frequent code path DBIter::Next() / DBIter::Prev() in a local data members and send them to Statistics when the iterator is destroyed
A better solution will be to have thread_local implementation for Statistics
New performance
```
readseq : 0.035 micros/op 28597881 ops/sec; 3163.7 MB/s
1,851,568,819 stalled-cycles-frontend # 31.29% frontend cycles idle [49.86%]
884,929,823 stalled-cycles-backend # 14.95% backend cycles idle [50.21%]
readreverse : 0.071 micros/op 14077393 ops/sec; 1557.3 MB/s
3,239,575,993 stalled-cycles-frontend # 27.36% frontend cycles idle [49.96%]
1,558,253,983 stalled-cycles-backend # 13.16% backend cycles idle [50.14%]
```
Existing performance
```
readreverse : 0.174 micros/op 5732342 ops/sec; 634.1 MB/s
20,570,209,389 stalled-cycles-frontend # 70.71% frontend cycles idle [50.01%]
18,422,816,837 stalled-cycles-backend # 63.33% backend cycles idle [50.04%]
readseq : 0.119 micros/op 8400537 ops/sec; 929.3 MB/s
15,634,225,844 stalled-cycles-frontend # 79.07% frontend cycles idle [49.96%]
14,227,427,453 stalled-cycles-backend # 71.95% backend cycles idle [50.09%]
```
Test Plan: unit tests
Reviewers: yhchiang, sdong, igor
Reviewed By: sdong
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D55107
2016-03-12 03:01:12 +00:00
|
|
|
};
|
|
|
|
|
2017-07-24 17:28:17 +00:00
|
|
|
DBIter(Env* _env, const ReadOptions& read_options,
|
2017-04-10 18:13:23 +00:00
|
|
|
const ImmutableCFOptions& cf_options, const Comparator* cmp,
|
2015-10-12 22:06:38 +00:00
|
|
|
InternalIterator* iter, SequenceNumber s, bool arena_mode,
|
2017-10-10 00:05:34 +00:00
|
|
|
uint64_t max_sequential_skip_in_iterations,
|
|
|
|
ReadCallback* read_callback, bool allow_blob)
|
In DB::NewIterator(), try to allocate the whole iterator tree in an arena
Summary:
In this patch, try to allocate the whole iterator tree starting from DBIter from an arena
1. ArenaWrappedDBIter is created when serves as the entry point of an iterator tree, with an arena in it.
2. Add an option to create iterator from arena for following iterators: DBIter, MergingIterator, MemtableIterator, all mem table's iterators, all table reader's iterators and two level iterator.
3. MergeIteratorBuilder is created to incrementally build the tree of internal iterators. It is passed to mem table list and version set and add iterators to it.
Limitations:
(1) Only DB::NewIterator() without tailing uses the arena. Other cases, including readonly DB and compactions are still from malloc
(2) Two level iterator itself is allocated in arena, but not iterators inside it.
Test Plan: make all check
Reviewers: ljin, haobo
Reviewed By: haobo
Subscribers: leveldb, dhruba, yhchiang, igor
Differential Revision: https://reviews.facebook.net/D18513
2014-06-02 23:38:00 +00:00
|
|
|
: arena_mode_(arena_mode),
|
2017-07-24 17:28:17 +00:00
|
|
|
env_(_env),
|
2017-04-10 18:13:23 +00:00
|
|
|
logger_(cf_options.info_log),
|
2011-03-18 22:37:00 +00:00
|
|
|
user_comparator_(cmp),
|
2017-04-10 18:13:23 +00:00
|
|
|
merge_operator_(cf_options.merge_operator),
|
2011-03-18 22:37:00 +00:00
|
|
|
iter_(iter),
|
|
|
|
sequence_(s),
|
2011-03-25 20:27:43 +00:00
|
|
|
direction_(kForward),
|
2013-03-21 22:59:47 +00:00
|
|
|
valid_(false),
|
2013-05-28 21:00:10 +00:00
|
|
|
current_entry_is_merged_(false),
|
2017-04-10 18:13:23 +00:00
|
|
|
statistics_(cf_options.statistics),
|
2017-10-27 04:00:17 +00:00
|
|
|
num_internal_keys_skipped_(0),
|
2017-10-27 00:14:04 +00:00
|
|
|
iterate_lower_bound_(read_options.iterate_lower_bound),
|
2017-04-10 18:13:23 +00:00
|
|
|
iterate_upper_bound_(read_options.iterate_upper_bound),
|
|
|
|
prefix_same_as_start_(read_options.prefix_same_as_start),
|
|
|
|
pin_thru_lifetime_(read_options.pin_data),
|
|
|
|
total_order_seek_(read_options.total_order_seek),
|
|
|
|
range_del_agg_(cf_options.internal_comparator, s,
|
2017-10-03 16:08:07 +00:00
|
|
|
true /* collapse_deletions */),
|
2017-10-10 00:05:34 +00:00
|
|
|
read_callback_(read_callback),
|
2017-10-27 04:00:17 +00:00
|
|
|
allow_blob_(allow_blob),
|
Added support for differential snapshots
Summary:
The motivation for this PR is to add to RocksDB support for differential (incremental) snapshots, as snapshot of the DB changes between two points in time (one can think of it as diff between to sequence numbers, or the diff D which can be thought of as an SST file or just set of KVs that can be applied to sequence number S1 to get the database to the state at sequence number S2).
This feature would be useful for various distributed storages layers built on top of RocksDB, as it should help reduce resources (time and network bandwidth) needed to recover and rebuilt DB instances as replicas in the context of distributed storages.
From the API standpoint that would like client app requesting iterator between (start seqnum) and current DB state, and reading the "diff".
This is a very draft PR for initial review in the discussion on the approach, i'm going to rework some parts and keep updating the PR.
For now, what's done here according to initial discussions:
Preserving deletes:
- We want to be able to optionally preserve recent deletes for some defined period of time, so that if a delete came in recently and might need to be included in the next incremental snapshot it would't get dropped by a compaction. This is done by adding new param to Options (preserve deletes flag) and new variable to DB Impl where we keep track of the sequence number after which we don't want to drop tombstones, even if they are otherwise eligible for deletion.
- I also added a new API call for clients to be able to advance this cutoff seqnum after which we drop deletes; i assume it's more flexible to let clients control this, since otherwise we'd need to keep some kind of timestamp < -- > seqnum mapping inside the DB, which sounds messy and painful to support. Clients could make use of it by periodically calling GetLatestSequenceNumber(), noting the timestamp, doing some calculation and figuring out by how much we need to advance the cutoff seqnum.
- Compaction codepath in compaction_iterator.cc has been modified to avoid dropping tombstones with seqnum > cutoff seqnum.
Iterator changes:
- couple params added to ReadOptions, to optionally allow client to request internal keys instead of user keys (so that client can get the latest value of a key, be it delete marker or a put), as well as min timestamp and min seqnum.
TableCache changes:
- I modified table_cache code to be able to quickly exclude SST files from iterators heep if creation_time on the file is less then iter_start_ts as passed in ReadOptions. That would help a lot in some DB settings (like reading very recent data only or using FIFO compactions), but not so much for universal compaction with more or less long iterator time span.
What's left:
- Still looking at how to best plug that inside DBIter codepath. So far it seems that FindNextUserKeyInternal only parses values as UserKeys, and iter->key() call generally returns user key. Can we add new API to DBIter as internal_key(), and modify this internal method to optionally set saved_key_ to point to the full internal key? I don't need to store actual seqnum there, but I do need to store type.
Closes https://github.com/facebook/rocksdb/pull/2999
Differential Revision: D6175602
Pulled By: mikhail-antonov
fbshipit-source-id: c779a6696ee2d574d86c69cec866a3ae095aa900
2017-11-02 01:43:29 +00:00
|
|
|
is_blob_(false),
|
|
|
|
start_seqnum_(read_options.iter_start_seqnum) {
|
2014-07-28 19:05:36 +00:00
|
|
|
RecordTick(statistics_, NO_ITERATORS);
|
2017-04-10 18:13:23 +00:00
|
|
|
prefix_extractor_ = cf_options.prefix_extractor;
|
2014-09-08 22:04:34 +00:00
|
|
|
max_skip_ = max_sequential_skip_in_iterations;
|
2017-04-10 18:13:23 +00:00
|
|
|
max_skippable_internal_keys_ = read_options.max_skippable_internal_keys;
|
2016-04-26 19:41:07 +00:00
|
|
|
if (pin_thru_lifetime_) {
|
|
|
|
pinned_iters_mgr_.StartPinning();
|
|
|
|
}
|
|
|
|
if (iter_) {
|
|
|
|
iter_->SetPinnedItersMgr(&pinned_iters_mgr_);
|
|
|
|
}
|
2011-03-18 22:37:00 +00:00
|
|
|
}
|
|
|
|
virtual ~DBIter() {
|
2016-05-03 04:46:30 +00:00
|
|
|
// Release pinned data if any
|
2016-08-11 18:54:17 +00:00
|
|
|
if (pinned_iters_mgr_.PinningEnabled()) {
|
|
|
|
pinned_iters_mgr_.ReleasePinnedData();
|
|
|
|
}
|
2017-10-19 17:48:47 +00:00
|
|
|
// Compiler warning issue filed:
|
|
|
|
// https://github.com/facebook/rocksdb/issues/3013
|
|
|
|
RecordTick(statistics_, NO_ITERATORS, uint64_t(-1));
|
2017-11-21 05:12:55 +00:00
|
|
|
ResetInternalKeysSkippedCounter();
|
Aggregate hot Iterator counters in LocalStatistics (DBIter::Next perf regression)
Summary:
This patch bump the counters in the frequent code path DBIter::Next() / DBIter::Prev() in a local data members and send them to Statistics when the iterator is destroyed
A better solution will be to have thread_local implementation for Statistics
New performance
```
readseq : 0.035 micros/op 28597881 ops/sec; 3163.7 MB/s
1,851,568,819 stalled-cycles-frontend # 31.29% frontend cycles idle [49.86%]
884,929,823 stalled-cycles-backend # 14.95% backend cycles idle [50.21%]
readreverse : 0.071 micros/op 14077393 ops/sec; 1557.3 MB/s
3,239,575,993 stalled-cycles-frontend # 27.36% frontend cycles idle [49.96%]
1,558,253,983 stalled-cycles-backend # 13.16% backend cycles idle [50.14%]
```
Existing performance
```
readreverse : 0.174 micros/op 5732342 ops/sec; 634.1 MB/s
20,570,209,389 stalled-cycles-frontend # 70.71% frontend cycles idle [50.01%]
18,422,816,837 stalled-cycles-backend # 63.33% backend cycles idle [50.04%]
readseq : 0.119 micros/op 8400537 ops/sec; 929.3 MB/s
15,634,225,844 stalled-cycles-frontend # 79.07% frontend cycles idle [49.96%]
14,227,427,453 stalled-cycles-backend # 71.95% backend cycles idle [50.09%]
```
Test Plan: unit tests
Reviewers: yhchiang, sdong, igor
Reviewed By: sdong
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D55107
2016-03-12 03:01:12 +00:00
|
|
|
local_stats_.BumpGlobalStatistics(statistics_);
|
In DB::NewIterator(), try to allocate the whole iterator tree in an arena
Summary:
In this patch, try to allocate the whole iterator tree starting from DBIter from an arena
1. ArenaWrappedDBIter is created when serves as the entry point of an iterator tree, with an arena in it.
2. Add an option to create iterator from arena for following iterators: DBIter, MergingIterator, MemtableIterator, all mem table's iterators, all table reader's iterators and two level iterator.
3. MergeIteratorBuilder is created to incrementally build the tree of internal iterators. It is passed to mem table list and version set and add iterators to it.
Limitations:
(1) Only DB::NewIterator() without tailing uses the arena. Other cases, including readonly DB and compactions are still from malloc
(2) Two level iterator itself is allocated in arena, but not iterators inside it.
Test Plan: make all check
Reviewers: ljin, haobo
Reviewed By: haobo
Subscribers: leveldb, dhruba, yhchiang, igor
Differential Revision: https://reviews.facebook.net/D18513
2014-06-02 23:38:00 +00:00
|
|
|
if (!arena_mode_) {
|
|
|
|
delete iter_;
|
|
|
|
} else {
|
2015-10-12 22:06:38 +00:00
|
|
|
iter_->~InternalIterator();
|
In DB::NewIterator(), try to allocate the whole iterator tree in an arena
Summary:
In this patch, try to allocate the whole iterator tree starting from DBIter from an arena
1. ArenaWrappedDBIter is created when serves as the entry point of an iterator tree, with an arena in it.
2. Add an option to create iterator from arena for following iterators: DBIter, MergingIterator, MemtableIterator, all mem table's iterators, all table reader's iterators and two level iterator.
3. MergeIteratorBuilder is created to incrementally build the tree of internal iterators. It is passed to mem table list and version set and add iterators to it.
Limitations:
(1) Only DB::NewIterator() without tailing uses the arena. Other cases, including readonly DB and compactions are still from malloc
(2) Two level iterator itself is allocated in arena, but not iterators inside it.
Test Plan: make all check
Reviewers: ljin, haobo
Reviewed By: haobo
Subscribers: leveldb, dhruba, yhchiang, igor
Differential Revision: https://reviews.facebook.net/D18513
2014-06-02 23:38:00 +00:00
|
|
|
}
|
|
|
|
}
|
2015-10-12 22:06:38 +00:00
|
|
|
virtual void SetIter(InternalIterator* iter) {
|
In DB::NewIterator(), try to allocate the whole iterator tree in an arena
Summary:
In this patch, try to allocate the whole iterator tree starting from DBIter from an arena
1. ArenaWrappedDBIter is created when serves as the entry point of an iterator tree, with an arena in it.
2. Add an option to create iterator from arena for following iterators: DBIter, MergingIterator, MemtableIterator, all mem table's iterators, all table reader's iterators and two level iterator.
3. MergeIteratorBuilder is created to incrementally build the tree of internal iterators. It is passed to mem table list and version set and add iterators to it.
Limitations:
(1) Only DB::NewIterator() without tailing uses the arena. Other cases, including readonly DB and compactions are still from malloc
(2) Two level iterator itself is allocated in arena, but not iterators inside it.
Test Plan: make all check
Reviewers: ljin, haobo
Reviewed By: haobo
Subscribers: leveldb, dhruba, yhchiang, igor
Differential Revision: https://reviews.facebook.net/D18513
2014-06-02 23:38:00 +00:00
|
|
|
assert(iter_ == nullptr);
|
|
|
|
iter_ = iter;
|
2016-04-26 19:41:07 +00:00
|
|
|
iter_->SetPinnedItersMgr(&pinned_iters_mgr_);
|
2011-03-18 22:37:00 +00:00
|
|
|
}
|
2016-11-04 18:53:38 +00:00
|
|
|
virtual RangeDelAggregator* GetRangeDelAggregator() {
|
|
|
|
return &range_del_agg_;
|
|
|
|
}
|
|
|
|
|
2015-02-26 19:28:41 +00:00
|
|
|
virtual bool Valid() const override { return valid_; }
|
|
|
|
virtual Slice key() const override {
|
2011-03-18 22:37:00 +00:00
|
|
|
assert(valid_);
|
Added support for differential snapshots
Summary:
The motivation for this PR is to add to RocksDB support for differential (incremental) snapshots, as snapshot of the DB changes between two points in time (one can think of it as diff between to sequence numbers, or the diff D which can be thought of as an SST file or just set of KVs that can be applied to sequence number S1 to get the database to the state at sequence number S2).
This feature would be useful for various distributed storages layers built on top of RocksDB, as it should help reduce resources (time and network bandwidth) needed to recover and rebuilt DB instances as replicas in the context of distributed storages.
From the API standpoint that would like client app requesting iterator between (start seqnum) and current DB state, and reading the "diff".
This is a very draft PR for initial review in the discussion on the approach, i'm going to rework some parts and keep updating the PR.
For now, what's done here according to initial discussions:
Preserving deletes:
- We want to be able to optionally preserve recent deletes for some defined period of time, so that if a delete came in recently and might need to be included in the next incremental snapshot it would't get dropped by a compaction. This is done by adding new param to Options (preserve deletes flag) and new variable to DB Impl where we keep track of the sequence number after which we don't want to drop tombstones, even if they are otherwise eligible for deletion.
- I also added a new API call for clients to be able to advance this cutoff seqnum after which we drop deletes; i assume it's more flexible to let clients control this, since otherwise we'd need to keep some kind of timestamp < -- > seqnum mapping inside the DB, which sounds messy and painful to support. Clients could make use of it by periodically calling GetLatestSequenceNumber(), noting the timestamp, doing some calculation and figuring out by how much we need to advance the cutoff seqnum.
- Compaction codepath in compaction_iterator.cc has been modified to avoid dropping tombstones with seqnum > cutoff seqnum.
Iterator changes:
- couple params added to ReadOptions, to optionally allow client to request internal keys instead of user keys (so that client can get the latest value of a key, be it delete marker or a put), as well as min timestamp and min seqnum.
TableCache changes:
- I modified table_cache code to be able to quickly exclude SST files from iterators heep if creation_time on the file is less then iter_start_ts as passed in ReadOptions. That would help a lot in some DB settings (like reading very recent data only or using FIFO compactions), but not so much for universal compaction with more or less long iterator time span.
What's left:
- Still looking at how to best plug that inside DBIter codepath. So far it seems that FindNextUserKeyInternal only parses values as UserKeys, and iter->key() call generally returns user key. Can we add new API to DBIter as internal_key(), and modify this internal method to optionally set saved_key_ to point to the full internal key? I don't need to store actual seqnum there, but I do need to store type.
Closes https://github.com/facebook/rocksdb/pull/2999
Differential Revision: D6175602
Pulled By: mikhail-antonov
fbshipit-source-id: c779a6696ee2d574d86c69cec866a3ae095aa900
2017-11-02 01:43:29 +00:00
|
|
|
if(start_seqnum_ > 0) {
|
|
|
|
return saved_key_.GetInternalKey();
|
|
|
|
} else {
|
|
|
|
return saved_key_.GetUserKey();
|
|
|
|
}
|
|
|
|
|
2011-03-18 22:37:00 +00:00
|
|
|
}
|
2015-02-26 19:28:41 +00:00
|
|
|
virtual Slice value() const override {
|
2011-03-18 22:37:00 +00:00
|
|
|
assert(valid_);
|
2016-05-03 04:46:30 +00:00
|
|
|
if (current_entry_is_merged_) {
|
Introduce FullMergeV2 (eliminate memcpy from merge operators)
Summary:
This diff update the code to pin the merge operator operands while the merge operation is done, so that we can eliminate the memcpy cost, to do that we need a new public API for FullMerge that replace the std::deque<std::string> with std::vector<Slice>
This diff is stacked on top of D56493 and D56511
In this diff we
- Update FullMergeV2 arguments to be encapsulated in MergeOperationInput and MergeOperationOutput which will make it easier to add new arguments in the future
- Replace std::deque<std::string> with std::vector<Slice> to pass operands
- Replace MergeContext std::deque with std::vector (based on a simple benchmark I ran https://gist.github.com/IslamAbdelRahman/78fc86c9ab9f52b1df791e58943fb187)
- Allow FullMergeV2 output to be an existing operand
```
[Everything in Memtable | 10K operands | 10 KB each | 1 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=10000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 0.607 micros/op 1648235 ops/sec; 16121.2 MB/s
readseq : 0.478 micros/op 2091546 ops/sec; 20457.2 MB/s
readseq : 0.252 micros/op 3972081 ops/sec; 38850.5 MB/s
readseq : 0.237 micros/op 4218328 ops/sec; 41259.0 MB/s
readseq : 0.247 micros/op 4043927 ops/sec; 39553.2 MB/s
[master]
readseq : 3.935 micros/op 254140 ops/sec; 2485.7 MB/s
readseq : 3.722 micros/op 268657 ops/sec; 2627.7 MB/s
readseq : 3.149 micros/op 317605 ops/sec; 3106.5 MB/s
readseq : 3.125 micros/op 320024 ops/sec; 3130.1 MB/s
readseq : 4.075 micros/op 245374 ops/sec; 2400.0 MB/s
```
```
[Everything in Memtable | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=1000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 3.472 micros/op 288018 ops/sec; 2817.1 MB/s
readseq : 2.304 micros/op 434027 ops/sec; 4245.2 MB/s
readseq : 1.163 micros/op 859845 ops/sec; 8410.0 MB/s
readseq : 1.192 micros/op 838926 ops/sec; 8205.4 MB/s
readseq : 1.250 micros/op 800000 ops/sec; 7824.7 MB/s
[master]
readseq : 24.025 micros/op 41623 ops/sec; 407.1 MB/s
readseq : 18.489 micros/op 54086 ops/sec; 529.0 MB/s
readseq : 18.693 micros/op 53495 ops/sec; 523.2 MB/s
readseq : 23.621 micros/op 42335 ops/sec; 414.1 MB/s
readseq : 18.775 micros/op 53262 ops/sec; 521.0 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 1 operand per key]
[FullMergeV2]
$ DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
readseq : 14.741 micros/op 67837 ops/sec; 663.5 MB/s
readseq : 1.029 micros/op 971446 ops/sec; 9501.6 MB/s
readseq : 0.974 micros/op 1026229 ops/sec; 10037.4 MB/s
readseq : 0.965 micros/op 1036080 ops/sec; 10133.8 MB/s
readseq : 0.943 micros/op 1060657 ops/sec; 10374.2 MB/s
[master]
readseq : 16.735 micros/op 59755 ops/sec; 584.5 MB/s
readseq : 3.029 micros/op 330151 ops/sec; 3229.2 MB/s
readseq : 3.136 micros/op 318883 ops/sec; 3119.0 MB/s
readseq : 3.065 micros/op 326245 ops/sec; 3191.0 MB/s
readseq : 3.014 micros/op 331813 ops/sec; 3245.4 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10-operands-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
[FullMergeV2]
readseq : 24.325 micros/op 41109 ops/sec; 402.1 MB/s
readseq : 1.470 micros/op 680272 ops/sec; 6653.7 MB/s
readseq : 1.231 micros/op 812347 ops/sec; 7945.5 MB/s
readseq : 1.091 micros/op 916590 ops/sec; 8965.1 MB/s
readseq : 1.109 micros/op 901713 ops/sec; 8819.6 MB/s
[master]
readseq : 27.257 micros/op 36687 ops/sec; 358.8 MB/s
readseq : 4.443 micros/op 225073 ops/sec; 2201.4 MB/s
readseq : 5.830 micros/op 171526 ops/sec; 1677.7 MB/s
readseq : 4.173 micros/op 239635 ops/sec; 2343.8 MB/s
readseq : 4.150 micros/op 240963 ops/sec; 2356.8 MB/s
```
Test Plan: COMPILE_WITH_ASAN=1 make check -j64
Reviewers: yhchiang, andrewkr, sdong
Reviewed By: sdong
Subscribers: lovro, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D57075
2016-07-20 16:49:03 +00:00
|
|
|
// If pinned_value_ is set then the result of merge operator is one of
|
|
|
|
// the merge operands and we should return it.
|
|
|
|
return pinned_value_.data() ? pinned_value_ : saved_value_;
|
2016-05-03 04:46:30 +00:00
|
|
|
} else if (direction_ == kReverse) {
|
|
|
|
return pinned_value_;
|
|
|
|
} else {
|
|
|
|
return iter_->value();
|
|
|
|
}
|
2011-03-18 22:37:00 +00:00
|
|
|
}
|
2015-02-26 19:28:41 +00:00
|
|
|
virtual Status status() const override {
|
2011-03-18 22:37:00 +00:00
|
|
|
if (status_.ok()) {
|
|
|
|
return iter_->status();
|
|
|
|
} else {
|
|
|
|
return status_;
|
|
|
|
}
|
|
|
|
}
|
2017-10-03 16:08:07 +00:00
|
|
|
bool IsBlob() const {
|
|
|
|
assert(valid_ && (allow_blob_ || !is_blob_));
|
|
|
|
return is_blob_;
|
|
|
|
}
|
2016-02-27 01:13:39 +00:00
|
|
|
|
|
|
|
virtual Status GetProperty(std::string prop_name,
|
|
|
|
std::string* prop) override {
|
|
|
|
if (prop == nullptr) {
|
|
|
|
return Status::InvalidArgument("prop is nullptr");
|
|
|
|
}
|
2016-03-03 21:18:56 +00:00
|
|
|
if (prop_name == "rocksdb.iterator.super-version-number") {
|
2016-03-01 02:38:03 +00:00
|
|
|
// First try to pass the value returned from inner iterator.
|
2017-07-24 17:28:17 +00:00
|
|
|
return iter_->GetProperty(prop_name, prop);
|
2016-03-01 02:38:03 +00:00
|
|
|
} else if (prop_name == "rocksdb.iterator.is-key-pinned") {
|
2016-02-27 01:13:39 +00:00
|
|
|
if (valid_) {
|
2016-04-26 19:41:07 +00:00
|
|
|
*prop = (pin_thru_lifetime_ && saved_key_.IsKeyPinned()) ? "1" : "0";
|
2016-02-27 01:13:39 +00:00
|
|
|
} else {
|
|
|
|
*prop = "Iterator is not valid.";
|
|
|
|
}
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
return Status::InvalidArgument("Undentified property.");
|
2015-12-16 20:08:30 +00:00
|
|
|
}
|
2011-03-18 22:37:00 +00:00
|
|
|
|
2015-02-26 19:28:41 +00:00
|
|
|
virtual void Next() override;
|
|
|
|
virtual void Prev() override;
|
|
|
|
virtual void Seek(const Slice& target) override;
|
2016-09-28 01:20:57 +00:00
|
|
|
virtual void SeekForPrev(const Slice& target) override;
|
2015-02-26 19:28:41 +00:00
|
|
|
virtual void SeekToFirst() override;
|
|
|
|
virtual void SeekToLast() override;
|
2017-07-24 17:28:17 +00:00
|
|
|
Env* env() { return env_; }
|
|
|
|
void set_sequence(uint64_t s) { sequence_ = s; }
|
|
|
|
void set_valid(bool v) { valid_ = v; }
|
2011-03-18 22:37:00 +00:00
|
|
|
|
2011-03-25 20:27:43 +00:00
|
|
|
private:
|
2016-10-11 20:54:26 +00:00
|
|
|
void ReverseToForward();
|
2015-08-04 23:50:40 +00:00
|
|
|
void ReverseToBackward();
|
2014-07-15 23:10:18 +00:00
|
|
|
void PrevInternal();
|
|
|
|
void FindParseableKey(ParsedInternalKey* ikey, Direction direction);
|
|
|
|
bool FindValueForCurrentKey();
|
|
|
|
bool FindValueForCurrentKeyUsingSeek();
|
|
|
|
void FindPrevUserKey();
|
|
|
|
void FindNextUserKey();
|
2016-04-28 23:48:03 +00:00
|
|
|
inline void FindNextUserEntry(bool skipping, bool prefix_check);
|
|
|
|
void FindNextUserEntryInternal(bool skipping, bool prefix_check);
|
2011-03-25 20:27:43 +00:00
|
|
|
bool ParseKey(ParsedInternalKey* key);
|
2013-03-21 22:59:47 +00:00
|
|
|
void MergeValuesNewToOld();
|
2017-03-30 18:57:05 +00:00
|
|
|
bool TooManyInternalKeysSkipped(bool increment = true);
|
2017-10-10 00:05:34 +00:00
|
|
|
bool IsVisible(SequenceNumber sequence);
|
2011-03-25 20:27:43 +00:00
|
|
|
|
2016-05-03 04:46:30 +00:00
|
|
|
// Temporarily pin the blocks that we encounter until ReleaseTempPinnedData()
|
|
|
|
// is called
|
|
|
|
void TempPinData() {
|
|
|
|
if (!pin_thru_lifetime_) {
|
|
|
|
pinned_iters_mgr_.StartPinning();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Release blocks pinned by TempPinData()
|
|
|
|
void ReleaseTempPinnedData() {
|
2016-08-11 18:54:17 +00:00
|
|
|
if (!pin_thru_lifetime_ && pinned_iters_mgr_.PinningEnabled()) {
|
|
|
|
pinned_iters_mgr_.ReleasePinnedData();
|
2016-05-03 04:46:30 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2011-03-25 20:27:43 +00:00
|
|
|
inline void ClearSavedValue() {
|
|
|
|
if (saved_value_.capacity() > 1048576) {
|
|
|
|
std::string empty;
|
|
|
|
swap(empty, saved_value_);
|
|
|
|
} else {
|
|
|
|
saved_value_.clear();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-03-30 18:57:05 +00:00
|
|
|
inline void ResetInternalKeysSkippedCounter() {
|
2017-11-21 05:12:55 +00:00
|
|
|
local_stats_.skip_count_ += num_internal_keys_skipped_;
|
|
|
|
if (valid_) {
|
|
|
|
local_stats_.skip_count_--;
|
|
|
|
}
|
2017-03-30 18:57:05 +00:00
|
|
|
num_internal_keys_skipped_ = 0;
|
|
|
|
}
|
|
|
|
|
2014-09-04 17:48:24 +00:00
|
|
|
const SliceTransform* prefix_extractor_;
|
In DB::NewIterator(), try to allocate the whole iterator tree in an arena
Summary:
In this patch, try to allocate the whole iterator tree starting from DBIter from an arena
1. ArenaWrappedDBIter is created when serves as the entry point of an iterator tree, with an arena in it.
2. Add an option to create iterator from arena for following iterators: DBIter, MergingIterator, MemtableIterator, all mem table's iterators, all table reader's iterators and two level iterator.
