|
|
|
// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
|
|
|
|
// This source code is licensed under both the GPLv2 (found in the
|
|
|
|
// COPYING file in the root directory) and Apache 2.0 License
|
|
|
|
// (found in the LICENSE.Apache file in the root directory).
|
|
|
|
//
|
|
|
|
// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
|
|
|
|
// Use of this source code is governed by a BSD-style license that can be
|
|
|
|
// found in the LICENSE file. See the AUTHORS file for names of contributors.
|
|
|
|
#include "table/block_based/block_based_table_reader.h"
|
|
|
|
|
|
|
|
#include <algorithm>
|
|
|
|
#include <array>
|
|
|
|
#include <limits>
|
|
|
|
#include <string>
|
|
|
|
#include <utility>
|
|
|
|
#include <vector>
|
|
|
|
|
|
|
|
#include "db/dbformat.h"
|
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
9 years ago
|
|
|
#include "db/pinned_iterators_manager.h"
|
|
|
|
|
|
|
|
#include "file/file_prefetch_buffer.h"
|
|
|
|
#include "file/random_access_file_reader.h"
|
|
|
|
|
|
|
|
#include "rocksdb/cache.h"
|
|
|
|
#include "rocksdb/comparator.h"
|
|
|
|
#include "rocksdb/env.h"
|
|
|
|
#include "rocksdb/filter_policy.h"
|
|
|
|
#include "rocksdb/iterator.h"
|
|
|
|
#include "rocksdb/options.h"
|
|
|
|
#include "rocksdb/statistics.h"
|
|
|
|
#include "rocksdb/table.h"
|
|
|
|
#include "rocksdb/table_properties.h"
|
|
|
|
|
|
|
|
#include "table/block_based/block.h"
|
|
|
|
#include "table/block_based/block_based_filter_block.h"
|
|
|
|
#include "table/block_based/block_based_table_factory.h"
|
|
|
|
#include "table/block_based/block_prefix_index.h"
|
|
|
|
#include "table/block_based/filter_block.h"
|
|
|
|
#include "table/block_based/full_filter_block.h"
|
|
|
|
#include "table/block_based/partitioned_filter_block.h"
|
|
|
|
#include "table/block_fetcher.h"
|
|
|
|
#include "table/format.h"
|
|
|
|
#include "table/get_context.h"
|
|
|
|
#include "table/internal_iterator.h"
|
|
|
|
#include "table/meta_blocks.h"
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
#include "table/multiget_context.h"
|
|
|
|
#include "table/persistent_cache_helper.h"
|
|
|
|
#include "table/sst_file_writer_collectors.h"
|
|
|
|
#include "table/two_level_iterator.h"
|
|
|
|
|
|
|
|
#include "monitoring/perf_context_imp.h"
|
|
|
|
#include "test_util/sync_point.h"
|
|
|
|
#include "util/coding.h"
|
|
|
|
#include "util/crc32c.h"
|
|
|
|
#include "util/stop_watch.h"
|
|
|
|
#include "util/string_util.h"
|
|
|
|
#include "util/xxhash.h"
|
|
|
|
|
|
|
|
namespace rocksdb {
|
|
|
|
|
|
|
|
extern const uint64_t kBlockBasedTableMagicNumber;
|
|
|
|
extern const std::string kHashIndexPrefixesBlock;
|
|
|
|
extern const std::string kHashIndexPrefixesMetadataBlock;
|
|
|
|
|
|
|
|
typedef BlockBasedTable::IndexReader IndexReader;
|
|
|
|
|
|
|
|
// Found that 256 KB readahead size provides the best performance, based on
|
|
|
|
// experiments, for auto readahead. Experiment data is in PR #3282.
|
|
|
|
const size_t BlockBasedTable::kMaxAutoReadaheadSize = 256 * 1024;
|
|
|
|
|
|
|
|
BlockBasedTable::~BlockBasedTable() {
|
|
|
|
delete rep_;
|
|
|
|
}
|
|
|
|
|
|
|
|
std::atomic<uint64_t> BlockBasedTable::next_cache_key_id_(0);
|
|
|
|
|
|
|
|
template <typename TBlocklike>
|
|
|
|
class BlocklikeTraits;
|
|
|
|
|
|
|
|
template <>
|
|
|
|
class BlocklikeTraits<BlockContents> {
|
|
|
|
public:
|
|
|
|
static BlockContents* Create(BlockContents&& contents,
|
|
|
|
SequenceNumber /* global_seqno */,
|
|
|
|
size_t /* read_amp_bytes_per_bit */,
|
|
|
|
Statistics* /* statistics */,
|
|
|
|
bool /* using_zstd */,
|
|
|
|
const FilterPolicy* /* filter_policy */) {
|
|
|
|
return new BlockContents(std::move(contents));
|
|
|
|
}
|
|
|
|
|
|
|
|
static uint32_t GetNumRestarts(const BlockContents& /* contents */) {
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
template <>
|
|
|
|
class BlocklikeTraits<ParsedFullFilterBlock> {
|
|
|
|
public:
|
|
|
|
static ParsedFullFilterBlock* Create(BlockContents&& contents,
|
|
|
|
SequenceNumber /* global_seqno */,
|
|
|
|
size_t /* read_amp_bytes_per_bit */,
|
|
|
|
Statistics* /* statistics */,
|
|
|
|
bool /* using_zstd */,
|
|
|
|
const FilterPolicy* filter_policy) {
|
|
|
|
return new ParsedFullFilterBlock(filter_policy, std::move(contents));
|
|
|
|
}
|
|
|
|
|
|
|
|
static uint32_t GetNumRestarts(const ParsedFullFilterBlock& /* block */) {
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
template <>
|
|
|
|
class BlocklikeTraits<Block> {
|
|
|
|
public:
|
|
|
|
static Block* Create(BlockContents&& contents, SequenceNumber global_seqno,
|
|
|
|
size_t read_amp_bytes_per_bit, Statistics* statistics,
|
|
|
|
bool /* using_zstd */,
|
|
|
|
const FilterPolicy* /* filter_policy */) {
|
|
|
|
return new Block(std::move(contents), global_seqno, read_amp_bytes_per_bit,
|
|
|
|
statistics);
|
|
|
|
}
|
|
|
|
|
|
|
|
static uint32_t GetNumRestarts(const Block& block) {
|
|
|
|
return block.NumRestarts();
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
template <>
|
|
|
|
class BlocklikeTraits<UncompressionDict> {
|
|
|
|
public:
|
|
|
|
static UncompressionDict* Create(BlockContents&& contents,
|
|
|
|
SequenceNumber /* global_seqno */,
|
|
|
|
size_t /* read_amp_bytes_per_bit */,
|
|
|
|
Statistics* /* statistics */,
|
|
|
|
bool using_zstd,
|
|
|
|
const FilterPolicy* /* filter_policy */) {
|
|
|
|
return new UncompressionDict(contents.data, std::move(contents.allocation),
|
|
|
|
using_zstd);
|
|
|
|
}
|
|
|
|
|
|
|
|
static uint32_t GetNumRestarts(const UncompressionDict& /* dict */) {
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
// Read the block identified by "handle" from "file".
|
|
|
|
// The only relevant option is options.verify_checksums for now.
|
|
|
|
// On failure return non-OK.
|
|
|
|
// On success fill *result and return OK - caller owns *result
|
|
|
|
// @param uncompression_dict Data for presetting the compression library's
|
Shared dictionary compression using reference block
Summary:
This adds a new metablock containing a shared dictionary that is used
to compress all data blocks in the SST file. The size of the shared dictionary
is configurable in CompressionOptions and defaults to 0. It's currently only
used for zlib/lz4/lz4hc, but the block will be stored in the SST regardless of
the compression type if the user chooses a nonzero dictionary size.
During compaction, computes the dictionary by randomly sampling the first
output file in each subcompaction. It pre-computes the intervals to sample
by assuming the output file will have the maximum allowable length. In case
the file is smaller, some of the pre-computed sampling intervals can be beyond
end-of-file, in which case we skip over those samples and the dictionary will
be a bit smaller. After the dictionary is generated using the first file in a
subcompaction, it is loaded into the compression library before writing each
block in each subsequent file of that subcompaction.
On the read path, gets the dictionary from the metablock, if it exists. Then,
loads that dictionary into the compression library before reading each block.
Test Plan: new unit test
Reviewers: yhchiang, IslamAbdelRahman, cyan, sdong
Reviewed By: sdong
Subscribers: andrewkr, yoshinorim, kradhakrishnan, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D52287
9 years ago
|
|
|
// dictionary.
|
|
|
|
template <typename TBlocklike>
|
|
|
|
Status ReadBlockFromFile(
|
|
|
|
RandomAccessFileReader* file, FilePrefetchBuffer* prefetch_buffer,
|
|
|
|
const Footer& footer, const ReadOptions& options, const BlockHandle& handle,
|
|
|
|
std::unique_ptr<TBlocklike>* result, const ImmutableCFOptions& ioptions,
|
|
|
|
bool do_uncompress, bool maybe_compressed, BlockType block_type,
|
|
|
|
const UncompressionDict& uncompression_dict,
|
|
|
|
const PersistentCacheOptions& cache_options, SequenceNumber global_seqno,
|
|
|
|
size_t read_amp_bytes_per_bit, MemoryAllocator* memory_allocator,
|
|
|
|
bool for_compaction, bool using_zstd, const FilterPolicy* filter_policy) {
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
assert(result);
|
|
|
|
|
|
|
|
BlockContents contents;
|
|
|
|
BlockFetcher block_fetcher(
|
|
|
|
file, prefetch_buffer, footer, options, handle, &contents, ioptions,
|
|
|
|
do_uncompress, maybe_compressed, block_type, uncompression_dict,
|
|
|
|
cache_options, memory_allocator, nullptr, for_compaction);
|
|
|
|
Status s = block_fetcher.ReadBlockContents();
|
|
|
|
if (s.ok()) {
|
|
|
|
result->reset(BlocklikeTraits<TBlocklike>::Create(
|
|
|
|
std::move(contents), global_seqno, read_amp_bytes_per_bit,
|
|
|
|
ioptions.statistics, using_zstd, filter_policy));
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
}
|
|
|
|
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
inline MemoryAllocator* GetMemoryAllocator(
|
|
|
|
const BlockBasedTableOptions& table_options) {
|
|
|
|
return table_options.block_cache.get()
|
|
|
|
? table_options.block_cache->memory_allocator()
|
|
|
|
: nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
inline MemoryAllocator* GetMemoryAllocatorForCompressedBlock(
|
|
|
|
const BlockBasedTableOptions& table_options) {
|
|
|
|
return table_options.block_cache_compressed.get()
|
|
|
|
? table_options.block_cache_compressed->memory_allocator()
|
|
|
|
: nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Delete the entry resided in the cache.
|
|
|
|
template <class Entry>
|
|
|
|
void DeleteCachedEntry(const Slice& /*key*/, void* value) {
|
|
|
|
auto entry = reinterpret_cast<Entry*>(value);
|
|
|
|
delete entry;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Release the cached entry and decrement its ref count.
|
|
|
|
void ForceReleaseCachedEntry(void* arg, void* h) {
|
|
|
|
Cache* cache = reinterpret_cast<Cache*>(arg);
|
|
|
|
Cache::Handle* handle = reinterpret_cast<Cache::Handle*>(h);
|
|
|
|
cache->Release(handle, true /* force_erase */);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Release the cached entry and decrement its ref count.
|
|
|
|
// Do not force erase
|
|
|
|
void ReleaseCachedEntry(void* arg, void* h) {
|
|
|
|
Cache* cache = reinterpret_cast<Cache*>(arg);
|
|
|
|
Cache::Handle* handle = reinterpret_cast<Cache::Handle*>(h);
|
|
|
|
cache->Release(handle, false /* force_erase */);
|
|
|
|
}
|
|
|
|
|
|
|
|
// For hash based index, return true if prefix_extractor and
|
|
|
|
// prefix_extractor_block mismatch, false otherwise. This flag will be used
|
|
|
|
// as total_order_seek via NewIndexIterator
|
|
|
|
bool PrefixExtractorChanged(const TableProperties* table_properties,
|
|
|
|
const SliceTransform* prefix_extractor) {
|
|
|
|
// BlockBasedTableOptions::kHashSearch requires prefix_extractor to be set.
|
|
|
|
// Turn off hash index in prefix_extractor is not set; if prefix_extractor
|
|
|
|
// is set but prefix_extractor_block is not set, also disable hash index
|
|
|
|
if (prefix_extractor == nullptr || table_properties == nullptr ||
|
|
|
|
table_properties->prefix_extractor_name.empty()) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
// prefix_extractor and prefix_extractor_block are both non-empty
|
|
|
|
if (table_properties->prefix_extractor_name.compare(
|
|
|
|
prefix_extractor->Name()) != 0) {
|
|
|
|
return true;
|
|
|
|
} else {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
CacheAllocationPtr CopyBufferToHeap(MemoryAllocator* allocator, Slice& buf) {
|
|
|
|
CacheAllocationPtr heap_buf;
|
|
|
|
heap_buf = AllocateBlock(buf.size(), allocator);
|
|
|
|
memcpy(heap_buf.get(), buf.data(), buf.size());
|
|
|
|
return heap_buf;
|
|
|
|
}
|
|
|
|
|
|
|
|
} // namespace
|
|
|
|
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
// Encapsulates common functionality for the various index reader
|
|
|
|
// implementations. Provides access to the index block regardless of whether
|
|
|
|
// it is owned by the reader or stored in the cache, or whether it is pinned
|
|
|
|
// in the cache or not.
|
|
|
|
class BlockBasedTable::IndexReaderCommon : public BlockBasedTable::IndexReader {
|
|
|
|
public:
|
|
|
|
IndexReaderCommon(const BlockBasedTable* t,
|
|
|
|
CachableEntry<Block>&& index_block)
|
|
|
|
: table_(t), index_block_(std::move(index_block)) {
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
assert(table_ != nullptr);
|
|
|
|
}
|
|
|
|
|
|
|
|
protected:
|
|
|
|
static Status ReadIndexBlock(const BlockBasedTable* table,
|
|
|
|
FilePrefetchBuffer* prefetch_buffer,
|
|
|
|
const ReadOptions& read_options, bool use_cache,
|
|
|
|
GetContext* get_context,
|
|
|
|
BlockCacheLookupContext* lookup_context,
|
|
|
|
CachableEntry<Block>* index_block);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
|
|
|
const BlockBasedTable* table() const { return table_; }
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
|
|
|
const InternalKeyComparator* internal_comparator() const {
|
|
|
|
assert(table_ != nullptr);
|
|
|
|
assert(table_->get_rep() != nullptr);
|
|
|
|
|
|
|
|
return &table_->get_rep()->internal_comparator;
|
|
|
|
}
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
bool index_has_first_key() const {
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
assert(table_ != nullptr);
|
|
|
|
assert(table_->get_rep() != nullptr);
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
return table_->get_rep()->index_has_first_key;
|
|
|
|
}
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
bool index_key_includes_seq() const {
|
|
|
|
assert(table_ != nullptr);
|
|
|
|
assert(table_->get_rep() != nullptr);
|
|
|
|
return table_->get_rep()->index_key_includes_seq;
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
}
|
|
|
|
|
|
|
|
bool index_value_is_full() const {
|
|
|
|
assert(table_ != nullptr);
|
|
|
|
assert(table_->get_rep() != nullptr);
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
return table_->get_rep()->index_value_is_full;
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
}
|
|
|
|
|
|
|
|
bool cache_index_blocks() const {
|
|
|
|
assert(table_ != nullptr);
|
|
|
|
assert(table_->get_rep() != nullptr);
|
|
|
|
return table_->get_rep()->table_options.cache_index_and_filter_blocks;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status GetOrReadIndexBlock(bool no_io, GetContext* get_context,
|
|
|
|
BlockCacheLookupContext* lookup_context,
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
CachableEntry<Block>* index_block) const;
|
|
|
|
|
|
|
|
size_t ApproximateIndexBlockMemoryUsage() const {
|
|
|
|
assert(!index_block_.GetOwnValue() || index_block_.GetValue() != nullptr);
|
|
|
|
return index_block_.GetOwnValue()
|
|
|
|
? index_block_.GetValue()->ApproximateMemoryUsage()
|
|
|
|
: 0;
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
}
|
|
|
|
|
|
|
|
private:
|
|
|
|
const BlockBasedTable* table_;
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
CachableEntry<Block> index_block_;
|
|
|
|
};
|
|
|
|
|
|
|
|
Status BlockBasedTable::IndexReaderCommon::ReadIndexBlock(
|
|
|
|
const BlockBasedTable* table, FilePrefetchBuffer* prefetch_buffer,
|
|
|
|
const ReadOptions& read_options, bool use_cache, GetContext* get_context,
|
|
|
|
BlockCacheLookupContext* lookup_context,
|
|
|
|
CachableEntry<Block>* index_block) {
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
PERF_TIMER_GUARD(read_index_block_nanos);
|
|
|
|
|
|
|
|
assert(table != nullptr);
|
|
|
|
assert(index_block != nullptr);
|
|
|
|
assert(index_block->IsEmpty());
|
|
|
|
|
|
|
|
const Rep* const rep = table->get_rep();
|
|
|
|
assert(rep != nullptr);
|
|
|
|
|
|
|
|
const Status s = table->RetrieveBlock(
|
|
|
|
prefetch_buffer, read_options, rep->footer.index_handle(),
|
|
|
|
UncompressionDict::GetEmptyDict(), index_block, BlockType::kIndex,
|
|
|
|
get_context, lookup_context, /* for_compaction */ false, use_cache);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::IndexReaderCommon::GetOrReadIndexBlock(
|
|
|
|
bool no_io, GetContext* get_context,
|
|
|
|
BlockCacheLookupContext* lookup_context,
|
|
|
|
CachableEntry<Block>* index_block) const {
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
assert(index_block != nullptr);
|
|
|
|
|
|
|
|
if (!index_block_.IsEmpty()) {
|
|
|
|
index_block->SetUnownedValue(index_block_.GetValue());
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
ReadOptions read_options;
|
|
|
|
if (no_io) {
|
|
|
|
read_options.read_tier = kBlockCacheTier;
|
|
|
|
}
|
|
|
|
|
|
|
|
return ReadIndexBlock(table_, /*prefetch_buffer=*/nullptr, read_options,
|
|
|
|
cache_index_blocks(), get_context, lookup_context,
|
|
|
|
index_block);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
}
|
|
|
|
|
|
|
|
// Index that allows binary search lookup in a two-level index structure.