3. MergeIteratorBuilder is created to incrementally build the tree of internal iterators. It is passed to mem table list and version set and add iterators to it.
Limitations:
(1) Only DB::NewIterator() without tailing uses the arena. Other cases, including readonly DB and compactions are still from malloc
(2) Two level iterator itself is allocated in arena, but not iterators inside it.
Test Plan: make all check
Reviewers: ljin, haobo
Reviewed By: haobo
Subscribers: leveldb, dhruba, yhchiang, igor
Differential Revision: https://reviews.facebook.net/D18513
2014-06-02 23:38:00 +00:00
|
|
|
bool arena_mode_;
|
2011-03-18 22:37:00 +00:00
|
|
|
Env* const env_;
|
2013-12-03 19:17:58 +00:00
|
|
|
Logger* logger_;
|
2011-03-18 22:37:00 +00:00
|
|
|
const Comparator* const user_comparator_;
|
2016-06-13 23:17:26 +00:00
|
|
|
const MergeOperator* const merge_operator_;
|
2015-10-12 22:06:38 +00:00
|
|
|
InternalIterator* iter_;
|
2017-07-24 17:28:17 +00:00
|
|
|
SequenceNumber sequence_;
|
2011-03-25 20:27:43 +00:00
|
|
|
|
2011-03-18 22:37:00 +00:00
|
|
|
Status status_;
|
2014-07-15 23:10:18 +00:00
|
|
|
IterKey saved_key_;
|
2017-09-15 00:43:41 +00:00
|
|
|
// Reusable internal key data structure. This is only used inside one function
|
|
|
|
// and should not be used across functions. Reusing this object can reduce
|
|
|
|
// overhead of calling construction of the function if creating it each time.
|
|
|
|
ParsedInternalKey ikey_;
|
2014-07-15 23:10:18 +00:00
|
|
|
std::string saved_value_;
|
2016-05-03 04:46:30 +00:00
|
|
|
Slice pinned_value_;
|
2011-03-25 20:27:43 +00:00
|
|
|
Direction direction_;
|
2011-03-18 22:37:00 +00:00
|
|
|
bool valid_;
|
2013-03-21 22:59:47 +00:00
|
|
|
bool current_entry_is_merged_;
|
2016-10-11 20:54:26 +00:00
|
|
|
// for prefix seek mode to support prev()
|
2013-11-22 22:14:05 +00:00
|
|
|
Statistics* statistics_;
|
2013-07-28 18:53:08 +00:00
|
|
|
uint64_t max_skip_;
|
2017-03-30 18:57:05 +00:00
|
|
|
uint64_t max_skippable_internal_keys_;
|
|
|
|
uint64_t num_internal_keys_skipped_;
|
2017-10-27 00:14:04 +00:00
|
|
|
const Slice* iterate_lower_bound_;
|
2014-09-04 17:48:24 +00:00
|
|
|
const Slice* iterate_upper_bound_;
|
2016-04-28 23:48:03 +00:00
|
|
|
IterKey prefix_start_buf_;
|
|
|
|
Slice prefix_start_key_;
|
|
|
|
const bool prefix_same_as_start_;
|
2016-04-26 19:41:07 +00:00
|
|
|
// Means that we will pin all data blocks we read as long the Iterator
|
|
|
|
// is not deleted, will be true if ReadOptions::pin_data is true
|
|
|
|
const bool pin_thru_lifetime_;
|
2016-10-11 20:54:26 +00:00
|
|
|
const bool total_order_seek_;
|
2016-01-07 15:59:14 +00:00
|
|
|
// List of operands for merge operator.
|
Eliminate std::deque initialization while iterating over merge operands
Summary:
This patch is similar to D52563, When we iterate over a DB with merge operands we keep creating std::queue to store the operands, optimize this by reusing merge_operands_ data member
Before the patch
```
./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq" --db="/dev/shm/bench_merge_memcpy_on_the_fly/" --merge_operator="put" --merge_keys=10000 --num=10000
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 3.757 micros/op 266141 ops/sec; 29.4 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.413 micros/op 2423538 ops/sec; 268.1 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.451 micros/op 2219071 ops/sec; 245.5 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.420 micros/op 2382039 ops/sec; 263.5 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.408 micros/op 2452017 ops/sec; 271.3 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 3.947 micros/op 253376 ops/sec; 28.0 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.441 micros/op 2266473 ops/sec; 250.7 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.471 micros/op 2122033 ops/sec; 234.8 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.440 micros/op 2271407 ops/sec; 251.3 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.429 micros/op 2331471 ops/sec; 257.9 MB/s
```
with the patch
```
./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq" --db="/dev/shm/bench_merge_memcpy_on_the_fly/" --merge_operator="put" --merge_keys=10000 --num=10000
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 4.080 micros/op 245092 ops/sec; 27.1 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.308 micros/op 3241843 ops/sec; 358.6 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.312 micros/op 3200408 ops/sec; 354.0 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.332 micros/op 3013962 ops/sec; 333.4 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.300 micros/op 3328017 ops/sec; 368.2 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 3.973 micros/op 251705 ops/sec; 27.8 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.320 micros/op 3123752 ops/sec; 345.6 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.335 micros/op 2986641 ops/sec; 330.4 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.339 micros/op 2950047 ops/sec; 326.4 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.319 micros/op 3131565 ops/sec; 346.4 MB/s
```
Test Plan: make check -j64
Reviewers: yhchiang, andrewkr, sdong
Reviewed By: sdong
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D56031
2016-04-01 22:48:55 +00:00
|
|
|
MergeContext merge_context_;
|
2016-11-04 18:53:38 +00:00
|
|
|
RangeDelAggregator range_del_agg_;
|
Aggregate hot Iterator counters in LocalStatistics (DBIter::Next perf regression)
Summary:
This patch bump the counters in the frequent code path DBIter::Next() / DBIter::Prev() in a local data members and send them to Statistics when the iterator is destroyed
A better solution will be to have thread_local implementation for Statistics
New performance
```
readseq : 0.035 micros/op 28597881 ops/sec; 3163.7 MB/s
1,851,568,819 stalled-cycles-frontend # 31.29% frontend cycles idle [49.86%]
884,929,823 stalled-cycles-backend # 14.95% backend cycles idle [50.21%]
readreverse : 0.071 micros/op 14077393 ops/sec; 1557.3 MB/s
3,239,575,993 stalled-cycles-frontend # 27.36% frontend cycles idle [49.96%]
1,558,253,983 stalled-cycles-backend # 13.16% backend cycles idle [50.14%]
```
Existing performance
```
readreverse : 0.174 micros/op 5732342 ops/sec; 634.1 MB/s
20,570,209,389 stalled-cycles-frontend # 70.71% frontend cycles idle [50.01%]
18,422,816,837 stalled-cycles-backend # 63.33% backend cycles idle [50.04%]
readseq : 0.119 micros/op 8400537 ops/sec; 929.3 MB/s
15,634,225,844 stalled-cycles-frontend # 79.07% frontend cycles idle [49.96%]
14,227,427,453 stalled-cycles-backend # 71.95% backend cycles idle [50.09%]
```
Test Plan: unit tests
Reviewers: yhchiang, sdong, igor
Reviewed By: sdong
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D55107
2016-03-12 03:01:12 +00:00
|
|
|
LocalStatistics local_stats_;
|
2016-04-26 19:41:07 +00:00
|
|
|
PinnedIteratorsManager pinned_iters_mgr_;
|
2017-10-10 00:05:34 +00:00
|
|
|
ReadCallback* read_callback_;
|
2017-10-03 16:08:07 +00:00
|
|
|
bool allow_blob_;
|
|
|
|
bool is_blob_;
|
Added support for differential snapshots
Summary:
The motivation for this PR is to add to RocksDB support for differential (incremental) snapshots, as snapshot of the DB changes between two points in time (one can think of it as diff between to sequence numbers, or the diff D which can be thought of as an SST file or just set of KVs that can be applied to sequence number S1 to get the database to the state at sequence number S2).
This feature would be useful for various distributed storages layers built on top of RocksDB, as it should help reduce resources (time and network bandwidth) needed to recover and rebuilt DB instances as replicas in the context of distributed storages.
From the API standpoint that would like client app requesting iterator between (start seqnum) and current DB state, and reading the "diff".
This is a very draft PR for initial review in the discussion on the approach, i'm going to rework some parts and keep updating the PR.
For now, what's done here according to initial discussions:
Preserving deletes:
- We want to be able to optionally preserve recent deletes for some defined period of time, so that if a delete came in recently and might need to be included in the next incremental snapshot it would't get dropped by a compaction. This is done by adding new param to Options (preserve deletes flag) and new variable to DB Impl where we keep track of the sequence number after which we don't want to drop tombstones, even if they are otherwise eligible for deletion.
- I also added a new API call for clients to be able to advance this cutoff seqnum after which we drop deletes; i assume it's more flexible to let clients control this, since otherwise we'd need to keep some kind of timestamp < -- > seqnum mapping inside the DB, which sounds messy and painful to support. Clients could make use of it by periodically calling GetLatestSequenceNumber(), noting the timestamp, doing some calculation and figuring out by how much we need to advance the cutoff seqnum.
- Compaction codepath in compaction_iterator.cc has been modified to avoid dropping tombstones with seqnum > cutoff seqnum.
Iterator changes:
- couple params added to ReadOptions, to optionally allow client to request internal keys instead of user keys (so that client can get the latest value of a key, be it delete marker or a put), as well as min timestamp and min seqnum.
TableCache changes:
- I modified table_cache code to be able to quickly exclude SST files from iterators heep if creation_time on the file is less then iter_start_ts as passed in ReadOptions. That would help a lot in some DB settings (like reading very recent data only or using FIFO compactions), but not so much for universal compaction with more or less long iterator time span.
What's left:
- Still looking at how to best plug that inside DBIter codepath. So far it seems that FindNextUserKeyInternal only parses values as UserKeys, and iter->key() call generally returns user key. Can we add new API to DBIter as internal_key(), and modify this internal method to optionally set saved_key_ to point to the full internal key? I don't need to store actual seqnum there, but I do need to store type.
Closes https://github.com/facebook/rocksdb/pull/2999
Differential Revision: D6175602
Pulled By: mikhail-antonov
fbshipit-source-id: c779a6696ee2d574d86c69cec866a3ae095aa900
2017-11-02 01:43:29 +00:00
|
|
|
// for diff snapshots we want the lower bound on the seqnum;
|
|
|
|
// if this value > 0 iterator will return internal keys
|
|
|
|
SequenceNumber start_seqnum_;
|
2011-03-18 22:37:00 +00:00
|
|
|
|
|
|
|
// No copying allowed
|
|
|
|
DBIter(const DBIter&);
|
|
|
|
void operator=(const DBIter&);
|
|
|
|
};
|
|
|
|
|
|
|
|
inline bool DBIter::ParseKey(ParsedInternalKey* ikey) {
|
|
|
|
if (!ParseInternalKey(iter_->key(), ikey)) {
|
|
|
|
status_ = Status::Corruption("corrupted internal key in DBIter");
|
2017-03-16 02:22:52 +00:00
|
|
|
ROCKS_LOG_ERROR(logger_, "corrupted internal key in DBIter: %s",
|
|
|
|
iter_->key().ToString(true).c_str());
|
2011-03-18 22:37:00 +00:00
|
|
|
return false;
|
|
|
|
} else {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2011-03-25 20:27:43 +00:00
|
|
|
void DBIter::Next() {
|
|
|
|
assert(valid_);
|
|
|
|
|
Introduce FullMergeV2 (eliminate memcpy from merge operators)
Summary:
This diff update the code to pin the merge operator operands while the merge operation is done, so that we can eliminate the memcpy cost, to do that we need a new public API for FullMerge that replace the std::deque<std::string> with std::vector<Slice>
This diff is stacked on top of D56493 and D56511
In this diff we
- Update FullMergeV2 arguments to be encapsulated in MergeOperationInput and MergeOperationOutput which will make it easier to add new arguments in the future
- Replace std::deque<std::string> with std::vector<Slice> to pass operands
- Replace MergeContext std::deque with std::vector (based on a simple benchmark I ran https://gist.github.com/IslamAbdelRahman/78fc86c9ab9f52b1df791e58943fb187)
- Allow FullMergeV2 output to be an existing operand
```
[Everything in Memtable | 10K operands | 10 KB each | 1 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=10000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 0.607 micros/op 1648235 ops/sec; 16121.2 MB/s
readseq : 0.478 micros/op 2091546 ops/sec; 20457.2 MB/s
readseq : 0.252 micros/op 3972081 ops/sec; 38850.5 MB/s
readseq : 0.237 micros/op 4218328 ops/sec; 41259.0 MB/s
readseq : 0.247 micros/op 4043927 ops/sec; 39553.2 MB/s
[master]
readseq : 3.935 micros/op 254140 ops/sec; 2485.7 MB/s
readseq : 3.722 micros/op 268657 ops/sec; 2627.7 MB/s
readseq : 3.149 micros/op 317605 ops/sec; 3106.5 MB/s
readseq : 3.125 micros/op 320024 ops/sec; 3130.1 MB/s
readseq : 4.075 micros/op 245374 ops/sec; 2400.0 MB/s
```
```
[Everything in Memtable | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=1000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 3.472 micros/op 288018 ops/sec; 2817.1 MB/s
readseq : 2.304 micros/op 434027 ops/sec; 4245.2 MB/s
readseq : 1.163 micros/op 859845 ops/sec; 8410.0 MB/s
readseq : 1.192 micros/op 838926 ops/sec; 8205.4 MB/s
readseq : 1.250 micros/op 800000 ops/sec; 7824.7 MB/s
[master]
readseq : 24.025 micros/op 41623 ops/sec; 407.1 MB/s
readseq : 18.489 micros/op 54086 ops/sec; 529.0 MB/s
readseq : 18.693 micros/op 53495 ops/sec; 523.2 MB/s
readseq : 23.621 micros/op 42335 ops/sec; 414.1 MB/s
readseq : 18.775 micros/op 53262 ops/sec; 521.0 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 1 operand per key]
[FullMergeV2]
$ DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
readseq : 14.741 micros/op 67837 ops/sec; 663.5 MB/s
readseq : 1.029 micros/op 971446 ops/sec; 9501.6 MB/s
readseq : 0.974 micros/op 1026229 ops/sec; 10037.4 MB/s
readseq : 0.965 micros/op 1036080 ops/sec; 10133.8 MB/s
readseq : 0.943 micros/op 1060657 ops/sec; 10374.2 MB/s
[master]
readseq : 16.735 micros/op 59755 ops/sec; 584.5 MB/s
readseq : 3.029 micros/op 330151 ops/sec; 3229.2 MB/s
readseq : 3.136 micros/op 318883 ops/sec; 3119.0 MB/s
readseq : 3.065 micros/op 326245 ops/sec; 3191.0 MB/s
readseq : 3.014 micros/op 331813 ops/sec; 3245.4 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10-operands-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
[FullMergeV2]
readseq : 24.325 micros/op 41109 ops/sec; 402.1 MB/s
readseq : 1.470 micros/op 680272 ops/sec; 6653.7 MB/s
readseq : 1.231 micros/op 812347 ops/sec; 7945.5 MB/s
readseq : 1.091 micros/op 916590 ops/sec; 8965.1 MB/s
readseq : 1.109 micros/op 901713 ops/sec; 8819.6 MB/s
[master]
readseq : 27.257 micros/op 36687 ops/sec; 358.8 MB/s
readseq : 4.443 micros/op 225073 ops/sec; 2201.4 MB/s
readseq : 5.830 micros/op 171526 ops/sec; 1677.7 MB/s
readseq : 4.173 micros/op 239635 ops/sec; 2343.8 MB/s
readseq : 4.150 micros/op 240963 ops/sec; 2356.8 MB/s
```
Test Plan: COMPILE_WITH_ASAN=1 make check -j64
Reviewers: yhchiang, andrewkr, sdong
Reviewed By: sdong
Subscribers: lovro, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D57075
2016-07-20 16:49:03 +00:00
|
|
|
// Release temporarily pinned blocks from last operation
|
|
|
|
ReleaseTempPinnedData();
|
2017-03-30 18:57:05 +00:00
|
|
|
ResetInternalKeysSkippedCounter();
|
2014-07-15 23:10:18 +00:00
|
|
|
if (direction_ == kReverse) {
|
2016-10-11 20:54:26 +00:00
|
|
|
ReverseToForward();
|
2015-10-07 00:48:27 +00:00
|
|
|
} else if (iter_->Valid() && !current_entry_is_merged_) {
|
|
|
|
// If the current value is not a merge, the iter position is the
|
|
|
|
// current key, which is already returned. We can safely issue a
|
|
|
|
// Next() without checking the current key.
|
|
|
|
// If the current key is a merge, very likely iter already points
|
|
|
|
// to the next internal position.
|
|
|
|
iter_->Next();
|
2015-12-01 05:32:59 +00:00
|
|
|
PERF_COUNTER_ADD(internal_key_skipped_count, 1);
|
2011-03-18 22:37:00 +00:00
|
|
|
}
|
2011-03-25 20:27:43 +00:00
|
|
|
|
Aggregate hot Iterator counters in LocalStatistics (DBIter::Next perf regression)
Summary:
This patch bump the counters in the frequent code path DBIter::Next() / DBIter::Prev() in a local data members and send them to Statistics when the iterator is destroyed
A better solution will be to have thread_local implementation for Statistics
New performance
```
readseq : 0.035 micros/op 28597881 ops/sec; 3163.7 MB/s
1,851,568,819 stalled-cycles-frontend # 31.29% frontend cycles idle [49.86%]
884,929,823 stalled-cycles-backend # 14.95% backend cycles idle [50.21%]
readreverse : 0.071 micros/op 14077393 ops/sec; 1557.3 MB/s
3,239,575,993 stalled-cycles-frontend # 27.36% frontend cycles idle [49.96%]
1,558,253,983 stalled-cycles-backend # 13.16% backend cycles idle [50.14%]
```
Existing performance
```
readreverse : 0.174 micros/op 5732342 ops/sec; 634.1 MB/s
20,570,209,389 stalled-cycles-frontend # 70.71% frontend cycles idle [50.01%]
18,422,816,837 stalled-cycles-backend # 63.33% backend cycles idle [50.04%]
readseq : 0.119 micros/op 8400537 ops/sec; 929.3 MB/s
15,634,225,844 stalled-cycles-frontend # 79.07% frontend cycles idle [49.96%]
14,227,427,453 stalled-cycles-backend # 71.95% backend cycles idle [50.09%]
```
Test Plan: unit tests
Reviewers: yhchiang, sdong, igor
Reviewed By: sdong
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D55107
2016-03-12 03:01:12 +00:00
|
|
|
if (statistics_ != nullptr) {
|
|
|
|
local_stats_.next_count_++;
|
|
|
|
}
|
2015-10-07 00:48:27 +00:00
|
|
|
// Now we point to the next internal position, for both of merge and
|
|
|
|
// not merge cases.
|
2013-03-21 22:59:47 +00:00
|
|
|
if (!iter_->Valid()) {
|
|
|
|
valid_ = false;
|
|
|
|
return;
|
|
|
|
}
|
2016-04-28 23:48:03 +00:00
|
|
|
FindNextUserEntry(true /* skipping the current user key */, prefix_same_as_start_);
|
Aggregate hot Iterator counters in LocalStatistics (DBIter::Next perf regression)
Summary:
This patch bump the counters in the frequent code path DBIter::Next() / DBIter::Prev() in a local data members and send them to Statistics when the iterator is destroyed
A better solution will be to have thread_local implementation for Statistics
New performance
```
readseq : 0.035 micros/op 28597881 ops/sec; 3163.7 MB/s
1,851,568,819 stalled-cycles-frontend # 31.29% frontend cycles idle [49.86%]
884,929,823 stalled-cycles-backend # 14.95% backend cycles idle [50.21%]
readreverse : 0.071 micros/op 14077393 ops/sec; 1557.3 MB/s
3,239,575,993 stalled-cycles-frontend # 27.36% frontend cycles idle [49.96%]
1,558,253,983 stalled-cycles-backend # 13.16% backend cycles idle [50.14%]
```
Existing performance
```
readreverse : 0.174 micros/op 5732342 ops/sec; 634.1 MB/s
20,570,209,389 stalled-cycles-frontend # 70.71% frontend cycles idle [50.01%]
18,422,816,837 stalled-cycles-backend # 63.33% backend cycles idle [50.04%]
readseq : 0.119 micros/op 8400537 ops/sec; 929.3 MB/s
15,634,225,844 stalled-cycles-frontend # 79.07% frontend cycles idle [49.96%]
14,227,427,453 stalled-cycles-backend # 71.95% backend cycles idle [50.09%]
```
Test Plan: unit tests
Reviewers: yhchiang, sdong, igor
Reviewed By: sdong
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D55107
2016-03-12 03:01:12 +00:00
|
|
|
if (statistics_ != nullptr && valid_) {
|
|
|
|
local_stats_.next_found_count_++;
|
|
|
|
local_stats_.bytes_read_ += (key().size() + value().size());
|
|
|
|
}
|
2011-03-18 22:37:00 +00:00
|
|
|
}
|
|
|
|
|
2013-03-21 22:59:47 +00:00
|
|
|
// PRE: saved_key_ has the current user key if skipping
|
|
|
|
// POST: saved_key_ should have the next user key if valid_,
|
|
|
|
// if the current entry is a result of merge
|
|
|
|
// current_entry_is_merged_ => true
|
|
|
|
// saved_value_ => the merged value
|
|
|
|
//
|
|
|
|
// NOTE: In between, saved_key_ can point to a user key that has
|
2016-11-28 18:12:28 +00:00
|
|
|
// a delete marker or a sequence number higher than sequence_
|
2017-03-09 01:12:04 +00:00
|
|
|
// saved_key_ MUST have a proper user_key before calling this function
|
2016-04-28 23:48:03 +00:00
|
|
|
//
|
|
|
|
// The prefix_check parameter controls whether we check the iterated
|
|
|
|
// keys against the prefix of the seeked key. Set to false when
|
|
|
|
// performing a seek without a key (e.g. SeekToFirst). Set to
|
|
|
|
// prefix_same_as_start_ for other iterations.
|
|
|
|
inline void DBIter::FindNextUserEntry(bool skipping, bool prefix_check) {
|
2014-08-22 22:28:58 +00:00
|
|
|
PERF_TIMER_GUARD(find_next_user_entry_time);
|
2016-04-28 23:48:03 +00:00
|
|
|
FindNextUserEntryInternal(skipping, prefix_check);
|
2013-11-18 19:32:54 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// Actual implementation of DBIter::FindNextUserEntry()
|
2016-04-28 23:48:03 +00:00
|
|
|
void DBIter::FindNextUserEntryInternal(bool skipping, bool prefix_check) {
|
2011-03-25 20:27:43 +00:00
|
|
|
// Loop until we hit an acceptable entry to yield
|
|
|
|
assert(iter_->Valid());
|
|
|
|
assert(direction_ == kForward);
|
2013-03-21 22:59:47 +00:00
|
|
|
current_entry_is_merged_ = false;
|
2016-11-28 18:12:28 +00:00
|
|
|
|
|
|
|
// How many times in a row we have skipped an entry with user key less than
|
|
|
|
// or equal to saved_key_. We could skip these entries either because
|
|
|
|
// sequence numbers were too high or because skipping = true.
|
|
|
|
// What saved_key_ contains throughout this method:
|
|
|
|
// - if skipping : saved_key_ contains the key that we need to skip,
|
|
|
|
// and we haven't seen any keys greater than that,
|
|
|
|
// - if num_skipped > 0 : saved_key_ contains the key that we have skipped
|
|
|
|
// num_skipped times, and we haven't seen any keys
|
|
|
|
// greater than that,
|
|
|
|
// - none of the above : saved_key_ can contain anything, it doesn't matter.
|
2013-07-28 18:53:08 +00:00
|
|
|
uint64_t num_skipped = 0;
|
2016-11-28 18:12:28 +00:00
|
|
|
|
2017-10-03 16:08:07 +00:00
|
|
|
is_blob_ = false;
|
|
|
|
|
2011-03-25 20:27:43 +00:00
|
|
|
do {
|
2017-09-15 00:43:41 +00:00
|
|
|
if (!ParseKey(&ikey_)) {
|
2016-11-28 18:12:28 +00:00
|
|
|
// Skip corrupted keys.
|
|
|
|
iter_->Next();
|
|
|
|
continue;
|
|
|
|
}
|
2014-09-04 17:48:24 +00:00
|
|
|
|
2016-11-28 18:12:28 +00:00
|
|
|
if (iterate_upper_bound_ != nullptr &&
|
2017-09-15 00:43:41 +00:00
|
|
|
user_comparator_->Compare(ikey_.user_key, *iterate_upper_bound_) >= 0) {
|
2016-11-28 18:12:28 +00:00
|
|
|
break;
|
|
|
|
}
|
2016-04-28 23:48:03 +00:00
|
|
|
|
2016-11-28 18:12:28 +00:00
|
|
|
if (prefix_extractor_ && prefix_check &&
|
2017-09-15 00:43:41 +00:00
|
|
|
prefix_extractor_->Transform(ikey_.user_key)
|
|
|
|
.compare(prefix_start_key_) != 0) {
|
2016-11-28 18:12:28 +00:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2017-03-30 18:57:05 +00:00
|
|
|
if (TooManyInternalKeysSkipped()) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2017-10-10 00:05:34 +00:00
|
|
|
if (IsVisible(ikey_.sequence)) {
|
2017-09-15 00:43:41 +00:00
|
|
|
if (skipping && user_comparator_->Compare(ikey_.user_key,
|
|
|
|
saved_key_.GetUserKey()) <= 0) {
|
2016-11-28 18:12:28 +00:00
|
|
|
num_skipped++; // skip this entry
|
|
|
|
PERF_COUNTER_ADD(internal_key_skipped_count, 1);
|
|
|
|
} else {
|
|
|
|
num_skipped = 0;
|
2017-09-15 00:43:41 +00:00
|
|
|
switch (ikey_.type) {
|
2016-11-28 18:12:28 +00:00
|
|
|
case kTypeDeletion:
|
|
|
|
case kTypeSingleDeletion:
|
|
|
|
// Arrange to skip all upcoming entries for this key since
|
|
|
|
// they are hidden by this deletion.
|
Added support for differential snapshots
Summary:
The motivation for this PR is to add to RocksDB support for differential (incremental) snapshots, as snapshot of the DB changes between two points in time (one can think of it as diff between to sequence numbers, or the diff D which can be thought of as an SST file or just set of KVs that can be applied to sequence number S1 to get the database to the state at sequence number S2).
This feature would be useful for various distributed storages layers built on top of RocksDB, as it should help reduce resources (time and network bandwidth) needed to recover and rebuilt DB instances as replicas in the context of distributed storages.
From the API standpoint that would like client app requesting iterator between (start seqnum) and current DB state, and reading the "diff".
This is a very draft PR for initial review in the discussion on the approach, i'm going to rework some parts and keep updating the PR.
For now, what's done here according to initial discussions:
Preserving deletes:
- We want to be able to optionally preserve recent deletes for some defined period of time, so that if a delete came in recently and might need to be included in the next incremental snapshot it would't get dropped by a compaction. This is done by adding new param to Options (preserve deletes flag) and new variable to DB Impl where we keep track of the sequence number after which we don't want to drop tombstones, even if they are otherwise eligible for deletion.
- I also added a new API call for clients to be able to advance this cutoff seqnum after which we drop deletes; i assume it's more flexible to let clients control this, since otherwise we'd need to keep some kind of timestamp < -- > seqnum mapping inside the DB, which sounds messy and painful to support. Clients could make use of it by periodically calling GetLatestSequenceNumber(), noting the timestamp, doing some calculation and figuring out by how much we need to advance the cutoff seqnum.
- Compaction codepath in compaction_iterator.cc has been modified to avoid dropping tombstones with seqnum > cutoff seqnum.
Iterator changes:
- couple params added to ReadOptions, to optionally allow client to request internal keys instead of user keys (so that client can get the latest value of a key, be it delete marker or a put), as well as min timestamp and min seqnum.
TableCache changes:
- I modified table_cache code to be able to quickly exclude SST files from iterators heep if creation_time on the file is less then iter_start_ts as passed in ReadOptions. That would help a lot in some DB settings (like reading very recent data only or using FIFO compactions), but not so much for universal compaction with more or less long iterator time span.
What's left:
- Still looking at how to best plug that inside DBIter codepath. So far it seems that FindNextUserKeyInternal only parses values as UserKeys, and iter->key() call generally returns user key. Can we add new API to DBIter as internal_key(), and modify this internal method to optionally set saved_key_ to point to the full internal key? I don't need to store actual seqnum there, but I do need to store type.
Closes https://github.com/facebook/rocksdb/pull/2999
Differential Revision: D6175602
Pulled By: mikhail-antonov
fbshipit-source-id: c779a6696ee2d574d86c69cec866a3ae095aa900
2017-11-02 01:43:29 +00:00
|
|
|
// if iterartor specified start_seqnum we
|
|
|
|
// 1) return internal key, including the type
|
|
|
|
// 2) return ikey only if ikey.seqnum >= start_seqnum_
|
|
|
|
// not that if deletion seqnum is < start_seqnum_ we
|
|
|
|
// just skip it like in normal iterator.
|
|
|
|
if (start_seqnum_ > 0 && ikey_.sequence >= start_seqnum_) {
|
|
|
|
saved_key_.SetInternalKey(ikey_);
|
|
|
|
valid_=true;
|
|
|
|
return;
|
|
|
|
} else {
|
|
|
|
saved_key_.SetUserKey(
|
2017-09-15 00:43:41 +00:00
|
|
|
ikey_.user_key,
|
|
|
|
!pin_thru_lifetime_ || !iter_->IsKeyPinned() /* copy */);
|
Added support for differential snapshots
Summary:
The motivation for this PR is to add to RocksDB support for differential (incremental) snapshots, as snapshot of the DB changes between two points in time (one can think of it as diff between to sequence numbers, or the diff D which can be thought of as an SST file or just set of KVs that can be applied to sequence number S1 to get the database to the state at sequence number S2).
This feature would be useful for various distributed storages layers built on top of RocksDB, as it should help reduce resources (time and network bandwidth) needed to recover and rebuilt DB instances as replicas in the context of distributed storages.
From the API standpoint that would like client app requesting iterator between (start seqnum) and current DB state, and reading the "diff".
This is a very draft PR for initial review in the discussion on the approach, i'm going to rework some parts and keep updating the PR.
For now, what's done here according to initial discussions:
Preserving deletes:
- We want to be able to optionally preserve recent deletes for some defined period of time, so that if a delete came in recently and might need to be included in the next incremental snapshot it would't get dropped by a compaction. This is done by adding new param to Options (preserve deletes flag) and new variable to DB Impl where we keep track of the sequence number after which we don't want to drop tombstones, even if they are otherwise eligible for deletion.
- I also added a new API call for clients to be able to advance this cutoff seqnum after which we drop deletes; i assume it's more flexible to let clients control this, since otherwise we'd need to keep some kind of timestamp < -- > seqnum mapping inside the DB, which sounds messy and painful to support. Clients could make use of it by periodically calling GetLatestSequenceNumber(), noting the timestamp, doing some calculation and figuring out by how much we need to advance the cutoff seqnum.
- Compaction codepath in compaction_iterator.cc has been modified to avoid dropping tombstones with seqnum > cutoff seqnum.
Iterator changes:
- couple params added to ReadOptions, to optionally allow client to request internal keys instead of user keys (so that client can get the latest value of a key, be it delete marker or a put), as well as min timestamp and min seqnum.
TableCache changes:
- I modified table_cache code to be able to quickly exclude SST files from iterators heep if creation_time on the file is less then iter_start_ts as passed in ReadOptions. That would help a lot in some DB settings (like reading very recent data only or using FIFO compactions), but not so much for universal compaction with more or less long iterator time span.
What's left:
- Still looking at how to best plug that inside DBIter codepath. So far it seems that FindNextUserKeyInternal only parses values as UserKeys, and iter->key() call generally returns user key. Can we add new API to DBIter as internal_key(), and modify this internal method to optionally set saved_key_ to point to the full internal key? I don't need to store actual seqnum there, but I do need to store type.
Closes https://github.com/facebook/rocksdb/pull/2999
Differential Revision: D6175602
Pulled By: mikhail-antonov
fbshipit-source-id: c779a6696ee2d574d86c69cec866a3ae095aa900
2017-11-02 01:43:29 +00:00
|
|
|
skipping = true;
|
|
|
|
PERF_COUNTER_ADD(internal_delete_skipped_count, 1);
|
|
|
|
}
|
2016-11-28 18:12:28 +00:00
|
|
|
break;
|
|
|
|
case kTypeValue:
|
2017-10-03 16:08:07 +00:00
|
|
|
case kTypeBlobIndex:
|
Added support for differential snapshots
Summary:
The motivation for this PR is to add to RocksDB support for differential (incremental) snapshots, as snapshot of the DB changes between two points in time (one can think of it as diff between to sequence numbers, or the diff D which can be thought of as an SST file or just set of KVs that can be applied to sequence number S1 to get the database to the state at sequence number S2).
This feature would be useful for various distributed storages layers built on top of RocksDB, as it should help reduce resources (time and network bandwidth) needed to recover and rebuilt DB instances as replicas in the context of distributed storages.
From the API standpoint that would like client app requesting iterator between (start seqnum) and current DB state, and reading the "diff".
This is a very draft PR for initial review in the discussion on the approach, i'm going to rework some parts and keep updating the PR.
For now, what's done here according to initial discussions:
Preserving deletes:
- We want to be able to optionally preserve recent deletes for some defined period of time, so that if a delete came in recently and might need to be included in the next incremental snapshot it would't get dropped by a compaction. This is done by adding new param to Options (preserve deletes flag) and new variable to DB Impl where we keep track of the sequence number after which we don't want to drop tombstones, even if they are otherwise eligible for deletion.
- I also added a new API call for clients to be able to advance this cutoff seqnum after which we drop deletes; i assume it's more flexible to let clients control this, since otherwise we'd need to keep some kind of timestamp < -- > seqnum mapping inside the DB, which sounds messy and painful to support. Clients could make use of it by periodically calling GetLatestSequenceNumber(), noting the timestamp, doing some calculation and figuring out by how much we need to advance the cutoff seqnum.
- Compaction codepath in compaction_iterator.cc has been modified to avoid dropping tombstones with seqnum > cutoff seqnum.
Iterator changes:
- couple params added to ReadOptions, to optionally allow client to request internal keys instead of user keys (so that client can get the latest value of a key, be it delete marker or a put), as well as min timestamp and min seqnum.
TableCache changes:
- I modified table_cache code to be able to quickly exclude SST files from iterators heep if creation_time on the file is less then iter_start_ts as passed in ReadOptions. That would help a lot in some DB settings (like reading very recent data only or using FIFO compactions), but not so much for universal compaction with more or less long iterator time span.
What's left:
- Still looking at how to best plug that inside DBIter codepath. So far it seems that FindNextUserKeyInternal only parses values as UserKeys, and iter->key() call generally returns user key. Can we add new API to DBIter as internal_key(), and modify this internal method to optionally set saved_key_ to point to the full internal key? I don't need to store actual seqnum there, but I do need to store type.
Closes https://github.com/facebook/rocksdb/pull/2999
Differential Revision: D6175602
Pulled By: mikhail-antonov
fbshipit-source-id: c779a6696ee2d574d86c69cec866a3ae095aa900
2017-11-02 01:43:29 +00:00
|
|
|
if (start_seqnum_ > 0) {
|
|
|
|
// we are taking incremental snapshot here
|
|
|
|
// incremental snapshots aren't supported on DB with range deletes
|
|
|
|
assert(!(
|
|
|
|
(ikey_.type == kTypeBlobIndex) && (start_seqnum_ > 0)
|
|
|
|
));
|
|
|
|
if (ikey_.sequence >= start_seqnum_) {
|
|
|
|
saved_key_.SetInternalKey(ikey_);
|
2017-10-03 16:08:07 +00:00
|
|
|
valid_ = true;
|
Added support for differential snapshots
Summary:
The motivation for this PR is to add to RocksDB support for differential (incremental) snapshots, as snapshot of the DB changes between two points in time (one can think of it as diff between to sequence numbers, or the diff D which can be thought of as an SST file or just set of KVs that can be applied to sequence number S1 to get the database to the state at sequence number S2).
This feature would be useful for various distributed storages layers built on top of RocksDB, as it should help reduce resources (time and network bandwidth) needed to recover and rebuilt DB instances as replicas in the context of distributed storages.
From the API standpoint that would like client app requesting iterator between (start seqnum) and current DB state, and reading the "diff".
This is a very draft PR for initial review in the discussion on the approach, i'm going to rework some parts and keep updating the PR.
For now, what's done here according to initial discussions:
Preserving deletes:
- We want to be able to optionally preserve recent deletes for some defined period of time, so that if a delete came in recently and might need to be included in the next incremental snapshot it would't get dropped by a compaction. This is done by adding new param to Options (preserve deletes flag) and new variable to DB Impl where we keep track of the sequence number after which we don't want to drop tombstones, even if they are otherwise eligible for deletion.
- I also added a new API call for clients to be able to advance this cutoff seqnum after which we drop deletes; i assume it's more flexible to let clients control this, since otherwise we'd need to keep some kind of timestamp < -- > seqnum mapping inside the DB, which sounds messy and painful to support. Clients could make use of it by periodically calling GetLatestSequenceNumber(), noting the timestamp, doing some calculation and figuring out by how much we need to advance the cutoff seqnum.
- Compaction codepath in compaction_iterator.cc has been modified to avoid dropping tombstones with seqnum > cutoff seqnum.
Iterator changes:
- couple params added to ReadOptions, to optionally allow client to request internal keys instead of user keys (so that client can get the latest value of a key, be it delete marker or a put), as well as min timestamp and min seqnum.
TableCache changes:
- I modified table_cache code to be able to quickly exclude SST files from iterators heep if creation_time on the file is less then iter_start_ts as passed in ReadOptions. That would help a lot in some DB settings (like reading very recent data only or using FIFO compactions), but not so much for universal compaction with more or less long iterator time span.
What's left:
- Still looking at how to best plug that inside DBIter codepath. So far it seems that FindNextUserKeyInternal only parses values as UserKeys, and iter->key() call generally returns user key. Can we add new API to DBIter as internal_key(), and modify this internal method to optionally set saved_key_ to point to the full internal key? I don't need to store actual seqnum there, but I do need to store type.
Closes https://github.com/facebook/rocksdb/pull/2999
Differential Revision: D6175602
Pulled By: mikhail-antonov
fbshipit-source-id: c779a6696ee2d574d86c69cec866a3ae095aa900
2017-11-02 01:43:29 +00:00
|
|
|
return;
|
|
|
|
} else {
|
|
|
|
// this key and all previous versions shouldn't be included,
|
|
|
|
// skipping
|
|
|
|
saved_key_.SetUserKey(ikey_.user_key,
|
|
|
|
!pin_thru_lifetime_ || !iter_->IsKeyPinned() /* copy */);
|
|
|
|
skipping = true;
|
2017-10-03 16:08:07 +00:00
|
|
|
}
|
2016-11-28 18:12:28 +00:00
|
|
|
} else {
|
Added support for differential snapshots
Summary:
The motivation for this PR is to add to RocksDB support for differential (incremental) snapshots, as snapshot of the DB changes between two points in time (one can think of it as diff between to sequence numbers, or the diff D which can be thought of as an SST file or just set of KVs that can be applied to sequence number S1 to get the database to the state at sequence number S2).
This feature would be useful for various distributed storages layers built on top of RocksDB, as it should help reduce resources (time and network bandwidth) needed to recover and rebuilt DB instances as replicas in the context of distributed storages.
From the API standpoint that would like client app requesting iterator between (start seqnum) and current DB state, and reading the "diff".
This is a very draft PR for initial review in the discussion on the approach, i'm going to rework some parts and keep updating the PR.
For now, what's done here according to initial discussions:
Preserving deletes:
- We want to be able to optionally preserve recent deletes for some defined period of time, so that if a delete came in recently and might need to be included in the next incremental snapshot it would't get dropped by a compaction. This is done by adding new param to Options (preserve deletes flag) and new variable to DB Impl where we keep track of the sequence number after which we don't want to drop tombstones, even if they are otherwise eligible for deletion.
- I also added a new API call for clients to be able to advance this cutoff seqnum after which we drop deletes; i assume it's more flexible to let clients control this, since otherwise we'd need to keep some kind of timestamp < -- > seqnum mapping inside the DB, which sounds messy and painful to support. Clients could make use of it by periodically calling GetLatestSequenceNumber(), noting the timestamp, doing some calculation and figuring out by how much we need to advance the cutoff seqnum.
- Compaction codepath in compaction_iterator.cc has been modified to avoid dropping tombstones with seqnum > cutoff seqnum.
Iterator changes:
- couple params added to ReadOptions, to optionally allow client to request internal keys instead of user keys (so that client can get the latest value of a key, be it delete marker or a put), as well as min timestamp and min seqnum.
TableCache changes:
- I modified table_cache code to be able to quickly exclude SST files from iterators heep if creation_time on the file is less then iter_start_ts as passed in ReadOptions. That would help a lot in some DB settings (like reading very recent data only or using FIFO compactions), but not so much for universal compaction with more or less long iterator time span.
What's left:
- Still looking at how to best plug that inside DBIter codepath. So far it seems that FindNextUserKeyInternal only parses values as UserKeys, and iter->key() call generally returns user key. Can we add new API to DBIter as internal_key(), and modify this internal method to optionally set saved_key_ to point to the full internal key? I don't need to store actual seqnum there, but I do need to store type.
Closes https://github.com/facebook/rocksdb/pull/2999
Differential Revision: D6175602
Pulled By: mikhail-antonov
fbshipit-source-id: c779a6696ee2d574d86c69cec866a3ae095aa900
2017-11-02 01:43:29 +00:00
|
|
|
saved_key_.SetUserKey(
|
|
|
|
ikey_.user_key,
|
|
|
|
!pin_thru_lifetime_ || !iter_->IsKeyPinned() /* copy */);
|
|
|
|
if (range_del_agg_.ShouldDelete(
|
|
|
|
ikey_, RangeDelAggregator::RangePositioningMode::
|
|
|
|
kForwardTraversal)) {
|
|
|
|
// Arrange to skip all upcoming entries for this key since
|
|
|
|
// they are hidden by this deletion.
|
|
|
|
skipping = true;
|
|
|
|
num_skipped = 0;
|
|
|
|
PERF_COUNTER_ADD(internal_delete_skipped_count, 1);
|
|
|
|
} else if (ikey_.type == kTypeBlobIndex) {
|
|
|
|
if (!allow_blob_) {
|
|
|
|
ROCKS_LOG_ERROR(logger_, "Encounter unexpected blob index.");
|
|
|
|
status_ = Status::NotSupported(
|
|
|
|
"Encounter unexpected blob index. Please open DB with "
|
|
|
|
"rocksdb::blob_db::BlobDB instead.");
|
|
|
|
valid_ = false;
|
|
|
|
} else {
|
|
|
|
is_blob_ = true;
|
|
|
|
valid_ = true;
|
|
|
|
}
|
|
|
|
return;
|
|
|
|
} else {
|
|
|
|
valid_ = true;
|
|
|
|
return;
|
|
|
|
}
|
2016-11-28 18:12:28 +00:00
|
|
|
}
|
|
|
|
break;
|
|
|
|
case kTypeMerge:
|
2017-04-04 21:17:16 +00:00
|
|
|
saved_key_.SetUserKey(
|
2017-09-15 00:43:41 +00:00
|
|
|
ikey_.user_key,
|
|
|
|
!pin_thru_lifetime_ || !iter_->IsKeyPinned() /* copy */);
|
Maintain position in range deletions map
Summary:
When deletion-collapsing mode is enabled (i.e., for DBIter/CompactionIterator), we maintain position in the tombstone maps across calls to ShouldDelete(). Since iterators often access keys sequentially (or reverse-sequentially), scanning forward/backward from the last position can be faster than binary-searching the map for every key.
- When Next() is invoked on an iterator, we use kForwardTraversal to scan forwards, if needed, until arriving at the range deletion containing the next key.
- Similarly for Prev(), we use kBackwardTraversal to scan backwards in the range deletion map.
- When the iterator seeks, we use kBinarySearch for repositioning
- After tombstones are added or before the first ShouldDelete() invocation, the current position is set to invalid, which forces kBinarySearch to be used.
- Non-iterator users (i.e., Get()) use kFullScan, which has the same behavior as before---scan the whole map for every key passed to ShouldDelete().
Closes https://github.com/facebook/rocksdb/pull/1701
Differential Revision: D4350318
Pulled By: ajkr
fbshipit-source-id: 5129b76
2017-01-05 18:22:46 +00:00
|
|
|
if (range_del_agg_.ShouldDelete(
|
2017-09-15 00:43:41 +00:00
|
|
|
ikey_, RangeDelAggregator::RangePositioningMode::
|
|
|
|
kForwardTraversal)) {
|
2016-11-28 18:12:28 +00:00
|
|
|
// Arrange to skip all upcoming entries for this key since
|
|
|
|
// they are hidden by this deletion.
|
|
|
|
skipping = true;
|
|
|
|
num_skipped = 0;
|
|
|
|
PERF_COUNTER_ADD(internal_delete_skipped_count, 1);
|
|
|
|
} else {
|
|
|
|
// By now, we are sure the current ikey is going to yield a
|
|
|
|
// value
|
|
|
|
current_entry_is_merged_ = true;
|
|
|
|
valid_ = true;
|
|
|
|
MergeValuesNewToOld(); // Go to a different state machine
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
assert(false);
|
|
|
|
break;
|
2013-03-21 22:59:47 +00:00
|
|
|
}
|
2011-03-25 20:27:43 +00:00
|
|
|
}
|
2016-11-28 18:12:28 +00:00
|
|
|
} else {
|
|
|
|
// This key was inserted after our snapshot was taken.
|
|
|
|
PERF_COUNTER_ADD(internal_recent_skipped_count, 1);
|
|
|
|
|
|
|
|
// Here saved_key_ may contain some old key, or the default empty key, or
|
|
|
|
// key assigned by some random other method. We don't care.
|
2017-09-15 00:43:41 +00:00
|
|
|
if (user_comparator_->Compare(ikey_.user_key, saved_key_.GetUserKey()) <=
|
2017-04-04 21:17:16 +00:00
|
|
|
0) {
|
2016-11-28 18:12:28 +00:00
|
|
|
num_skipped++;
|
|
|
|
} else {
|
2017-04-04 21:17:16 +00:00
|
|
|
saved_key_.SetUserKey(
|
2017-09-15 00:43:41 +00:00
|
|
|
ikey_.user_key,
|
2017-04-04 21:17:16 +00:00
|
|
|
!iter_->IsKeyPinned() || !pin_thru_lifetime_ /* copy */);
|
2016-11-28 18:12:28 +00:00
|
|
|
skipping = false;
|
|
|
|
num_skipped = 0;
|
|
|
|
}
|
2011-03-18 22:37:00 +00:00
|
|
|
}
|
2016-11-28 18:12:28 +00:00
|
|
|
|
|
|
|
// If we have sequentially iterated via numerous equal keys, then it's
|
|
|
|
// better to seek so that we can avoid too many key comparisons.
|
|
|
|
if (num_skipped > max_skip_) {
|
2013-07-28 18:53:08 +00:00
|
|
|
num_skipped = 0;
|
|
|
|
std::string last_key;
|
2016-11-28 18:12:28 +00:00
|
|
|
if (skipping) {
|
|
|
|
// We're looking for the next user-key but all we see are the same
|
|
|
|
// user-key with decreasing sequence numbers. Fast forward to
|
|
|
|
// sequence number 0 and type deletion (the smallest type).
|
2017-04-04 21:17:16 +00:00
|
|
|
AppendInternalKey(&last_key, ParsedInternalKey(saved_key_.GetUserKey(),
|
|
|
|
0, kTypeDeletion));
|
2016-11-28 18:12:28 +00:00
|
|
|
// Don't set skipping = false because we may still see more user-keys
|
|
|
|
// equal to saved_key_.
|
|
|
|
} else {
|
|
|
|
// We saw multiple entries with this user key and sequence numbers
|
|
|
|
// higher than sequence_. Fast forward to sequence_.
|
|
|
|
// Note that this only covers a case when a higher key was overwritten
|
|
|
|
// many times since our snapshot was taken, not the case when a lot of
|
|
|
|
// different keys were inserted after our snapshot was taken.
|
|
|
|
AppendInternalKey(&last_key,
|
2017-04-04 21:17:16 +00:00
|
|
|
ParsedInternalKey(saved_key_.GetUserKey(), sequence_,
|
2016-11-28 18:12:28 +00:00
|
|
|
kValueTypeForSeek));
|
|
|
|
}
|
2013-07-28 18:53:08 +00:00
|
|
|
iter_->Seek(last_key);
|
|
|
|
RecordTick(statistics_, NUMBER_OF_RESEEKS_IN_ITERATION);
|
|
|
|
} else {
|
|
|
|
iter_->Next();
|
|
|
|
}
|
2011-03-25 20:27:43 +00:00
|
|
|
} while (iter_->Valid());
|
|
|
|
valid_ = false;
|
2011-03-18 22:37:00 +00:00
|
|
|
}
|
|
|
|
|
2013-03-21 22:59:47 +00:00
|
|
|
// Merge values of the same user key starting from the current iter_ position
|
|
|
|
// Scan from the newer entries to older entries.