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
class PartitionIndexReader : public BlockBasedTable::IndexReaderCommon {
|
|
|
|
public:
|
|
|
|
// Read the partition index from the file and create an instance for
|
|
|
|
// `PartitionIndexReader`.
|
|
|
|
// On success, index_reader will be populated; otherwise it will remain
|
|
|
|
// unmodified.
|
|
|
|
static Status Create(const BlockBasedTable* table,
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
FilePrefetchBuffer* prefetch_buffer, bool use_cache,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
bool prefetch, bool pin,
|
|
|
|
BlockCacheLookupContext* lookup_context,
|
|
|
|
std::unique_ptr<IndexReader>* index_reader) {
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
assert(table != nullptr);
|
|
|
|
assert(table->get_rep());
|
|
|
|
assert(!pin || prefetch);
|
|
|
|
assert(index_reader != nullptr);
|
|
|
|
|
|
|
|
CachableEntry<Block> index_block;
|
|
|
|
if (prefetch || !use_cache) {
|
|
|
|
const Status s =
|
|
|
|
ReadIndexBlock(table, prefetch_buffer, ReadOptions(), use_cache,
|
|
|
|
/*get_context=*/nullptr, lookup_context, &index_block);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
if (use_cache && !pin) {
|
|
|
|
index_block.Reset();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
index_reader->reset(
|
|
|
|
new PartitionIndexReader(table, std::move(index_block)));
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
// return a two-level iterator: first level is on the partition index
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
InternalIteratorBase<IndexValue>* NewIterator(
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
const ReadOptions& read_options, bool /* disable_prefix_seek */,
|
|
|
|
IndexBlockIter* iter, GetContext* get_context,
|
|
|
|
BlockCacheLookupContext* lookup_context) override {
|
|
|
|
const bool no_io = (read_options.read_tier == kBlockCacheTier);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
CachableEntry<Block> index_block;
|
|
|
|
const Status s =
|
|
|
|
GetOrReadIndexBlock(no_io, get_context, lookup_context, &index_block);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
if (!s.ok()) {
|
|
|
|
if (iter != nullptr) {
|
|
|
|
iter->Invalidate(s);
|
|
|
|
return iter;
|
|
|
|
}
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
return NewErrorInternalIterator<IndexValue>(s);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
}
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
InternalIteratorBase<IndexValue>* it = nullptr;
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
|
|
|
Statistics* kNullStats = nullptr;
|
|
|
|
// Filters are already checked before seeking the index
|
|
|
|
if (!partition_map_.empty()) {
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
// We don't return pinned data from index blocks, so no need
|
|
|
|
// to set `block_contents_pinned`.
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
it = NewTwoLevelIterator(
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
new BlockBasedTable::PartitionedIndexIteratorState(table(),
|
|
|
|
&partition_map_),
|
|
|
|
index_block.GetValue()->NewIndexIterator(
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
internal_comparator(), internal_comparator()->user_comparator(),
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
nullptr, kNullStats, true, index_has_first_key(),
|
|
|
|
index_key_includes_seq(), index_value_is_full()));
|
|
|
|
} else {
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
ReadOptions ro;
|
|
|
|
ro.fill_cache = read_options.fill_cache;
|
|
|
|
// We don't return pinned data from index blocks, so no need
|
|
|
|
// to set `block_contents_pinned`.
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
it = new BlockBasedTableIterator<IndexBlockIter, IndexValue>(
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
table(), ro, *internal_comparator(),
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
index_block.GetValue()->NewIndexIterator(
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
internal_comparator(), internal_comparator()->user_comparator(),
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
nullptr, kNullStats, true, index_has_first_key(),
|
|
|
|
index_key_includes_seq(), index_value_is_full()),
|
|
|
|
false, true, /* prefix_extractor */ nullptr, BlockType::kIndex,
|
|
|
|
lookup_context ? lookup_context->caller
|
|
|
|
: TableReaderCaller::kUncategorized);
|
|
|
|
}
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
|
|
|
assert(it != nullptr);
|
|
|
|
index_block.TransferTo(it);
|
|
|
|
|
|
|
|
return it;
|
|
|
|
|
|
|
|
// TODO(myabandeh): Update TwoLevelIterator to be able to make use of
|
|
|
|
// on-stack BlockIter while the state is on heap. Currentlly it assumes
|
|
|
|
// the first level iter is always on heap and will attempt to delete it
|
|
|
|
// in its destructor.
|
|
|
|
}
|
|
|
|
|
|
|
|
void CacheDependencies(bool pin) override {
|
|
|
|
// Before read partitions, prefetch them to avoid lots of IOs
|
|
|
|
BlockCacheLookupContext lookup_context{TableReaderCaller::kPrefetch};
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
const BlockBasedTable::Rep* rep = table()->rep_;
|
|
|
|
IndexBlockIter biter;
|
|
|
|
BlockHandle handle;
|
|
|
|
Statistics* kNullStats = nullptr;
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
|
|
|
CachableEntry<Block> index_block;
|
|
|
|
Status s = GetOrReadIndexBlock(false /* no_io */, nullptr /* get_context */,
|
|
|
|
&lookup_context, &index_block);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
if (!s.ok()) {
|
|
|
|
ROCKS_LOG_WARN(rep->ioptions.info_log,
|
|
|
|
"Error retrieving top-level index block while trying to "
|
|
|
|
"cache index partitions: %s",
|
|
|
|
s.ToString().c_str());
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
// We don't return pinned data from index blocks, so no need
|
|
|
|
// to set `block_contents_pinned`.
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
index_block.GetValue()->NewIndexIterator(
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
internal_comparator(), internal_comparator()->user_comparator(), &biter,
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
kNullStats, true, index_has_first_key(), index_key_includes_seq(),
|
|
|
|
index_value_is_full());
|
|
|
|
// Index partitions are assumed to be consecuitive. Prefetch them all.
|
|
|
|
// Read the first block offset
|
|
|
|
biter.SeekToFirst();
|
|
|
|
if (!biter.Valid()) {
|
|
|
|
// Empty index.
|
|
|
|
return;
|
|
|
|
}
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
handle = biter.value().handle;
|
|
|
|
uint64_t prefetch_off = handle.offset();
|
|
|
|
|
|
|
|
// Read the last block's offset
|
|
|
|
biter.SeekToLast();
|
|
|
|
if (!biter.Valid()) {
|
|
|
|
// Empty index.
|
|
|
|
return;
|
|
|
|
}
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
handle = biter.value().handle;
|
|
|
|
uint64_t last_off = handle.offset() + handle.size() + kBlockTrailerSize;
|
|
|
|
uint64_t prefetch_len = last_off - prefetch_off;
|
|
|
|
std::unique_ptr<FilePrefetchBuffer> prefetch_buffer;
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
auto& file = rep->file;
|
|
|
|
prefetch_buffer.reset(new FilePrefetchBuffer());
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
s = prefetch_buffer->Prefetch(file.get(), prefetch_off,
|
|
|
|
static_cast<size_t>(prefetch_len));
|
|
|
|
|
|
|
|
// After prefetch, read the partitions one by one
|
|
|
|
biter.SeekToFirst();
|
|
|
|
auto ro = ReadOptions();
|
|
|
|
for (; biter.Valid(); biter.Next()) {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
handle = biter.value().handle;
|
|
|
|
CachableEntry<Block> block;
|
|
|
|
// TODO: Support counter batch update for partitioned index and
|
|
|
|
// filter blocks
|
|
|
|
s = table()->MaybeReadBlockAndLoadToCache(
|
|
|
|
prefetch_buffer.get(), ro, handle, UncompressionDict::GetEmptyDict(),
|
|
|
|
&block, BlockType::kIndex, /*get_context=*/nullptr, &lookup_context,
|
|
|
|
/*contents=*/nullptr);
|
|
|
|
|
|
|
|
assert(s.ok() || block.GetValue() == nullptr);
|
|
|
|
if (s.ok() && block.GetValue() != nullptr) {
|
|
|
|
if (block.IsCached()) {
|
|
|
|
if (pin) {
|
|
|
|
partition_map_[handle.offset()] = std::move(block);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
size_t ApproximateMemoryUsage() const override {
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
size_t usage = ApproximateIndexBlockMemoryUsage();
|
|
|
|
#ifdef ROCKSDB_MALLOC_USABLE_SIZE
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
usage += malloc_usable_size(const_cast<PartitionIndexReader*>(this));
|
|
|
|
#else
|
|
|
|
usage += sizeof(*this);
|
|
|
|
#endif // ROCKSDB_MALLOC_USABLE_SIZE
|
|
|
|
// TODO(myabandeh): more accurate estimate of partition_map_ mem usage
|
|
|
|
return usage;
|
|
|
|
}
|
|
|
|
|
|
|
|
private:
|
|
|
|
PartitionIndexReader(const BlockBasedTable* t,
|
|
|
|
CachableEntry<Block>&& index_block)
|
|
|
|
: IndexReaderCommon(t, std::move(index_block)) {}
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
|
|
|
std::unordered_map<uint64_t, CachableEntry<Block>> partition_map_;
|
|
|
|
};
|
|
|
|
|
|
|
|
// Index that allows binary search lookup for the first key of each block.
|
|
|
|
// This class can be viewed as a thin wrapper for `Block` class which already
|
|
|
|
// supports binary search.
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
class BinarySearchIndexReader : public BlockBasedTable::IndexReaderCommon {
|
|
|
|
public:
|
|
|
|
// Read index from the file and create an intance for
|
|
|
|
// `BinarySearchIndexReader`.
|
|
|
|
// On success, index_reader will be populated; otherwise it will remain
|
|
|
|
// unmodified.
|
|
|
|
static Status Create(const BlockBasedTable* table,
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
FilePrefetchBuffer* prefetch_buffer, bool use_cache,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
bool prefetch, bool pin,
|
|
|
|
BlockCacheLookupContext* lookup_context,
|
|
|
|
std::unique_ptr<IndexReader>* index_reader) {
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
assert(table != nullptr);
|
|
|
|
assert(table->get_rep());
|
|
|
|
assert(!pin || prefetch);
|
|
|
|
assert(index_reader != nullptr);
|
|
|
|
|
|
|
|
CachableEntry<Block> index_block;
|
|
|
|
if (prefetch || !use_cache) {
|
|
|
|
const Status s =
|
|
|
|
ReadIndexBlock(table, prefetch_buffer, ReadOptions(), use_cache,
|
|
|
|
/*get_context=*/nullptr, lookup_context, &index_block);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
if (use_cache && !pin) {
|
|
|
|
index_block.Reset();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
index_reader->reset(
|
|
|
|
new BinarySearchIndexReader(table, std::move(index_block)));
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
InternalIteratorBase<IndexValue>* NewIterator(
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
const ReadOptions& read_options, bool /* disable_prefix_seek */,
|
|
|
|
IndexBlockIter* iter, GetContext* get_context,
|
|
|
|
BlockCacheLookupContext* lookup_context) override {
|
|
|
|
const bool no_io = (read_options.read_tier == kBlockCacheTier);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
CachableEntry<Block> index_block;
|
|
|
|
const Status s =
|
|
|
|
GetOrReadIndexBlock(no_io, get_context, lookup_context, &index_block);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
if (!s.ok()) {
|
|
|
|
if (iter != nullptr) {
|
|
|
|
iter->Invalidate(s);
|
|
|
|
return iter;
|
|
|
|
}
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
return NewErrorInternalIterator<IndexValue>(s);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
}
|
|
|
|
|
|
|
|
Statistics* kNullStats = nullptr;
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
// We don't return pinned data from index blocks, so no need
|
|
|
|
// to set `block_contents_pinned`.
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
auto it = index_block.GetValue()->NewIndexIterator(
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
internal_comparator(), internal_comparator()->user_comparator(), iter,
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
kNullStats, true, index_has_first_key(), index_key_includes_seq(),
|
|
|
|
index_value_is_full());
|
|
|
|
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
assert(it != nullptr);
|
|
|
|
index_block.TransferTo(it);
|
|
|
|
|
|
|
|
return it;
|
|
|
|
}
|
|
|
|
|
|
|
|
size_t ApproximateMemoryUsage() const override {
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
size_t usage = ApproximateIndexBlockMemoryUsage();
|
|
|
|
#ifdef ROCKSDB_MALLOC_USABLE_SIZE
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
usage += malloc_usable_size(const_cast<BinarySearchIndexReader*>(this));
|
|
|
|
#else
|
|
|
|
usage += sizeof(*this);
|
|
|
|
#endif // ROCKSDB_MALLOC_USABLE_SIZE
|
|
|
|
return usage;
|
|
|
|
}
|
|
|
|
|
|
|
|
private:
|
|
|
|
BinarySearchIndexReader(const BlockBasedTable* t,
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
CachableEntry<Block>&& index_block)
|
|
|
|
: IndexReaderCommon(t, std::move(index_block)) {}
|
|
|
|
};
|
|
|
|
|
|
|
|
// Index that leverages an internal hash table to quicken the lookup for a given
|
|
|
|
// key.
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
class HashIndexReader : public BlockBasedTable::IndexReaderCommon {
|
|
|
|
public:
|
|
|
|
static Status Create(const BlockBasedTable* table,
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
FilePrefetchBuffer* prefetch_buffer,
|
|
|
|
InternalIterator* meta_index_iter, bool use_cache,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
bool prefetch, bool pin,
|
|
|
|
BlockCacheLookupContext* lookup_context,
|
|
|
|
std::unique_ptr<IndexReader>* index_reader) {
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
assert(table != nullptr);
|
|
|
|
assert(index_reader != nullptr);
|
|
|
|
assert(!pin || prefetch);
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
const BlockBasedTable::Rep* rep = table->get_rep();
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
assert(rep != nullptr);
|
|
|
|
|
|
|
|
CachableEntry<Block> index_block;
|
|
|
|
if (prefetch || !use_cache) {
|
|
|
|
const Status s =
|
|
|
|
ReadIndexBlock(table, prefetch_buffer, ReadOptions(), use_cache,
|
|
|
|
/*get_context=*/nullptr, lookup_context, &index_block);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
if (use_cache && !pin) {
|
|
|
|
index_block.Reset();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Note, failure to create prefix hash index does not need to be a
|
|
|
|
// hard error. We can still fall back to the original binary search index.
|
|
|
|
// So, Create will succeed regardless, from this point on.
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
index_reader->reset(new HashIndexReader(table, std::move(index_block)));
|
|
|
|
|
|
|
|
// Get prefixes block
|
|
|
|
BlockHandle prefixes_handle;
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
Status s = FindMetaBlock(meta_index_iter, kHashIndexPrefixesBlock,
|
|
|
|
&prefixes_handle);
|
|
|
|
if (!s.ok()) {
|
|
|
|
// TODO: log error
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
// Get index metadata block
|
|
|
|
BlockHandle prefixes_meta_handle;
|
|
|
|
s = FindMetaBlock(meta_index_iter, kHashIndexPrefixesMetadataBlock,
|
|
|
|
&prefixes_meta_handle);
|
|
|
|
if (!s.ok()) {
|
|
|
|
// TODO: log error
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
RandomAccessFileReader* const file = rep->file.get();
|
|
|
|
const Footer& footer = rep->footer;
|
|
|
|
const ImmutableCFOptions& ioptions = rep->ioptions;
|
|
|
|
const PersistentCacheOptions& cache_options = rep->persistent_cache_options;
|
|
|
|
MemoryAllocator* const memory_allocator =
|
|
|
|
GetMemoryAllocator(rep->table_options);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
|
|
|
// Read contents for the blocks
|
|
|
|
BlockContents prefixes_contents;
|
|
|
|
BlockFetcher prefixes_block_fetcher(
|
|
|
|
file, prefetch_buffer, footer, ReadOptions(), prefixes_handle,
|
|
|
|
&prefixes_contents, ioptions, true /*decompress*/,
|
|
|
|
true /*maybe_compressed*/, BlockType::kHashIndexPrefixes,
|
|
|
|
UncompressionDict::GetEmptyDict(), cache_options, memory_allocator);
|
|
|
|
s = prefixes_block_fetcher.ReadBlockContents();
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
BlockContents prefixes_meta_contents;
|
|
|
|
BlockFetcher prefixes_meta_block_fetcher(
|
|
|
|
file, prefetch_buffer, footer, ReadOptions(), prefixes_meta_handle,
|
|
|
|
&prefixes_meta_contents, ioptions, true /*decompress*/,
|
|
|
|
true /*maybe_compressed*/, BlockType::kHashIndexMetadata,
|
|
|
|
UncompressionDict::GetEmptyDict(), cache_options, memory_allocator);
|
|
|
|
s = prefixes_meta_block_fetcher.ReadBlockContents();
|
|
|
|
if (!s.ok()) {
|
|
|
|
// TODO: log error
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
BlockPrefixIndex* prefix_index = nullptr;
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
s = BlockPrefixIndex::Create(rep->internal_prefix_transform.get(),
|
|
|
|
prefixes_contents.data,
|
|
|
|
prefixes_meta_contents.data, &prefix_index);
|
|
|
|
// TODO: log error
|
|
|
|
if (s.ok()) {
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
HashIndexReader* const hash_index_reader =
|
|
|
|
static_cast<HashIndexReader*>(index_reader->get());
|
|
|
|
hash_index_reader->prefix_index_.reset(prefix_index);
|
|
|
|
}
|
|
|
|
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
InternalIteratorBase<IndexValue>* NewIterator(
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
const ReadOptions& read_options, bool disable_prefix_seek,
|
|
|
|
IndexBlockIter* iter, GetContext* get_context,
|
|
|
|
BlockCacheLookupContext* lookup_context) override {
|
|
|
|
const bool no_io = (read_options.read_tier == kBlockCacheTier);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
CachableEntry<Block> index_block;
|
|
|
|
const Status s =
|
|
|
|
GetOrReadIndexBlock(no_io, get_context, lookup_context, &index_block);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
if (!s.ok()) {
|
|
|
|
if (iter != nullptr) {
|
|
|
|
iter->Invalidate(s);
|
|
|
|
return iter;
|
|
|
|
}
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
return NewErrorInternalIterator<IndexValue>(s);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
}
|
|
|
|
|
|
|
|
Statistics* kNullStats = nullptr;
|
|
|
|
const bool total_order_seek =
|
|
|
|
read_options.total_order_seek || disable_prefix_seek;
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
// We don't return pinned data from index blocks, so no need
|
|
|
|
// to set `block_contents_pinned`.