|
|
|
|
// PRE: iter_->key() points to the first merge type entry
|
|
|
|
// saved_key_ stores the user key
|
|
|
|
// POST: saved_value_ has the merged value for the user key
|
|
|
|
// iter_ points to the next entry (or invalid)
|
|
|
|
void DBIter::MergeValuesNewToOld() {
|
2016-06-13 23:17:26 +00:00
|
|
|
if (!merge_operator_) {
|
2017-03-16 02:22:52 +00:00
|
|
|
ROCKS_LOG_ERROR(logger_, "Options::merge_operator is null.");
|
2016-06-13 23:17:26 +00:00
|
|
|
status_ = Status::InvalidArgument("merge_operator_ must be set.");
|
2014-12-04 19:11:11 +00:00
|
|
|
valid_ = false;
|
|
|
|
return;
|
2013-08-19 18:42:47 +00:00
|
|
|
}
|
2013-03-21 22:59:47 +00:00
|
|
|
|
Introduce FullMergeV2 (eliminate memcpy from merge operators)
Summary:
This diff update the code to pin the merge operator operands while the merge operation is done, so that we can eliminate the memcpy cost, to do that we need a new public API for FullMerge that replace the std::deque<std::string> with std::vector<Slice>
This diff is stacked on top of D56493 and D56511
In this diff we
- Update FullMergeV2 arguments to be encapsulated in MergeOperationInput and MergeOperationOutput which will make it easier to add new arguments in the future
- Replace std::deque<std::string> with std::vector<Slice> to pass operands
- Replace MergeContext std::deque with std::vector (based on a simple benchmark I ran https://gist.github.com/IslamAbdelRahman/78fc86c9ab9f52b1df791e58943fb187)
- Allow FullMergeV2 output to be an existing operand
```
[Everything in Memtable | 10K operands | 10 KB each | 1 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=10000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 0.607 micros/op 1648235 ops/sec; 16121.2 MB/s
readseq : 0.478 micros/op 2091546 ops/sec; 20457.2 MB/s
readseq : 0.252 micros/op 3972081 ops/sec; 38850.5 MB/s
readseq : 0.237 micros/op 4218328 ops/sec; 41259.0 MB/s
readseq : 0.247 micros/op 4043927 ops/sec; 39553.2 MB/s
[master]
readseq : 3.935 micros/op 254140 ops/sec; 2485.7 MB/s
readseq : 3.722 micros/op 268657 ops/sec; 2627.7 MB/s
readseq : 3.149 micros/op 317605 ops/sec; 3106.5 MB/s
readseq : 3.125 micros/op 320024 ops/sec; 3130.1 MB/s
readseq : 4.075 micros/op 245374 ops/sec; 2400.0 MB/s
```
```
[Everything in Memtable | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=1000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 3.472 micros/op 288018 ops/sec; 2817.1 MB/s
readseq : 2.304 micros/op 434027 ops/sec; 4245.2 MB/s
readseq : 1.163 micros/op 859845 ops/sec; 8410.0 MB/s
readseq : 1.192 micros/op 838926 ops/sec; 8205.4 MB/s
readseq : 1.250 micros/op 800000 ops/sec; 7824.7 MB/s
[master]
readseq : 24.025 micros/op 41623 ops/sec; 407.1 MB/s
readseq : 18.489 micros/op 54086 ops/sec; 529.0 MB/s
readseq : 18.693 micros/op 53495 ops/sec; 523.2 MB/s
readseq : 23.621 micros/op 42335 ops/sec; 414.1 MB/s
readseq : 18.775 micros/op 53262 ops/sec; 521.0 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 1 operand per key]
[FullMergeV2]
$ DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
readseq : 14.741 micros/op 67837 ops/sec; 663.5 MB/s
readseq : 1.029 micros/op 971446 ops/sec; 9501.6 MB/s
readseq : 0.974 micros/op 1026229 ops/sec; 10037.4 MB/s
readseq : 0.965 micros/op 1036080 ops/sec; 10133.8 MB/s
readseq : 0.943 micros/op 1060657 ops/sec; 10374.2 MB/s
[master]
readseq : 16.735 micros/op 59755 ops/sec; 584.5 MB/s
readseq : 3.029 micros/op 330151 ops/sec; 3229.2 MB/s
readseq : 3.136 micros/op 318883 ops/sec; 3119.0 MB/s
readseq : 3.065 micros/op 326245 ops/sec; 3191.0 MB/s
readseq : 3.014 micros/op 331813 ops/sec; 3245.4 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10-operands-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
[FullMergeV2]
readseq : 24.325 micros/op 41109 ops/sec; 402.1 MB/s
readseq : 1.470 micros/op 680272 ops/sec; 6653.7 MB/s
readseq : 1.231 micros/op 812347 ops/sec; 7945.5 MB/s
readseq : 1.091 micros/op 916590 ops/sec; 8965.1 MB/s
readseq : 1.109 micros/op 901713 ops/sec; 8819.6 MB/s
[master]
readseq : 27.257 micros/op 36687 ops/sec; 358.8 MB/s
readseq : 4.443 micros/op 225073 ops/sec; 2201.4 MB/s
readseq : 5.830 micros/op 171526 ops/sec; 1677.7 MB/s
readseq : 4.173 micros/op 239635 ops/sec; 2343.8 MB/s
readseq : 4.150 micros/op 240963 ops/sec; 2356.8 MB/s
```
Test Plan: COMPILE_WITH_ASAN=1 make check -j64
Reviewers: yhchiang, andrewkr, sdong
Reviewed By: sdong
Subscribers: lovro, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D57075
2016-07-20 16:49:03 +00:00
|
|
|
// Temporarily pin the blocks that hold merge operands
|
|
|
|
TempPinData();
|
Eliminate std::deque initialization while iterating over merge operands
Summary:
This patch is similar to D52563, When we iterate over a DB with merge operands we keep creating std::queue to store the operands, optimize this by reusing merge_operands_ data member
Before the patch
```
./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq" --db="/dev/shm/bench_merge_memcpy_on_the_fly/" --merge_operator="put" --merge_keys=10000 --num=10000
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 3.757 micros/op 266141 ops/sec; 29.4 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.413 micros/op 2423538 ops/sec; 268.1 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.451 micros/op 2219071 ops/sec; 245.5 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.420 micros/op 2382039 ops/sec; 263.5 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.408 micros/op 2452017 ops/sec; 271.3 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 3.947 micros/op 253376 ops/sec; 28.0 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.441 micros/op 2266473 ops/sec; 250.7 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.471 micros/op 2122033 ops/sec; 234.8 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.440 micros/op 2271407 ops/sec; 251.3 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.429 micros/op 2331471 ops/sec; 257.9 MB/s
```
with the patch
```
./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq" --db="/dev/shm/bench_merge_memcpy_on_the_fly/" --merge_operator="put" --merge_keys=10000 --num=10000
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 4.080 micros/op 245092 ops/sec; 27.1 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.308 micros/op 3241843 ops/sec; 358.6 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.312 micros/op 3200408 ops/sec; 354.0 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.332 micros/op 3013962 ops/sec; 333.4 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.300 micros/op 3328017 ops/sec; 368.2 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 3.973 micros/op 251705 ops/sec; 27.8 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.320 micros/op 3123752 ops/sec; 345.6 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.335 micros/op 2986641 ops/sec; 330.4 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.339 micros/op 2950047 ops/sec; 326.4 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.319 micros/op 3131565 ops/sec; 346.4 MB/s
```
Test Plan: make check -j64
Reviewers: yhchiang, andrewkr, sdong
Reviewed By: sdong
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D56031
2016-04-01 22:48:55 +00:00
|
|
|
merge_context_.Clear();
|
[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 03:14:32 +00:00
|
|
|
// Start the merge process by pushing the first operand
|
Introduce FullMergeV2 (eliminate memcpy from merge operators)
Summary:
This diff update the code to pin the merge operator operands while the merge operation is done, so that we can eliminate the memcpy cost, to do that we need a new public API for FullMerge that replace the std::deque<std::string> with std::vector<Slice>
This diff is stacked on top of D56493 and D56511
In this diff we
- Update FullMergeV2 arguments to be encapsulated in MergeOperationInput and MergeOperationOutput which will make it easier to add new arguments in the future
- Replace std::deque<std::string> with std::vector<Slice> to pass operands
- Replace MergeContext std::deque with std::vector (based on a simple benchmark I ran https://gist.github.com/IslamAbdelRahman/78fc86c9ab9f52b1df791e58943fb187)
- Allow FullMergeV2 output to be an existing operand
```
[Everything in Memtable | 10K operands | 10 KB each | 1 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=10000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 0.607 micros/op 1648235 ops/sec; 16121.2 MB/s
readseq : 0.478 micros/op 2091546 ops/sec; 20457.2 MB/s
readseq : 0.252 micros/op 3972081 ops/sec; 38850.5 MB/s
readseq : 0.237 micros/op 4218328 ops/sec; 41259.0 MB/s
readseq : 0.247 micros/op 4043927 ops/sec; 39553.2 MB/s
[master]
readseq : 3.935 micros/op 254140 ops/sec; 2485.7 MB/s
readseq : 3.722 micros/op 268657 ops/sec; 2627.7 MB/s
readseq : 3.149 micros/op 317605 ops/sec; 3106.5 MB/s
readseq : 3.125 micros/op 320024 ops/sec; 3130.1 MB/s
readseq : 4.075 micros/op 245374 ops/sec; 2400.0 MB/s
```
```
[Everything in Memtable | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=1000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 3.472 micros/op 288018 ops/sec; 2817.1 MB/s
readseq : 2.304 micros/op 434027 ops/sec; 4245.2 MB/s
readseq : 1.163 micros/op 859845 ops/sec; 8410.0 MB/s
readseq : 1.192 micros/op 838926 ops/sec; 8205.4 MB/s
readseq : 1.250 micros/op 800000 ops/sec; 7824.7 MB/s
[master]
readseq : 24.025 micros/op 41623 ops/sec; 407.1 MB/s
readseq : 18.489 micros/op 54086 ops/sec; 529.0 MB/s
readseq : 18.693 micros/op 53495 ops/sec; 523.2 MB/s
readseq : 23.621 micros/op 42335 ops/sec; 414.1 MB/s
readseq : 18.775 micros/op 53262 ops/sec; 521.0 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 1 operand per key]
[FullMergeV2]
$ DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
readseq : 14.741 micros/op 67837 ops/sec; 663.5 MB/s
readseq : 1.029 micros/op 971446 ops/sec; 9501.6 MB/s
readseq : 0.974 micros/op 1026229 ops/sec; 10037.4 MB/s
readseq : 0.965 micros/op 1036080 ops/sec; 10133.8 MB/s
readseq : 0.943 micros/op 1060657 ops/sec; 10374.2 MB/s
[master]
readseq : 16.735 micros/op 59755 ops/sec; 584.5 MB/s
readseq : 3.029 micros/op 330151 ops/sec; 3229.2 MB/s
readseq : 3.136 micros/op 318883 ops/sec; 3119.0 MB/s
readseq : 3.065 micros/op 326245 ops/sec; 3191.0 MB/s
readseq : 3.014 micros/op 331813 ops/sec; 3245.4 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10-operands-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
[FullMergeV2]
readseq : 24.325 micros/op 41109 ops/sec; 402.1 MB/s
readseq : 1.470 micros/op 680272 ops/sec; 6653.7 MB/s
readseq : 1.231 micros/op 812347 ops/sec; 7945.5 MB/s
readseq : 1.091 micros/op 916590 ops/sec; 8965.1 MB/s
readseq : 1.109 micros/op 901713 ops/sec; 8819.6 MB/s
[master]
readseq : 27.257 micros/op 36687 ops/sec; 358.8 MB/s
readseq : 4.443 micros/op 225073 ops/sec; 2201.4 MB/s
readseq : 5.830 micros/op 171526 ops/sec; 1677.7 MB/s
readseq : 4.173 micros/op 239635 ops/sec; 2343.8 MB/s
readseq : 4.150 micros/op 240963 ops/sec; 2356.8 MB/s
```
Test Plan: COMPILE_WITH_ASAN=1 make check -j64
Reviewers: yhchiang, andrewkr, sdong
Reviewed By: sdong
Subscribers: lovro, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D57075
2016-07-20 16:49:03 +00:00
|
|
|
merge_context_.PushOperand(iter_->value(),
|
|
|
|
iter_->IsValuePinned() /* operand_pinned */);
|
2018-02-03 05:16:26 +00:00
|
|
|
TEST_SYNC_POINT("DBIter::MergeValuesNewToOld:PushedFirstOperand");
|
2013-03-21 22:59:47 +00:00
|
|
|
|
|
|
|
ParsedInternalKey ikey;
|
2016-12-16 19:00:03 +00:00
|
|
|
Status s;
|
2013-03-21 22:59:47 +00:00
|
|
|
for (iter_->Next(); iter_->Valid(); iter_->Next()) {
|
2018-02-03 05:16:26 +00:00
|
|
|
TEST_SYNC_POINT("DBIter::MergeValuesNewToOld:SteppedToNextOperand");
|
2013-03-21 22:59:47 +00:00
|
|
|
if (!ParseKey(&ikey)) {
|
|
|
|
// skip corrupted key
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2017-04-04 21:17:16 +00:00
|
|
|
if (!user_comparator_->Equal(ikey.user_key, saved_key_.GetUserKey())) {
|
2013-03-21 22:59:47 +00:00
|
|
|
// hit the next user key, stop right here
|
|
|
|
break;
|
2016-11-04 18:53:38 +00:00
|
|
|
} else if (kTypeDeletion == ikey.type || kTypeSingleDeletion == ikey.type ||
|
Maintain position in range deletions map
Summary:
When deletion-collapsing mode is enabled (i.e., for DBIter/CompactionIterator), we maintain position in the tombstone maps across calls to ShouldDelete(). Since iterators often access keys sequentially (or reverse-sequentially), scanning forward/backward from the last position can be faster than binary-searching the map for every key.
- When Next() is invoked on an iterator, we use kForwardTraversal to scan forwards, if needed, until arriving at the range deletion containing the next key.
- Similarly for Prev(), we use kBackwardTraversal to scan backwards in the range deletion map.
- When the iterator seeks, we use kBinarySearch for repositioning
- After tombstones are added or before the first ShouldDelete() invocation, the current position is set to invalid, which forces kBinarySearch to be used.
- Non-iterator users (i.e., Get()) use kFullScan, which has the same behavior as before---scan the whole map for every key passed to ShouldDelete().
Closes https://github.com/facebook/rocksdb/pull/1701
Differential Revision: D4350318
Pulled By: ajkr
fbshipit-source-id: 5129b76
2017-01-05 18:22:46 +00:00
|
|
|
range_del_agg_.ShouldDelete(
|
|
|
|
ikey, RangeDelAggregator::RangePositioningMode::
|
|
|
|
kForwardTraversal)) {
|
2013-03-21 22:59:47 +00:00
|
|
|
// hit a delete with the same user key, stop right here
|
|
|
|
// iter_ is positioned after delete
|
|
|
|
iter_->Next();
|
|
|
|
break;
|
Support for SingleDelete()
Summary:
This patch fixes #7460559. It introduces SingleDelete as a new database
operation. This operation can be used to delete keys that were never
overwritten (no put following another put of the same key). If an overwritten
key is single deleted the behavior is undefined. Single deletion of a
non-existent key has no effect but multiple consecutive single deletions are
not allowed (see limitations).
In contrast to the conventional Delete() operation, the deletion entry is
removed along with the value when the two are lined up in a compaction. Note:
The semantics are similar to @igor's prototype that allowed to have this
behavior on the granularity of a column family (
https://reviews.facebook.net/D42093 ). This new patch, however, is more
aggressive when it comes to removing tombstones: It removes the SingleDelete
together with the value whenever there is no snapshot between them while the
older patch only did this when the sequence number of the deletion was older
than the earliest snapshot.
Most of the complex additions are in the Compaction Iterator, all other changes
should be relatively straightforward. The patch also includes basic support for
single deletions in db_stress and db_bench.
Limitations:
- Not compatible with cuckoo hash tables
- Single deletions cannot be used in combination with merges and normal
deletions on the same key (other keys are not affected by this)
- Consecutive single deletions are currently not allowed (and older version of
this patch supported this so it could be resurrected if needed)
Test Plan: make all check
Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor
Reviewed By: igor
Subscribers: maykov, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 18:42:56 +00:00
|
|
|
} else if (kTypeValue == ikey.type) {
|
[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 03:14:32 +00:00
|
|
|
// hit a put, merge the put value with operands and store the
|
|
|
|
// final result in saved_value_. We are done!
|
|
|
|
// ignore corruption if there is any.
|
2014-11-06 19:14:28 +00:00
|
|
|
const Slice val = iter_->value();
|
2016-12-16 19:00:03 +00:00
|
|
|
s = MergeHelper::TimedFullMerge(
|
|
|
|
merge_operator_, ikey.user_key, &val, merge_context_.GetOperands(),
|
2017-05-31 14:27:40 +00:00
|
|
|
&saved_value_, logger_, statistics_, env_, &pinned_value_, true);
|
2016-12-16 19:00:03 +00:00
|
|
|
if (!s.ok()) {
|
2017-12-06 19:54:45 +00:00
|
|
|
valid_ = false;
|
2016-12-16 19:00:03 +00:00
|
|
|
status_ = s;
|
|
|
|
}
|
2013-03-21 22:59:47 +00:00
|
|
|
// iter_ is positioned after put
|
|
|
|
iter_->Next();
|
|
|
|
return;
|
Support for SingleDelete()
Summary:
This patch fixes #7460559. It introduces SingleDelete as a new database
operation. This operation can be used to delete keys that were never
overwritten (no put following another put of the same key). If an overwritten
key is single deleted the behavior is undefined. Single deletion of a
non-existent key has no effect but multiple consecutive single deletions are
not allowed (see limitations).
In contrast to the conventional Delete() operation, the deletion entry is
removed along with the value when the two are lined up in a compaction. Note:
The semantics are similar to @igor's prototype that allowed to have this
behavior on the granularity of a column family (
https://reviews.facebook.net/D42093 ). This new patch, however, is more
aggressive when it comes to removing tombstones: It removes the SingleDelete
together with the value whenever there is no snapshot between them while the
older patch only did this when the sequence number of the deletion was older
than the earliest snapshot.
Most of the complex additions are in the Compaction Iterator, all other changes
should be relatively straightforward. The patch also includes basic support for
single deletions in db_stress and db_bench.
Limitations:
- Not compatible with cuckoo hash tables
- Single deletions cannot be used in combination with merges and normal
deletions on the same key (other keys are not affected by this)
- Consecutive single deletions are currently not allowed (and older version of
this patch supported this so it could be resurrected if needed)
Test Plan: make all check
Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor
Reviewed By: igor
Subscribers: maykov, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 18:42:56 +00:00
|
|
|
} else if (kTypeMerge == ikey.type) {
|
[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 03:14:32 +00:00
|
|
|
// hit a merge, add the value as an operand and run associative merge.
|
|
|
|
// when complete, add result to operands and continue.
|
Introduce FullMergeV2 (eliminate memcpy from merge operators)
Summary:
This diff update the code to pin the merge operator operands while the merge operation is done, so that we can eliminate the memcpy cost, to do that we need a new public API for FullMerge that replace the std::deque<std::string> with std::vector<Slice>
This diff is stacked on top of D56493 and D56511
In this diff we
- Update FullMergeV2 arguments to be encapsulated in MergeOperationInput and MergeOperationOutput which will make it easier to add new arguments in the future
- Replace std::deque<std::string> with std::vector<Slice> to pass operands
- Replace MergeContext std::deque with std::vector (based on a simple benchmark I ran https://gist.github.com/IslamAbdelRahman/78fc86c9ab9f52b1df791e58943fb187)
- Allow FullMergeV2 output to be an existing operand
```
[Everything in Memtable | 10K operands | 10 KB each | 1 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=10000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 0.607 micros/op 1648235 ops/sec; 16121.2 MB/s
readseq : 0.478 micros/op 2091546 ops/sec; 20457.2 MB/s
readseq : 0.252 micros/op 3972081 ops/sec; 38850.5 MB/s
readseq : 0.237 micros/op 4218328 ops/sec; 41259.0 MB/s
readseq : 0.247 micros/op 4043927 ops/sec; 39553.2 MB/s
[master]
readseq : 3.935 micros/op 254140 ops/sec; 2485.7 MB/s
readseq : 3.722 micros/op 268657 ops/sec; 2627.7 MB/s
readseq : 3.149 micros/op 317605 ops/sec; 3106.5 MB/s
readseq : 3.125 micros/op 320024 ops/sec; 3130.1 MB/s
readseq : 4.075 micros/op 245374 ops/sec; 2400.0 MB/s
```
```
[Everything in Memtable | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=1000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 3.472 micros/op 288018 ops/sec; 2817.1 MB/s
readseq : 2.304 micros/op 434027 ops/sec; 4245.2 MB/s
readseq : 1.163 micros/op 859845 ops/sec; 8410.0 MB/s
readseq : 1.192 micros/op 838926 ops/sec; 8205.4 MB/s
readseq : 1.250 micros/op 800000 ops/sec; 7824.7 MB/s
[master]
readseq : 24.025 micros/op 41623 ops/sec; 407.1 MB/s
readseq : 18.489 micros/op 54086 ops/sec; 529.0 MB/s
readseq : 18.693 micros/op 53495 ops/sec; 523.2 MB/s
readseq : 23.621 micros/op 42335 ops/sec; 414.1 MB/s
readseq : 18.775 micros/op 53262 ops/sec; 521.0 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 1 operand per key]
[FullMergeV2]
$ DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
readseq : 14.741 micros/op 67837 ops/sec; 663.5 MB/s
readseq : 1.029 micros/op 971446 ops/sec; 9501.6 MB/s
readseq : 0.974 micros/op 1026229 ops/sec; 10037.4 MB/s
readseq : 0.965 micros/op 1036080 ops/sec; 10133.8 MB/s
readseq : 0.943 micros/op 1060657 ops/sec; 10374.2 MB/s
[master]
readseq : 16.735 micros/op 59755 ops/sec; 584.5 MB/s
readseq : 3.029 micros/op 330151 ops/sec; 3229.2 MB/s
readseq : 3.136 micros/op 318883 ops/sec; 3119.0 MB/s
readseq : 3.065 micros/op 326245 ops/sec; 3191.0 MB/s
readseq : 3.014 micros/op 331813 ops/sec; 3245.4 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10-operands-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
[FullMergeV2]
readseq : 24.325 micros/op 41109 ops/sec; 402.1 MB/s
readseq : 1.470 micros/op 680272 ops/sec; 6653.7 MB/s
readseq : 1.231 micros/op 812347 ops/sec; 7945.5 MB/s
readseq : 1.091 micros/op 916590 ops/sec; 8965.1 MB/s
readseq : 1.109 micros/op 901713 ops/sec; 8819.6 MB/s
[master]
readseq : 27.257 micros/op 36687 ops/sec; 358.8 MB/s
readseq : 4.443 micros/op 225073 ops/sec; 2201.4 MB/s
readseq : 5.830 micros/op 171526 ops/sec; 1677.7 MB/s
readseq : 4.173 micros/op 239635 ops/sec; 2343.8 MB/s
readseq : 4.150 micros/op 240963 ops/sec; 2356.8 MB/s
```
Test Plan: COMPILE_WITH_ASAN=1 make check -j64
Reviewers: yhchiang, andrewkr, sdong
Reviewed By: sdong
Subscribers: lovro, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D57075
2016-07-20 16:49:03 +00:00
|
|
|
merge_context_.PushOperand(iter_->value(),
|
|
|
|
iter_->IsValuePinned() /* operand_pinned */);
|
2016-11-28 18:12:28 +00:00
|
|
|
PERF_COUNTER_ADD(internal_merge_count, 1);
|
2017-10-03 16:08:07 +00:00
|
|
|
} else if (kTypeBlobIndex == ikey.type) {
|
|
|
|
if (!allow_blob_) {
|
|
|
|
ROCKS_LOG_ERROR(logger_, "Encounter unexpected blob index.");
|
|
|
|
status_ = Status::NotSupported(
|
|
|
|
"Encounter unexpected blob index. Please open DB with "
|
|
|
|
"rocksdb::blob_db::BlobDB instead.");
|
|
|
|
} else {
|
|
|
|
status_ =
|
|
|
|
Status::NotSupported("Blob DB does not support merge operator.");
|
|
|
|
}
|
|
|
|
valid_ = false;
|
|
|
|
return;
|
Support for SingleDelete()
Summary:
This patch fixes #7460559. It introduces SingleDelete as a new database
operation. This operation can be used to delete keys that were never
overwritten (no put following another put of the same key). If an overwritten
key is single deleted the behavior is undefined. Single deletion of a
non-existent key has no effect but multiple consecutive single deletions are
not allowed (see limitations).
In contrast to the conventional Delete() operation, the deletion entry is
removed along with the value when the two are lined up in a compaction. Note:
The semantics are similar to @igor's prototype that allowed to have this
behavior on the granularity of a column family (
https://reviews.facebook.net/D42093 ). This new patch, however, is more
aggressive when it comes to removing tombstones: It removes the SingleDelete
together with the value whenever there is no snapshot between them while the
older patch only did this when the sequence number of the deletion was older
than the earliest snapshot.
Most of the complex additions are in the Compaction Iterator, all other changes
should be relatively straightforward. The patch also includes basic support for
single deletions in db_stress and db_bench.
Limitations:
- Not compatible with cuckoo hash tables
- Single deletions cannot be used in combination with merges and normal
deletions on the same key (other keys are not affected by this)
- Consecutive single deletions are currently not allowed (and older version of
this patch supported this so it could be resurrected if needed)
Test Plan: make all check
Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor
Reviewed By: igor
Subscribers: maykov, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 18:42:56 +00:00
|
|
|
} else {
|
|
|
|
assert(false);
|
2013-03-21 22:59:47 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-06-13 23:17:26 +00:00
|
|
|
// we either exhausted all internal keys under this user key, or hit
|
|
|
|
// a deletion marker.
|
|
|
|
// feed null as the existing value to the merge operator, such that
|
|
|
|
// client can differentiate this scenario and do things accordingly.
|
2017-04-04 21:17:16 +00:00
|
|
|
s = MergeHelper::TimedFullMerge(merge_operator_, saved_key_.GetUserKey(),
|
|
|
|
nullptr, merge_context_.GetOperands(),
|
|
|
|
&saved_value_, logger_, statistics_, env_,
|
2017-05-31 14:27:40 +00:00
|
|
|
&pinned_value_, true);
|
2016-12-16 19:00:03 +00:00
|
|
|
if (!s.ok()) {
|
2017-12-06 19:54:45 +00:00
|
|
|
valid_ = false;
|
2016-12-16 19:00:03 +00:00
|
|
|
status_ = s;
|
|
|
|
}
|
2013-03-21 22:59:47 +00:00
|
|
|
}
|
|
|
|
|
2011-03-25 20:27:43 +00:00
|
|
|
void DBIter::Prev() {
|
|
|
|
assert(valid_);
|
Introduce FullMergeV2 (eliminate memcpy from merge operators)
Summary:
This diff update the code to pin the merge operator operands while the merge operation is done, so that we can eliminate the memcpy cost, to do that we need a new public API for FullMerge that replace the std::deque<std::string> with std::vector<Slice>
This diff is stacked on top of D56493 and D56511
In this diff we
- Update FullMergeV2 arguments to be encapsulated in MergeOperationInput and MergeOperationOutput which will make it easier to add new arguments in the future
- Replace std::deque<std::string> with std::vector<Slice> to pass operands
- Replace MergeContext std::deque with std::vector (based on a simple benchmark I ran https://gist.github.com/IslamAbdelRahman/78fc86c9ab9f52b1df791e58943fb187)
- Allow FullMergeV2 output to be an existing operand
```
[Everything in Memtable | 10K operands | 10 KB each | 1 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=10000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 0.607 micros/op 1648235 ops/sec; 16121.2 MB/s
readseq : 0.478 micros/op 2091546 ops/sec; 20457.2 MB/s
readseq : 0.252 micros/op 3972081 ops/sec; 38850.5 MB/s
readseq : 0.237 micros/op 4218328 ops/sec; 41259.0 MB/s
readseq : 0.247 micros/op 4043927 ops/sec; 39553.2 MB/s
[master]
readseq : 3.935 micros/op 254140 ops/sec; 2485.7 MB/s
readseq : 3.722 micros/op 268657 ops/sec; 2627.7 MB/s
readseq : 3.149 micros/op 317605 ops/sec; 3106.5 MB/s
readseq : 3.125 micros/op 320024 ops/sec; 3130.1 MB/s
readseq : 4.075 micros/op 245374 ops/sec; 2400.0 MB/s
```
```
[Everything in Memtable | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=1000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 3.472 micros/op 288018 ops/sec; 2817.1 MB/s
readseq : 2.304 micros/op 434027 ops/sec; 4245.2 MB/s
readseq : 1.163 micros/op 859845 ops/sec; 8410.0 MB/s
readseq : 1.192 micros/op 838926 ops/sec; 8205.4 MB/s
readseq : 1.250 micros/op 800000 ops/sec; 7824.7 MB/s
[master]
readseq : 24.025 micros/op 41623 ops/sec; 407.1 MB/s
readseq : 18.489 micros/op 54086 ops/sec; 529.0 MB/s
readseq : 18.693 micros/op 53495 ops/sec; 523.2 MB/s
readseq : 23.621 micros/op 42335 ops/sec; 414.1 MB/s
readseq : 18.775 micros/op 53262 ops/sec; 521.0 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 1 operand per key]
[FullMergeV2]
$ DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
readseq : 14.741 micros/op 67837 ops/sec; 663.5 MB/s
readseq : 1.029 micros/op 971446 ops/sec; 9501.6 MB/s
readseq : 0.974 micros/op 1026229 ops/sec; 10037.4 MB/s
readseq : 0.965 micros/op 1036080 ops/sec; 10133.8 MB/s
readseq : 0.943 micros/op 1060657 ops/sec; 10374.2 MB/s
[master]
readseq : 16.735 micros/op 59755 ops/sec; 584.5 MB/s
readseq : 3.029 micros/op 330151 ops/sec; 3229.2 MB/s
readseq : 3.136 micros/op 318883 ops/sec; 3119.0 MB/s
readseq : 3.065 micros/op 326245 ops/sec; 3191.0 MB/s
readseq : 3.014 micros/op 331813 ops/sec; 3245.4 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10-operands-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
[FullMergeV2]
readseq : 24.325 micros/op 41109 ops/sec; 402.1 MB/s
readseq : 1.470 micros/op 680272 ops/sec; 6653.7 MB/s
readseq : 1.231 micros/op 812347 ops/sec; 7945.5 MB/s
readseq : 1.091 micros/op 916590 ops/sec; 8965.1 MB/s
readseq : 1.109 micros/op 901713 ops/sec; 8819.6 MB/s
[master]
readseq : 27.257 micros/op 36687 ops/sec; 358.8 MB/s
readseq : 4.443 micros/op 225073 ops/sec; 2201.4 MB/s
readseq : 5.830 micros/op 171526 ops/sec; 1677.7 MB/s
readseq : 4.173 micros/op 239635 ops/sec; 2343.8 MB/s
readseq : 4.150 micros/op 240963 ops/sec; 2356.8 MB/s
```
Test Plan: COMPILE_WITH_ASAN=1 make check -j64
Reviewers: yhchiang, andrewkr, sdong
Reviewed By: sdong
Subscribers: lovro, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D57075
2016-07-20 16:49:03 +00:00
|
|
|
ReleaseTempPinnedData();
|
2017-03-30 18:57:05 +00:00
|
|
|
ResetInternalKeysSkippedCounter();
|
2014-07-15 23:10:18 +00:00
|
|
|
if (direction_ == kForward) {
|
2015-08-04 23:50:40 +00:00
|
|
|
ReverseToBackward();
|
2014-07-15 23:10:18 +00:00
|
|
|
}
|
|
|
|
PrevInternal();
|
2015-09-11 18:37:44 +00:00
|
|
|
if (statistics_ != nullptr) {
|
Aggregate hot Iterator counters in LocalStatistics (DBIter::Next perf regression)
Summary:
This patch bump the counters in the frequent code path DBIter::Next() / DBIter::Prev() in a local data members and send them to Statistics when the iterator is destroyed
A better solution will be to have thread_local implementation for Statistics
New performance
```
readseq : 0.035 micros/op 28597881 ops/sec; 3163.7 MB/s
1,851,568,819 stalled-cycles-frontend # 31.29% frontend cycles idle [49.86%]
884,929,823 stalled-cycles-backend # 14.95% backend cycles idle [50.21%]
readreverse : 0.071 micros/op 14077393 ops/sec; 1557.3 MB/s
3,239,575,993 stalled-cycles-frontend # 27.36% frontend cycles idle [49.96%]
1,558,253,983 stalled-cycles-backend # 13.16% backend cycles idle [50.14%]
```
Existing performance
```
readreverse : 0.174 micros/op 5732342 ops/sec; 634.1 MB/s
20,570,209,389 stalled-cycles-frontend # 70.71% frontend cycles idle [50.01%]
18,422,816,837 stalled-cycles-backend # 63.33% backend cycles idle [50.04%]
readseq : 0.119 micros/op 8400537 ops/sec; 929.3 MB/s
15,634,225,844 stalled-cycles-frontend # 79.07% frontend cycles idle [49.96%]
14,227,427,453 stalled-cycles-backend # 71.95% backend cycles idle [50.09%]
```
Test Plan: unit tests
Reviewers: yhchiang, sdong, igor
Reviewed By: sdong
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D55107
2016-03-12 03:01:12 +00:00
|
|
|
local_stats_.prev_count_++;
|
2015-09-11 18:37:44 +00:00
|
|
|
if (valid_) {
|
Aggregate hot Iterator counters in LocalStatistics (DBIter::Next perf regression)
Summary:
This patch bump the counters in the frequent code path DBIter::Next() / DBIter::Prev() in a local data members and send them to Statistics when the iterator is destroyed
A better solution will be to have thread_local implementation for Statistics
New performance
```
readseq : 0.035 micros/op 28597881 ops/sec; 3163.7 MB/s
1,851,568,819 stalled-cycles-frontend # 31.29% frontend cycles idle [49.86%]
884,929,823 stalled-cycles-backend # 14.95% backend cycles idle [50.21%]
readreverse : 0.071 micros/op 14077393 ops/sec; 1557.3 MB/s
3,239,575,993 stalled-cycles-frontend # 27.36% frontend cycles idle [49.96%]
1,558,253,983 stalled-cycles-backend # 13.16% backend cycles idle [50.14%]
```
Existing performance
```
readreverse : 0.174 micros/op 5732342 ops/sec; 634.1 MB/s
20,570,209,389 stalled-cycles-frontend # 70.71% frontend cycles idle [50.01%]
18,422,816,837 stalled-cycles-backend # 63.33% backend cycles idle [50.04%]
readseq : 0.119 micros/op 8400537 ops/sec; 929.3 MB/s
15,634,225,844 stalled-cycles-frontend # 79.07% frontend cycles idle [49.96%]
14,227,427,453 stalled-cycles-backend # 71.95% backend cycles idle [50.09%]
```
Test Plan: unit tests
Reviewers: yhchiang, sdong, igor
Reviewed By: sdong
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D55107
2016-03-12 03:01:12 +00:00
|
|
|
local_stats_.prev_found_count_++;
|
|
|
|
local_stats_.bytes_read_ += (key().size() + value().size());
|
2015-09-11 18:37:44 +00:00
|
|
|
}
|
|
|
|
}
|
2014-07-15 23:10:18 +00:00
|
|
|
}
|
2011-03-18 22:37:00 +00:00
|
|
|
|
2016-10-11 20:54:26 +00:00
|
|
|
void DBIter::ReverseToForward() {
|
|
|
|
if (prefix_extractor_ != nullptr && !total_order_seek_) {
|
|
|
|
IterKey last_key;
|
|
|
|
last_key.SetInternalKey(ParsedInternalKey(
|
2017-04-04 21:17:16 +00:00
|
|
|
saved_key_.GetUserKey(), kMaxSequenceNumber, kValueTypeForSeek));
|
|
|
|
iter_->Seek(last_key.GetInternalKey());
|
2016-10-11 20:54:26 +00:00
|
|
|
}
|
|
|
|
FindNextUserKey();
|
|
|
|
direction_ = kForward;
|
|
|
|
if (!iter_->Valid()) {
|
|
|
|
iter_->SeekToFirst();
|
Maintain position in range deletions map
Summary:
When deletion-collapsing mode is enabled (i.e., for DBIter/CompactionIterator), we maintain position in the tombstone maps across calls to ShouldDelete(). Since iterators often access keys sequentially (or reverse-sequentially), scanning forward/backward from the last position can be faster than binary-searching the map for every key.
- When Next() is invoked on an iterator, we use kForwardTraversal to scan forwards, if needed, until arriving at the range deletion containing the next key.
- Similarly for Prev(), we use kBackwardTraversal to scan backwards in the range deletion map.
- When the iterator seeks, we use kBinarySearch for repositioning
- After tombstones are added or before the first ShouldDelete() invocation, the current position is set to invalid, which forces kBinarySearch to be used.
- Non-iterator users (i.e., Get()) use kFullScan, which has the same behavior as before---scan the whole map for every key passed to ShouldDelete().
Closes https://github.com/facebook/rocksdb/pull/1701
Differential Revision: D4350318
Pulled By: ajkr
fbshipit-source-id: 5129b76
2017-01-05 18:22:46 +00:00
|
|
|
range_del_agg_.InvalidateTombstoneMapPositions();
|
2016-10-11 20:54:26 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-08-04 23:50:40 +00:00
|
|
|
void DBIter::ReverseToBackward() {
|
2016-10-11 20:54:26 +00:00
|
|
|
if (prefix_extractor_ != nullptr && !total_order_seek_) {
|
|
|
|
IterKey last_key;
|
2017-04-04 21:17:16 +00:00
|
|
|
last_key.SetInternalKey(ParsedInternalKey(saved_key_.GetUserKey(), 0,
|
|
|
|
kValueTypeForSeekForPrev));
|
|
|
|
iter_->SeekForPrev(last_key.GetInternalKey());
|
2016-10-11 20:54:26 +00:00
|
|
|
}
|
2015-08-04 23:50:40 +00:00
|
|
|
if (current_entry_is_merged_) {
|
|
|
|
// Not placed in the same key. Need to call Prev() until finding the
|
|
|
|
// previous key.
|
|
|
|
if (!iter_->Valid()) {
|
|
|
|
iter_->SeekToLast();
|
Maintain position in range deletions map
Summary:
When deletion-collapsing mode is enabled (i.e., for DBIter/CompactionIterator), we maintain position in the tombstone maps across calls to ShouldDelete(). Since iterators often access keys sequentially (or reverse-sequentially), scanning forward/backward from the last position can be faster than binary-searching the map for every key.
- When Next() is invoked on an iterator, we use kForwardTraversal to scan forwards, if needed, until arriving at the range deletion containing the next key.
- Similarly for Prev(), we use kBackwardTraversal to scan backwards in the range deletion map.
- When the iterator seeks, we use kBinarySearch for repositioning
- After tombstones are added or before the first ShouldDelete() invocation, the current position is set to invalid, which forces kBinarySearch to be used.
- Non-iterator users (i.e., Get()) use kFullScan, which has the same behavior as before---scan the whole map for every key passed to ShouldDelete().
Closes https://github.com/facebook/rocksdb/pull/1701
Differential Revision: D4350318
Pulled By: ajkr
fbshipit-source-id: 5129b76
2017-01-05 18:22:46 +00:00
|
|
|
range_del_agg_.InvalidateTombstoneMapPositions();
|
2015-08-04 23:50:40 +00:00
|
|
|
}
|
|
|
|
ParsedInternalKey ikey;
|
|
|
|
FindParseableKey(&ikey, kReverse);
|
|
|
|
while (iter_->Valid() &&
|
2017-04-04 21:17:16 +00:00
|
|
|
user_comparator_->Compare(ikey.user_key, saved_key_.GetUserKey()) >
|
|
|
|
0) {
|
2017-06-28 22:36:11 +00:00
|
|
|
assert(ikey.sequence != kMaxSequenceNumber);
|
2017-10-10 00:05:34 +00:00
|
|
|
if (!IsVisible(ikey.sequence)) {
|
2016-11-28 18:12:28 +00:00
|
|
|
PERF_COUNTER_ADD(internal_recent_skipped_count, 1);
|
|
|
|
} else {
|
|
|
|
PERF_COUNTER_ADD(internal_key_skipped_count, 1);
|
|
|
|
}
|
2015-08-04 23:50:40 +00:00
|
|
|
iter_->Prev();
|
|
|
|
FindParseableKey(&ikey, kReverse);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#ifndef NDEBUG
|
|
|
|
if (iter_->Valid()) {
|
|
|
|
ParsedInternalKey ikey;
|
|
|
|
assert(ParseKey(&ikey));
|
2017-04-04 21:17:16 +00:00
|
|
|
assert(user_comparator_->Compare(ikey.user_key, saved_key_.GetUserKey()) <=
|
|
|
|
0);
|
2015-08-04 23:50:40 +00:00
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
|
|
|
FindPrevUserKey();
|
|
|
|
direction_ = kReverse;
|
|
|
|
}
|
|
|
|
|
2014-07-15 23:10:18 +00:00
|
|
|
void DBIter::PrevInternal() {
|
|
|
|
if (!iter_->Valid()) {
|
|
|
|
valid_ = false;
|
|
|
|
return;
|
2013-03-21 22:59:47 +00:00
|
|
|
}
|
|
|
|
|
2014-07-15 23:10:18 +00:00
|
|
|
ParsedInternalKey ikey;
|
|
|
|
|
|
|
|
while (iter_->Valid()) {
|
2017-04-04 21:17:16 +00:00
|
|
|
saved_key_.SetUserKey(
|
|
|
|
ExtractUserKey(iter_->key()),
|
|
|
|
!iter_->IsKeyPinned() || !pin_thru_lifetime_ /* copy */);
|
2017-03-30 18:57:05 +00:00
|
|
|
|
2017-10-25 22:08:01 +00:00
|
|
|
if (prefix_extractor_ && prefix_same_as_start_ &&
|
|
|
|
prefix_extractor_->Transform(saved_key_.GetUserKey())
|
|
|
|
.compare(prefix_start_key_) != 0) {
|
|
|
|
// Current key does not have the same prefix as start
|
|
|
|
valid_ = false;
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2017-10-27 00:14:04 +00:00
|
|
|
if (iterate_lower_bound_ != nullptr &&
|
|
|
|
user_comparator_->Compare(saved_key_.GetUserKey(),
|
|
|
|
*iterate_lower_bound_) < 0) {
|
|
|
|
// We've iterated earlier than the user-specified lower bound.
|
|
|
|
valid_ = false;
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2014-07-15 23:10:18 +00:00
|
|
|
if (FindValueForCurrentKey()) {
|
2011-03-18 22:37:00 +00:00
|
|
|
if (!iter_->Valid()) {
|
|
|
|
return;
|
|
|
|
}
|
2014-07-15 23:10:18 +00:00
|
|
|
FindParseableKey(&ikey, kReverse);
|
2017-04-04 21:17:16 +00:00
|
|
|
if (user_comparator_->Equal(ikey.user_key, saved_key_.GetUserKey())) {
|
2014-07-15 23:10:18 +00:00
|
|
|
FindPrevUserKey();
|
2011-03-25 20:27:43 +00:00
|
|
|
}
|
2014-07-15 23:10:18 +00:00
|
|
|
return;
|
2011-03-18 22:37:00 +00:00
|
|
|
}
|
2017-03-30 18:57:05 +00:00
|
|
|
|
|
|
|
if (TooManyInternalKeysSkipped(false)) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2014-07-15 23:10:18 +00:00
|
|
|
if (!iter_->Valid()) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
FindParseableKey(&ikey, kReverse);
|
2017-04-04 21:17:16 +00:00
|
|
|
if (user_comparator_->Equal(ikey.user_key, saved_key_.GetUserKey())) {
|
2014-07-15 23:10:18 +00:00
|
|
|
FindPrevUserKey();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
// We haven't found any key - iterator is not valid
|
2016-09-28 01:20:57 +00:00
|
|
|
// Or the prefix is different than start prefix
|
2016-09-29 20:19:58 +00:00
|
|
|
assert(!iter_->Valid());
|
2014-07-15 23:10:18 +00:00
|
|
|
valid_ = false;
|
2011-03-18 22:37:00 +00:00
|
|
|
}
|
|
|
|
|
2014-07-15 23:10:18 +00:00
|
|
|
// This function checks, if the entry with biggest sequence_number <= sequence_
|
Support for SingleDelete()
Summary:
This patch fixes #7460559. It introduces SingleDelete as a new database
operation. This operation can be used to delete keys that were never
overwritten (no put following another put of the same key). If an overwritten
key is single deleted the behavior is undefined. Single deletion of a
non-existent key has no effect but multiple consecutive single deletions are
not allowed (see limitations).
In contrast to the conventional Delete() operation, the deletion entry is
removed along with the value when the two are lined up in a compaction. Note:
The semantics are similar to @igor's prototype that allowed to have this
behavior on the granularity of a column family (
https://reviews.facebook.net/D42093 ). This new patch, however, is more
aggressive when it comes to removing tombstones: It removes the SingleDelete
together with the value whenever there is no snapshot between them while the
older patch only did this when the sequence number of the deletion was older
than the earliest snapshot.
Most of the complex additions are in the Compaction Iterator, all other changes
should be relatively straightforward. The patch also includes basic support for
single deletions in db_stress and db_bench.
Limitations:
- Not compatible with cuckoo hash tables
- Single deletions cannot be used in combination with merges and normal
deletions on the same key (other keys are not affected by this)
- Consecutive single deletions are currently not allowed (and older version of
this patch supported this so it could be resurrected if needed)
Test Plan: make all check
Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor
Reviewed By: igor
Subscribers: maykov, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 18:42:56 +00:00
|
|
|
// is non kTypeDeletion or kTypeSingleDeletion. If it's not, we save value in
|
|
|
|
// saved_value_
|
2014-07-15 23:10:18 +00:00
|
|
|
bool DBIter::FindValueForCurrentKey() {
|
|
|
|
assert(iter_->Valid());
|
Eliminate std::deque initialization while iterating over merge operands
Summary:
This patch is similar to D52563, When we iterate over a DB with merge operands we keep creating std::queue to store the operands, optimize this by reusing merge_operands_ data member
Before the patch
```
./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq" --db="/dev/shm/bench_merge_memcpy_on_the_fly/" --merge_operator="put" --merge_keys=10000 --num=10000
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 3.757 micros/op 266141 ops/sec; 29.4 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.413 micros/op 2423538 ops/sec; 268.1 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.451 micros/op 2219071 ops/sec; 245.5 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.420 micros/op 2382039 ops/sec; 263.5 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.408 micros/op 2452017 ops/sec; 271.3 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 3.947 micros/op 253376 ops/sec; 28.0 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.441 micros/op 2266473 ops/sec; 250.7 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.471 micros/op 2122033 ops/sec; 234.8 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.440 micros/op 2271407 ops/sec; 251.3 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.429 micros/op 2331471 ops/sec; 257.9 MB/s
```
with the patch
```
./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq" --db="/dev/shm/bench_merge_memcpy_on_the_fly/" --merge_operator="put" --merge_keys=10000 --num=10000
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 4.080 micros/op 245092 ops/sec; 27.1 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.308 micros/op 3241843 ops/sec; 358.6 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.312 micros/op 3200408 ops/sec; 354.0 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.332 micros/op 3013962 ops/sec; 333.4 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.300 micros/op 3328017 ops/sec; 368.2 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 3.973 micros/op 251705 ops/sec; 27.8 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.320 micros/op 3123752 ops/sec; 345.6 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.335 micros/op 2986641 ops/sec; 330.4 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.339 micros/op 2950047 ops/sec; 326.4 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.319 micros/op 3131565 ops/sec; 346.4 MB/s
```
Test Plan: make check -j64
Reviewers: yhchiang, andrewkr, sdong
Reviewed By: sdong
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D56031
2016-04-01 22:48:55 +00:00
|
|
|
merge_context_.Clear();
|
2016-05-03 04:46:30 +00:00
|
|
|
current_entry_is_merged_ = false;
|
Support for SingleDelete()
Summary:
This patch fixes #7460559. It introduces SingleDelete as a new database
operation. This operation can be used to delete keys that were never
overwritten (no put following another put of the same key). If an overwritten
key is single deleted the behavior is undefined. Single deletion of a
non-existent key has no effect but multiple consecutive single deletions are
not allowed (see limitations).
In contrast to the conventional Delete() operation, the deletion entry is
removed along with the value when the two are lined up in a compaction. Note:
The semantics are similar to @igor's prototype that allowed to have this
behavior on the granularity of a column family (
https://reviews.facebook.net/D42093 ). This new patch, however, is more
aggressive when it comes to removing tombstones: It removes the SingleDelete
together with the value whenever there is no snapshot between them while the
older patch only did this when the sequence number of the deletion was older
than the earliest snapshot.
Most of the complex additions are in the Compaction Iterator, all other changes
should be relatively straightforward. The patch also includes basic support for
single deletions in db_stress and db_bench.
Limitations:
- Not compatible with cuckoo hash tables
- Single deletions cannot be used in combination with merges and normal
deletions on the same key (other keys are not affected by this)
- Consecutive single deletions are currently not allowed (and older version of
this patch supported this so it could be resurrected if needed)
Test Plan: make all check
Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor
Reviewed By: igor
Subscribers: maykov, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 18:42:56 +00:00
|
|
|
// last entry before merge (could be kTypeDeletion, kTypeSingleDeletion or
|
|
|
|
// kTypeValue)
|
2014-07-15 23:10:18 +00:00
|
|
|
ValueType last_not_merge_type = kTypeDeletion;
|
|
|
|
ValueType last_key_entry_type = kTypeDeletion;
|
2011-03-18 22:37:00 +00:00
|
|
|
|
2014-07-15 23:10:18 +00:00
|
|
|
ParsedInternalKey ikey;
|
|
|
|
FindParseableKey(&ikey, kReverse);
|
|
|
|
|
Introduce FullMergeV2 (eliminate memcpy from merge operators)
Summary:
This diff update the code to pin the merge operator operands while the merge operation is done, so that we can eliminate the memcpy cost, to do that we need a new public API for FullMerge that replace the std::deque<std::string> with std::vector<Slice>
This diff is stacked on top of D56493 and D56511
In this diff we
- Update FullMergeV2 arguments to be encapsulated in MergeOperationInput and MergeOperationOutput which will make it easier to add new arguments in the future
- Replace std::deque<std::string> with std::vector<Slice> to pass operands
- Replace MergeContext std::deque with std::vector (based on a simple benchmark I ran https://gist.github.com/IslamAbdelRahman/78fc86c9ab9f52b1df791e58943fb187)
- Allow FullMergeV2 output to be an existing operand
```
[Everything in Memtable | 10K operands | 10 KB each | 1 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=10000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 0.607 micros/op 1648235 ops/sec; 16121.2 MB/s
readseq : 0.478 micros/op 2091546 ops/sec; 20457.2 MB/s
readseq : 0.252 micros/op 3972081 ops/sec; 38850.5 MB/s
readseq : 0.237 micros/op 4218328 ops/sec; 41259.0 MB/s
readseq : 0.247 micros/op 4043927 ops/sec; 39553.2 MB/s
[master]
readseq : 3.935 micros/op 254140 ops/sec; 2485.7 MB/s
readseq : 3.722 micros/op 268657 ops/sec; 2627.7 MB/s
readseq : 3.149 micros/op 317605 ops/sec; 3106.5 MB/s
readseq : 3.125 micros/op 320024 ops/sec; 3130.1 MB/s
readseq : 4.075 micros/op 245374 ops/sec; 2400.0 MB/s
```
```
[Everything in Memtable | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=1000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 3.472 micros/op 288018 ops/sec; 2817.1 MB/s
readseq : 2.304 micros/op 434027 ops/sec; 4245.2 MB/s
readseq : 1.163 micros/op 859845 ops/sec; 8410.0 MB/s
readseq : 1.192 micros/op 838926 ops/sec; 8205.4 MB/s
readseq : 1.250 micros/op 800000 ops/sec; 7824.7 MB/s
[master]
readseq : 24.025 micros/op 41623 ops/sec; 407.1 MB/s
readseq : 18.489 micros/op 54086 ops/sec; 529.0 MB/s
readseq : 18.693 micros/op 53495 ops/sec; 523.2 MB/s
readseq : 23.621 micros/op 42335 ops/sec; 414.1 MB/s
readseq : 18.775 micros/op 53262 ops/sec; 521.0 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 1 operand per key]
[FullMergeV2]
$ DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
readseq : 14.741 micros/op 67837 ops/sec; 663.5 MB/s
readseq : 1.029 micros/op 971446 ops/sec; 9501.6 MB/s
readseq : 0.974 micros/op 1026229 ops/sec; 10037.4 MB/s
readseq : 0.965 micros/op 1036080 ops/sec; 10133.8 MB/s
readseq : 0.943 micros/op 1060657 ops/sec; 10374.2 MB/s
[master]
readseq : 16.735 micros/op 59755 ops/sec; 584.5 MB/s
readseq : 3.029 micros/op 330151 ops/sec; 3229.2 MB/s
readseq : 3.136 micros/op 318883 ops/sec; 3119.0 MB/s
readseq : 3.065 micros/op 326245 ops/sec; 3191.0 MB/s
readseq : 3.014 micros/op 331813 ops/sec; 3245.4 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10-operands-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
[FullMergeV2]
readseq : 24.325 micros/op 41109 ops/sec; 402.1 MB/s
readseq : 1.470 micros/op 680272 ops/sec; 6653.7 MB/s
readseq : 1.231 micros/op 812347 ops/sec; 7945.5 MB/s
readseq : 1.091 micros/op 916590 ops/sec; 8965.1 MB/s
readseq : 1.109 micros/op 901713 ops/sec; 8819.6 MB/s
[master]
readseq : 27.257 micros/op 36687 ops/sec; 358.8 MB/s
readseq : 4.443 micros/op 225073 ops/sec; 2201.4 MB/s
readseq : 5.830 micros/op 171526 ops/sec; 1677.7 MB/s
readseq : 4.173 micros/op 239635 ops/sec; 2343.8 MB/s
readseq : 4.150 micros/op 240963 ops/sec; 2356.8 MB/s
```
Test Plan: COMPILE_WITH_ASAN=1 make check -j64
Reviewers: yhchiang, andrewkr, sdong
Reviewed By: sdong
Subscribers: lovro, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D57075
2016-07-20 16:49:03 +00:00
|
|
|
// Temporarily pin blocks that hold (merge operands / the value)
|
|
|
|
ReleaseTempPinnedData();
|
|
|
|
TempPinData();
|
2014-07-15 23:10:18 +00:00
|
|
|
size_t num_skipped = 0;
|
2017-10-10 00:05:34 +00:00
|
|
|
while (iter_->Valid() && IsVisible(ikey.sequence) &&
|
2017-04-04 21:17:16 +00:00
|
|
|
user_comparator_->Equal(ikey.user_key, saved_key_.GetUserKey())) {
|
2017-03-30 18:57:05 +00:00
|
|
|
if (TooManyInternalKeysSkipped()) {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2014-07-15 23:10:18 +00:00
|
|
|
// We iterate too much: let's use Seek() to avoid too much key comparisons
|
|
|
|
if (num_skipped >= max_skip_) {
|
|
|
|
return FindValueForCurrentKeyUsingSeek();
|
|
|
|
}
|
|
|
|
|
|
|
|
last_key_entry_type = ikey.type;
|
|
|
|
switch (last_key_entry_type) {
|
|
|
|
case kTypeValue:
|
2017-10-03 16:08:07 +00:00
|
|
|
case kTypeBlobIndex:
|
Maintain position in range deletions map
Summary:
When deletion-collapsing mode is enabled (i.e., for DBIter/CompactionIterator), we maintain position in the tombstone maps across calls to ShouldDelete(). Since iterators often access keys sequentially (or reverse-sequentially), scanning forward/backward from the last position can be faster than binary-searching the map for every key.
- When Next() is invoked on an iterator, we use kForwardTraversal to scan forwards, if needed, until arriving at the range deletion containing the next key.
- Similarly for Prev(), we use kBackwardTraversal to scan backwards in the range deletion map.
- When the iterator seeks, we use kBinarySearch for repositioning
- After tombstones are added or before the first ShouldDelete() invocation, the current position is set to invalid, which forces kBinarySearch to be used.