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
auto it = index_block.GetValue()->NewIndexIterator(
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
internal_comparator(), internal_comparator()->user_comparator(), iter,
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
kNullStats, total_order_seek, index_has_first_key(),
|
|
|
|
index_key_includes_seq(), index_value_is_full(),
|
|
|
|
false /* block_contents_pinned */, prefix_index_.get());
|
|
|
|
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
assert(it != nullptr);
|
|
|
|
index_block.TransferTo(it);
|
|
|
|
|
|
|
|
return it;
|
|
|
|
}
|
|
|
|
|
|
|
|
size_t ApproximateMemoryUsage() const override {
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
size_t usage = ApproximateIndexBlockMemoryUsage();
|
|
|
|
#ifdef ROCKSDB_MALLOC_USABLE_SIZE
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
usage += malloc_usable_size(const_cast<HashIndexReader*>(this));
|
|
|
|
#else
|
|
|
|
if (prefix_index_) {
|
|
|
|
usage += prefix_index_->ApproximateMemoryUsage();
|
|
|
|
}
|
|
|
|
usage += sizeof(*this);
|
|
|
|
#endif // ROCKSDB_MALLOC_USABLE_SIZE
|
|
|
|
return usage;
|
|
|
|
}
|
|
|
|
|
|
|
|
private:
|
|
|
|
HashIndexReader(const BlockBasedTable* t, CachableEntry<Block>&& index_block)
|
|
|
|
: IndexReaderCommon(t, std::move(index_block)) {}
|
|
|
|
|
|
|
|
std::unique_ptr<BlockPrefixIndex> prefix_index_;
|
|
|
|
};
|
|
|
|
|
|
|
|
void BlockBasedTable::UpdateCacheHitMetrics(BlockType block_type,
|
|
|
|
GetContext* get_context,
|
|
|
|
size_t usage) const {
|
|
|
|
Statistics* const statistics = rep_->ioptions.statistics;
|
|
|
|
|
|
|
|
PERF_COUNTER_ADD(block_cache_hit_count, 1);
|
|
|
|
PERF_COUNTER_BY_LEVEL_ADD(block_cache_hit_count, 1,
|
|
|
|
static_cast<uint32_t>(rep_->level));
|
|
|
|
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_hit;
|
|
|
|
get_context->get_context_stats_.num_cache_bytes_read += usage;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_HIT);
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_BYTES_READ, usage);
|
|
|
|
}
|
|
|
|
|
|
|
|
switch (block_type) {
|
|
|
|
case BlockType::kFilter:
|
|
|
|
PERF_COUNTER_ADD(block_cache_filter_hit_count, 1);
|
|
|
|
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_filter_hit;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_FILTER_HIT);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
case BlockType::kCompressionDictionary:
|
|
|
|
// TODO: introduce perf counter for compression dictionary hit count
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_compression_dict_hit;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_COMPRESSION_DICT_HIT);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
case BlockType::kIndex:
|
|
|
|
PERF_COUNTER_ADD(block_cache_index_hit_count, 1);
|
|
|
|
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_index_hit;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_INDEX_HIT);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
default:
|
|
|
|
// TODO: introduce dedicated tickers/statistics/counters
|
|
|
|
// for range tombstones
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_data_hit;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_DATA_HIT);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void BlockBasedTable::UpdateCacheMissMetrics(BlockType block_type,
|
|
|
|
GetContext* get_context) const {
|
|
|
|
Statistics* const statistics = rep_->ioptions.statistics;
|
|
|
|
|
|
|
|
// TODO: introduce aggregate (not per-level) block cache miss count
|
|
|
|
PERF_COUNTER_BY_LEVEL_ADD(block_cache_miss_count, 1,
|
|
|
|
static_cast<uint32_t>(rep_->level));
|
|
|
|
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_miss;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_MISS);
|
|
|
|
}
|
|
|
|
|
|
|
|
// TODO: introduce perf counters for misses per block type
|
|
|
|
switch (block_type) {
|
|
|
|
case BlockType::kFilter:
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_filter_miss;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_FILTER_MISS);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
case BlockType::kCompressionDictionary:
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_compression_dict_miss;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_COMPRESSION_DICT_MISS);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
case BlockType::kIndex:
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_index_miss;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_INDEX_MISS);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
default:
|
|
|
|
// TODO: introduce dedicated tickers/statistics/counters
|
|
|
|
// for range tombstones
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_data_miss;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_DATA_MISS);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void BlockBasedTable::UpdateCacheInsertionMetrics(BlockType block_type,
|
|
|
|
GetContext* get_context,
|
|
|
|
size_t usage) const {
|
|
|
|
Statistics* const statistics = rep_->ioptions.statistics;
|
|
|
|
|
|
|
|
// TODO: introduce perf counters for block cache insertions
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_add;
|
|
|
|
get_context->get_context_stats_.num_cache_bytes_write += usage;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_ADD);
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_BYTES_WRITE, usage);
|
|
|
|
}
|
|
|
|
|
|
|
|
switch (block_type) {
|
|
|
|
case BlockType::kFilter:
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_filter_add;
|
|
|
|
get_context->get_context_stats_.num_cache_filter_bytes_insert += usage;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_FILTER_ADD);
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_FILTER_BYTES_INSERT, usage);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
case BlockType::kCompressionDictionary:
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_compression_dict_add;
|
|
|
|
get_context->get_context_stats_
|
|
|
|
.num_cache_compression_dict_bytes_insert += usage;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_COMPRESSION_DICT_ADD);
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_COMPRESSION_DICT_BYTES_INSERT,
|
|
|
|
usage);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
case BlockType::kIndex:
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_index_add;
|
|
|
|
get_context->get_context_stats_.num_cache_index_bytes_insert += usage;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_INDEX_ADD);
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_INDEX_BYTES_INSERT, usage);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
default:
|
|
|
|
// TODO: introduce dedicated tickers/statistics/counters
|
|
|
|
// for range tombstones
|
|
|
|
if (get_context) {
|
|
|
|
++get_context->get_context_stats_.num_cache_data_add;
|
|
|
|
get_context->get_context_stats_.num_cache_data_bytes_insert += usage;
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_DATA_ADD);
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_DATA_BYTES_INSERT, usage);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
Cache::Handle* BlockBasedTable::GetEntryFromCache(
|
|
|
|
Cache* block_cache, const Slice& key, BlockType block_type,
|
|
|
|
GetContext* get_context) const {
|
|
|
|
auto cache_handle = block_cache->Lookup(key, rep_->ioptions.statistics);
|
|
|
|
|
|
|
|
if (cache_handle != nullptr) {
|
|
|
|
UpdateCacheHitMetrics(block_type, get_context,
|
|
|
|
block_cache->GetUsage(cache_handle));
|
|
|
|
} else {
|
|
|
|
UpdateCacheMissMetrics(block_type, get_context);
|
|
|
|
}
|
|
|
|
|
|
|
|
return cache_handle;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Helper function to setup the cache key's prefix for the Table.
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
void BlockBasedTable::SetupCacheKeyPrefix(Rep* rep) {
|
|
|
|
assert(kMaxCacheKeyPrefixSize >= 10);
|
|
|
|
rep->cache_key_prefix_size = 0;
|
|
|
|
rep->compressed_cache_key_prefix_size = 0;
|
|
|
|
if (rep->table_options.block_cache != nullptr) {
|
Move rate_limiter, write buffering, most perf context instrumentation and most random kill out of Env
Summary: We want to keep Env a think layer for better portability. Less platform dependent codes should be moved out of Env. In this patch, I create a wrapper of file readers and writers, and put rate limiting, write buffering, as well as most perf context instrumentation and random kill out of Env. It will make it easier to maintain multiple Env in the future.
Test Plan: Run all existing unit tests.
Reviewers: anthony, kradhakrishnan, IslamAbdelRahman, yhchiang, igor
Reviewed By: igor
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D42321
10 years ago
|
|
|
GenerateCachePrefix(rep->table_options.block_cache.get(), rep->file->file(),
|
|
|
|
&rep->cache_key_prefix[0], &rep->cache_key_prefix_size);
|
|
|
|
}
|
|
|
|
if (rep->table_options.persistent_cache != nullptr) {
|
|
|
|
GenerateCachePrefix(/*cache=*/nullptr, rep->file->file(),
|
|
|
|
&rep->persistent_cache_key_prefix[0],
|
|
|
|
&rep->persistent_cache_key_prefix_size);
|
|
|
|
}
|
|
|
|
if (rep->table_options.block_cache_compressed != nullptr) {
|
|
|
|
GenerateCachePrefix(rep->table_options.block_cache_compressed.get(),
|
Move rate_limiter, write buffering, most perf context instrumentation and most random kill out of Env
Summary: We want to keep Env a think layer for better portability. Less platform dependent codes should be moved out of Env. In this patch, I create a wrapper of file readers and writers, and put rate limiting, write buffering, as well as most perf context instrumentation and random kill out of Env. It will make it easier to maintain multiple Env in the future.
Test Plan: Run all existing unit tests.
Reviewers: anthony, kradhakrishnan, IslamAbdelRahman, yhchiang, igor
Reviewed By: igor
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D42321
10 years ago
|
|
|
rep->file->file(), &rep->compressed_cache_key_prefix[0],
|
|
|
|
&rep->compressed_cache_key_prefix_size);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void BlockBasedTable::GenerateCachePrefix(Cache* cc, RandomAccessFile* file,
|
|
|
|
char* buffer, size_t* size) {
|
|
|
|
// generate an id from the file
|
|
|
|
*size = file->GetUniqueId(buffer, kMaxCacheKeyPrefixSize);
|
|
|
|
|
|
|
|
// If the prefix wasn't generated or was too long,
|
|
|
|
// create one from the cache.
|
|
|
|
if (cc != nullptr && *size == 0) {
|
|
|
|
char* end = EncodeVarint64(buffer, cc->NewId());
|
|
|
|
*size = static_cast<size_t>(end - buffer);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void BlockBasedTable::GenerateCachePrefix(Cache* cc, WritableFile* file,
|
|
|
|
char* buffer, size_t* size) {
|
|
|
|
// generate an id from the file
|
|
|
|
*size = file->GetUniqueId(buffer, kMaxCacheKeyPrefixSize);
|
|
|
|
|
|
|
|
// If the prefix wasn't generated or was too long,
|
|
|
|
// create one from the cache.
|
|
|
|
if (cc != nullptr && *size == 0) {
|
|
|
|
char* end = EncodeVarint64(buffer, cc->NewId());
|
|
|
|
*size = static_cast<size_t>(end - buffer);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
// Return True if table_properties has `user_prop_name` has a `true` value
|
|
|
|
// or it doesn't contain this property (for backward compatible).
|
|
|
|
bool IsFeatureSupported(const TableProperties& table_properties,
|
|
|
|
const std::string& user_prop_name, Logger* info_log) {
|
|
|
|
auto& props = table_properties.user_collected_properties;
|
|
|
|
auto pos = props.find(user_prop_name);
|
|
|
|
// Older version doesn't have this value set. Skip this check.
|
|
|
|
if (pos != props.end()) {
|
|
|
|
if (pos->second == kPropFalse) {
|
|
|
|
return false;
|
|
|
|
} else if (pos->second != kPropTrue) {
|
|
|
|
ROCKS_LOG_WARN(info_log, "Property %s has invalidate value %s",
|
|
|
|
user_prop_name.c_str(), pos->second.c_str());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Caller has to ensure seqno is not nullptr.
|
|
|
|
Status GetGlobalSequenceNumber(const TableProperties& table_properties,
|
|
|
|
SequenceNumber largest_seqno,
|
|
|
|
SequenceNumber* seqno) {
|
|
|
|
const auto& props = table_properties.user_collected_properties;
|
|
|
|
const auto version_pos = props.find(ExternalSstFilePropertyNames::kVersion);
|
|
|
|
const auto seqno_pos = props.find(ExternalSstFilePropertyNames::kGlobalSeqno);
|
|
|
|
|
|
|
|
*seqno = kDisableGlobalSequenceNumber;
|
|
|
|
if (version_pos == props.end()) {
|
|
|
|
if (seqno_pos != props.end()) {
|
|
|
|
std::array<char, 200> msg_buf;
|
|
|
|
// This is not an external sst file, global_seqno is not supported.
|
|
|
|
snprintf(
|
|
|
|
msg_buf.data(), msg_buf.max_size(),
|
|
|
|
"A non-external sst file have global seqno property with value %s",
|
|
|
|
seqno_pos->second.c_str());
|
|
|
|
return Status::Corruption(msg_buf.data());
|
|
|
|
}
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
uint32_t version = DecodeFixed32(version_pos->second.c_str());
|
|
|
|
if (version < 2) {
|
|
|
|
if (seqno_pos != props.end() || version != 1) {
|
|
|
|
std::array<char, 200> msg_buf;
|
|
|
|
// This is a v1 external sst file, global_seqno is not supported.
|
|
|
|
snprintf(msg_buf.data(), msg_buf.max_size(),
|
|
|
|
"An external sst file with version %u have global seqno "
|
|
|
|
"property with value %s",
|
|
|
|
version, seqno_pos->second.c_str());
|
|
|
|
return Status::Corruption(msg_buf.data());
|
|
|
|
}
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
// Since we have a plan to deprecate global_seqno, we do not return failure
|
|
|
|
// if seqno_pos == props.end(). We rely on version_pos to detect whether the
|
|
|
|
// SST is external.
|
|
|
|
SequenceNumber global_seqno(0);
|
|
|
|
if (seqno_pos != props.end()) {
|
|
|
|
global_seqno = DecodeFixed64(seqno_pos->second.c_str());
|
|
|
|
}
|
|
|
|
// SstTableReader open table reader with kMaxSequenceNumber as largest_seqno
|
|
|
|
// to denote it is unknown.
|
|
|
|
if (largest_seqno < kMaxSequenceNumber) {
|
|
|
|
if (global_seqno == 0) {
|
|
|
|
global_seqno = largest_seqno;
|
|
|
|
}
|
|
|
|
if (global_seqno != largest_seqno) {
|
|
|
|
std::array<char, 200> msg_buf;
|
|
|
|
snprintf(
|
|
|
|
msg_buf.data(), msg_buf.max_size(),
|
|
|
|
"An external sst file with version %u have global seqno property "
|
|
|
|
"with value %s, while largest seqno in the file is %llu",
|
|
|
|
version, seqno_pos->second.c_str(),
|
|
|
|
static_cast<unsigned long long>(largest_seqno));
|
|
|
|
return Status::Corruption(msg_buf.data());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
*seqno = global_seqno;
|
|
|
|
|
|
|
|
if (global_seqno > kMaxSequenceNumber) {
|
|
|
|
std::array<char, 200> msg_buf;
|
|
|
|
snprintf(msg_buf.data(), msg_buf.max_size(),
|
|
|
|
"An external sst file with version %u have global seqno property "
|
|
|
|
"with value %llu, which is greater than kMaxSequenceNumber",
|
|
|
|
version, static_cast<unsigned long long>(global_seqno));
|
|
|
|
return Status::Corruption(msg_buf.data());
|
|
|
|
}
|
|
|
|
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
} // namespace
|
|
|
|
|
|
|
|
Slice BlockBasedTable::GetCacheKey(const char* cache_key_prefix,
|
|
|
|
size_t cache_key_prefix_size,
|
|
|
|
const BlockHandle& handle, char* cache_key) {
|
|
|
|
assert(cache_key != nullptr);
|
|
|
|
assert(cache_key_prefix_size != 0);
|
|
|
|
assert(cache_key_prefix_size <= kMaxCacheKeyPrefixSize);
|
|
|
|
memcpy(cache_key, cache_key_prefix, cache_key_prefix_size);
|
|
|
|
char* end =
|
|
|
|
EncodeVarint64(cache_key + cache_key_prefix_size, handle.offset());
|
|
|
|
return Slice(cache_key, static_cast<size_t>(end - cache_key));
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::Open(
|
|
|
|
const ImmutableCFOptions& ioptions, const EnvOptions& env_options,
|
|
|
|
const BlockBasedTableOptions& table_options,
|
|
|
|
const InternalKeyComparator& internal_comparator,
|
|
|
|
std::unique_ptr<RandomAccessFileReader>&& file, uint64_t file_size,
|
|
|
|
std::unique_ptr<TableReader>* table_reader,
|
|
|
|
const SliceTransform* prefix_extractor,
|
|
|
|
const bool prefetch_index_and_filter_in_cache, const bool skip_filters,
|
|
|
|
const int level, const bool immortal_table,
|
|
|
|
const SequenceNumber largest_seqno, TailPrefetchStats* tail_prefetch_stats,
|
|
|
|
BlockCacheTracer* const block_cache_tracer) {
|
|
|
|
table_reader->reset();
|
|
|
|
|
|
|
|
Status s;
|
|
|
|
Footer footer;
|
|
|
|
std::unique_ptr<FilePrefetchBuffer> prefetch_buffer;
|
|
|
|
|
|
|
|
// prefetch both index and filters, down to all partitions
|
|
|
|
const bool prefetch_all = prefetch_index_and_filter_in_cache || level == 0;
|
|
|
|
const bool preload_all = !table_options.cache_index_and_filter_blocks;
|
|
|
|
|
|
|
|
s = PrefetchTail(file.get(), file_size, tail_prefetch_stats, prefetch_all,
|
|
|
|
preload_all, &prefetch_buffer);
|
|
|
|
|
|
|
|
// Read in the following order:
|
|
|
|
// 1. Footer
|
|
|
|
// 2. [metaindex block]
|
|
|
|
// 3. [meta block: properties]
|
|
|
|
// 4. [meta block: range deletion tombstone]
|
|
|
|
// 5. [meta block: compression dictionary]
|
|
|
|
// 6. [meta block: index]
|
|
|
|
// 7. [meta block: filter]
|
|
|
|
s = ReadFooterFromFile(file.get(), prefetch_buffer.get(), file_size, &footer,
|
|
|
|
kBlockBasedTableMagicNumber);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
if (!BlockBasedTableSupportedVersion(footer.version())) {
|
|
|
|
return Status::Corruption(
|
|
|
|
"Unknown Footer version. Maybe this file was created with newer "
|
|
|
|
"version of RocksDB?");
|
|
|
|
}
|
|
|
|
|
|
|
|
// We've successfully read the footer. We are ready to serve requests.
|
|
|
|
// Better not mutate rep_ after the creation. eg. internal_prefix_transform
|
|
|
|
// raw pointer will be used to create HashIndexReader, whose reset may
|
|
|
|
// access a dangling pointer.
|
|
|
|
BlockCacheLookupContext lookup_context{TableReaderCaller::kPrefetch};
|
Skip bottom-level filter block caching when hit-optimized
Summary:
When Get() or NewIterator() trigger file loads, skip caching the filter block if
(1) optimize_filters_for_hits is set and (2) the file is on the bottommost
level. Also skip checking filters under the same conditions, which means that
for a preloaded file or a file that was trivially-moved to the bottom level, its
filter block will eventually expire from the cache.