- Non-iterator users (i.e., Get()) use kFullScan, which has the same behavior as before---scan the whole map for every key passed to ShouldDelete().
Closes https://github.com/facebook/rocksdb/pull/1701
Differential Revision: D4350318
Pulled By: ajkr
fbshipit-source-id: 5129b76
2017-01-05 18:22:46 +00:00
|
|
|
if (range_del_agg_.ShouldDelete(
|
|
|
|
ikey,
|
|
|
|
RangeDelAggregator::RangePositioningMode::kBackwardTraversal)) {
|
2016-11-04 18:53:38 +00:00
|
|
|
last_key_entry_type = kTypeRangeDeletion;
|
|
|
|
PERF_COUNTER_ADD(internal_delete_skipped_count, 1);
|
|
|
|
} else {
|
|
|
|
assert(iter_->IsValuePinned());
|
|
|
|
pinned_value_ = iter_->value();
|
|
|
|
}
|
Eliminate std::deque initialization while iterating over merge operands
Summary:
This patch is similar to D52563, When we iterate over a DB with merge operands we keep creating std::queue to store the operands, optimize this by reusing merge_operands_ data member
Before the patch
```
./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq" --db="/dev/shm/bench_merge_memcpy_on_the_fly/" --merge_operator="put" --merge_keys=10000 --num=10000
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 3.757 micros/op 266141 ops/sec; 29.4 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.413 micros/op 2423538 ops/sec; 268.1 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.451 micros/op 2219071 ops/sec; 245.5 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.420 micros/op 2382039 ops/sec; 263.5 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.408 micros/op 2452017 ops/sec; 271.3 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 3.947 micros/op 253376 ops/sec; 28.0 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.441 micros/op 2266473 ops/sec; 250.7 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.471 micros/op 2122033 ops/sec; 234.8 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.440 micros/op 2271407 ops/sec; 251.3 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.429 micros/op 2331471 ops/sec; 257.9 MB/s
```
with the patch
```
./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq" --db="/dev/shm/bench_merge_memcpy_on_the_fly/" --merge_operator="put" --merge_keys=10000 --num=10000
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 4.080 micros/op 245092 ops/sec; 27.1 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.308 micros/op 3241843 ops/sec; 358.6 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.312 micros/op 3200408 ops/sec; 354.0 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.332 micros/op 3013962 ops/sec; 333.4 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.300 micros/op 3328017 ops/sec; 368.2 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 3.973 micros/op 251705 ops/sec; 27.8 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.320 micros/op 3123752 ops/sec; 345.6 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.335 micros/op 2986641 ops/sec; 330.4 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.339 micros/op 2950047 ops/sec; 326.4 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.319 micros/op 3131565 ops/sec; 346.4 MB/s
```
Test Plan: make check -j64
Reviewers: yhchiang, andrewkr, sdong
Reviewed By: sdong
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D56031
2016-04-01 22:48:55 +00:00
|
|
|
merge_context_.Clear();
|
2016-11-04 18:53:38 +00:00
|
|
|
last_not_merge_type = last_key_entry_type;
|
2014-07-15 23:10:18 +00:00
|
|
|
break;
|
|
|
|
case kTypeDeletion:
|
Support for SingleDelete()
Summary:
This patch fixes #7460559. It introduces SingleDelete as a new database
operation. This operation can be used to delete keys that were never
overwritten (no put following another put of the same key). If an overwritten
key is single deleted the behavior is undefined. Single deletion of a
non-existent key has no effect but multiple consecutive single deletions are
not allowed (see limitations).
In contrast to the conventional Delete() operation, the deletion entry is
removed along with the value when the two are lined up in a compaction. Note:
The semantics are similar to @igor's prototype that allowed to have this
behavior on the granularity of a column family (
https://reviews.facebook.net/D42093 ). This new patch, however, is more
aggressive when it comes to removing tombstones: It removes the SingleDelete
together with the value whenever there is no snapshot between them while the
older patch only did this when the sequence number of the deletion was older
than the earliest snapshot.
Most of the complex additions are in the Compaction Iterator, all other changes
should be relatively straightforward. The patch also includes basic support for
single deletions in db_stress and db_bench.
Limitations:
- Not compatible with cuckoo hash tables
- Single deletions cannot be used in combination with merges and normal
deletions on the same key (other keys are not affected by this)
- Consecutive single deletions are currently not allowed (and older version of
this patch supported this so it could be resurrected if needed)
Test Plan: make all check
Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor
Reviewed By: igor
Subscribers: maykov, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 18:42:56 +00:00
|
|
|
case kTypeSingleDeletion:
|
Eliminate std::deque initialization while iterating over merge operands
Summary:
This patch is similar to D52563, When we iterate over a DB with merge operands we keep creating std::queue to store the operands, optimize this by reusing merge_operands_ data member
Before the patch
```
./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq" --db="/dev/shm/bench_merge_memcpy_on_the_fly/" --merge_operator="put" --merge_keys=10000 --num=10000
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 3.757 micros/op 266141 ops/sec; 29.4 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.413 micros/op 2423538 ops/sec; 268.1 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.451 micros/op 2219071 ops/sec; 245.5 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.420 micros/op 2382039 ops/sec; 263.5 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.408 micros/op 2452017 ops/sec; 271.3 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 3.947 micros/op 253376 ops/sec; 28.0 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.441 micros/op 2266473 ops/sec; 250.7 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.471 micros/op 2122033 ops/sec; 234.8 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.440 micros/op 2271407 ops/sec; 251.3 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.429 micros/op 2331471 ops/sec; 257.9 MB/s
```
with the patch
```
./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq" --db="/dev/shm/bench_merge_memcpy_on_the_fly/" --merge_operator="put" --merge_keys=10000 --num=10000
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 4.080 micros/op 245092 ops/sec; 27.1 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.308 micros/op 3241843 ops/sec; 358.6 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.312 micros/op 3200408 ops/sec; 354.0 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.332 micros/op 3013962 ops/sec; 333.4 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.300 micros/op 3328017 ops/sec; 368.2 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 3.973 micros/op 251705 ops/sec; 27.8 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.320 micros/op 3123752 ops/sec; 345.6 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.335 micros/op 2986641 ops/sec; 330.4 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.339 micros/op 2950047 ops/sec; 326.4 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.319 micros/op 3131565 ops/sec; 346.4 MB/s
```
Test Plan: make check -j64
Reviewers: yhchiang, andrewkr, sdong
Reviewed By: sdong
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D56031
2016-04-01 22:48:55 +00:00
|
|
|
merge_context_.Clear();
|
Support for SingleDelete()
Summary:
This patch fixes #7460559. It introduces SingleDelete as a new database
operation. This operation can be used to delete keys that were never
overwritten (no put following another put of the same key). If an overwritten
key is single deleted the behavior is undefined. Single deletion of a
non-existent key has no effect but multiple consecutive single deletions are
not allowed (see limitations).
In contrast to the conventional Delete() operation, the deletion entry is
removed along with the value when the two are lined up in a compaction. Note:
The semantics are similar to @igor's prototype that allowed to have this
behavior on the granularity of a column family (
https://reviews.facebook.net/D42093 ). This new patch, however, is more
aggressive when it comes to removing tombstones: It removes the SingleDelete
together with the value whenever there is no snapshot between them while the
older patch only did this when the sequence number of the deletion was older
than the earliest snapshot.
Most of the complex additions are in the Compaction Iterator, all other changes
should be relatively straightforward. The patch also includes basic support for
single deletions in db_stress and db_bench.
Limitations:
- Not compatible with cuckoo hash tables
- Single deletions cannot be used in combination with merges and normal
deletions on the same key (other keys are not affected by this)
- Consecutive single deletions are currently not allowed (and older version of
this patch supported this so it could be resurrected if needed)
Test Plan: make all check
Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor
Reviewed By: igor
Subscribers: maykov, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 18:42:56 +00:00
|
|
|
last_not_merge_type = last_key_entry_type;
|
2014-09-04 17:48:24 +00:00
|
|
|
PERF_COUNTER_ADD(internal_delete_skipped_count, 1);
|
2014-07-15 23:10:18 +00:00
|
|
|
break;
|
|
|
|
case kTypeMerge:
|
Maintain position in range deletions map
Summary:
When deletion-collapsing mode is enabled (i.e., for DBIter/CompactionIterator), we maintain position in the tombstone maps across calls to ShouldDelete(). Since iterators often access keys sequentially (or reverse-sequentially), scanning forward/backward from the last position can be faster than binary-searching the map for every key.
- When Next() is invoked on an iterator, we use kForwardTraversal to scan forwards, if needed, until arriving at the range deletion containing the next key.
- Similarly for Prev(), we use kBackwardTraversal to scan backwards in the range deletion map.
- When the iterator seeks, we use kBinarySearch for repositioning
- After tombstones are added or before the first ShouldDelete() invocation, the current position is set to invalid, which forces kBinarySearch to be used.
- Non-iterator users (i.e., Get()) use kFullScan, which has the same behavior as before---scan the whole map for every key passed to ShouldDelete().
Closes https://github.com/facebook/rocksdb/pull/1701
Differential Revision: D4350318
Pulled By: ajkr
fbshipit-source-id: 5129b76
2017-01-05 18:22:46 +00:00
|
|
|
if (range_del_agg_.ShouldDelete(
|
|
|
|
ikey,
|
|
|
|
RangeDelAggregator::RangePositioningMode::kBackwardTraversal)) {
|
2016-11-04 18:53:38 +00:00
|
|
|
merge_context_.Clear();
|
|
|
|
last_key_entry_type = kTypeRangeDeletion;
|
|
|
|
last_not_merge_type = last_key_entry_type;
|
|
|
|
PERF_COUNTER_ADD(internal_delete_skipped_count, 1);
|
|
|
|
} else {
|
|
|
|
assert(merge_operator_ != nullptr);
|
|
|
|
merge_context_.PushOperandBack(
|
|
|
|
iter_->value(), iter_->IsValuePinned() /* operand_pinned */);
|
2016-11-28 18:12:28 +00:00
|
|
|
PERF_COUNTER_ADD(internal_merge_count, 1);
|
2016-11-04 18:53:38 +00:00
|
|
|
}
|
2014-07-15 23:10:18 +00:00
|
|
|
break;
|
|
|
|
default:
|
|
|
|
assert(false);
|
|
|
|
}
|
|
|
|
|
2014-09-04 17:48:24 +00:00
|
|
|
PERF_COUNTER_ADD(internal_key_skipped_count, 1);
|
2017-04-04 21:17:16 +00:00
|
|
|
assert(user_comparator_->Equal(ikey.user_key, saved_key_.GetUserKey()));
|
2014-07-15 23:10:18 +00:00
|
|
|
iter_->Prev();
|
|
|
|
++num_skipped;
|
|
|
|
FindParseableKey(&ikey, kReverse);
|
|
|
|
}
|
|
|
|
|
2016-12-16 19:00:03 +00:00
|
|
|
Status s;
|
2017-10-03 16:08:07 +00:00
|
|
|
is_blob_ = false;
|
2014-07-15 23:10:18 +00:00
|
|
|
switch (last_key_entry_type) {
|
|
|
|
case kTypeDeletion:
|
Support for SingleDelete()
Summary:
This patch fixes #7460559. It introduces SingleDelete as a new database
operation. This operation can be used to delete keys that were never
overwritten (no put following another put of the same key). If an overwritten
key is single deleted the behavior is undefined. Single deletion of a
non-existent key has no effect but multiple consecutive single deletions are
not allowed (see limitations).
In contrast to the conventional Delete() operation, the deletion entry is
removed along with the value when the two are lined up in a compaction. Note:
The semantics are similar to @igor's prototype that allowed to have this
behavior on the granularity of a column family (
https://reviews.facebook.net/D42093 ). This new patch, however, is more
aggressive when it comes to removing tombstones: It removes the SingleDelete
together with the value whenever there is no snapshot between them while the
older patch only did this when the sequence number of the deletion was older
than the earliest snapshot.
Most of the complex additions are in the Compaction Iterator, all other changes
should be relatively straightforward. The patch also includes basic support for
single deletions in db_stress and db_bench.
Limitations:
- Not compatible with cuckoo hash tables
- Single deletions cannot be used in combination with merges and normal
deletions on the same key (other keys are not affected by this)
- Consecutive single deletions are currently not allowed (and older version of
this patch supported this so it could be resurrected if needed)
Test Plan: make all check
Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor
Reviewed By: igor
Subscribers: maykov, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 18:42:56 +00:00
|
|
|
case kTypeSingleDeletion:
|
2016-11-04 18:53:38 +00:00
|
|
|
case kTypeRangeDeletion:
|
2014-07-15 23:10:18 +00:00
|
|
|
valid_ = false;
|
|
|
|
return false;
|
|
|
|
case kTypeMerge:
|
2016-05-03 04:46:30 +00:00
|
|
|
current_entry_is_merged_ = true;
|
2016-10-18 23:07:10 +00:00
|
|
|
if (last_not_merge_type == kTypeDeletion ||
|
2016-11-04 18:53:38 +00:00
|
|
|
last_not_merge_type == kTypeSingleDeletion ||
|
|
|
|
last_not_merge_type == kTypeRangeDeletion) {
|
2017-04-04 21:17:16 +00:00
|
|
|
s = MergeHelper::TimedFullMerge(
|
|
|
|
merge_operator_, saved_key_.GetUserKey(), nullptr,
|
|
|
|
merge_context_.GetOperands(), &saved_value_, logger_, statistics_,
|
2017-05-31 14:27:40 +00:00
|
|
|
env_, &pinned_value_, true);
|
2017-10-03 16:08:07 +00:00
|
|
|
} else if (last_not_merge_type == kTypeBlobIndex) {
|
|
|
|
if (!allow_blob_) {
|
|
|
|
ROCKS_LOG_ERROR(logger_, "Encounter unexpected blob index.");
|
|
|
|
status_ = Status::NotSupported(
|
|
|
|
"Encounter unexpected blob index. Please open DB with "
|
|
|
|
"rocksdb::blob_db::BlobDB instead.");
|
|
|
|
} else {
|
|
|
|
status_ =
|
|
|
|
Status::NotSupported("Blob DB does not support merge operator.");
|
|
|
|
}
|
|
|
|
valid_ = false;
|
|
|
|
return true;
|
2013-07-28 18:53:08 +00:00
|
|
|
} else {
|
2014-07-15 23:10:18 +00:00
|
|
|
assert(last_not_merge_type == kTypeValue);
|
2016-12-16 19:00:03 +00:00
|
|
|
s = MergeHelper::TimedFullMerge(
|
2017-04-04 21:17:16 +00:00
|
|
|
merge_operator_, saved_key_.GetUserKey(), &pinned_value_,
|
2016-12-16 19:00:03 +00:00
|
|
|
merge_context_.GetOperands(), &saved_value_, logger_, statistics_,
|
2017-05-31 14:27:40 +00:00
|
|
|
env_, &pinned_value_, true);
|
2013-07-28 18:53:08 +00:00
|
|
|
}
|
2014-07-15 23:10:18 +00:00
|
|
|
break;
|
|
|
|
case kTypeValue:
|
|
|
|
// do nothing - we've already has value in saved_value_
|
|
|
|
break;
|
2017-10-03 16:08:07 +00:00
|
|
|
case kTypeBlobIndex:
|
|
|
|
if (!allow_blob_) {
|
|
|
|
ROCKS_LOG_ERROR(logger_, "Encounter unexpected blob index.");
|
|
|
|
status_ = Status::NotSupported(
|
|
|
|
"Encounter unexpected blob index. Please open DB with "
|
|
|
|
"rocksdb::blob_db::BlobDB instead.");
|
|
|
|
valid_ = false;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
is_blob_ = true;
|
|
|
|
break;
|
2014-07-15 23:10:18 +00:00
|
|
|
default:
|
|
|
|
assert(false);
|
|
|
|
break;
|
2011-03-25 20:27:43 +00:00
|
|
|
}
|
2017-12-06 19:54:45 +00:00
|
|
|
if (s.ok()) {
|
|
|
|
valid_ = true;
|
|
|
|
} else {
|
|
|
|
valid_ = false;
|
2016-12-16 19:00:03 +00:00
|
|
|
status_ = s;
|
|
|
|
}
|
2014-07-15 23:10:18 +00:00
|
|
|
return true;
|
|
|
|
}
|
2011-03-18 22:37:00 +00:00
|
|
|
|
2014-07-15 23:10:18 +00:00
|
|
|
// This function is used in FindValueForCurrentKey.
|
|
|
|
// We use Seek() function instead of Prev() to find necessary value
|
|
|
|
bool DBIter::FindValueForCurrentKeyUsingSeek() {
|
Introduce FullMergeV2 (eliminate memcpy from merge operators)
Summary:
This diff update the code to pin the merge operator operands while the merge operation is done, so that we can eliminate the memcpy cost, to do that we need a new public API for FullMerge that replace the std::deque<std::string> with std::vector<Slice>
This diff is stacked on top of D56493 and D56511
In this diff we
- Update FullMergeV2 arguments to be encapsulated in MergeOperationInput and MergeOperationOutput which will make it easier to add new arguments in the future
- Replace std::deque<std::string> with std::vector<Slice> to pass operands
- Replace MergeContext std::deque with std::vector (based on a simple benchmark I ran https://gist.github.com/IslamAbdelRahman/78fc86c9ab9f52b1df791e58943fb187)
- Allow FullMergeV2 output to be an existing operand
```
[Everything in Memtable | 10K operands | 10 KB each | 1 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=10000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 0.607 micros/op 1648235 ops/sec; 16121.2 MB/s
readseq : 0.478 micros/op 2091546 ops/sec; 20457.2 MB/s
readseq : 0.252 micros/op 3972081 ops/sec; 38850.5 MB/s
readseq : 0.237 micros/op 4218328 ops/sec; 41259.0 MB/s
readseq : 0.247 micros/op 4043927 ops/sec; 39553.2 MB/s
[master]
readseq : 3.935 micros/op 254140 ops/sec; 2485.7 MB/s
readseq : 3.722 micros/op 268657 ops/sec; 2627.7 MB/s
readseq : 3.149 micros/op 317605 ops/sec; 3106.5 MB/s
readseq : 3.125 micros/op 320024 ops/sec; 3130.1 MB/s
readseq : 4.075 micros/op 245374 ops/sec; 2400.0 MB/s
```
```
[Everything in Memtable | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=1000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 3.472 micros/op 288018 ops/sec; 2817.1 MB/s
readseq : 2.304 micros/op 434027 ops/sec; 4245.2 MB/s
readseq : 1.163 micros/op 859845 ops/sec; 8410.0 MB/s
readseq : 1.192 micros/op 838926 ops/sec; 8205.4 MB/s
readseq : 1.250 micros/op 800000 ops/sec; 7824.7 MB/s
[master]
readseq : 24.025 micros/op 41623 ops/sec; 407.1 MB/s
readseq : 18.489 micros/op 54086 ops/sec; 529.0 MB/s
readseq : 18.693 micros/op 53495 ops/sec; 523.2 MB/s
readseq : 23.621 micros/op 42335 ops/sec; 414.1 MB/s
readseq : 18.775 micros/op 53262 ops/sec; 521.0 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 1 operand per key]
[FullMergeV2]
$ DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
readseq : 14.741 micros/op 67837 ops/sec; 663.5 MB/s
readseq : 1.029 micros/op 971446 ops/sec; 9501.6 MB/s
readseq : 0.974 micros/op 1026229 ops/sec; 10037.4 MB/s
readseq : 0.965 micros/op 1036080 ops/sec; 10133.8 MB/s
readseq : 0.943 micros/op 1060657 ops/sec; 10374.2 MB/s
[master]
readseq : 16.735 micros/op 59755 ops/sec; 584.5 MB/s
readseq : 3.029 micros/op 330151 ops/sec; 3229.2 MB/s
readseq : 3.136 micros/op 318883 ops/sec; 3119.0 MB/s
readseq : 3.065 micros/op 326245 ops/sec; 3191.0 MB/s
readseq : 3.014 micros/op 331813 ops/sec; 3245.4 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10-operands-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
[FullMergeV2]
readseq : 24.325 micros/op 41109 ops/sec; 402.1 MB/s
readseq : 1.470 micros/op 680272 ops/sec; 6653.7 MB/s
readseq : 1.231 micros/op 812347 ops/sec; 7945.5 MB/s
readseq : 1.091 micros/op 916590 ops/sec; 8965.1 MB/s
readseq : 1.109 micros/op 901713 ops/sec; 8819.6 MB/s
[master]
readseq : 27.257 micros/op 36687 ops/sec; 358.8 MB/s
readseq : 4.443 micros/op 225073 ops/sec; 2201.4 MB/s
readseq : 5.830 micros/op 171526 ops/sec; 1677.7 MB/s
readseq : 4.173 micros/op 239635 ops/sec; 2343.8 MB/s
readseq : 4.150 micros/op 240963 ops/sec; 2356.8 MB/s
```
Test Plan: COMPILE_WITH_ASAN=1 make check -j64
Reviewers: yhchiang, andrewkr, sdong
Reviewed By: sdong
Subscribers: lovro, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D57075
2016-07-20 16:49:03 +00:00
|
|
|
// FindValueForCurrentKey will enable pinning before calling
|
|
|
|
// FindValueForCurrentKeyUsingSeek()
|
|
|
|
assert(pinned_iters_mgr_.PinningEnabled());
|
2014-07-15 23:10:18 +00:00
|
|
|
std::string last_key;
|
2017-04-04 21:17:16 +00:00
|
|
|
AppendInternalKey(&last_key, ParsedInternalKey(saved_key_.GetUserKey(),
|
|
|
|
sequence_, kValueTypeForSeek));
|
2014-07-15 23:10:18 +00:00
|
|
|
iter_->Seek(last_key);
|
|
|
|
RecordTick(statistics_, NUMBER_OF_RESEEKS_IN_ITERATION);
|
|
|
|
|
|
|
|
// assume there is at least one parseable key for this user key
|
|
|
|
ParsedInternalKey ikey;
|
|
|
|
FindParseableKey(&ikey, kForward);
|
2018-01-24 00:47:41 +00:00
|
|
|
assert(iter_->Valid());
|
|
|
|
assert(user_comparator_->Equal(ikey.user_key, saved_key_.GetUserKey()));
|
|
|
|
|
|
|
|
// In case read_callback presents, the value we seek to may not be visible.
|
|
|
|
// Seek for the next value that's visible.
|
|
|
|
while (!IsVisible(ikey.sequence)) {
|
|
|
|
iter_->Next();
|
|
|
|
FindParseableKey(&ikey, kForward);
|
|
|
|
assert(iter_->Valid());
|
|
|
|
assert(user_comparator_->Equal(ikey.user_key, saved_key_.GetUserKey()));
|
|
|
|
}
|
2014-07-15 23:10:18 +00:00
|
|
|
|
2016-11-04 18:53:38 +00:00
|
|
|
if (ikey.type == kTypeDeletion || ikey.type == kTypeSingleDeletion ||
|
Maintain position in range deletions map
Summary:
When deletion-collapsing mode is enabled (i.e., for DBIter/CompactionIterator), we maintain position in the tombstone maps across calls to ShouldDelete(). Since iterators often access keys sequentially (or reverse-sequentially), scanning forward/backward from the last position can be faster than binary-searching the map for every key.
- When Next() is invoked on an iterator, we use kForwardTraversal to scan forwards, if needed, until arriving at the range deletion containing the next key.
- Similarly for Prev(), we use kBackwardTraversal to scan backwards in the range deletion map.
- When the iterator seeks, we use kBinarySearch for repositioning
- After tombstones are added or before the first ShouldDelete() invocation, the current position is set to invalid, which forces kBinarySearch to be used.
- Non-iterator users (i.e., Get()) use kFullScan, which has the same behavior as before---scan the whole map for every key passed to ShouldDelete().
Closes https://github.com/facebook/rocksdb/pull/1701
Differential Revision: D4350318
Pulled By: ajkr
fbshipit-source-id: 5129b76
2017-01-05 18:22:46 +00:00
|
|
|
range_del_agg_.ShouldDelete(
|
|
|
|
ikey, RangeDelAggregator::RangePositioningMode::kBackwardTraversal)) {
|
2011-03-25 20:27:43 +00:00
|
|
|
valid_ = false;
|
2014-07-15 23:10:18 +00:00
|
|
|
return false;
|
|
|
|
}
|
2017-10-03 16:08:07 +00:00
|
|
|
if (ikey.type == kTypeBlobIndex && !allow_blob_) {
|
|
|
|
ROCKS_LOG_ERROR(logger_, "Encounter unexpected blob index.");
|
|
|
|
status_ = Status::NotSupported(
|
|
|
|
"Encounter unexpected blob index. Please open DB with "
|
|
|
|
"rocksdb::blob_db::BlobDB instead.");
|
|
|
|
valid_ = false;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
if (ikey.type == kTypeValue || ikey.type == kTypeBlobIndex) {
|
2016-11-04 18:53:38 +00:00
|
|
|
assert(iter_->IsValuePinned());
|
|
|
|
pinned_value_ = iter_->value();
|
|
|
|
valid_ = true;
|
|
|
|
return true;
|
|
|
|
}
|
2014-07-15 23:10:18 +00:00
|
|
|
|
|
|
|
// kTypeMerge. We need to collect all kTypeMerge values and save them
|
|
|
|
// in operands
|
2016-05-03 04:46:30 +00:00
|
|
|
current_entry_is_merged_ = true;
|
Eliminate std::deque initialization while iterating over merge operands
Summary:
This patch is similar to D52563, When we iterate over a DB with merge operands we keep creating std::queue to store the operands, optimize this by reusing merge_operands_ data member
Before the patch
```
./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq" --db="/dev/shm/bench_merge_memcpy_on_the_fly/" --merge_operator="put" --merge_keys=10000 --num=10000
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 3.757 micros/op 266141 ops/sec; 29.4 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.413 micros/op 2423538 ops/sec; 268.1 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.451 micros/op 2219071 ops/sec; 245.5 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.420 micros/op 2382039 ops/sec; 263.5 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.408 micros/op 2452017 ops/sec; 271.3 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 3.947 micros/op 253376 ops/sec; 28.0 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.441 micros/op 2266473 ops/sec; 250.7 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.471 micros/op 2122033 ops/sec; 234.8 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.440 micros/op 2271407 ops/sec; 251.3 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.429 micros/op 2331471 ops/sec; 257.9 MB/s
```
with the patch
```
./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq" --db="/dev/shm/bench_merge_memcpy_on_the_fly/" --merge_operator="put" --merge_keys=10000 --num=10000
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 4.080 micros/op 245092 ops/sec; 27.1 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.308 micros/op 3241843 ops/sec; 358.6 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.312 micros/op 3200408 ops/sec; 354.0 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.332 micros/op 3013962 ops/sec; 333.4 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.300 micros/op 3328017 ops/sec; 368.2 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
mergerandom : 3.973 micros/op 251705 ops/sec; 27.8 MB/s ( updates:10000)
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.320 micros/op 3123752 ops/sec; 345.6 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.335 micros/op 2986641 ops/sec; 330.4 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.339 micros/op 2950047 ops/sec; 326.4 MB/s
DB path: [/dev/shm/bench_merge_memcpy_on_the_fly/]
readseq : 0.319 micros/op 3131565 ops/sec; 346.4 MB/s
```
Test Plan: make check -j64
Reviewers: yhchiang, andrewkr, sdong
Reviewed By: sdong
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D56031
2016-04-01 22:48:55 +00:00
|
|
|
merge_context_.Clear();
|
Maintain position in range deletions map
Summary:
When deletion-collapsing mode is enabled (i.e., for DBIter/CompactionIterator), we maintain position in the tombstone maps across calls to ShouldDelete(). Since iterators often access keys sequentially (or reverse-sequentially), scanning forward/backward from the last position can be faster than binary-searching the map for every key.
- When Next() is invoked on an iterator, we use kForwardTraversal to scan forwards, if needed, until arriving at the range deletion containing the next key.
- Similarly for Prev(), we use kBackwardTraversal to scan backwards in the range deletion map.
- When the iterator seeks, we use kBinarySearch for repositioning
- After tombstones are added or before the first ShouldDelete() invocation, the current position is set to invalid, which forces kBinarySearch to be used.
- Non-iterator users (i.e., Get()) use kFullScan, which has the same behavior as before---scan the whole map for every key passed to ShouldDelete().
Closes https://github.com/facebook/rocksdb/pull/1701
Differential Revision: D4350318
Pulled By: ajkr
fbshipit-source-id: 5129b76
2017-01-05 18:22:46 +00:00
|
|
|
while (
|
|
|
|
iter_->Valid() &&
|
2017-04-04 21:17:16 +00:00
|
|
|
user_comparator_->Equal(ikey.user_key, saved_key_.GetUserKey()) &&
|
Maintain position in range deletions map
Summary:
When deletion-collapsing mode is enabled (i.e., for DBIter/CompactionIterator), we maintain position in the tombstone maps across calls to ShouldDelete(). Since iterators often access keys sequentially (or reverse-sequentially), scanning forward/backward from the last position can be faster than binary-searching the map for every key.
- When Next() is invoked on an iterator, we use kForwardTraversal to scan forwards, if needed, until arriving at the range deletion containing the next key.
- Similarly for Prev(), we use kBackwardTraversal to scan backwards in the range deletion map.
- When the iterator seeks, we use kBinarySearch for repositioning
- After tombstones are added or before the first ShouldDelete() invocation, the current position is set to invalid, which forces kBinarySearch to be used.
- Non-iterator users (i.e., Get()) use kFullScan, which has the same behavior as before---scan the whole map for every key passed to ShouldDelete().
Closes https://github.com/facebook/rocksdb/pull/1701
Differential Revision: D4350318
Pulled By: ajkr
fbshipit-source-id: 5129b76
2017-01-05 18:22:46 +00:00
|
|
|
ikey.type == kTypeMerge &&
|
|
|
|
!range_del_agg_.ShouldDelete(
|
|
|
|
ikey, RangeDelAggregator::RangePositioningMode::kBackwardTraversal)) {
|
Introduce FullMergeV2 (eliminate memcpy from merge operators)
Summary:
This diff update the code to pin the merge operator operands while the merge operation is done, so that we can eliminate the memcpy cost, to do that we need a new public API for FullMerge that replace the std::deque<std::string> with std::vector<Slice>
This diff is stacked on top of D56493 and D56511
In this diff we
- Update FullMergeV2 arguments to be encapsulated in MergeOperationInput and MergeOperationOutput which will make it easier to add new arguments in the future
- Replace std::deque<std::string> with std::vector<Slice> to pass operands
- Replace MergeContext std::deque with std::vector (based on a simple benchmark I ran https://gist.github.com/IslamAbdelRahman/78fc86c9ab9f52b1df791e58943fb187)
- Allow FullMergeV2 output to be an existing operand
```
[Everything in Memtable | 10K operands | 10 KB each | 1 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=10000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 0.607 micros/op 1648235 ops/sec; 16121.2 MB/s
readseq : 0.478 micros/op 2091546 ops/sec; 20457.2 MB/s
readseq : 0.252 micros/op 3972081 ops/sec; 38850.5 MB/s
readseq : 0.237 micros/op 4218328 ops/sec; 41259.0 MB/s
readseq : 0.247 micros/op 4043927 ops/sec; 39553.2 MB/s
[master]
readseq : 3.935 micros/op 254140 ops/sec; 2485.7 MB/s
readseq : 3.722 micros/op 268657 ops/sec; 2627.7 MB/s
readseq : 3.149 micros/op 317605 ops/sec; 3106.5 MB/s
readseq : 3.125 micros/op 320024 ops/sec; 3130.1 MB/s
readseq : 4.075 micros/op 245374 ops/sec; 2400.0 MB/s
```
```
[Everything in Memtable | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=1000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 3.472 micros/op 288018 ops/sec; 2817.1 MB/s
readseq : 2.304 micros/op 434027 ops/sec; 4245.2 MB/s
readseq : 1.163 micros/op 859845 ops/sec; 8410.0 MB/s
readseq : 1.192 micros/op 838926 ops/sec; 8205.4 MB/s
readseq : 1.250 micros/op 800000 ops/sec; 7824.7 MB/s
[master]
readseq : 24.025 micros/op 41623 ops/sec; 407.1 MB/s
readseq : 18.489 micros/op 54086 ops/sec; 529.0 MB/s
readseq : 18.693 micros/op 53495 ops/sec; 523.2 MB/s
readseq : 23.621 micros/op 42335 ops/sec; 414.1 MB/s
readseq : 18.775 micros/op 53262 ops/sec; 521.0 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 1 operand per key]
[FullMergeV2]
$ DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
readseq : 14.741 micros/op 67837 ops/sec; 663.5 MB/s
readseq : 1.029 micros/op 971446 ops/sec; 9501.6 MB/s
readseq : 0.974 micros/op 1026229 ops/sec; 10037.4 MB/s
readseq : 0.965 micros/op 1036080 ops/sec; 10133.8 MB/s
readseq : 0.943 micros/op 1060657 ops/sec; 10374.2 MB/s
[master]
readseq : 16.735 micros/op 59755 ops/sec; 584.5 MB/s
readseq : 3.029 micros/op 330151 ops/sec; 3229.2 MB/s
readseq : 3.136 micros/op 318883 ops/sec; 3119.0 MB/s
readseq : 3.065 micros/op 326245 ops/sec; 3191.0 MB/s
readseq : 3.014 micros/op 331813 ops/sec; 3245.4 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10-operands-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
[FullMergeV2]
readseq : 24.325 micros/op 41109 ops/sec; 402.1 MB/s
readseq : 1.470 micros/op 680272 ops/sec; 6653.7 MB/s
readseq : 1.231 micros/op 812347 ops/sec; 7945.5 MB/s
readseq : 1.091 micros/op 916590 ops/sec; 8965.1 MB/s
readseq : 1.109 micros/op 901713 ops/sec; 8819.6 MB/s
[master]
readseq : 27.257 micros/op 36687 ops/sec; 358.8 MB/s
readseq : 4.443 micros/op 225073 ops/sec; 2201.4 MB/s
readseq : 5.830 micros/op 171526 ops/sec; 1677.7 MB/s
readseq : 4.173 micros/op 239635 ops/sec; 2343.8 MB/s
readseq : 4.150 micros/op 240963 ops/sec; 2356.8 MB/s
```
Test Plan: COMPILE_WITH_ASAN=1 make check -j64
Reviewers: yhchiang, andrewkr, sdong
Reviewed By: sdong
Subscribers: lovro, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D57075
2016-07-20 16:49:03 +00:00
|
|
|
merge_context_.PushOperand(iter_->value(),
|
|
|
|
iter_->IsValuePinned() /* operand_pinned */);
|
2016-11-28 18:12:28 +00:00
|
|
|
PERF_COUNTER_ADD(internal_merge_count, 1);
|
2014-07-15 23:10:18 +00:00
|
|
|
iter_->Next();
|
|
|
|
FindParseableKey(&ikey, kForward);
|
|
|
|
}
|
|
|
|
|
2016-12-16 19:00:03 +00:00
|
|
|
Status s;
|
2014-07-15 23:10:18 +00:00
|
|
|
if (!iter_->Valid() ||
|
2017-04-04 21:17:16 +00:00
|
|
|
!user_comparator_->Equal(ikey.user_key, saved_key_.GetUserKey()) ||
|
2016-11-04 18:53:38 +00:00
|
|
|
ikey.type == kTypeDeletion || ikey.type == kTypeSingleDeletion ||
|
Maintain position in range deletions map
Summary:
When deletion-collapsing mode is enabled (i.e., for DBIter/CompactionIterator), we maintain position in the tombstone maps across calls to ShouldDelete(). Since iterators often access keys sequentially (or reverse-sequentially), scanning forward/backward from the last position can be faster than binary-searching the map for every key.
- When Next() is invoked on an iterator, we use kForwardTraversal to scan forwards, if needed, until arriving at the range deletion containing the next key.
- Similarly for Prev(), we use kBackwardTraversal to scan backwards in the range deletion map.
- When the iterator seeks, we use kBinarySearch for repositioning
- After tombstones are added or before the first ShouldDelete() invocation, the current position is set to invalid, which forces kBinarySearch to be used.
- Non-iterator users (i.e., Get()) use kFullScan, which has the same behavior as before---scan the whole map for every key passed to ShouldDelete().
Closes https://github.com/facebook/rocksdb/pull/1701
Differential Revision: D4350318
Pulled By: ajkr
fbshipit-source-id: 5129b76
2017-01-05 18:22:46 +00:00
|
|
|
range_del_agg_.ShouldDelete(
|
|
|
|
ikey, RangeDelAggregator::RangePositioningMode::kBackwardTraversal)) {
|
2017-04-04 21:17:16 +00:00
|
|
|
s = MergeHelper::TimedFullMerge(merge_operator_, saved_key_.GetUserKey(),
|
2016-12-16 19:00:03 +00:00
|
|
|
nullptr, merge_context_.GetOperands(),
|
|
|
|
&saved_value_, logger_, statistics_, env_,
|
2017-05-31 14:27:40 +00:00
|
|
|
&pinned_value_, true);
|
2014-07-15 23:10:18 +00:00
|
|
|
// Make iter_ valid and point to saved_key_
|
|
|
|
if (!iter_->Valid() ||
|
2017-04-04 21:17:16 +00:00
|
|
|
!user_comparator_->Equal(ikey.user_key, saved_key_.GetUserKey())) {
|
2014-07-15 23:10:18 +00:00
|
|
|
iter_->Seek(last_key);
|
|
|
|
RecordTick(statistics_, NUMBER_OF_RESEEKS_IN_ITERATION);
|
|
|
|
}
|
2017-12-06 19:54:45 +00:00
|
|
|
if (s.ok()) {
|
|
|
|
valid_ = true;
|
|
|
|
} else {
|
|
|
|
valid_ = false;
|
2016-12-16 19:00:03 +00:00
|
|
|
status_ = s;
|
|
|
|
}
|
2014-07-15 23:10:18 +00:00
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2014-11-06 19:14:28 +00:00
|
|
|
const Slice& val = iter_->value();
|
2017-04-04 21:17:16 +00:00
|
|
|
s = MergeHelper::TimedFullMerge(merge_operator_, saved_key_.GetUserKey(),
|
|
|
|
&val, merge_context_.GetOperands(),
|
|
|
|
&saved_value_, logger_, statistics_, env_,
|
2017-05-31 14:27:40 +00:00
|
|
|
&pinned_value_, true);
|
2017-12-06 19:54:45 +00:00
|
|
|
if (s.ok()) {
|
|
|
|
valid_ = true;
|
|
|
|
} else {
|
|
|
|
valid_ = false;
|
2016-12-16 19:00:03 +00:00
|
|
|
status_ = s;
|
|
|
|
}
|
2014-07-15 23:10:18 +00:00
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Used in Next to change directions
|
|
|
|
// Go to next user key
|
|
|
|
// Don't use Seek(),
|
|
|
|
// because next user key will be very close
|
|
|
|
void DBIter::FindNextUserKey() {
|
|
|
|
if (!iter_->Valid()) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
ParsedInternalKey ikey;
|
|
|
|
FindParseableKey(&ikey, kForward);
|
|
|
|
while (iter_->Valid() &&
|
2017-04-04 21:17:16 +00:00
|
|
|
!user_comparator_->Equal(ikey.user_key, saved_key_.GetUserKey())) {
|
2014-07-15 23:10:18 +00:00
|
|
|
iter_->Next();
|
|
|
|
FindParseableKey(&ikey, kForward);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Go to previous user_key
|
|
|
|
void DBIter::FindPrevUserKey() {
|
|
|
|
if (!iter_->Valid()) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
size_t num_skipped = 0;
|
|
|
|
ParsedInternalKey ikey;
|
|
|
|
FindParseableKey(&ikey, kReverse);
|
DBIter to out extra keys with higher sequence numbers when changing direction from forward to backward
Summary:
When DBIter changes iterating direction from forward to backward, it might see some much larger keys with higher sequence ID. With this commit, these rows will be actively filtered out. It should fix existing disabled tests in db_iter_test.
This may not be a perfect fix, but it introduces least impact on existing codes, in order to be safe.
Test Plan:
Enable existing tests and make sure they pass. Add a new test DBIterWithMergeIterTest.InnerMergeIteratorDataRace8.
Also run all existing tests.
Reviewers: yhchiang, rven, anthony, IslamAbdelRahman, kradhakrishnan, igor
Reviewed By: igor
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D45567
2015-08-25 20:40:52 +00:00
|
|
|
int cmp;
|
2017-04-04 21:17:16 +00:00
|
|
|
while (iter_->Valid() &&
|
|
|
|
((cmp = user_comparator_->Compare(ikey.user_key,
|
|
|
|
saved_key_.GetUserKey())) == 0 ||
|
2017-10-10 00:05:34 +00:00
|
|
|
(cmp > 0 && !IsVisible(ikey.sequence)))) {
|
2017-03-30 18:57:05 +00:00
|
|
|
if (TooManyInternalKeysSkipped()) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
DBIter to out extra keys with higher sequence numbers when changing direction from forward to backward
Summary:
When DBIter changes iterating direction from forward to backward, it might see some much larger keys with higher sequence ID. With this commit, these rows will be actively filtered out. It should fix existing disabled tests in db_iter_test.
This may not be a perfect fix, but it introduces least impact on existing codes, in order to be safe.
Test Plan:
Enable existing tests and make sure they pass. Add a new test DBIterWithMergeIterTest.InnerMergeIteratorDataRace8.
Also run all existing tests.
Reviewers: yhchiang, rven, anthony, IslamAbdelRahman, kradhakrishnan, igor
Reviewed By: igor
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D45567
2015-08-25 20:40:52 +00:00
|
|
|
if (cmp == 0) {
|
|
|
|
if (num_skipped >= max_skip_) {
|
|
|
|
num_skipped = 0;
|
|
|
|
IterKey last_key;
|
|
|
|
last_key.SetInternalKey(ParsedInternalKey(
|
2017-04-04 21:17:16 +00:00
|
|
|
saved_key_.GetUserKey(), kMaxSequenceNumber, kValueTypeForSeek));
|
|
|
|
iter_->Seek(last_key.GetInternalKey());
|
DBIter to out extra keys with higher sequence numbers when changing direction from forward to backward
Summary:
When DBIter changes iterating direction from forward to backward, it might see some much larger keys with higher sequence ID. With this commit, these rows will be actively filtered out. It should fix existing disabled tests in db_iter_test.
This may not be a perfect fix, but it introduces least impact on existing codes, in order to be safe.
Test Plan:
Enable existing tests and make sure they pass. Add a new test DBIterWithMergeIterTest.InnerMergeIteratorDataRace8.
Also run all existing tests.
Reviewers: yhchiang, rven, anthony, IslamAbdelRahman, kradhakrishnan, igor
Reviewed By: igor
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D45567
2015-08-25 20:40:52 +00:00
|
|
|
RecordTick(statistics_, NUMBER_OF_RESEEKS_IN_ITERATION);
|
|
|
|
} else {
|
|
|
|
++num_skipped;
|
|
|
|
}
|
2014-07-15 23:10:18 +00:00
|
|
|
}
|
2017-06-28 22:36:11 +00:00
|
|
|
assert(ikey.sequence != kMaxSequenceNumber);
|
2017-10-10 00:05:34 +00:00
|
|
|
if (!IsVisible(ikey.sequence)) {
|
2016-11-28 18:12:28 +00:00
|
|
|
PERF_COUNTER_ADD(internal_recent_skipped_count, 1);
|
|
|
|
} else {
|
|
|
|
PERF_COUNTER_ADD(internal_key_skipped_count, 1);
|
|
|
|
}
|
2014-07-15 23:10:18 +00:00
|
|
|
iter_->Prev();
|
|
|
|
FindParseableKey(&ikey, kReverse);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-03-30 18:57:05 +00:00
|
|
|
bool DBIter::TooManyInternalKeysSkipped(bool increment) {
|
|
|
|
if ((max_skippable_internal_keys_ > 0) &&
|
|
|
|
(num_internal_keys_skipped_ > max_skippable_internal_keys_)) {
|
|
|
|
valid_ = false;
|
|
|
|
status_ = Status::Incomplete("Too many internal keys skipped.");
|
|
|
|
return true;
|
|
|
|
} else if (increment) {
|
|
|
|
num_internal_keys_skipped_++;
|
|
|
|
}
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2017-10-10 00:05:34 +00:00
|
|
|
bool DBIter::IsVisible(SequenceNumber sequence) {
|
|
|
|
return sequence <= sequence_ &&
|
|
|
|
(read_callback_ == nullptr || read_callback_->IsCommitted(sequence));
|
|
|
|
}
|
|
|
|
|
2014-07-15 23:10:18 +00:00
|
|
|
// Skip all unparseable keys
|
|
|
|
void DBIter::FindParseableKey(ParsedInternalKey* ikey, Direction direction) {
|
|
|
|
while (iter_->Valid() && !ParseKey(ikey)) {
|
|
|
|
if (direction == kReverse) {
|
|
|
|
iter_->Prev();
|
|
|
|
} else {
|
|
|
|
iter_->Next();
|
|
|
|
}
|
2011-03-25 20:27:43 +00:00
|
|
|
}
|
|
|
|
}
|
2011-03-18 22:37:00 +00:00
|
|
|
|
2011-03-25 20:27:43 +00:00
|
|
|
void DBIter::Seek(const Slice& target) {
|
2014-08-13 22:56:37 +00:00
|
|
|
StopWatch sw(env_, statistics_, DB_SEEK);
|
2016-05-03 04:46:30 +00:00
|
|
|
ReleaseTempPinnedData();
|
2017-03-30 18:57:05 +00:00
|
|
|
ResetInternalKeysSkippedCounter();
|
2014-04-01 21:45:30 +00:00
|
|
|
saved_key_.Clear();
|
|
|
|
saved_key_.SetInternalKey(target, sequence_);
|
2014-08-22 22:28:58 +00:00
|
|
|
|
2017-11-30 06:48:55 +00:00
|
|
|
if (iterate_lower_bound_ != nullptr &&
|
|
|
|
user_comparator_->Compare(saved_key_.GetUserKey(),
|
|
|
|
*iterate_lower_bound_) < 0) {
|
|
|
|
saved_key_.Clear();
|
|
|
|
saved_key_.SetInternalKey(*iterate_lower_bound_, sequence_);
|
|
|
|
}
|
|
|
|
|
2014-08-22 22:28:58 +00:00
|
|
|
{
|
|
|
|
PERF_TIMER_GUARD(seek_internal_seek_time);
|
2017-04-04 21:17:16 +00:00
|
|
|
iter_->Seek(saved_key_.GetInternalKey());
|
Maintain position in range deletions map
Summary:
When deletion-collapsing mode is enabled (i.e., for DBIter/CompactionIterator), we maintain position in the tombstone maps across calls to ShouldDelete(). Since iterators often access keys sequentially (or reverse-sequentially), scanning forward/backward from the last position can be faster than binary-searching the map for every key.
- When Next() is invoked on an iterator, we use kForwardTraversal to scan forwards, if needed, until arriving at the range deletion containing the next key.
- Similarly for Prev(), we use kBackwardTraversal to scan backwards in the range deletion map.
- When the iterator seeks, we use kBinarySearch for repositioning
- After tombstones are added or before the first ShouldDelete() invocation, the current position is set to invalid, which forces kBinarySearch to be used.
- Non-iterator users (i.e., Get()) use kFullScan, which has the same behavior as before---scan the whole map for every key passed to ShouldDelete().
Closes https://github.com/facebook/rocksdb/pull/1701
Differential Revision: D4350318
Pulled By: ajkr
fbshipit-source-id: 5129b76
2017-01-05 18:22:46 +00:00
|
|
|
range_del_agg_.InvalidateTombstoneMapPositions();
|
2014-08-22 22:28:58 +00:00
|
|
|
}
|
2015-09-11 18:37:44 +00:00
|
|
|
RecordTick(statistics_, NUMBER_DB_SEEK);
|
2011-03-25 20:27:43 +00:00
|
|
|
if (iter_->Valid()) {
|
2016-04-28 23:48:03 +00:00
|
|
|
if (prefix_extractor_ && prefix_same_as_start_) {
|
|
|
|
prefix_start_key_ = prefix_extractor_->Transform(target);
|
|
|
|
}
|
2013-11-20 00:37:34 +00:00
|
|
|
direction_ = kForward;
|
|
|
|
ClearSavedValue();
|
2016-04-28 23:48:03 +00:00
|
|
|
FindNextUserEntry(false /* not skipping */, prefix_same_as_start_);
|
|
|
|
if (!valid_) {
|
|
|
|
prefix_start_key_.clear();
|
|
|
|
}
|
2015-09-11 18:37:44 +00:00
|
|
|
if (statistics_ != nullptr) {
|
|
|
|
if (valid_) {
|
2017-11-21 05:12:55 +00:00
|
|
|
// Decrement since we don't want to count this key as skipped
|
2015-09-11 18:37:44 +00:00
|
|
|
RecordTick(statistics_, NUMBER_DB_SEEK_FOUND);
|
|
|
|
RecordTick(statistics_, ITER_BYTES_READ, key().size() + value().size());
|
2017-08-18 18:40:36 +00:00
|
|
|
PERF_COUNTER_ADD(iter_read_bytes, key().size() + value().size());
|
2015-09-11 18:37:44 +00:00
|
|
|
}
|
|
|
|
}
|
2011-03-25 20:27:43 +00:00
|
|
|
} else {
|
|
|
|
valid_ = false;
|
|
|
|
}
|
2016-10-11 20:54:26 +00:00
|
|
|
|
2015-11-05 21:24:05 +00:00
|
|
|
if (valid_ && prefix_extractor_ && prefix_same_as_start_) {
|
2017-04-04 21:17:16 +00:00
|
|
|
prefix_start_buf_.SetUserKey(prefix_start_key_);
|
|
|
|
prefix_start_key_ = prefix_start_buf_.GetUserKey();
|
2015-11-05 21:24:05 +00:00
|
|
|
}
|
2011-03-25 20:27:43 +00:00
|
|
|
}
|
2011-03-18 22:37:00 +00:00
|
|
|
|
2016-09-28 01:20:57 +00:00
|
|
|
void DBIter::SeekForPrev(const Slice& target) {
|
|
|
|
StopWatch sw(env_, statistics_, DB_SEEK);
|
|
|
|
ReleaseTempPinnedData();
|
2017-03-30 18:57:05 +00:00
|
|
|
ResetInternalKeysSkippedCounter();
|
2016-09-28 01:20:57 +00:00
|
|
|
saved_key_.Clear();
|
|
|
|
// now saved_key is used to store internal key.
|
|
|
|
saved_key_.SetInternalKey(target, 0 /* sequence_number */,
|
|
|
|
kValueTypeForSeekForPrev);
|
|
|
|
|
2017-11-30 06:48:55 +00:00
|
|
|
if (iterate_upper_bound_ != nullptr &&
|
|
|
|
user_comparator_->Compare(saved_key_.GetUserKey(),
|
|
|
|
*iterate_upper_bound_) >= 0) {
|
|
|
|
saved_key_.Clear();
|
|
|
|
saved_key_.SetInternalKey(*iterate_upper_bound_, kMaxSequenceNumber);
|
|
|
|
}
|
|
|
|
|
2016-09-28 01:20:57 +00:00
|
|
|
{
|
|
|
|
PERF_TIMER_GUARD(seek_internal_seek_time);
|
2017-04-04 21:17:16 +00:00
|
|
|
iter_->SeekForPrev(saved_key_.GetInternalKey());
|
Maintain position in range deletions map
Summary:
When deletion-collapsing mode is enabled (i.e., for DBIter/CompactionIterator), we maintain position in the tombstone maps across calls to ShouldDelete(). Since iterators often access keys sequentially (or reverse-sequentially), scanning forward/backward from the last position can be faster than binary-searching the map for every key.
- When Next() is invoked on an iterator, we use kForwardTraversal to scan forwards, if needed, until arriving at the range deletion containing the next key.
- Similarly for Prev(), we use kBackwardTraversal to scan backwards in the range deletion map.
- When the iterator seeks, we use kBinarySearch for repositioning
- After tombstones are added or before the first ShouldDelete() invocation, the current position is set to invalid, which forces kBinarySearch to be used.
- Non-iterator users (i.e., Get()) use kFullScan, which has the same behavior as before---scan the whole map for every key passed to ShouldDelete().
Closes https://github.com/facebook/rocksdb/pull/1701
Differential Revision: D4350318
Pulled By: ajkr
fbshipit-source-id: 5129b76
2017-01-05 18:22:46 +00:00
|
|
|
range_del_agg_.InvalidateTombstoneMapPositions();
|
2016-09-28 01:20:57 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
RecordTick(statistics_, NUMBER_DB_SEEK);
|
|
|
|
if (iter_->Valid()) {
|
|
|
|
if (prefix_extractor_ && prefix_same_as_start_) {
|
|
|
|
prefix_start_key_ = prefix_extractor_->Transform(target);
|
|
|
|
}
|
|
|
|
direction_ = kReverse;
|
|
|
|
ClearSavedValue();
|
|
|
|
PrevInternal();
|
|
|
|
if (!valid_) {
|
|
|
|
prefix_start_key_.clear();
|
|
|
|
}
|
|
|
|
if (statistics_ != nullptr) {
|
|
|
|
if (valid_) {
|
|
|
|
RecordTick(statistics_, NUMBER_DB_SEEK_FOUND);
|
|
|
|
RecordTick(statistics_, ITER_BYTES_READ, key().size() + value().size());
|
2017-08-18 18:40:36 +00:00
|
|
|
PERF_COUNTER_ADD(iter_read_bytes, key().size() + value().size());
|
2016-09-28 01:20:57 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
valid_ = false;
|
|
|
|
}
|
|
|
|
if (valid_ && prefix_extractor_ && prefix_same_as_start_) {
|
2017-04-04 21:17:16 +00:00
|
|
|
prefix_start_buf_.SetUserKey(prefix_start_key_);
|
|
|
|
prefix_start_key_ = prefix_start_buf_.GetUserKey();
|
2016-09-28 01:20:57 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2011-03-25 20:27:43 +00:00
|
|
|
void DBIter::SeekToFirst() {
|
2014-08-08 16:44:14 +00:00
|
|
|
// Don't use iter_::Seek() if we set a prefix extractor
|
2015-08-11 18:46:15 +00:00
|
|
|
// because prefix seek will be used.
|
2014-09-04 17:48:24 +00:00
|
|
|
if (prefix_extractor_ != nullptr) {
|
2014-08-08 16:44:14 +00:00
|
|
|
max_skip_ = std::numeric_limits<uint64_t>::max();
|
|
|
|
}
|
2017-10-27 00:14:04 +00:00
|
|
|
if (iterate_lower_bound_ != nullptr) {
|
|
|
|
Seek(*iterate_lower_bound_);
|
|
|
|
return;
|
|
|
|
}
|
2011-03-25 20:27:43 +00:00
|
|
|
direction_ = kForward;
|
2016-05-03 04:46:30 +00:00
|
|
|
ReleaseTempPinnedData();
|
2017-03-30 18:57:05 +00:00
|
|
|
ResetInternalKeysSkippedCounter();
|
2011-03-25 20:27:43 +00:00
|
|
|
ClearSavedValue();
|
2014-08-22 22:28:58 +00:00
|
|
|
|
|
|
|
{
|
|
|
|
PERF_TIMER_GUARD(seek_internal_seek_time);
|
|
|
|
iter_->SeekToFirst();
|
Maintain position in range deletions map
Summary:
When deletion-collapsing mode is enabled (i.e., for DBIter/CompactionIterator), we maintain position in the tombstone maps across calls to ShouldDelete(). Since iterators often access keys sequentially (or reverse-sequentially), scanning forward/backward from the last position can be faster than binary-searching the map for every key.