- added parameters/instance variables in various places in order to propagate the config ("skip_filters") from version_set to block_based_table_reader
- in BlockBasedTable::Rep, this optimization prevents filter from being loaded when the file is opened simply by setting filter_policy = nullptr
- in BlockBasedTable::Get/BlockBasedTable::NewIterator, this optimization prevents filter from being used (even if it was loaded already) by setting filter = nullptr
Test Plan:
updated unit test:
$ ./db_test --gtest_filter=DBTest.OptimizeFiltersForHits
will also run 'make check'
Reviewers: sdong, igor, paultuckfield, anthony, rven, kradhakrishnan, IslamAbdelRahman, yhchiang
Reviewed By: yhchiang
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D51633
9 years ago
|
|
|
Rep* rep = new BlockBasedTable::Rep(ioptions, env_options, table_options,
|
|
|
|
internal_comparator, skip_filters, level,
|
|
|
|
immortal_table);
|
|
|
|
rep->file = std::move(file);
|
|
|
|
rep->footer = footer;
|
|
|
|
rep->hash_index_allow_collision = table_options.hash_index_allow_collision;
|
|
|
|
// We need to wrap data with internal_prefix_transform to make sure it can
|
|
|
|
// handle prefix correctly.
|
|
|
|
rep->internal_prefix_transform.reset(
|
|
|
|
new InternalKeySliceTransform(prefix_extractor));
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
SetupCacheKeyPrefix(rep);
|
|
|
|
std::unique_ptr<BlockBasedTable> new_table(
|
|
|
|
new BlockBasedTable(rep, block_cache_tracer));
|
|
|
|
|
|
|
|
// page cache options
|
|
|
|
rep->persistent_cache_options =
|
|
|
|
PersistentCacheOptions(rep->table_options.persistent_cache,
|
|
|
|
std::string(rep->persistent_cache_key_prefix,
|
|
|
|
rep->persistent_cache_key_prefix_size),
|
|
|
|
rep->ioptions.statistics);
|
|
|
|
|
|
|
|
// Meta-blocks are not dictionary compressed. Explicitly set the dictionary
|
|
|
|
// handle to null, otherwise it may be seen as uninitialized during the below
|
|
|
|
// meta-block reads.
|
|
|
|
rep->compression_dict_handle = BlockHandle::NullBlockHandle();
|
|
|
|
|
|
|
|
// Read metaindex
|
|
|
|
std::unique_ptr<Block> meta;
|
|
|
|
std::unique_ptr<InternalIterator> meta_iter;
|
|
|
|
s = new_table->ReadMetaBlock(prefetch_buffer.get(), &meta, &meta_iter);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
// Populates table_properties and some fields that depend on it,
|
|
|
|
// such as index_type.
|
|
|
|
s = new_table->ReadPropertiesBlock(prefetch_buffer.get(), meta_iter.get(),
|
|
|
|
largest_seqno);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
s = new_table->ReadRangeDelBlock(prefetch_buffer.get(), meta_iter.get(),
|
|
|
|
internal_comparator, &lookup_context);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
s = new_table->PrefetchIndexAndFilterBlocks(
|
|
|
|
prefetch_buffer.get(), meta_iter.get(), new_table.get(), prefetch_all,
|
|
|
|
table_options, level, &lookup_context);
|
|
|
|
|
|
|
|
if (s.ok()) {
|
|
|
|
// Update tail prefetch stats
|
|
|
|
assert(prefetch_buffer.get() != nullptr);
|
|
|
|
if (tail_prefetch_stats != nullptr) {
|
|
|
|
assert(prefetch_buffer->min_offset_read() < file_size);
|
|
|
|
tail_prefetch_stats->RecordEffectiveSize(
|
|
|
|
static_cast<size_t>(file_size) - prefetch_buffer->min_offset_read());
|
|
|
|
}
|
|
|
|
|
|
|
|
*table_reader = std::move(new_table);
|
|
|
|
}
|
|
|
|
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::PrefetchTail(
|
|
|
|
RandomAccessFileReader* file, uint64_t file_size,
|
|
|
|
TailPrefetchStats* tail_prefetch_stats, const bool prefetch_all,
|
|
|
|
const bool preload_all,
|
|
|
|
std::unique_ptr<FilePrefetchBuffer>* prefetch_buffer) {
|
|
|
|
size_t tail_prefetch_size = 0;
|
|
|
|
if (tail_prefetch_stats != nullptr) {
|
|
|
|
// Multiple threads may get a 0 (no history) when running in parallel,
|
|
|
|
// but it will get cleared after the first of them finishes.
|
|
|
|
tail_prefetch_size = tail_prefetch_stats->GetSuggestedPrefetchSize();
|
|
|
|
}
|
|
|
|
if (tail_prefetch_size == 0) {
|
|
|
|
// Before read footer, readahead backwards to prefetch data. Do more
|
|
|
|
// readahead if we're going to read index/filter.
|
|
|
|
// TODO: This may incorrectly select small readahead in case partitioned
|
|
|
|
// index/filter is enabled and top-level partition pinning is enabled.
|
|
|
|
// That's because we need to issue readahead before we read the properties,
|
|
|
|
// at which point we don't yet know the index type.
|
|
|
|
tail_prefetch_size = prefetch_all || preload_all ? 512 * 1024 : 4 * 1024;
|
|
|
|
}
|
|
|
|
size_t prefetch_off;
|
|
|
|
size_t prefetch_len;
|
|
|
|
if (file_size < tail_prefetch_size) {
|
|
|
|
prefetch_off = 0;
|
|
|
|
prefetch_len = static_cast<size_t>(file_size);
|
|
|
|
} else {
|
|
|
|
prefetch_off = static_cast<size_t>(file_size - tail_prefetch_size);
|
|
|
|
prefetch_len = tail_prefetch_size;
|
|
|
|
}
|
|
|
|
TEST_SYNC_POINT_CALLBACK("BlockBasedTable::Open::TailPrefetchLen",
|
|
|
|
&tail_prefetch_size);
|
|
|
|
Status s;
|
|
|
|
// TODO should not have this special logic in the future.
|
|
|
|
if (!file->use_direct_io()) {
|
|
|
|
prefetch_buffer->reset(new FilePrefetchBuffer(nullptr, 0, 0, false, true));
|
|
|
|
s = file->Prefetch(prefetch_off, prefetch_len);
|
|
|
|
} else {
|
|
|
|
prefetch_buffer->reset(new FilePrefetchBuffer(nullptr, 0, 0, true, true));
|
|
|
|
s = (*prefetch_buffer)->Prefetch(file, prefetch_off, prefetch_len);
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status VerifyChecksum(const ChecksumType type, const char* buf, size_t len,
|
|
|
|
uint32_t expected) {
|
|
|
|
Status s;
|
|
|
|
uint32_t actual = 0;
|
|
|
|
switch (type) {
|
|
|
|
case kNoChecksum:
|
|
|
|
break;
|
|
|
|
case kCRC32c:
|
|
|
|
expected = crc32c::Unmask(expected);
|
|
|
|
actual = crc32c::Value(buf, len);
|
|
|
|
break;
|
|
|
|
case kxxHash:
|
|
|
|
actual = XXH32(buf, static_cast<int>(len), 0);
|
|
|
|
break;
|
|
|
|
case kxxHash64:
|
|
|
|
actual = static_cast<uint32_t>(XXH64(buf, static_cast<int>(len), 0) &
|
|
|
|
uint64_t{0xffffffff});
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
s = Status::Corruption("unknown checksum type");
|
|
|
|
}
|
|
|
|
if (s.ok() && actual != expected) {
|
|
|
|
s = Status::Corruption("properties block checksum mismatched");
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::TryReadPropertiesWithGlobalSeqno(
|
|
|
|
FilePrefetchBuffer* prefetch_buffer, const Slice& handle_value,
|
|
|
|
TableProperties** table_properties) {
|
|
|
|
assert(table_properties != nullptr);
|
|
|
|
// If this is an external SST file ingested with write_global_seqno set to
|
|
|
|
// true, then we expect the checksum mismatch because checksum was written
|
|
|
|
// by SstFileWriter, but its global seqno in the properties block may have
|
|
|
|
// been changed during ingestion. In this case, we read the properties
|
|
|
|
// block, copy it to a memory buffer, change the global seqno to its
|
|
|
|
// original value, i.e. 0, and verify the checksum again.
|
|
|
|
BlockHandle props_block_handle;
|
|
|
|
CacheAllocationPtr tmp_buf;
|
|
|
|
Status s = ReadProperties(handle_value, rep_->file.get(), prefetch_buffer,
|
|
|
|
rep_->footer, rep_->ioptions, table_properties,
|
|
|
|
false /* verify_checksum */, &props_block_handle,
|
|
|
|
&tmp_buf, false /* compression_type_missing */,
|
|
|
|
nullptr /* memory_allocator */);
|
|
|
|
if (s.ok() && tmp_buf) {
|
|
|
|
const auto seqno_pos_iter =
|
|
|
|
(*table_properties)
|
|
|
|
->properties_offsets.find(
|
|
|
|
ExternalSstFilePropertyNames::kGlobalSeqno);
|
|
|
|
size_t block_size = static_cast<size_t>(props_block_handle.size());
|
|
|
|
if (seqno_pos_iter != (*table_properties)->properties_offsets.end()) {
|
|
|
|
uint64_t global_seqno_offset = seqno_pos_iter->second;
|
|
|
|
EncodeFixed64(
|
|
|
|
tmp_buf.get() + global_seqno_offset - props_block_handle.offset(), 0);
|
|
|
|
}
|
|
|
|
uint32_t value = DecodeFixed32(tmp_buf.get() + block_size + 1);
|
|
|
|
s = rocksdb::VerifyChecksum(rep_->footer.checksum(), tmp_buf.get(),
|
|
|
|
block_size + 1, value);
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::ReadPropertiesBlock(
|
|
|
|
FilePrefetchBuffer* prefetch_buffer, InternalIterator* meta_iter,
|
|
|
|
const SequenceNumber largest_seqno) {
|
RocksDB 2.8 to be able to read files generated by 2.6
Summary:
From 2.6 to 2.7, property block name is renamed from rocksdb.stats to rocksdb.properties. Older properties were not able to be loaded. In 2.8, we seem to have added some logic that uses property block without checking null pointers, which create segment faults.
In this patch, we fix it by:
(1) try rocksdb.stats if rocksdb.properties is not found
(2) add some null checking before consuming rep->table_properties
Test Plan: make sure a file generated in 2.7 couldn't be opened now can be opened.
Reviewers: haobo, igor, yhchiang
Reviewed By: igor
CC: ljin, xjin, dhruba, kailiu, leveldb
Differential Revision: https://reviews.facebook.net/D17961
11 years ago
|
|
|
bool found_properties_block = true;
|
|
|
|
Status s;
|
|
|
|
s = SeekToPropertiesBlock(meta_iter, &found_properties_block);
|
RocksDB 2.8 to be able to read files generated by 2.6
Summary:
From 2.6 to 2.7, property block name is renamed from rocksdb.stats to rocksdb.properties. Older properties were not able to be loaded. In 2.8, we seem to have added some logic that uses property block without checking null pointers, which create segment faults.
In this patch, we fix it by:
(1) try rocksdb.stats if rocksdb.properties is not found
(2) add some null checking before consuming rep->table_properties
Test Plan: make sure a file generated in 2.7 couldn't be opened now can be opened.
Reviewers: haobo, igor, yhchiang
Reviewed By: igor
CC: ljin, xjin, dhruba, kailiu, leveldb
Differential Revision: https://reviews.facebook.net/D17961
11 years ago
|
|
|
|
|
|
|
if (!s.ok()) {
|
|
|
|
ROCKS_LOG_WARN(rep_->ioptions.info_log,
|
|
|
|
"Error when seeking to properties block from file: %s",
|
|
|
|
s.ToString().c_str());
|
|
|
|
} else if (found_properties_block) {
|
|
|
|
s = meta_iter->status();
|
|
|
|
TableProperties* table_properties = nullptr;
|
|
|
|
if (s.ok()) {
|
|
|
|
s = ReadProperties(
|
|
|
|
meta_iter->value(), rep_->file.get(), prefetch_buffer, rep_->footer,
|
|
|
|
rep_->ioptions, &table_properties, true /* verify_checksum */,
|
|
|
|
nullptr /* ret_block_handle */, nullptr /* ret_block_contents */,
|
|
|
|
false /* compression_type_missing */, nullptr /* memory_allocator */);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (s.IsCorruption()) {
|
|
|
|
s = TryReadPropertiesWithGlobalSeqno(prefetch_buffer, meta_iter->value(),
|
|
|
|
&table_properties);
|
|
|
|
}
|
|
|
|
std::unique_ptr<TableProperties> props_guard;
|
|
|
|
if (table_properties != nullptr) {
|
|
|
|
props_guard.reset(table_properties);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!s.ok()) {
|
|
|
|
ROCKS_LOG_WARN(rep_->ioptions.info_log,
|
|
|
|
"Encountered error while reading data from properties "
|
|
|
|
"block %s",
|
|
|
|
s.ToString().c_str());
|
|
|
|
} else {
|
|
|
|
assert(table_properties != nullptr);
|
|
|
|
rep_->table_properties.reset(props_guard.release());
|
|
|
|
rep_->blocks_maybe_compressed =
|
|
|
|
rep_->table_properties->compression_name !=
|
|
|
|
CompressionTypeToString(kNoCompression);
|
|
|
|
rep_->blocks_definitely_zstd_compressed =
|
|
|
|
(rep_->table_properties->compression_name ==
|
|
|
|
CompressionTypeToString(kZSTD) ||
|
|
|
|
rep_->table_properties->compression_name ==
|
|
|
|
CompressionTypeToString(kZSTDNotFinalCompression));
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
ROCKS_LOG_ERROR(rep_->ioptions.info_log,
|
|
|
|
"Cannot find Properties block from file.");
|
|
|
|
}
|
|
|
|
#ifndef ROCKSDB_LITE
|
|
|
|
if (rep_->table_properties) {
|
|
|
|
ParseSliceTransform(rep_->table_properties->prefix_extractor_name,
|
|
|
|
&(rep_->table_prefix_extractor));
|
|
|
|
}
|
|
|
|
#endif // ROCKSDB_LITE
|
|
|
|
|
|
|
|
// Read the table properties, if provided.
|
|
|
|
if (rep_->table_properties) {
|
|
|
|
rep_->whole_key_filtering &=
|
|
|
|
IsFeatureSupported(*(rep_->table_properties),
|
|
|
|
BlockBasedTablePropertyNames::kWholeKeyFiltering,
|
|
|
|
rep_->ioptions.info_log);
|
|
|
|
rep_->prefix_filtering &=
|
|
|
|
IsFeatureSupported(*(rep_->table_properties),
|
|
|
|
BlockBasedTablePropertyNames::kPrefixFiltering,
|
|
|
|
rep_->ioptions.info_log);
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
rep_->index_key_includes_seq =
|
|
|
|
rep_->table_properties->index_key_is_user_key == 0;
|
|
|
|
rep_->index_value_is_full =
|
|
|
|
rep_->table_properties->index_value_is_delta_encoded == 0;
|
|
|
|
|
|
|
|
// Update index_type with the true type.
|
|
|
|
// If table properties don't contain index type, we assume that the table
|
|
|
|
// is in very old format and has kBinarySearch index type.
|
|
|
|
auto& props = rep_->table_properties->user_collected_properties;
|
|
|
|
auto pos = props.find(BlockBasedTablePropertyNames::kIndexType);
|
|
|
|
if (pos != props.end()) {
|
|
|
|
rep_->index_type = static_cast<BlockBasedTableOptions::IndexType>(
|
|
|
|
DecodeFixed32(pos->second.c_str()));
|
|
|
|
}
|
|
|
|
|
|
|
|
rep_->index_has_first_key =
|
|
|
|
rep_->index_type == BlockBasedTableOptions::kBinarySearchWithFirstKey;
|
|
|
|
|
|
|
|
s = GetGlobalSequenceNumber(*(rep_->table_properties), largest_seqno,
|
|
|
|
&(rep_->global_seqno));
|
|
|
|
if (!s.ok()) {
|
|
|
|
ROCKS_LOG_ERROR(rep_->ioptions.info_log, "%s", s.ToString().c_str());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::ReadRangeDelBlock(
|
|
|
|
FilePrefetchBuffer* prefetch_buffer, InternalIterator* meta_iter,
|
|
|
|
const InternalKeyComparator& internal_comparator,
|
|
|
|
BlockCacheLookupContext* lookup_context) {
|
|
|
|
Status s;
|
|
|
|
bool found_range_del_block;
|
|
|
|
BlockHandle range_del_handle;
|
|
|
|
s = SeekToRangeDelBlock(meta_iter, &found_range_del_block, &range_del_handle);
|
|
|
|
if (!s.ok()) {
|
|
|
|
ROCKS_LOG_WARN(
|
|
|
|
rep_->ioptions.info_log,
|
|
|
|
"Error when seeking to range delete tombstones block from file: %s",
|
|
|
|
s.ToString().c_str());
|
|
|
|
} else if (found_range_del_block && !range_del_handle.IsNull()) {
|
Cache fragmented range tombstones in BlockBasedTableReader (#4493)
Summary:
This allows tombstone fragmenting to only be performed when the table is opened, and cached for subsequent accesses.
On the same DB used in #4449, running `readrandom` results in the following:
```
readrandom : 0.983 micros/op 1017076 ops/sec; 78.3 MB/s (63103 of 100000 found)
```
Now that Get performance in the presence of range tombstones is reasonable, I also compared the performance between a DB with range tombstones, "expanded" range tombstones (several point tombstones that cover the same keys the equivalent range tombstone would cover, a common workaround for DeleteRange), and no range tombstones. The created DBs had 5 million keys each, and DeleteRange was called at regular intervals (depending on the total number of range tombstones being written) after 4.5 million Puts. The table below summarizes the results of a `readwhilewriting` benchmark (in order to provide somewhat more realistic results):
```
Tombstones? | avg micros/op | stddev micros/op | avg ops/s | stddev ops/s
----------------- | ------------- | ---------------- | ------------ | ------------
None | 0.6186 | 0.04637 | 1,625,252.90 | 124,679.41
500 Expanded | 0.6019 | 0.03628 | 1,666,670.40 | 101,142.65
500 Unexpanded | 0.6435 | 0.03994 | 1,559,979.40 | 104,090.52
1k Expanded | 0.6034 | 0.04349 | 1,665,128.10 | 125,144.57
1k Unexpanded | 0.6261 | 0.03093 | 1,600,457.50 | 79,024.94
5k Expanded | 0.6163 | 0.05926 | 1,636,668.80 | 154,888.85
5k Unexpanded | 0.6402 | 0.04002 | 1,567,804.70 | 100,965.55
10k Expanded | 0.6036 | 0.05105 | 1,667,237.70 | 142,830.36
10k Unexpanded | 0.6128 | 0.02598 | 1,634,633.40 | 72,161.82
25k Expanded | 0.6198 | 0.04542 | 1,620,980.50 | 116,662.93
25k Unexpanded | 0.5478 | 0.0362 | 1,833,059.10 | 121,233.81
50k Expanded | 0.5104 | 0.04347 | 1,973,107.90 | 184,073.49
50k Unexpanded | 0.4528 | 0.03387 | 2,219,034.50 | 170,984.32
```
After a large enough quantity of range tombstones are written, range tombstone Gets can become faster than reading from an equivalent DB with several point tombstones.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4493
Differential Revision: D10842844
Pulled By: abhimadan
fbshipit-source-id: a7d44534f8120e6aabb65779d26c6b9df954c509
6 years ago
|
|
|
ReadOptions read_options;
|
|
|
|
std::unique_ptr<InternalIterator> iter(NewDataBlockIterator<DataBlockIter>(
|
|
|
|
read_options, range_del_handle,
|
|
|
|
/*input_iter=*/nullptr, BlockType::kRangeDeletion,
|
|
|
|
/*get_context=*/nullptr, lookup_context, Status(), prefetch_buffer));
|
|
|
|
assert(iter != nullptr);
|
|
|
|
s = iter->status();
|
Cache fragmented range tombstones in BlockBasedTableReader (#4493)
Summary:
This allows tombstone fragmenting to only be performed when the table is opened, and cached for subsequent accesses.
On the same DB used in #4449, running `readrandom` results in the following:
```
readrandom : 0.983 micros/op 1017076 ops/sec; 78.3 MB/s (63103 of 100000 found)
```
Now that Get performance in the presence of range tombstones is reasonable, I also compared the performance between a DB with range tombstones, "expanded" range tombstones (several point tombstones that cover the same keys the equivalent range tombstone would cover, a common workaround for DeleteRange), and no range tombstones. The created DBs had 5 million keys each, and DeleteRange was called at regular intervals (depending on the total number of range tombstones being written) after 4.5 million Puts. The table below summarizes the results of a `readwhilewriting` benchmark (in order to provide somewhat more realistic results):
```
Tombstones? | avg micros/op | stddev micros/op | avg ops/s | stddev ops/s
----------------- | ------------- | ---------------- | ------------ | ------------
None | 0.6186 | 0.04637 | 1,625,252.90 | 124,679.41
500 Expanded | 0.6019 | 0.03628 | 1,666,670.40 | 101,142.65
500 Unexpanded | 0.6435 | 0.03994 | 1,559,979.40 | 104,090.52
1k Expanded | 0.6034 | 0.04349 | 1,665,128.10 | 125,144.57
1k Unexpanded | 0.6261 | 0.03093 | 1,600,457.50 | 79,024.94
5k Expanded | 0.6163 | 0.05926 | 1,636,668.80 | 154,888.85
5k Unexpanded | 0.6402 | 0.04002 | 1,567,804.70 | 100,965.55
10k Expanded | 0.6036 | 0.05105 | 1,667,237.70 | 142,830.36
10k Unexpanded | 0.6128 | 0.02598 | 1,634,633.40 | 72,161.82
25k Expanded | 0.6198 | 0.04542 | 1,620,980.50 | 116,662.93
25k Unexpanded | 0.5478 | 0.0362 | 1,833,059.10 | 121,233.81
50k Expanded | 0.5104 | 0.04347 | 1,973,107.90 | 184,073.49
50k Unexpanded | 0.4528 | 0.03387 | 2,219,034.50 | 170,984.32
```
After a large enough quantity of range tombstones are written, range tombstone Gets can become faster than reading from an equivalent DB with several point tombstones.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4493
Differential Revision: D10842844
Pulled By: abhimadan
fbshipit-source-id: a7d44534f8120e6aabb65779d26c6b9df954c509
6 years ago
|
|
|
if (!s.ok()) {
|
|
|
|
ROCKS_LOG_WARN(
|
|
|
|
rep_->ioptions.info_log,
|
Cache fragmented range tombstones in BlockBasedTableReader (#4493)
Summary:
This allows tombstone fragmenting to only be performed when the table is opened, and cached for subsequent accesses.
On the same DB used in #4449, running `readrandom` results in the following:
```
readrandom : 0.983 micros/op 1017076 ops/sec; 78.3 MB/s (63103 of 100000 found)
```
Now that Get performance in the presence of range tombstones is reasonable, I also compared the performance between a DB with range tombstones, "expanded" range tombstones (several point tombstones that cover the same keys the equivalent range tombstone would cover, a common workaround for DeleteRange), and no range tombstones. The created DBs had 5 million keys each, and DeleteRange was called at regular intervals (depending on the total number of range tombstones being written) after 4.5 million Puts. The table below summarizes the results of a `readwhilewriting` benchmark (in order to provide somewhat more realistic results):
```
Tombstones? | avg micros/op | stddev micros/op | avg ops/s | stddev ops/s
----------------- | ------------- | ---------------- | ------------ | ------------
None | 0.6186 | 0.04637 | 1,625,252.90 | 124,679.41
500 Expanded | 0.6019 | 0.03628 | 1,666,670.40 | 101,142.65
500 Unexpanded | 0.6435 | 0.03994 | 1,559,979.40 | 104,090.52
1k Expanded | 0.6034 | 0.04349 | 1,665,128.10 | 125,144.57
1k Unexpanded | 0.6261 | 0.03093 | 1,600,457.50 | 79,024.94
5k Expanded | 0.6163 | 0.05926 | 1,636,668.80 | 154,888.85
5k Unexpanded | 0.6402 | 0.04002 | 1,567,804.70 | 100,965.55
10k Expanded | 0.6036 | 0.05105 | 1,667,237.70 | 142,830.36
10k Unexpanded | 0.6128 | 0.02598 | 1,634,633.40 | 72,161.82
25k Expanded | 0.6198 | 0.04542 | 1,620,980.50 | 116,662.93
25k Unexpanded | 0.5478 | 0.0362 | 1,833,059.10 | 121,233.81
50k Expanded | 0.5104 | 0.04347 | 1,973,107.90 | 184,073.49
50k Unexpanded | 0.4528 | 0.03387 | 2,219,034.50 | 170,984.32
```
After a large enough quantity of range tombstones are written, range tombstone Gets can become faster than reading from an equivalent DB with several point tombstones.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4493
Differential Revision: D10842844
Pulled By: abhimadan
fbshipit-source-id: a7d44534f8120e6aabb65779d26c6b9df954c509
6 years ago
|
|
|
"Encountered error while reading data from range del block %s",
|
|
|
|
s.ToString().c_str());
|
|
|
|
} else {
|
|
|
|
rep_->fragmented_range_dels =
|
|
|
|
std::make_shared<FragmentedRangeTombstoneList>(std::move(iter),
|
|
|
|
internal_comparator);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::PrefetchIndexAndFilterBlocks(
|
|
|
|
FilePrefetchBuffer* prefetch_buffer, InternalIterator* meta_iter,
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
BlockBasedTable* new_table, bool prefetch_all,
|
|
|
|
const BlockBasedTableOptions& table_options, const int level,
|
|
|
|
BlockCacheLookupContext* lookup_context) {
|
|
|
|
Status s;
|
|
|
|
|
|
|
|
// Find filter handle and filter type
|
|
|
|
if (rep_->filter_policy) {
|
|
|
|
for (auto filter_type :
|
|
|
|
{Rep::FilterType::kFullFilter, Rep::FilterType::kPartitionedFilter,
|
|
|
|
Rep::FilterType::kBlockFilter}) {
|
|
|
|
std::string prefix;
|
|
|
|
switch (filter_type) {
|
|
|
|
case Rep::FilterType::kFullFilter:
|
|
|
|
prefix = kFullFilterBlockPrefix;
|
|
|
|
break;
|
|
|
|
case Rep::FilterType::kPartitionedFilter:
|
|
|
|
prefix = kPartitionedFilterBlockPrefix;
|
|
|
|
break;
|
|
|
|
case Rep::FilterType::kBlockFilter:
|
|
|
|
prefix = kFilterBlockPrefix;
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
assert(0);
|
|
|
|
}
|
|
|
|
std::string filter_block_key = prefix;
|
|
|
|
filter_block_key.append(rep_->filter_policy->Name());
|
|
|
|
if (FindMetaBlock(meta_iter, filter_block_key, &rep_->filter_handle)
|
|
|
|
.ok()) {
|
|
|
|
rep_->filter_type = filter_type;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
// Find compression dictionary handle
|
|
|
|
bool found_compression_dict = false;
|
|
|
|
s = SeekToCompressionDictBlock(meta_iter, &found_compression_dict,
|
|
|
|
&rep_->compression_dict_handle);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
BlockBasedTableOptions::IndexType index_type = rep_->index_type;
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
|
|
|
const bool use_cache = table_options.cache_index_and_filter_blocks;
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
// pin both index and filters, down to all partitions
|
|
|
|
const bool pin_all =
|
|
|
|
rep_->table_options.pin_l0_filter_and_index_blocks_in_cache && level == 0;
|
|
|
|
|
|
|
|
// prefetch the first level of index
|
|
|
|
const bool prefetch_index =
|
|
|
|
prefetch_all ||
|
|
|
|
(table_options.pin_top_level_index_and_filter &&
|
|
|
|
index_type == BlockBasedTableOptions::kTwoLevelIndexSearch);
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
// pin the first level of index
|
|
|
|
const bool pin_index =
|
|
|
|
pin_all || (table_options.pin_top_level_index_and_filter &&
|
|
|
|
index_type == BlockBasedTableOptions::kTwoLevelIndexSearch);
|
|
|
|
|
|
|
|
std::unique_ptr<IndexReader> index_reader;
|
|
|
|
s = new_table->CreateIndexReader(prefetch_buffer, meta_iter, use_cache,
|
|
|
|
prefetch_index, pin_index, lookup_context,
|
|
|
|
&index_reader);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
rep_->index_reader = std::move(index_reader);
|
|
|
|
|
|
|
|
// The partitions of partitioned index are always stored in cache. They
|
|
|
|
// are hence follow the configuration for pin and prefetch regardless of
|
|
|
|
// the value of cache_index_and_filter_blocks
|
|
|
|
if (prefetch_all) {
|
|
|
|
rep_->index_reader->CacheDependencies(pin_all);
|
|
|
|
}
|
|
|
|
|
|
|
|
// prefetch the first level of filter
|
|
|
|
const bool prefetch_filter =
|
|
|
|
prefetch_all ||
|
|
|
|
(table_options.pin_top_level_index_and_filter &&
|
|
|
|
rep_->filter_type == Rep::FilterType::kPartitionedFilter);
|
|
|
|
// Partition fitlers cannot be enabled without partition indexes
|
|
|
|
assert(!prefetch_filter || prefetch_index);
|
|
|
|
// pin the first level of filter
|
|
|
|
const bool pin_filter =
|
|
|
|
pin_all || (table_options.pin_top_level_index_and_filter &&
|
|
|
|
rep_->filter_type == Rep::FilterType::kPartitionedFilter);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
if (rep_->filter_policy) {
|
|
|
|
auto filter = new_table->CreateFilterBlockReader(
|
|
|
|
prefetch_buffer, use_cache, prefetch_filter, pin_filter,
|
|
|
|
lookup_context);
|
|
|
|
if (filter) {
|
|
|
|
// Refer to the comment above about paritioned indexes always being cached
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
if (prefetch_all) {
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
filter->CacheDependencies(pin_all);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
}
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
|
|
|
|
rep_->filter = std::move(filter);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!rep_->compression_dict_handle.IsNull()) {
|
|
|
|
std::unique_ptr<UncompressionDictReader> uncompression_dict_reader;
|
|
|
|
s = UncompressionDictReader::Create(this, prefetch_buffer, use_cache,
|
|
|
|
prefetch_all, pin_all, lookup_context,
|
|
|
|
&uncompression_dict_reader);
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
|
|
|
|
rep_->uncompression_dict_reader = std::move(uncompression_dict_reader);
|
|
|
|
}
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
|
|
|
|
assert(s.ok());
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
void BlockBasedTable::SetupForCompaction() {
|
|
|
|
switch (rep_->ioptions.access_hint_on_compaction_start) {
|
|
|
|
case Options::NONE:
|
|
|
|
break;
|
|
|
|
case Options::NORMAL:
|
|
|
|
rep_->file->file()->Hint(RandomAccessFile::NORMAL);
|
|
|
|
break;
|
|
|
|
case Options::SEQUENTIAL:
|
|
|
|
rep_->file->file()->Hint(RandomAccessFile::SEQUENTIAL);
|
|
|
|
break;
|
|
|
|
case Options::WILLNEED:
|
|
|
|
rep_->file->file()->Hint(RandomAccessFile::WILLNEED);
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
assert(false);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
std::shared_ptr<const TableProperties> BlockBasedTable::GetTableProperties()
|
|
|
|
const {
|
|
|
|
return rep_->table_properties;
|
|
|
|
}
|
|
|
|
|
|
|
|
size_t BlockBasedTable::ApproximateMemoryUsage() const {
|
|
|
|
size_t usage = 0;
|
|
|
|
if (rep_->filter) {
|
|
|
|
usage += rep_->filter->ApproximateMemoryUsage();
|
|
|
|
}
|
|
|
|
if (rep_->index_reader) {
|
|
|
|
usage += rep_->index_reader->ApproximateMemoryUsage();
|
|
|
|
}
|
|
|
|
if (rep_->uncompression_dict_reader) {
|
|
|
|
usage += rep_->uncompression_dict_reader->ApproximateMemoryUsage();
|
|
|
|
}
|
|
|
|
return usage;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Load the meta-block from the file. On success, return the loaded meta block
|
|
|
|
// and its iterator.
|
|
|
|
Status BlockBasedTable::ReadMetaBlock(FilePrefetchBuffer* prefetch_buffer,
|
|
|
|
std::unique_ptr<Block>* meta_block,
|
|
|
|
std::unique_ptr<InternalIterator>* iter) {
|
|
|
|
// TODO(sanjay): Skip this if footer.metaindex_handle() size indicates
|
|
|
|
// it is an empty block.
|
|
|
|
std::unique_ptr<Block> meta;
|
|
|
|
Status s = ReadBlockFromFile(
|
|
|
|
rep_->file.get(), prefetch_buffer, rep_->footer, ReadOptions(),
|
|
|
|
rep_->footer.metaindex_handle(), &meta, rep_->ioptions,
|
|
|
|
true /* decompress */, true /*maybe_compressed*/, BlockType::kMetaIndex,
|
|
|
|
UncompressionDict::GetEmptyDict(), rep_->persistent_cache_options,
|
|
|
|
kDisableGlobalSequenceNumber, 0 /* read_amp_bytes_per_bit */,
|
|
|
|
GetMemoryAllocator(rep_->table_options), false /* for_compaction */,
|
|
|
|
rep_->blocks_definitely_zstd_compressed, nullptr /* filter_policy */);
|
|
|
|
|
|
|
|
if (!s.ok()) {
|
|
|
|
ROCKS_LOG_ERROR(rep_->ioptions.info_log,
|
|
|
|
"Encountered error while reading data from properties"
|
|
|
|
" block %s",
|
|
|
|
s.ToString().c_str());
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
*meta_block = std::move(meta);
|
|
|
|
// meta block uses bytewise comparator.
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
iter->reset(meta_block->get()->NewDataIterator(BytewiseComparator(),
|
|
|
|
BytewiseComparator()));
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
template <typename TBlocklike>
|
|
|
|
Status BlockBasedTable::GetDataBlockFromCache(
|
|
|
|
const Slice& block_cache_key, const Slice& compressed_block_cache_key,
|
|
|
|
Cache* block_cache, Cache* block_cache_compressed,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
const ReadOptions& read_options, CachableEntry<TBlocklike>* block,
|
|
|
|
const UncompressionDict& uncompression_dict, BlockType block_type,
|
|
|
|
GetContext* get_context) const {
|
|
|
|
const size_t read_amp_bytes_per_bit =
|
|
|
|
block_type == BlockType::kData
|
|
|
|
? rep_->table_options.read_amp_bytes_per_bit
|
|
|
|
: 0;
|
|
|
|
assert(block);
|
|
|
|
assert(block->IsEmpty());
|
|
|
|
|
|
|
|
Status s;
|
|
|
|
BlockContents* compressed_block = nullptr;
|
|
|
|
Cache::Handle* block_cache_compressed_handle = nullptr;
|
|
|
|
|
|
|
|
// Lookup uncompressed cache first
|
|
|
|
if (block_cache != nullptr) {
|
|
|
|
auto cache_handle = GetEntryFromCache(block_cache, block_cache_key,
|
|
|
|
block_type, get_context);
|
|
|
|
if (cache_handle != nullptr) {
|
|
|
|
block->SetCachedValue(
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
reinterpret_cast<TBlocklike*>(block_cache->Value(cache_handle)),
|
|
|
|
block_cache, cache_handle);
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// If not found, search from the compressed block cache.