- When Next() is invoked on an iterator, we use kForwardTraversal to scan forwards, if needed, until arriving at the range deletion containing the next key.
- Similarly for Prev(), we use kBackwardTraversal to scan backwards in the range deletion map.
- When the iterator seeks, we use kBinarySearch for repositioning
- After tombstones are added or before the first ShouldDelete() invocation, the current position is set to invalid, which forces kBinarySearch to be used.
- Non-iterator users (i.e., Get()) use kFullScan, which has the same behavior as before---scan the whole map for every key passed to ShouldDelete().
Closes https://github.com/facebook/rocksdb/pull/1701
Differential Revision: D4350318
Pulled By: ajkr
fbshipit-source-id: 5129b76
2017-01-05 18:22:46 +00:00
|
|
|
range_del_agg_.InvalidateTombstoneMapPositions();
|
2014-08-22 22:28:58 +00:00
|
|
|
}
|
|
|
|
|
2015-09-11 18:37:44 +00:00
|
|
|
RecordTick(statistics_, NUMBER_DB_SEEK);
|
2011-03-25 20:27:43 +00:00
|
|
|
if (iter_->Valid()) {
|
2017-04-04 21:17:16 +00:00
|
|
|
saved_key_.SetUserKey(
|
|
|
|
ExtractUserKey(iter_->key()),
|
|
|
|
!iter_->IsKeyPinned() || !pin_thru_lifetime_ /* copy */);
|
2016-04-28 23:48:03 +00:00
|
|
|
FindNextUserEntry(false /* not skipping */, false /* no prefix check */);
|
2015-09-11 18:37:44 +00:00
|
|
|
if (statistics_ != nullptr) {
|
|
|
|
if (valid_) {
|
|
|
|
RecordTick(statistics_, NUMBER_DB_SEEK_FOUND);
|
|
|
|
RecordTick(statistics_, ITER_BYTES_READ, key().size() + value().size());
|
2017-08-18 18:40:36 +00:00
|
|
|
PERF_COUNTER_ADD(iter_read_bytes, key().size() + value().size());
|
2015-09-11 18:37:44 +00:00
|
|
|
}
|
|
|
|
}
|
2011-03-25 20:27:43 +00:00
|
|
|
} else {
|
|
|
|
valid_ = false;
|
2011-03-18 22:37:00 +00:00
|
|
|
}
|
2015-11-05 21:24:05 +00:00
|
|
|
if (valid_ && prefix_extractor_ && prefix_same_as_start_) {
|
2017-04-04 21:17:16 +00:00
|
|
|
prefix_start_buf_.SetUserKey(
|
|
|
|
prefix_extractor_->Transform(saved_key_.GetUserKey()));
|
|
|
|
prefix_start_key_ = prefix_start_buf_.GetUserKey();
|
2015-11-05 21:24:05 +00:00
|
|
|
}
|
2011-03-18 22:37:00 +00:00
|
|
|
}
|
|
|
|
|
2011-03-25 20:27:43 +00:00
|
|
|
void DBIter::SeekToLast() {
|
2014-08-08 16:44:14 +00:00
|
|
|
// Don't use iter_::Seek() if we set a prefix extractor
|
2015-08-11 18:46:15 +00:00
|
|
|
// because prefix seek will be used.
|
2014-09-04 17:48:24 +00:00
|
|
|
if (prefix_extractor_ != nullptr) {
|
2014-08-08 16:44:14 +00:00
|
|
|
max_skip_ = std::numeric_limits<uint64_t>::max();
|
|
|
|
}
|
2011-03-25 20:27:43 +00:00
|
|
|
direction_ = kReverse;
|
2016-05-03 04:46:30 +00:00
|
|
|
ReleaseTempPinnedData();
|
2017-03-30 18:57:05 +00:00
|
|
|
ResetInternalKeysSkippedCounter();
|
2011-03-25 20:27:43 +00:00
|
|
|
ClearSavedValue();
|
2014-08-22 22:28:58 +00:00
|
|
|
|
|
|
|
{
|
|
|
|
PERF_TIMER_GUARD(seek_internal_seek_time);
|
|
|
|
iter_->SeekToLast();
|
Maintain position in range deletions map
Summary:
When deletion-collapsing mode is enabled (i.e., for DBIter/CompactionIterator), we maintain position in the tombstone maps across calls to ShouldDelete(). Since iterators often access keys sequentially (or reverse-sequentially), scanning forward/backward from the last position can be faster than binary-searching the map for every key.
- When Next() is invoked on an iterator, we use kForwardTraversal to scan forwards, if needed, until arriving at the range deletion containing the next key.
- Similarly for Prev(), we use kBackwardTraversal to scan backwards in the range deletion map.
- When the iterator seeks, we use kBinarySearch for repositioning
- After tombstones are added or before the first ShouldDelete() invocation, the current position is set to invalid, which forces kBinarySearch to be used.
- Non-iterator users (i.e., Get()) use kFullScan, which has the same behavior as before---scan the whole map for every key passed to ShouldDelete().
Closes https://github.com/facebook/rocksdb/pull/1701
Differential Revision: D4350318
Pulled By: ajkr
fbshipit-source-id: 5129b76
2017-01-05 18:22:46 +00:00
|
|
|
range_del_agg_.InvalidateTombstoneMapPositions();
|
2014-08-22 22:28:58 +00:00
|
|
|
}
|
2015-06-25 16:44:30 +00:00
|
|
|
// When the iterate_upper_bound is set to a value,
|
|
|
|
// it will seek to the last key before the
|
|
|
|
// ReadOptions.iterate_upper_bound
|
|
|
|
if (iter_->Valid() && iterate_upper_bound_ != nullptr) {
|
2016-10-11 20:54:26 +00:00
|
|
|
SeekForPrev(*iterate_upper_bound_);
|
Maintain position in range deletions map
Summary:
When deletion-collapsing mode is enabled (i.e., for DBIter/CompactionIterator), we maintain position in the tombstone maps across calls to ShouldDelete(). Since iterators often access keys sequentially (or reverse-sequentially), scanning forward/backward from the last position can be faster than binary-searching the map for every key.
- When Next() is invoked on an iterator, we use kForwardTraversal to scan forwards, if needed, until arriving at the range deletion containing the next key.
- Similarly for Prev(), we use kBackwardTraversal to scan backwards in the range deletion map.
- When the iterator seeks, we use kBinarySearch for repositioning
- After tombstones are added or before the first ShouldDelete() invocation, the current position is set to invalid, which forces kBinarySearch to be used.
- Non-iterator users (i.e., Get()) use kFullScan, which has the same behavior as before---scan the whole map for every key passed to ShouldDelete().
Closes https://github.com/facebook/rocksdb/pull/1701
Differential Revision: D4350318
Pulled By: ajkr
fbshipit-source-id: 5129b76
2017-01-05 18:22:46 +00:00
|
|
|
range_del_agg_.InvalidateTombstoneMapPositions();
|
2016-10-11 20:54:26 +00:00
|
|
|
if (!Valid()) {
|
|
|
|
return;
|
|
|
|
} else if (user_comparator_->Equal(*iterate_upper_bound_, key())) {
|
2017-11-21 05:12:55 +00:00
|
|
|
ReleaseTempPinnedData();
|
|
|
|
PrevInternal();
|
2015-06-25 16:44:30 +00:00
|
|
|
}
|
2016-10-11 20:54:26 +00:00
|
|
|
} else {
|
|
|
|
PrevInternal();
|
2015-06-25 16:44:30 +00:00
|
|
|
}
|
2015-09-11 18:37:44 +00:00
|
|
|
if (statistics_ != nullptr) {
|
|
|
|
RecordTick(statistics_, NUMBER_DB_SEEK);
|
|
|
|
if (valid_) {
|
|
|
|
RecordTick(statistics_, NUMBER_DB_SEEK_FOUND);
|
|
|
|
RecordTick(statistics_, ITER_BYTES_READ, key().size() + value().size());
|
2017-08-18 18:40:36 +00:00
|
|
|
PERF_COUNTER_ADD(iter_read_bytes, key().size() + value().size());
|
2015-09-11 18:37:44 +00:00
|
|
|
}
|
|
|
|
}
|
2015-11-05 21:24:05 +00:00
|
|
|
if (valid_ && prefix_extractor_ && prefix_same_as_start_) {
|
2017-04-04 21:17:16 +00:00
|
|
|
prefix_start_buf_.SetUserKey(
|
|
|
|
prefix_extractor_->Transform(saved_key_.GetUserKey()));
|
|
|
|
prefix_start_key_ = prefix_start_buf_.GetUserKey();
|
2015-11-05 21:24:05 +00:00
|
|
|
}
|
2011-03-25 20:27:43 +00:00
|
|
|
}
|
|
|
|
|
2017-04-10 18:13:23 +00:00
|
|
|
Iterator* NewDBIterator(Env* env, const ReadOptions& read_options,
|
|
|
|
const ImmutableCFOptions& cf_options,
|
|
|
|
const Comparator* user_key_comparator,
|
|
|
|
InternalIterator* internal_iter,
|
|
|
|
const SequenceNumber& sequence,
|
2017-10-03 16:08:07 +00:00
|
|
|
uint64_t max_sequential_skip_in_iterations,
|
2017-10-10 00:05:34 +00:00
|
|
|
ReadCallback* read_callback, bool allow_blob) {
|
|
|
|
DBIter* db_iter =
|
|
|
|
new DBIter(env, read_options, cf_options, user_key_comparator,
|
|
|
|
internal_iter, sequence, false,
|
|
|
|
max_sequential_skip_in_iterations, read_callback, allow_blob);
|
2015-12-16 20:08:30 +00:00
|
|
|
return db_iter;
|
In DB::NewIterator(), try to allocate the whole iterator tree in an arena
Summary:
In this patch, try to allocate the whole iterator tree starting from DBIter from an arena
1. ArenaWrappedDBIter is created when serves as the entry point of an iterator tree, with an arena in it.
2. Add an option to create iterator from arena for following iterators: DBIter, MergingIterator, MemtableIterator, all mem table's iterators, all table reader's iterators and two level iterator.
3. MergeIteratorBuilder is created to incrementally build the tree of internal iterators. It is passed to mem table list and version set and add iterators to it.
Limitations:
(1) Only DB::NewIterator() without tailing uses the arena. Other cases, including readonly DB and compactions are still from malloc
(2) Two level iterator itself is allocated in arena, but not iterators inside it.
Test Plan: make all check
Reviewers: ljin, haobo
Reviewed By: haobo
Subscribers: leveldb, dhruba, yhchiang, igor
Differential Revision: https://reviews.facebook.net/D18513
2014-06-02 23:38:00 +00:00
|
|
|
}
|
|
|
|
|
2014-06-03 19:28:58 +00:00
|
|
|
ArenaWrappedDBIter::~ArenaWrappedDBIter() { db_iter_->~DBIter(); }
|
In DB::NewIterator(), try to allocate the whole iterator tree in an arena
Summary:
In this patch, try to allocate the whole iterator tree starting from DBIter from an arena
1. ArenaWrappedDBIter is created when serves as the entry point of an iterator tree, with an arena in it.
2. Add an option to create iterator from arena for following iterators: DBIter, MergingIterator, MemtableIterator, all mem table's iterators, all table reader's iterators and two level iterator.
3. MergeIteratorBuilder is created to incrementally build the tree of internal iterators. It is passed to mem table list and version set and add iterators to it.
Limitations:
(1) Only DB::NewIterator() without tailing uses the arena. Other cases, including readonly DB and compactions are still from malloc
(2) Two level iterator itself is allocated in arena, but not iterators inside it.
Test Plan: make all check
Reviewers: ljin, haobo
Reviewed By: haobo
Subscribers: leveldb, dhruba, yhchiang, igor
Differential Revision: https://reviews.facebook.net/D18513
2014-06-02 23:38:00 +00:00
|
|
|
|
2016-11-04 18:53:38 +00:00
|
|
|
RangeDelAggregator* ArenaWrappedDBIter::GetRangeDelAggregator() {
|
|
|
|
return db_iter_->GetRangeDelAggregator();
|
|
|
|
}
|
|
|
|
|
2015-10-12 22:06:38 +00:00
|
|
|
void ArenaWrappedDBIter::SetIterUnderDBIter(InternalIterator* iter) {
|
In DB::NewIterator(), try to allocate the whole iterator tree in an arena
Summary:
In this patch, try to allocate the whole iterator tree starting from DBIter from an arena
1. ArenaWrappedDBIter is created when serves as the entry point of an iterator tree, with an arena in it.
2. Add an option to create iterator from arena for following iterators: DBIter, MergingIterator, MemtableIterator, all mem table's iterators, all table reader's iterators and two level iterator.
3. MergeIteratorBuilder is created to incrementally build the tree of internal iterators. It is passed to mem table list and version set and add iterators to it.
Limitations:
(1) Only DB::NewIterator() without tailing uses the arena. Other cases, including readonly DB and compactions are still from malloc
(2) Two level iterator itself is allocated in arena, but not iterators inside it.
Test Plan: make all check
Reviewers: ljin, haobo
Reviewed By: haobo
Subscribers: leveldb, dhruba, yhchiang, igor
Differential Revision: https://reviews.facebook.net/D18513
2014-06-02 23:38:00 +00:00
|
|
|
static_cast<DBIter*>(db_iter_)->SetIter(iter);
|
|
|
|
}
|
|
|
|
|
|
|
|
inline bool ArenaWrappedDBIter::Valid() const { return db_iter_->Valid(); }
|
|
|
|
inline void ArenaWrappedDBIter::SeekToFirst() { db_iter_->SeekToFirst(); }
|
|
|
|
inline void ArenaWrappedDBIter::SeekToLast() { db_iter_->SeekToLast(); }
|
|
|
|
inline void ArenaWrappedDBIter::Seek(const Slice& target) {
|
|
|
|
db_iter_->Seek(target);
|
|
|
|
}
|
2016-09-28 01:20:57 +00:00
|
|
|
inline void ArenaWrappedDBIter::SeekForPrev(const Slice& target) {
|
|
|
|
db_iter_->SeekForPrev(target);
|
|
|
|
}
|
In DB::NewIterator(), try to allocate the whole iterator tree in an arena
Summary:
In this patch, try to allocate the whole iterator tree starting from DBIter from an arena
1. ArenaWrappedDBIter is created when serves as the entry point of an iterator tree, with an arena in it.
2. Add an option to create iterator from arena for following iterators: DBIter, MergingIterator, MemtableIterator, all mem table's iterators, all table reader's iterators and two level iterator.
3. MergeIteratorBuilder is created to incrementally build the tree of internal iterators. It is passed to mem table list and version set and add iterators to it.
Limitations:
(1) Only DB::NewIterator() without tailing uses the arena. Other cases, including readonly DB and compactions are still from malloc
(2) Two level iterator itself is allocated in arena, but not iterators inside it.
Test Plan: make all check
Reviewers: ljin, haobo
Reviewed By: haobo
Subscribers: leveldb, dhruba, yhchiang, igor
Differential Revision: https://reviews.facebook.net/D18513
2014-06-02 23:38:00 +00:00
|
|
|
inline void ArenaWrappedDBIter::Next() { db_iter_->Next(); }
|
|
|
|
inline void ArenaWrappedDBIter::Prev() { db_iter_->Prev(); }
|
|
|
|
inline Slice ArenaWrappedDBIter::key() const { return db_iter_->key(); }
|
|
|
|
inline Slice ArenaWrappedDBIter::value() const { return db_iter_->value(); }
|
|
|
|
inline Status ArenaWrappedDBIter::status() const { return db_iter_->status(); }
|
2017-10-03 16:08:07 +00:00
|
|
|
bool ArenaWrappedDBIter::IsBlob() const { return db_iter_->IsBlob(); }
|
2016-02-27 01:13:39 +00:00
|
|
|
inline Status ArenaWrappedDBIter::GetProperty(std::string prop_name,
|
|
|
|
std::string* prop) {
|
2017-07-24 17:28:17 +00:00
|
|
|
if (prop_name == "rocksdb.iterator.super-version-number") {
|
|
|
|
// First try to pass the value returned from inner iterator.
|
|
|
|
if (!db_iter_->GetProperty(prop_name, prop).ok()) {
|
|
|
|
*prop = ToString(sv_number_);
|
|
|
|
}
|
|
|
|
return Status::OK();
|
|
|
|
}
|
2016-02-27 01:13:39 +00:00
|
|
|
return db_iter_->GetProperty(prop_name, prop);
|
|
|
|
}
|
2017-07-24 17:28:17 +00:00
|
|
|
|
|
|
|
void ArenaWrappedDBIter::Init(Env* env, const ReadOptions& read_options,
|
|
|
|
const ImmutableCFOptions& cf_options,
|
|
|
|
const SequenceNumber& sequence,
|
|
|
|
uint64_t max_sequential_skip_in_iteration,
|
2017-10-10 00:05:34 +00:00
|
|
|
uint64_t version_number,
|
2017-12-19 06:13:08 +00:00
|
|
|
ReadCallback* read_callback, bool allow_blob,
|
|
|
|
bool allow_refresh) {
|
2017-07-24 17:28:17 +00:00
|
|
|
auto mem = arena_.AllocateAligned(sizeof(DBIter));
|
|
|
|
db_iter_ = new (mem)
|
|
|
|
DBIter(env, read_options, cf_options, cf_options.user_comparator, nullptr,
|
2017-10-10 00:05:34 +00:00
|
|
|
sequence, true, max_sequential_skip_in_iteration, read_callback,
|
|
|
|
allow_blob);
|
2017-07-24 17:28:17 +00:00
|
|
|
sv_number_ = version_number;
|
2017-12-19 06:13:08 +00:00
|
|
|
allow_refresh_ = allow_refresh;
|
2017-07-24 17:28:17 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
Status ArenaWrappedDBIter::Refresh() {
|
2017-12-19 06:13:08 +00:00
|
|
|
if (cfd_ == nullptr || db_impl_ == nullptr || !allow_refresh_) {
|
2017-07-24 17:28:17 +00:00
|
|
|
return Status::NotSupported("Creating renew iterator is not allowed.");
|
|
|
|
}
|
|
|
|
assert(db_iter_ != nullptr);
|
2017-12-01 07:39:56 +00:00
|
|
|
// TODO(yiwu): For last_seq_same_as_publish_seq_==false, this is not the
|
|
|
|
// correct behavior. Will be corrected automatically when we take a snapshot
|
|
|
|
// here for the case of WritePreparedTxnDB.
|
2017-07-24 17:28:17 +00:00
|
|
|
SequenceNumber latest_seq = db_impl_->GetLatestSequenceNumber();
|
|
|
|
uint64_t cur_sv_number = cfd_->GetSuperVersionNumber();
|
|
|
|
if (sv_number_ != cur_sv_number) {
|
|
|
|
Env* env = db_iter_->env();
|
|
|
|
db_iter_->~DBIter();
|
|
|
|
arena_.~Arena();
|
|
|
|
new (&arena_) Arena();
|
|
|
|
|
|
|
|
SuperVersion* sv = cfd_->GetReferencedSuperVersion(db_impl_->mutex());
|
|
|
|
Init(env, read_options_, *(cfd_->ioptions()), latest_seq,
|
|
|
|
sv->mutable_cf_options.max_sequential_skip_in_iterations,
|
2017-12-19 06:13:08 +00:00
|
|
|
cur_sv_number, read_callback_, allow_blob_, allow_refresh_);
|
2017-07-24 17:28:17 +00:00
|
|
|
|
|
|
|
InternalIterator* internal_iter = db_impl_->NewInternalIterator(
|
|
|
|
read_options_, cfd_, sv, &arena_, db_iter_->GetRangeDelAggregator());
|
|
|
|
SetIterUnderDBIter(internal_iter);
|
|
|
|
} else {
|
|
|
|
db_iter_->set_sequence(latest_seq);
|
|
|
|
db_iter_->set_valid(false);
|
|
|
|
}
|
|
|
|
return Status::OK();
|
In DB::NewIterator(), try to allocate the whole iterator tree in an arena
Summary:
In this patch, try to allocate the whole iterator tree starting from DBIter from an arena
1. ArenaWrappedDBIter is created when serves as the entry point of an iterator tree, with an arena in it.
2. Add an option to create iterator from arena for following iterators: DBIter, MergingIterator, MemtableIterator, all mem table's iterators, all table reader's iterators and two level iterator.
3. MergeIteratorBuilder is created to incrementally build the tree of internal iterators. It is passed to mem table list and version set and add iterators to it.
Limitations:
(1) Only DB::NewIterator() without tailing uses the arena. Other cases, including readonly DB and compactions are still from malloc
(2) Two level iterator itself is allocated in arena, but not iterators inside it.
Test Plan: make all check
Reviewers: ljin, haobo
Reviewed By: haobo
Subscribers: leveldb, dhruba, yhchiang, igor
Differential Revision: https://reviews.facebook.net/D18513
2014-06-02 23:38:00 +00:00
|
|
|
}
|
2011-03-18 22:37:00 +00:00
|
|
|
|
In DB::NewIterator(), try to allocate the whole iterator tree in an arena
Summary:
In this patch, try to allocate the whole iterator tree starting from DBIter from an arena
1. ArenaWrappedDBIter is created when serves as the entry point of an iterator tree, with an arena in it.
2. Add an option to create iterator from arena for following iterators: DBIter, MergingIterator, MemtableIterator, all mem table's iterators, all table reader's iterators and two level iterator.
3. MergeIteratorBuilder is created to incrementally build the tree of internal iterators. It is passed to mem table list and version set and add iterators to it.
Limitations:
(1) Only DB::NewIterator() without tailing uses the arena. Other cases, including readonly DB and compactions are still from malloc
(2) Two level iterator itself is allocated in arena, but not iterators inside it.
Test Plan: make all check
Reviewers: ljin, haobo
Reviewed By: haobo
Subscribers: leveldb, dhruba, yhchiang, igor
Differential Revision: https://reviews.facebook.net/D18513
2014-06-02 23:38:00 +00:00
|
|
|
ArenaWrappedDBIter* NewArenaWrappedDbIterator(
|
2017-04-10 18:13:23 +00:00
|
|
|
Env* env, const ReadOptions& read_options,
|
2017-07-24 17:28:17 +00:00
|
|
|
const ImmutableCFOptions& cf_options, const SequenceNumber& sequence,
|
|
|
|
uint64_t max_sequential_skip_in_iterations, uint64_t version_number,
|
2017-10-10 00:05:34 +00:00
|
|
|
ReadCallback* read_callback, DBImpl* db_impl, ColumnFamilyData* cfd,
|
2017-12-19 06:13:08 +00:00
|
|
|
bool allow_blob, bool allow_refresh) {
|
In DB::NewIterator(), try to allocate the whole iterator tree in an arena
Summary:
In this patch, try to allocate the whole iterator tree starting from DBIter from an arena
1. ArenaWrappedDBIter is created when serves as the entry point of an iterator tree, with an arena in it.
2. Add an option to create iterator from arena for following iterators: DBIter, MergingIterator, MemtableIterator, all mem table's iterators, all table reader's iterators and two level iterator.
3. MergeIteratorBuilder is created to incrementally build the tree of internal iterators. It is passed to mem table list and version set and add iterators to it.
Limitations:
(1) Only DB::NewIterator() without tailing uses the arena. Other cases, including readonly DB and compactions are still from malloc
(2) Two level iterator itself is allocated in arena, but not iterators inside it.
Test Plan: make all check
Reviewers: ljin, haobo
Reviewed By: haobo
Subscribers: leveldb, dhruba, yhchiang, igor
Differential Revision: https://reviews.facebook.net/D18513
2014-06-02 23:38:00 +00:00
|
|
|
ArenaWrappedDBIter* iter = new ArenaWrappedDBIter();
|
2017-07-24 17:28:17 +00:00
|
|
|
iter->Init(env, read_options, cf_options, sequence,
|
2017-10-10 00:05:34 +00:00
|
|
|
max_sequential_skip_in_iterations, version_number, read_callback,
|
2017-12-19 06:13:08 +00:00
|
|
|
allow_blob, allow_refresh);
|
|
|
|
if (db_impl != nullptr && cfd != nullptr && allow_refresh) {
|
2017-10-10 00:05:34 +00:00
|
|
|
iter->StoreRefreshInfo(read_options, db_impl, cfd, read_callback,
|
|
|
|
allow_blob);
|
2017-07-24 17:28:17 +00:00
|
|
|
}
|
2014-09-04 17:48:24 +00:00
|
|
|
|
In DB::NewIterator(), try to allocate the whole iterator tree in an arena
Summary:
In this patch, try to allocate the whole iterator tree starting from DBIter from an arena
1. ArenaWrappedDBIter is created when serves as the entry point of an iterator tree, with an arena in it.
2. Add an option to create iterator from arena for following iterators: DBIter, MergingIterator, MemtableIterator, all mem table's iterators, all table reader's iterators and two level iterator.
3. MergeIteratorBuilder is created to incrementally build the tree of internal iterators. It is passed to mem table list and version set and add iterators to it.
Limitations:
(1) Only DB::NewIterator() without tailing uses the arena. Other cases, including readonly DB and compactions are still from malloc
(2) Two level iterator itself is allocated in arena, but not iterators inside it.
Test Plan: make all check
Reviewers: ljin, haobo
Reviewed By: haobo
Subscribers: leveldb, dhruba, yhchiang, igor
Differential Revision: https://reviews.facebook.net/D18513
2014-06-02 23:38:00 +00:00
|
|
|
return iter;
|
2011-03-18 22:37:00 +00:00
|
|
|
}
|
|
|
|
|
2013-10-04 04:49:15 +00:00
|
|
|
} // namespace rocksdb
|