|
|
|
|
assert(block->IsEmpty());
|
|
|
|
|
|
|
|
if (block_cache_compressed == nullptr) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
assert(!compressed_block_cache_key.empty());
|
|
|
|
block_cache_compressed_handle =
|
|
|
|
block_cache_compressed->Lookup(compressed_block_cache_key);
|
|
|
|
|
|
|
|
Statistics* statistics = rep_->ioptions.statistics;
|
|
|
|
|
|
|
|
// if we found in the compressed cache, then uncompress and insert into
|
|
|
|
// uncompressed cache
|
|
|
|
if (block_cache_compressed_handle == nullptr) {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_COMPRESSED_MISS);
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
// found compressed block
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_COMPRESSED_HIT);
|
|
|
|
compressed_block = reinterpret_cast<BlockContents*>(
|
|
|
|
block_cache_compressed->Value(block_cache_compressed_handle));
|
|
|
|
CompressionType compression_type = compressed_block->get_compression_type();
|
|
|
|
assert(compression_type != kNoCompression);
|
|
|
|
|
|
|
|
// Retrieve the uncompressed contents into a new buffer
|
|
|
|
BlockContents contents;
|
|
|
|
UncompressionContext context(compression_type);
|
|
|
|
UncompressionInfo info(context, uncompression_dict, compression_type);
|
|
|
|
s = UncompressBlockContents(
|
|
|
|
info, compressed_block->data.data(), compressed_block->data.size(),
|
|
|
|
&contents, rep_->table_options.format_version, rep_->ioptions,
|
|
|
|
GetMemoryAllocator(rep_->table_options));
|
|
|
|
|
|
|
|
// Insert uncompressed block into block cache
|
|
|
|
if (s.ok()) {
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
std::unique_ptr<TBlocklike> block_holder(
|
|
|
|
BlocklikeTraits<TBlocklike>::Create(
|
|
|
|
std::move(contents), rep_->get_global_seqno(block_type),
|
|
|
|
read_amp_bytes_per_bit, statistics,
|
|
|
|
rep_->blocks_definitely_zstd_compressed,
|
|
|
|
rep_->table_options.filter_policy.get())); // uncompressed block
|
|
|
|
|
|
|
|
if (block_cache != nullptr && block_holder->own_bytes() &&
|
|
|
|
read_options.fill_cache) {
|
|
|
|
size_t charge = block_holder->ApproximateMemoryUsage();
|
|
|
|
Cache::Handle* cache_handle = nullptr;
|
|
|
|
s = block_cache->Insert(block_cache_key, block_holder.get(), charge,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
&DeleteCachedEntry<TBlocklike>, &cache_handle);
|
|
|
|
if (s.ok()) {
|
|
|
|
assert(cache_handle != nullptr);
|
|
|
|
block->SetCachedValue(block_holder.release(), block_cache,
|
|
|
|
cache_handle);
|
|
|
|
|
|
|
|
UpdateCacheInsertionMetrics(block_type, get_context, charge);
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_ADD_FAILURES);
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
block->SetOwnedValue(block_holder.release());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Release hold on compressed cache entry
|
|
|
|
block_cache_compressed->Release(block_cache_compressed_handle);
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
template <typename TBlocklike>
|
|
|
|
Status BlockBasedTable::PutDataBlockToCache(
|
|
|
|
const Slice& block_cache_key, const Slice& compressed_block_cache_key,
|
|
|
|
Cache* block_cache, Cache* block_cache_compressed,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
CachableEntry<TBlocklike>* cached_block, BlockContents* raw_block_contents,
|
|
|
|
CompressionType raw_block_comp_type,
|
|
|
|
const UncompressionDict& uncompression_dict, SequenceNumber seq_no,
|
|
|
|
MemoryAllocator* memory_allocator, BlockType block_type,
|
|
|
|
GetContext* get_context) const {
|
|
|
|
const ImmutableCFOptions& ioptions = rep_->ioptions;
|
|
|
|
const uint32_t format_version = rep_->table_options.format_version;
|
|
|
|
const size_t read_amp_bytes_per_bit =
|
|
|
|
block_type == BlockType::kData
|
|
|
|
? rep_->table_options.read_amp_bytes_per_bit
|
|
|
|
: 0;
|
|
|
|
const Cache::Priority priority =
|
|
|
|
rep_->table_options.cache_index_and_filter_blocks_with_high_priority &&
|
|
|
|
(block_type == BlockType::kFilter ||
|
|
|
|
block_type == BlockType::kCompressionDictionary ||
|
|
|
|
block_type == BlockType::kIndex)
|
|
|
|
? Cache::Priority::HIGH
|
|
|
|
: Cache::Priority::LOW;
|
|
|
|
assert(cached_block);
|
|
|
|
assert(cached_block->IsEmpty());
|
|
|
|
|
|
|
|
Status s;
|
|
|
|
Statistics* statistics = ioptions.statistics;
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
std::unique_ptr<TBlocklike> block_holder;
|
|
|
|
if (raw_block_comp_type != kNoCompression) {
|
|
|
|
// Retrieve the uncompressed contents into a new buffer
|
|
|
|
BlockContents uncompressed_block_contents;
|
|
|
|
UncompressionContext context(raw_block_comp_type);
|
|
|
|
UncompressionInfo info(context, uncompression_dict, raw_block_comp_type);
|
|
|
|
s = UncompressBlockContents(info, raw_block_contents->data.data(),
|
|
|
|
raw_block_contents->data.size(),
|
|
|
|
&uncompressed_block_contents, format_version,
|
|
|
|
ioptions, memory_allocator);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
block_holder.reset(BlocklikeTraits<TBlocklike>::Create(
|
|
|
|
std::move(uncompressed_block_contents), seq_no, read_amp_bytes_per_bit,
|
|
|
|
statistics, rep_->blocks_definitely_zstd_compressed,
|
|
|
|
rep_->table_options.filter_policy.get()));
|
|
|
|
} else {
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
block_holder.reset(BlocklikeTraits<TBlocklike>::Create(
|
|
|
|
std::move(*raw_block_contents), seq_no, read_amp_bytes_per_bit,
|
|
|
|
statistics, rep_->blocks_definitely_zstd_compressed,
|
|
|
|
rep_->table_options.filter_policy.get()));
|
|
|
|
}
|
|
|
|
|
|
|
|
// Insert compressed block into compressed block cache.
|
|
|
|
// Release the hold on the compressed cache entry immediately.
|
|
|
|
if (block_cache_compressed != nullptr &&
|
|
|
|
raw_block_comp_type != kNoCompression && raw_block_contents != nullptr &&
|
|
|
|
raw_block_contents->own_bytes()) {
|
|
|
|
#ifndef NDEBUG
|
|
|
|
assert(raw_block_contents->is_raw_block);
|
|
|
|
#endif // NDEBUG
|
|
|
|
|
|
|
|
// We cannot directly put raw_block_contents because this could point to
|
|
|
|
// an object in the stack.
|
|
|
|
BlockContents* block_cont_for_comp_cache =
|
|
|
|
new BlockContents(std::move(*raw_block_contents));
|
|
|
|
s = block_cache_compressed->Insert(
|
|
|
|
compressed_block_cache_key, block_cont_for_comp_cache,
|
|
|
|
block_cont_for_comp_cache->ApproximateMemoryUsage(),
|
|
|
|
&DeleteCachedEntry<BlockContents>);
|
|
|
|
if (s.ok()) {
|
|
|
|
// Avoid the following code to delete this cached block.
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_COMPRESSED_ADD);
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_COMPRESSED_ADD_FAILURES);
|
|
|
|
delete block_cont_for_comp_cache;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// insert into uncompressed block cache
|
|
|
|
if (block_cache != nullptr && block_holder->own_bytes()) {
|
|
|
|
size_t charge = block_holder->ApproximateMemoryUsage();
|
|
|
|
Cache::Handle* cache_handle = nullptr;
|
|
|
|
s = block_cache->Insert(block_cache_key, block_holder.get(), charge,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
&DeleteCachedEntry<TBlocklike>, &cache_handle,
|
|
|
|
priority);
|
|
|
|
if (s.ok()) {
|
|
|
|
assert(cache_handle != nullptr);
|
|
|
|
cached_block->SetCachedValue(block_holder.release(), block_cache,
|
|
|
|
cache_handle);
|
|
|
|
|
|
|
|
UpdateCacheInsertionMetrics(block_type, get_context, charge);
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics, BLOCK_CACHE_ADD_FAILURES);
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
cached_block->SetOwnedValue(block_holder.release());
|
|
|
|
}
|
|
|
|
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
std::unique_ptr<FilterBlockReader> BlockBasedTable::CreateFilterBlockReader(
|
|
|
|
FilePrefetchBuffer* prefetch_buffer, bool use_cache, bool prefetch,
|
|
|
|
bool pin, BlockCacheLookupContext* lookup_context) {
|
|
|
|
auto& rep = rep_;
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
auto filter_type = rep->filter_type;
|
|
|
|
if (filter_type == Rep::FilterType::kNoFilter) {
|
|
|
|
return std::unique_ptr<FilterBlockReader>();
|
|
|
|
}
|
|
|
|
|
|
|
|
assert(rep->filter_policy);
|
|
|
|
|
|
|
|
switch (filter_type) {
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
case Rep::FilterType::kPartitionedFilter:
|
|
|
|
return PartitionedFilterBlockReader::Create(
|
|
|
|
this, prefetch_buffer, use_cache, prefetch, pin, lookup_context);
|
|
|
|
|
|
|
|
case Rep::FilterType::kBlockFilter:
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
return BlockBasedFilterBlockReader::Create(
|
|
|
|
this, prefetch_buffer, use_cache, prefetch, pin, lookup_context);
|
|
|
|
|
|
|
|
case Rep::FilterType::kFullFilter:
|
|
|
|
return FullFilterBlockReader::Create(this, prefetch_buffer, use_cache,
|
|
|
|
prefetch, pin, lookup_context);
|
|
|
|
|
|
|
|
default:
|
|
|
|
// filter_type is either kNoFilter (exited the function at the first if),
|
|
|
|
// or it must be covered in this switch block
|
|
|
|
assert(false);
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
return std::unique_ptr<FilterBlockReader>();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// disable_prefix_seek should be set to true when prefix_extractor found in SST
|
|
|
|
// differs from the one in mutable_cf_options and index type is HashBasedIndex
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
InternalIteratorBase<IndexValue>* BlockBasedTable::NewIndexIterator(
|
|
|
|
const ReadOptions& read_options, bool disable_prefix_seek,
|
|
|
|
IndexBlockIter* input_iter, GetContext* get_context,
|
|
|
|
BlockCacheLookupContext* lookup_context) const {
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
assert(rep_ != nullptr);
|
|
|
|
assert(rep_->index_reader != nullptr);
|
|
|
|
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
// We don't return pinned data from index blocks, so no need
|
|
|
|
// to set `block_contents_pinned`.
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
return rep_->index_reader->NewIterator(read_options, disable_prefix_seek,
|
|
|
|
input_iter, get_context,
|
|
|
|
lookup_context);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Convert an index iterator value (i.e., an encoded BlockHandle)
|
|
|
|
// into an iterator over the contents of the corresponding block.
|
|
|
|
// If input_iter is null, new a iterator
|
|
|
|
// If input_iter is not null, update this iter and return it
|
|
|
|
template <typename TBlockIter>
|
|
|
|
TBlockIter* BlockBasedTable::NewDataBlockIterator(
|
|
|
|
const ReadOptions& ro, const BlockHandle& handle, TBlockIter* input_iter,
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
BlockType block_type, GetContext* get_context,
|
|
|
|
BlockCacheLookupContext* lookup_context, Status s,
|
|
|
|
FilePrefetchBuffer* prefetch_buffer, bool for_compaction) const {
|
|
|
|
PERF_TIMER_GUARD(new_table_block_iter_nanos);
|
|
|
|
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
TBlockIter* iter = input_iter != nullptr ? input_iter : new TBlockIter;
|
|
|
|
if (!s.ok()) {
|
|
|
|
iter->Invalidate(s);
|
|
|
|
return iter;
|
|
|
|
}
|
|
|
|
|
|
|
|
CachableEntry<UncompressionDict> uncompression_dict;
|
|
|
|
if (rep_->uncompression_dict_reader) {
|
|
|
|
const bool no_io = (ro.read_tier == kBlockCacheTier);
|
|
|
|
s = rep_->uncompression_dict_reader->GetOrReadUncompressionDictionary(
|
|
|
|
prefetch_buffer, no_io, get_context, lookup_context,
|
|
|
|
&uncompression_dict);
|
|
|
|
if (!s.ok()) {
|
|
|
|
iter->Invalidate(s);
|
|
|
|
return iter;
|
|
|
|
}
|
|
|
|
}
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
|
|
|
const UncompressionDict& dict = uncompression_dict.GetValue()
|
|
|
|
? *uncompression_dict.GetValue()
|
|
|
|
: UncompressionDict::GetEmptyDict();
|
|
|
|
|
|
|
|
CachableEntry<Block> block;
|
|
|
|
s = RetrieveBlock(prefetch_buffer, ro, handle, dict, &block, block_type,
|
|
|
|
get_context, lookup_context, for_compaction,
|
|
|
|
/* use_cache */ true);
|
|
|
|
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
if (!s.ok()) {
|
|
|
|
assert(block.IsEmpty());
|
|
|
|
iter->Invalidate(s);
|
|
|
|
return iter;
|
|
|
|
}
|
|
|
|
|
|
|
|
assert(block.GetValue() != nullptr);
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
// Block contents are pinned and it is still pinned after the iterator
|
|
|
|
// is destroyed as long as cleanup functions are moved to another object,
|
|
|
|
// when:
|
|
|
|
// 1. block cache handle is set to be released in cleanup function, or
|
|
|
|
// 2. it's pointing to immortal source. If own_bytes is true then we are
|
|
|
|
// not reading data from the original source, whether immortal or not.
|
|
|
|
// Otherwise, the block is pinned iff the source is immortal.
|
|
|
|
const bool block_contents_pinned =
|
|
|
|
block.IsCached() ||
|
|
|
|
(!block.GetValue()->own_bytes() && rep_->immortal_table);
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
iter = InitBlockIterator<TBlockIter>(rep_, block.GetValue(), iter,
|
|
|
|
block_contents_pinned);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
|
|
|
if (!block.IsCached()) {
|
|
|
|
if (!ro.fill_cache && rep_->cache_key_prefix_size != 0) {
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
// insert a dummy record to block cache to track the memory usage
|
|
|
|
Cache* const block_cache = rep_->table_options.block_cache.get();
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
Cache::Handle* cache_handle = nullptr;
|
|
|
|
// There are two other types of cache keys: 1) SST cache key added in
|
|
|
|
// `MaybeReadBlockAndLoadToCache` 2) dummy cache key added in
|
|
|
|
// `write_buffer_manager`. Use longer prefix (41 bytes) to differentiate
|
|
|
|
// from SST cache key(31 bytes), and use non-zero prefix to
|
|
|
|
// differentiate from `write_buffer_manager`
|
|
|
|
const size_t kExtraCacheKeyPrefix = kMaxVarint64Length * 4 + 1;
|
|
|
|
char cache_key[kExtraCacheKeyPrefix + kMaxVarint64Length];
|
|
|
|
// Prefix: use rep_->cache_key_prefix padded by 0s
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
memset(cache_key, 0, kExtraCacheKeyPrefix + kMaxVarint64Length);
|
|
|
|
assert(rep_->cache_key_prefix_size != 0);
|
|
|
|
assert(rep_->cache_key_prefix_size <= kExtraCacheKeyPrefix);
|
|
|
|
memcpy(cache_key, rep_->cache_key_prefix, rep_->cache_key_prefix_size);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
char* end = EncodeVarint64(cache_key + kExtraCacheKeyPrefix,
|
|
|
|
next_cache_key_id_++);
|
|
|
|
assert(end - cache_key <=
|
|
|
|
static_cast<int>(kExtraCacheKeyPrefix + kMaxVarint64Length));
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
const Slice unique_key(cache_key, static_cast<size_t>(end - cache_key));
|
|
|
|
s = block_cache->Insert(unique_key, nullptr,
|
|
|
|
block.GetValue()->ApproximateMemoryUsage(),
|
|
|
|
nullptr, &cache_handle);
|
|
|
|
|
|
|
|
if (s.ok()) {
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
assert(cache_handle != nullptr);
|
|
|
|
iter->RegisterCleanup(&ForceReleaseCachedEntry, block_cache,
|
|
|
|
cache_handle);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
iter->SetCacheHandle(block.GetCacheHandle());
|
|
|
|
}
|
|
|
|
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
block.TransferTo(iter);
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
|
|
|
|
return iter;
|
|
|
|
}
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
template <>
|
|
|
|
DataBlockIter* BlockBasedTable::InitBlockIterator<DataBlockIter>(
|
|
|
|
const Rep* rep, Block* block, DataBlockIter* input_iter,
|
|
|
|
bool block_contents_pinned) {
|
|
|
|
return block->NewDataIterator(
|
|
|
|
&rep->internal_comparator, rep->internal_comparator.user_comparator(),
|
|
|
|
input_iter, rep->ioptions.statistics, block_contents_pinned);
|
|
|
|
}
|
|
|
|
|
|
|
|
template <>
|
|
|
|
IndexBlockIter* BlockBasedTable::InitBlockIterator<IndexBlockIter>(
|
|
|
|
const Rep* rep, Block* block, IndexBlockIter* input_iter,
|
|
|
|
bool block_contents_pinned) {
|
|
|
|
return block->NewIndexIterator(
|
|
|
|
&rep->internal_comparator, rep->internal_comparator.user_comparator(),
|
|
|
|
input_iter, rep->ioptions.statistics, /* total_order_seek */ true,
|
|
|
|
rep->index_has_first_key, rep->index_key_includes_seq,
|
|
|
|
rep->index_value_is_full, block_contents_pinned);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Convert an uncompressed data block (i.e CachableEntry<Block>)
|
|
|
|
// into an iterator over the contents of the corresponding block.
|
|
|
|
// If input_iter is null, new a iterator
|
|
|
|
// If input_iter is not null, update this iter and return it
|
|
|
|
template <typename TBlockIter>
|
|
|
|
TBlockIter* BlockBasedTable::NewDataBlockIterator(const ReadOptions& ro,
|
|
|
|
CachableEntry<Block>& block,
|
|
|
|
TBlockIter* input_iter,
|
|
|
|
Status s) const {
|
|
|
|
PERF_TIMER_GUARD(new_table_block_iter_nanos);
|
|
|
|
|
|
|
|
TBlockIter* iter = input_iter != nullptr ? input_iter : new TBlockIter;
|
|
|
|
if (!s.ok()) {
|
|
|
|
iter->Invalidate(s);
|
|
|
|
return iter;
|
|
|
|
}
|
|
|
|
|
|
|
|
assert(block.GetValue() != nullptr);
|
|
|
|
// Block contents are pinned and it is still pinned after the iterator
|
|
|
|
// is destroyed as long as cleanup functions are moved to another object,
|
|
|
|
// when:
|
|
|
|
// 1. block cache handle is set to be released in cleanup function, or
|
|
|
|
// 2. it's pointing to immortal source. If own_bytes is true then we are
|
|
|
|
// not reading data from the original source, whether immortal or not.
|
|
|
|
// Otherwise, the block is pinned iff the source is immortal.
|
|
|
|
const bool block_contents_pinned =
|
|
|
|
block.IsCached() ||
|
|
|
|
(!block.GetValue()->own_bytes() && rep_->immortal_table);
|
|
|
|
iter = InitBlockIterator<TBlockIter>(rep_, block.GetValue(), iter,
|
|
|
|
block_contents_pinned);
|
|
|
|
|
|
|
|
if (!block.IsCached()) {
|
|
|
|
if (!ro.fill_cache && rep_->cache_key_prefix_size != 0) {
|
|
|
|
// insert a dummy record to block cache to track the memory usage
|
|
|
|
Cache* const block_cache = rep_->table_options.block_cache.get();
|
|
|
|
Cache::Handle* cache_handle = nullptr;
|
|
|
|
// There are two other types of cache keys: 1) SST cache key added in
|
|
|
|
// `MaybeReadBlockAndLoadToCache` 2) dummy cache key added in
|
|
|
|
// `write_buffer_manager`. Use longer prefix (41 bytes) to differentiate
|
|
|
|
// from SST cache key(31 bytes), and use non-zero prefix to
|
|
|
|
// differentiate from `write_buffer_manager`
|
|
|
|
const size_t kExtraCacheKeyPrefix = kMaxVarint64Length * 4 + 1;
|
|
|
|
char cache_key[kExtraCacheKeyPrefix + kMaxVarint64Length];
|
|
|
|
// Prefix: use rep_->cache_key_prefix padded by 0s
|
|
|
|
memset(cache_key, 0, kExtraCacheKeyPrefix + kMaxVarint64Length);
|
|
|
|
assert(rep_->cache_key_prefix_size != 0);
|
|
|
|
assert(rep_->cache_key_prefix_size <= kExtraCacheKeyPrefix);
|
|
|
|
memcpy(cache_key, rep_->cache_key_prefix, rep_->cache_key_prefix_size);
|
|
|
|
char* end = EncodeVarint64(cache_key + kExtraCacheKeyPrefix,
|
|
|
|
next_cache_key_id_++);
|
|
|
|
assert(end - cache_key <=
|
|
|
|
static_cast<int>(kExtraCacheKeyPrefix + kMaxVarint64Length));
|
|
|
|
const Slice unique_key(cache_key, static_cast<size_t>(end - cache_key));
|
|
|
|
s = block_cache->Insert(unique_key, nullptr,
|
|
|
|
block.GetValue()->ApproximateMemoryUsage(),
|
|
|
|
nullptr, &cache_handle);
|
|
|
|
if (s.ok()) {
|
|
|
|
assert(cache_handle != nullptr);
|
|
|
|
iter->RegisterCleanup(&ForceReleaseCachedEntry, block_cache,
|
|
|
|
cache_handle);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
iter->SetCacheHandle(block.GetCacheHandle());
|
|
|
|
}
|
|
|
|
|
|
|
|
block.TransferTo(iter);
|
|
|
|
return iter;
|
|
|
|
}
|
|
|
|
|
|
|
|
// If contents is nullptr, this function looks up the block caches for the
|
|
|
|
// data block referenced by handle, and read the block from disk if necessary.
|
|
|
|
// If contents is non-null, it skips the cache lookup and disk read, since
|
|
|
|
// the caller has already read it. In both cases, if ro.fill_cache is true,
|
|
|
|
// it inserts the block into the block cache.
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
template <typename TBlocklike>
|
|
|
|
Status BlockBasedTable::MaybeReadBlockAndLoadToCache(
|
|
|
|
FilePrefetchBuffer* prefetch_buffer, const ReadOptions& ro,
|
|
|
|
const BlockHandle& handle, const UncompressionDict& uncompression_dict,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
CachableEntry<TBlocklike>* block_entry, BlockType block_type,
|
|
|
|
GetContext* get_context, BlockCacheLookupContext* lookup_context,
|
|
|
|
BlockContents* contents) const {
|
|
|
|
assert(block_entry != nullptr);
|
|
|
|
const bool no_io = (ro.read_tier == kBlockCacheTier);
|
|
|
|
Cache* block_cache = rep_->table_options.block_cache.get();
|
|
|
|
// No point to cache compressed blocks if it never goes away
|
|
|
|
Cache* block_cache_compressed =
|
|
|
|
rep_->immortal_table ? nullptr
|
|
|
|
: rep_->table_options.block_cache_compressed.get();
|
|
|
|
|
|
|
|
// First, try to get the block from the cache
|
|
|
|
//
|
|
|
|
// If either block cache is enabled, we'll try to read from it.
|
|
|
|
Status s;
|
|
|
|
char cache_key[kMaxCacheKeyPrefixSize + kMaxVarint64Length];
|
|
|
|
char compressed_cache_key[kMaxCacheKeyPrefixSize + kMaxVarint64Length];
|
|
|
|
Slice key /* key to the block cache */;
|
|
|
|
Slice ckey /* key to the compressed block cache */;
|
|
|
|
bool is_cache_hit = false;
|
|
|
|
if (block_cache != nullptr || block_cache_compressed != nullptr) {
|
|
|
|
// create key for block cache
|
|
|
|
if (block_cache != nullptr) {
|
|
|
|
key = GetCacheKey(rep_->cache_key_prefix, rep_->cache_key_prefix_size,
|
|
|
|
handle, cache_key);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (block_cache_compressed != nullptr) {
|
|
|
|
ckey = GetCacheKey(rep_->compressed_cache_key_prefix,
|
|
|
|
rep_->compressed_cache_key_prefix_size, handle,
|
|
|
|
compressed_cache_key);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!contents) {
|
|
|
|
s = GetDataBlockFromCache(key, ckey, block_cache, block_cache_compressed,
|
|
|
|
ro, block_entry, uncompression_dict, block_type,
|
|
|
|
get_context);
|
|
|
|
if (block_entry->GetValue()) {
|
|
|
|
// TODO(haoyu): Differentiate cache hit on uncompressed block cache and
|
|
|
|
// compressed block cache.
|
|
|
|
is_cache_hit = true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Can't find the block from the cache. If I/O is allowed, read from the
|
|
|
|
// file.
|
|
|
|
if (block_entry->GetValue() == nullptr && !no_io && ro.fill_cache) {
|
|
|
|
Statistics* statistics = rep_->ioptions.statistics;
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
const bool maybe_compressed =
|
|
|
|
block_type != BlockType::kFilter &&
|
|
|
|
block_type != BlockType::kCompressionDictionary &&
|
|
|
|
rep_->blocks_maybe_compressed;
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
const bool do_uncompress = maybe_compressed && !block_cache_compressed;
|
|
|
|
CompressionType raw_block_comp_type;
|
|
|
|
BlockContents raw_block_contents;
|
|
|
|
if (!contents) {
|
|
|
|
StopWatch sw(rep_->ioptions.env, statistics, READ_BLOCK_GET_MICROS);
|
|
|
|
BlockFetcher block_fetcher(
|
|
|
|
rep_->file.get(), prefetch_buffer, rep_->footer, ro, handle,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
&raw_block_contents, rep_->ioptions, do_uncompress,
|
|
|
|
maybe_compressed, block_type, uncompression_dict,
|
|
|
|
rep_->persistent_cache_options,
|
|
|
|
GetMemoryAllocator(rep_->table_options),
|
|
|
|
GetMemoryAllocatorForCompressedBlock(rep_->table_options));
|
|
|
|
s = block_fetcher.ReadBlockContents();
|
|
|
|
raw_block_comp_type = block_fetcher.get_compression_type();
|
|
|
|
contents = &raw_block_contents;
|
|
|
|
} else {
|
|
|
|
raw_block_comp_type = contents->get_compression_type();
|
|
|
|
}
|
|
|
|
|
|
|
|
if (s.ok()) {
|
|
|
|
SequenceNumber seq_no = rep_->get_global_seqno(block_type);
|
|
|
|
// If filling cache is allowed and a cache is configured, try to put the
|
|
|
|
// block to the cache.
|
|
|
|
s = PutDataBlockToCache(
|
|
|
|
key, ckey, block_cache, block_cache_compressed, block_entry,
|
|
|
|
contents, raw_block_comp_type, uncompression_dict, seq_no,
|
|
|
|
GetMemoryAllocator(rep_->table_options), block_type, get_context);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Fill lookup_context.
|
|
|
|
if (block_cache_tracer_ && block_cache_tracer_->is_tracing_enabled() &&
|
|
|
|
lookup_context) {
|
|
|
|
size_t usage = 0;
|
|
|
|
uint64_t nkeys = 0;
|
|
|
|
if (block_entry->GetValue()) {
|
|
|
|
// Approximate the number of keys in the block using restarts.
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
nkeys =
|
|
|
|
rep_->table_options.block_restart_interval *
|
|
|
|
BlocklikeTraits<TBlocklike>::GetNumRestarts(*block_entry->GetValue());
|
|
|
|
usage = block_entry->GetValue()->ApproximateMemoryUsage();
|
|
|
|
}
|
|
|
|
TraceType trace_block_type = TraceType::kTraceMax;
|
|
|
|
switch (block_type) {
|
|
|
|
case BlockType::kData:
|
|
|
|
trace_block_type = TraceType::kBlockTraceDataBlock;
|
|
|
|
break;
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
case BlockType::kFilter:
|
|
|
|
trace_block_type = TraceType::kBlockTraceFilterBlock;
|
|
|
|
break;
|
|
|
|
case BlockType::kCompressionDictionary:
|
|
|
|
trace_block_type = TraceType::kBlockTraceUncompressionDictBlock;
|
|
|
|
break;
|
|
|
|
case BlockType::kRangeDeletion:
|
|
|
|
trace_block_type = TraceType::kBlockTraceRangeDeletionBlock;
|
|
|
|
break;
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
case BlockType::kIndex:
|
|
|
|
trace_block_type = TraceType::kBlockTraceIndexBlock;
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
// This cannot happen.
|
|
|
|
assert(false);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
bool no_insert = no_io || !ro.fill_cache;
|
|
|
|
if (BlockCacheTraceHelper::IsGetOrMultiGetOnDataBlock(
|
|
|
|
trace_block_type, lookup_context->caller)) {
|
|
|
|
// Defer logging the access to Get() and MultiGet() to trace additional
|
|
|
|
// information, e.g., referenced_key_exist_in_block.
|
|
|
|
|
|
|
|
// Make a copy of the block key here since it will be logged later.
|
|
|
|
lookup_context->FillLookupContext(
|
|
|
|
is_cache_hit, no_insert, trace_block_type,
|
|
|
|
/*block_size=*/usage, /*block_key=*/key.ToString(), nkeys);
|
|
|
|
} else {
|
|
|
|
// Avoid making copy of block_key and cf_name when constructing the access
|
|
|
|
// record.
|
|
|
|
BlockCacheTraceRecord access_record(
|
|
|
|
rep_->ioptions.env->NowMicros(),
|
|
|
|
/*block_key=*/"", trace_block_type,
|
|
|
|
/*block_size=*/usage, rep_->cf_id_for_tracing(),
|
|
|
|
/*cf_name=*/"", rep_->level_for_tracing(),
|
|
|
|
rep_->sst_number_for_tracing(), lookup_context->caller, is_cache_hit,
|
|
|
|
no_insert, lookup_context->get_id,
|
|
|
|
lookup_context->get_from_user_specified_snapshot,
|
|
|
|
/*referenced_key=*/"");
|
|
|
|
block_cache_tracer_->WriteBlockAccess(access_record, key,
|
|
|
|
rep_->cf_name_for_tracing(),
|
|
|
|
lookup_context->referenced_key);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
assert(s.ok() || block_entry->GetValue() == nullptr);
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
// This function reads multiple data blocks from disk using Env::MultiRead()
|
|
|
|
// and optionally inserts them into the block cache. It uses the scratch
|
|
|
|
// buffer provided by the caller, which is contiguous. If scratch is a nullptr
|
|
|
|
// it allocates a separate buffer for each block. Typically, if the blocks
|
|
|
|
// need to be uncompressed and there is no compressed block cache, callers
|
|
|
|
// can allocate a temporary scratch buffer in order to minimize memory
|
|
|
|
// allocations.
|
|
|
|
// If options.fill_cache is true, it inserts the blocks into cache. If its
|
|
|
|
// false and scratch is non-null and the blocks are uncompressed, it copies
|
|
|
|
// the buffers to heap. In any case, the CachableEntry<Block> returned will
|
|
|
|
// own the data bytes.
|
|
|
|
// batch - A MultiGetRange with only those keys with unique data blocks not
|
|
|
|
// found in cache
|
|
|
|
// handles - A vector of block handles. Some of them me be NULL handles
|
|
|
|
// scratch - An optional contiguous buffer to read compressed blocks into
|
|
|
|
void BlockBasedTable::RetrieveMultipleBlocks(
|
|
|
|
const ReadOptions& options, const MultiGetRange* batch,
|
|
|
|
const autovector<BlockHandle, MultiGetContext::MAX_BATCH_SIZE>* handles,
|
|
|
|
autovector<Status, MultiGetContext::MAX_BATCH_SIZE>* statuses,
|
|
|
|
autovector<CachableEntry<Block>, MultiGetContext::MAX_BATCH_SIZE>* results,
|
|
|
|
char* scratch, const UncompressionDict& uncompression_dict) const {
|
|
|
|
RandomAccessFileReader* file = rep_->file.get();
|
|
|
|
const Footer& footer = rep_->footer;
|
|
|
|
const ImmutableCFOptions& ioptions = rep_->ioptions;
|
|
|
|
SequenceNumber global_seqno = rep_->get_global_seqno(BlockType::kData);
|
|
|
|
size_t read_amp_bytes_per_bit = rep_->table_options.read_amp_bytes_per_bit;
|
|
|
|
MemoryAllocator* memory_allocator = GetMemoryAllocator(rep_->table_options);
|
|
|
|
|
|
|
|
if (file->use_direct_io() || ioptions.allow_mmap_reads) {
|
|
|
|
size_t idx_in_batch = 0;
|
|
|
|
for (auto mget_iter = batch->begin(); mget_iter != batch->end();
|
|
|
|
++mget_iter, ++idx_in_batch) {
|
|
|
|
BlockCacheLookupContext lookup_data_block_context(
|
|
|
|
TableReaderCaller::kUserMultiGet);
|
|
|
|
const BlockHandle& handle = (*handles)[idx_in_batch];
|
|
|
|
if (handle.IsNull()) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
(*statuses)[idx_in_batch] =
|
|
|
|
RetrieveBlock(nullptr, options, handle, uncompression_dict,
|
|
|
|
&(*results)[idx_in_batch], BlockType::kData,
|
|
|
|
mget_iter->get_context, &lookup_data_block_context,
|
|
|
|
/* for_compaction */ false, /* use_cache */ true);
|
|
|
|
}
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
autovector<ReadRequest, MultiGetContext::MAX_BATCH_SIZE> read_reqs;
|
|
|
|
size_t buf_offset = 0;
|
|
|
|
size_t idx_in_batch = 0;
|
|
|
|
for (auto mget_iter = batch->begin(); mget_iter != batch->end();
|
|
|
|
++mget_iter, ++idx_in_batch) {
|
|
|
|
const BlockHandle& handle = (*handles)[idx_in_batch];
|
|
|
|
if (handle.IsNull()) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
ReadRequest req;
|
|
|
|
req.len = handle.size() + kBlockTrailerSize;
|
|
|
|
if (scratch == nullptr) {
|
|
|
|
req.scratch = new char[req.len];
|
|
|
|
} else {
|
|
|
|
req.scratch = scratch + buf_offset;
|
|
|
|
buf_offset += req.len;
|
|
|
|
}
|
|
|
|
req.offset = handle.offset();
|
|
|
|
req.status = Status::OK();
|
|
|
|
read_reqs.emplace_back(req);
|
|
|
|
}
|
|
|
|
|
|
|
|
file->MultiRead(&read_reqs[0], read_reqs.size());
|
|
|
|
|
|
|
|
size_t read_req_idx = 0;
|
|
|
|
idx_in_batch = 0;
|
|
|
|
for (auto mget_iter = batch->begin(); mget_iter != batch->end();
|
|
|
|
++mget_iter, ++idx_in_batch) {
|
|
|
|
const BlockHandle& handle = (*handles)[idx_in_batch];
|
|
|
|
|
|
|
|
if (handle.IsNull()) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
ReadRequest& req = read_reqs[read_req_idx++];
|
|
|
|
Status s = req.status;
|
|
|
|
if (s.ok()) {
|
|
|
|
if (req.result.size() != handle.size() + kBlockTrailerSize) {
|
|
|
|
s = Status::Corruption("truncated block read from " +
|
|
|
|
rep_->file->file_name() + " offset " +
|
|
|
|
ToString(handle.offset()) + ", expected " +
|
|
|
|
ToString(handle.size() + kBlockTrailerSize) +
|
|
|
|
" bytes, got " + ToString(req.result.size()));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
BlockContents raw_block_contents;
|
|
|
|
if (s.ok()) {
|
|
|
|
if (scratch == nullptr) {
|
|
|
|
// We allocated a buffer for this block. Give ownership of it to
|
|
|
|
// BlockContents so it can free the memory
|
|
|
|
assert(req.result.data() == req.scratch);
|
|
|
|
std::unique_ptr<char[]> raw_block(req.scratch);
|
|
|
|
raw_block_contents = BlockContents(std::move(raw_block), handle.size());
|
|
|
|
} else {
|
|
|
|
// We used the scratch buffer, so no need to free anything
|
|
|
|
raw_block_contents = BlockContents(Slice(req.scratch, handle.size()));
|
|
|
|
}
|
|
|
|
#ifndef NDEBUG
|
|
|
|
raw_block_contents.is_raw_block = true;
|
|
|
|
#endif
|
|
|
|
if (options.verify_checksums) {
|
|
|
|
PERF_TIMER_GUARD(block_checksum_time);
|
|
|
|
const char* data = req.result.data();
|
|
|
|
uint32_t expected = DecodeFixed32(data + handle.size() + 1);
|
|
|
|
s = rocksdb::VerifyChecksum(footer.checksum(), req.result.data(),
|
|
|
|
handle.size() + 1, expected);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (s.ok()) {
|
|
|
|
if (options.fill_cache) {
|
|
|
|
BlockCacheLookupContext lookup_data_block_context(
|
|
|
|
TableReaderCaller::kUserMultiGet);
|
|
|
|
CachableEntry<Block>* block_entry = &(*results)[idx_in_batch];
|
|
|
|
// MaybeReadBlockAndLoadToCache will insert into the block caches if
|
|
|
|
// necessary. Since we're passing the raw block contents, it will
|
|
|
|
// avoid looking up the block cache
|
|
|
|
s = MaybeReadBlockAndLoadToCache(
|
|
|
|
nullptr, options, handle, uncompression_dict, block_entry,
|
|
|
|
BlockType::kData, mget_iter->get_context,
|
|
|
|
&lookup_data_block_context, &raw_block_contents);
|
|
|
|
} else {
|
|
|
|
CompressionType compression_type =
|
|
|
|
raw_block_contents.get_compression_type();
|
|
|
|
BlockContents contents;
|
|
|
|
if (compression_type != kNoCompression) {
|
|
|
|
UncompressionContext context(compression_type);
|
|
|
|
UncompressionInfo info(context, uncompression_dict, compression_type);
|
|
|
|
s = UncompressBlockContents(info, req.result.data(), handle.size(),
|
|
|
|
&contents, footer.version(),
|
|
|
|
rep_->ioptions, memory_allocator);
|
|
|
|
} else {
|
|
|
|
if (scratch != nullptr) {
|
|
|
|
// If we used the scratch buffer, then the contents need to be
|
|
|
|
// copied to heap
|
|
|
|
Slice raw = Slice(req.result.data(), handle.size());
|
|
|
|
contents = BlockContents(
|
|
|
|
CopyBufferToHeap(GetMemoryAllocator(rep_->table_options), raw),
|
|
|
|
handle.size());
|
|
|
|
} else {
|
|
|
|
contents = std::move(raw_block_contents);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (s.ok()) {
|
|
|
|
(*results)[idx_in_batch].SetOwnedValue(
|
|
|
|
new Block(std::move(contents), global_seqno,
|
|
|
|
read_amp_bytes_per_bit, ioptions.statistics));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
(*statuses)[idx_in_batch] = s;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
template <typename TBlocklike>
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
Status BlockBasedTable::RetrieveBlock(
|
|
|
|
FilePrefetchBuffer* prefetch_buffer, const ReadOptions& ro,
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
const BlockHandle& handle, const UncompressionDict& uncompression_dict,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
CachableEntry<TBlocklike>* block_entry, BlockType block_type,
|
|
|
|
GetContext* get_context, BlockCacheLookupContext* lookup_context,
|
|
|
|
bool for_compaction, bool use_cache) const {
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
assert(block_entry);
|
|
|
|
assert(block_entry->IsEmpty());
|
|
|
|
|
|
|
|
Status s;
|
|
|
|
if (use_cache) {
|
|
|
|
s = MaybeReadBlockAndLoadToCache(prefetch_buffer, ro, handle,
|
|
|
|
uncompression_dict, block_entry,
|
|
|
|
block_type, get_context, lookup_context,
|
|
|
|
/*contents=*/nullptr);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (block_entry->GetValue() != nullptr) {
|
|
|
|
assert(s.ok());
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
return s;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
assert(block_entry->IsEmpty());
|
|
|
|
|
|
|
|
const bool no_io = ro.read_tier == kBlockCacheTier;
|
|
|
|
if (no_io) {
|
|
|
|
return Status::Incomplete("no blocking io");
|
|
|
|
}
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
const bool maybe_compressed =
|
|
|
|
block_type != BlockType::kFilter &&
|
|
|
|
block_type != BlockType::kCompressionDictionary &&
|
|
|
|
rep_->blocks_maybe_compressed;
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
const bool do_uncompress = maybe_compressed;
|
|
|
|
std::unique_ptr<TBlocklike> block;
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
|
|
|
{
|
|
|
|
StopWatch sw(rep_->ioptions.env, rep_->ioptions.statistics,
|
|
|
|
READ_BLOCK_GET_MICROS);
|
|
|
|
s = ReadBlockFromFile(
|
|
|
|
rep_->file.get(), prefetch_buffer, rep_->footer, ro, handle, &block,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
rep_->ioptions, do_uncompress, maybe_compressed, block_type,
|
|
|
|
uncompression_dict, rep_->persistent_cache_options,
|
|
|
|
rep_->get_global_seqno(block_type),
|
|
|
|
block_type == BlockType::kData
|
|
|
|
? rep_->table_options.read_amp_bytes_per_bit
|
|
|
|
: 0,
|
|
|
|
GetMemoryAllocator(rep_->table_options), for_compaction,
|
|
|
|
rep_->blocks_definitely_zstd_compressed,
|
|
|
|
rep_->table_options.filter_policy.get());
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
}
|
|
|
|
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
block_entry->SetOwnedValue(block.release());
|
|
|
|
|
|
|
|
assert(s.ok());
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
// Explicitly instantiate templates for both "blocklike" types we use.
|
|
|
|
// This makes it possible to keep the template definitions in the .cc file.
|
|
|
|
template Status BlockBasedTable::RetrieveBlock<BlockContents>(
|
|
|
|
FilePrefetchBuffer* prefetch_buffer, const ReadOptions& ro,
|
|
|
|
const BlockHandle& handle, const UncompressionDict& uncompression_dict,
|
|
|
|
CachableEntry<BlockContents>* block_entry, BlockType block_type,
|
|
|
|
GetContext* get_context, BlockCacheLookupContext* lookup_context,
|
|
|
|
bool for_compaction, bool use_cache) const;
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
|
|
|
|
template Status BlockBasedTable::RetrieveBlock<ParsedFullFilterBlock>(
|
|
|
|
FilePrefetchBuffer* prefetch_buffer, const ReadOptions& ro,
|
|
|
|
const BlockHandle& handle, const UncompressionDict& uncompression_dict,
|
|
|
|
CachableEntry<ParsedFullFilterBlock>* block_entry, BlockType block_type,
|
|
|
|
GetContext* get_context, BlockCacheLookupContext* lookup_context,
|
|
|
|
bool for_compaction, bool use_cache) const;
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
template Status BlockBasedTable::RetrieveBlock<Block>(
|
|
|
|
FilePrefetchBuffer* prefetch_buffer, const ReadOptions& ro,
|
|
|
|
const BlockHandle& handle, const UncompressionDict& uncompression_dict,
|
|
|
|
CachableEntry<Block>* block_entry, BlockType block_type,
|
|
|
|
GetContext* get_context, BlockCacheLookupContext* lookup_context,
|
|
|
|
bool for_compaction, bool use_cache) const;
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
|
|
|
|
template Status BlockBasedTable::RetrieveBlock<UncompressionDict>(
|
|
|
|
FilePrefetchBuffer* prefetch_buffer, const ReadOptions& ro,
|
|
|
|
const BlockHandle& handle, const UncompressionDict& uncompression_dict,
|
|
|
|
CachableEntry<UncompressionDict>* block_entry, BlockType block_type,
|
|
|
|
GetContext* get_context, BlockCacheLookupContext* lookup_context,
|
|
|
|
bool for_compaction, bool use_cache) const;
|
|
|
|
|
|
|
|
BlockBasedTable::PartitionedIndexIteratorState::PartitionedIndexIteratorState(
|
|
|
|
const BlockBasedTable* table,
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
std::unordered_map<uint64_t, CachableEntry<Block>>* block_map)
|
|
|
|
: table_(table), block_map_(block_map) {}
|
|
|
|
|
|
|
|
InternalIteratorBase<IndexValue>*
|
|
|
|
BlockBasedTable::PartitionedIndexIteratorState::NewSecondaryIterator(
|
|
|
|
const BlockHandle& handle) {
|
|
|
|
// Return a block iterator on the index partition
|
|
|
|
auto block = block_map_->find(handle.offset());
|
|
|
|
// This is a possible scenario since block cache might not have had space
|
|
|
|
// for the partition
|
|
|
|
if (block != block_map_->end()) {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
const Rep* rep = table_->get_rep();
|
|
|
|
assert(rep);
|
|
|
|
|
|
|
|
Statistics* kNullStats = nullptr;
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
// We don't return pinned data from index blocks, so no need
|
|
|
|
// to set `block_contents_pinned`.
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
return block->second.GetValue()->NewIndexIterator(
|
|
|
|
&rep->internal_comparator, rep->internal_comparator.user_comparator(),
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
nullptr, kNullStats, true, rep->index_has_first_key,
|
|
|
|
rep->index_key_includes_seq, rep->index_value_is_full);
|
|
|
|
}
|
|
|
|
// Create an empty iterator
|
|
|
|
return new IndexBlockIter();
|
|
|
|
}
|
|
|
|
|
|
|
|
// This will be broken if the user specifies an unusual implementation
|
|
|
|
// of Options.comparator, or if the user specifies an unusual
|
|
|
|
// definition of prefixes in BlockBasedTableOptions.filter_policy.
|
|
|
|
// In particular, we require the following three properties:
|
|
|
|
//
|
|
|
|
// 1) key.starts_with(prefix(key))
|
|
|
|
// 2) Compare(prefix(key), key) <= 0.
|
|
|
|
// 3) If Compare(key1, key2) <= 0, then Compare(prefix(key1), prefix(key2)) <= 0
|
|
|
|
//
|
|
|
|
// Otherwise, this method guarantees no I/O will be incurred.
|
|
|
|
//
|
|
|
|
// REQUIRES: this method shouldn't be called while the DB lock is held.
|
|
|
|
bool BlockBasedTable::PrefixMayMatch(
|
|
|
|
const Slice& internal_key, const ReadOptions& read_options,
|
|
|
|
const SliceTransform* options_prefix_extractor,
|
|
|
|
const bool need_upper_bound_check,
|
|
|
|
BlockCacheLookupContext* lookup_context) const {
|
|
|
|
if (!rep_->filter_policy) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
const SliceTransform* prefix_extractor;
|
|
|
|
|
|
|
|
if (rep_->table_prefix_extractor == nullptr) {
|
|
|
|
if (need_upper_bound_check) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
prefix_extractor = options_prefix_extractor;
|
|
|
|
} else {
|
|
|
|
prefix_extractor = rep_->table_prefix_extractor.get();
|
|
|
|
}
|
|
|
|
auto user_key = ExtractUserKey(internal_key);
|
|
|
|
if (!prefix_extractor->InDomain(user_key)) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool may_match = true;
|
|
|
|
Status s;
|
|
|
|
|
|
|
|
// First, try check with full filter
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
FilterBlockReader* const filter = rep_->filter.get();
|
|
|
|
bool filter_checked = true;
|
|
|
|
if (filter != nullptr) {
|
|
|
|
if (!filter->IsBlockBased()) {
|
|
|
|
const Slice* const const_ikey_ptr = &internal_key;
|
|
|
|
may_match = filter->RangeMayExist(
|
|
|
|
read_options.iterate_upper_bound, user_key, prefix_extractor,
|
|
|
|
rep_->internal_comparator.user_comparator(), const_ikey_ptr,
|
|
|
|
&filter_checked, need_upper_bound_check, lookup_context);
|
|
|
|
} else {
|
|
|
|
// if prefix_extractor changed for block based filter, skip filter
|
|
|
|
if (need_upper_bound_check) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
auto prefix = prefix_extractor->Transform(user_key);
|
|
|
|
InternalKey internal_key_prefix(prefix, kMaxSequenceNumber, kTypeValue);
|
|
|
|
auto internal_prefix = internal_key_prefix.Encode();
|
|
|
|
|
|
|
|
// To prevent any io operation in this method, we set `read_tier` to make
|
|
|
|
// sure we always read index or filter only when they have already been
|
|
|
|
// loaded to memory.
|
|
|
|
ReadOptions no_io_read_options;
|
|
|
|
no_io_read_options.read_tier = kBlockCacheTier;
|
|
|
|
|
|
|
|
// Then, try find it within each block
|
|
|
|
// we already know prefix_extractor and prefix_extractor_name must match
|
|
|
|
// because `CheckPrefixMayMatch` first checks `check_filter_ == true`
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
std::unique_ptr<InternalIteratorBase<IndexValue>> iiter(NewIndexIterator(
|
|
|
|
no_io_read_options,
|
|
|
|
/*need_upper_bound_check=*/false, /*input_iter=*/nullptr,
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
/*get_context=*/nullptr, lookup_context));
|
|
|
|
iiter->Seek(internal_prefix);
|
|
|
|
|
|
|
|
if (!iiter->Valid()) {
|
|
|
|
// we're past end of file
|
|
|
|
// if it's incomplete, it means that we avoided I/O
|
|
|
|
// and we're not really sure that we're past the end
|
|
|
|
// of the file
|
|
|
|
may_match = iiter->status().IsIncomplete();
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
} else if ((rep_->index_key_includes_seq ? ExtractUserKey(iiter->key())
|
|
|
|
: iiter->key())
|
|
|
|
.starts_with(ExtractUserKey(internal_prefix))) {
|
|
|
|
// we need to check for this subtle case because our only
|
|
|
|
// guarantee is that "the key is a string >= last key in that data
|
|
|
|
// block" according to the doc/table_format.txt spec.
|
|
|
|
//
|
|
|
|
// Suppose iiter->key() starts with the desired prefix; it is not
|
|
|
|
// necessarily the case that the corresponding data block will
|
|
|
|
// contain the prefix, since iiter->key() need not be in the
|
|
|
|
// block. However, the next data block may contain the prefix, so
|
|
|
|
// we return true to play it safe.
|
|
|
|
may_match = true;
|
|
|
|
} else if (filter->IsBlockBased()) {
|
|
|
|
// iiter->key() does NOT start with the desired prefix. Because
|
|
|
|
// Seek() finds the first key that is >= the seek target, this
|
|
|
|
// means that iiter->key() > prefix. Thus, any data blocks coming
|
|
|
|
// after the data block corresponding to iiter->key() cannot
|
|
|
|
// possibly contain the key. Thus, the corresponding data block
|
|
|
|
// is the only on could potentially contain the prefix.
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
BlockHandle handle = iiter->value().handle;
|
|
|
|
may_match = filter->PrefixMayMatch(
|
|
|
|
prefix, prefix_extractor, handle.offset(), /*no_io=*/false,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
/*const_key_ptr=*/nullptr, /*get_context=*/nullptr, lookup_context);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (filter_checked) {
|
|
|
|
Statistics* statistics = rep_->ioptions.statistics;
|
|
|
|
RecordTick(statistics, BLOOM_FILTER_PREFIX_CHECKED);
|
|
|
|
if (!may_match) {
|
|
|
|
RecordTick(statistics, BLOOM_FILTER_PREFIX_USEFUL);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return may_match;
|
|
|
|
}
|
|
|
|
|
|
|
|
template <class TBlockIter, typename TValue>
|
|
|
|
void BlockBasedTableIterator<TBlockIter, TValue>::Seek(const Slice& target) {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
SeekImpl(&target);
|
|
|
|
}
|
|
|
|
|
|
|
|
template <class TBlockIter, typename TValue>
|
|
|
|
void BlockBasedTableIterator<TBlockIter, TValue>::SeekToFirst() {
|
|
|
|
SeekImpl(nullptr);
|
|
|
|
}
|
|
|
|
|
|
|
|
template <class TBlockIter, typename TValue>
|
|
|
|
void BlockBasedTableIterator<TBlockIter, TValue>::SeekImpl(
|
|
|
|
const Slice* target) {
|
|
|
|
is_out_of_bound_ = false;
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
is_at_first_key_from_index_ = false;
|
|
|
|
if (target && !CheckPrefixMayMatch(*target)) {
|
|
|
|
ResetDataIter();
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool need_seek_index = true;
|
|
|
|
if (block_iter_points_to_real_block_ && block_iter_.Valid()) {
|
|
|
|
// Reseek.
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
prev_block_offset_ = index_iter_->value().handle.offset();
|
|
|
|
|
|
|
|
if (target) {
|
|
|
|
// We can avoid an index seek if:
|
|
|
|
// 1. The new seek key is larger than the current key
|
|
|
|
// 2. The new seek key is within the upper bound of the block
|
|
|
|
// Since we don't necessarily know the internal key for either
|
|
|
|
// the current key or the upper bound, we check user keys and
|
|
|
|
// exclude the equality case. Considering internal keys can
|
|
|
|
// improve for the boundary cases, but it would complicate the
|
|
|
|
// code.
|
|
|
|
if (user_comparator_.Compare(ExtractUserKey(*target),
|
|
|
|
block_iter_.user_key()) > 0 &&
|
|
|
|
user_comparator_.Compare(ExtractUserKey(*target),
|
|
|
|
index_iter_->user_key()) < 0) {
|
|
|
|
need_seek_index = false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (need_seek_index) {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
if (target) {
|
|
|
|
index_iter_->Seek(*target);
|
|
|
|
} else {
|
|
|
|
index_iter_->SeekToFirst();
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!index_iter_->Valid()) {
|
|
|
|
ResetDataIter();
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
IndexValue v = index_iter_->value();
|
|
|
|
const bool same_block = block_iter_points_to_real_block_ &&
|
|
|
|
v.handle.offset() == prev_block_offset_;
|
|
|
|
|
|
|
|
// TODO(kolmike): Remove the != kBlockCacheTier condition.
|
|
|
|
if (!v.first_internal_key.empty() && !same_block &&
|
|
|
|
(!target || icomp_.Compare(*target, v.first_internal_key) <= 0) &&
|
|
|
|
read_options_.read_tier != kBlockCacheTier) {
|
|
|
|
// Index contains the first key of the block, and it's >= target.
|
|
|
|
// We can defer reading the block.
|
|
|
|
is_at_first_key_from_index_ = true;
|
|
|
|
// ResetDataIter() will invalidate block_iter_. Thus, there is no need to
|
|
|
|
// call CheckDataBlockWithinUpperBound() to check for iterate_upper_bound
|
|
|
|
// as that will be done later when the data block is actually read.
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
ResetDataIter();
|
|
|
|
} else {
|
|
|
|
// Need to use the data block.
|
|
|
|
if (!same_block) {
|
|
|
|
InitDataBlock();
|
|
|
|
} else {
|
|
|
|
// When the user does a reseek, the iterate_upper_bound might have
|
|
|
|
// changed. CheckDataBlockWithinUpperBound() needs to be called
|
|
|
|
// explicitly if the reseek ends up in the same data block.
|
|
|
|
// If the reseek ends up in a different block, InitDataBlock() will do
|
|
|
|
// the iterator upper bound check.
|
|
|
|
CheckDataBlockWithinUpperBound();
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
}
|
|
|
|
|
|
|
|
if (target) {
|
|
|
|
block_iter_.Seek(*target);
|
|
|
|
} else {
|
|
|
|
block_iter_.SeekToFirst();
|
|
|
|
}
|
|
|
|
FindKeyForward();
|
|
|
|
}
|
|
|
|
|
|
|
|
CheckOutOfBound();
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
|
|
|
|
if (target) {
|
|
|
|
assert(!Valid() || ((block_type_ == BlockType::kIndex &&
|
|
|
|
!table_->get_rep()->index_key_includes_seq)
|
|
|
|
? (user_comparator_.Compare(ExtractUserKey(*target),
|
|
|
|
key()) <= 0)
|
|
|
|
: (icomp_.Compare(*target, key()) <= 0)));
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
template <class TBlockIter, typename TValue>
|
|
|
|
void BlockBasedTableIterator<TBlockIter, TValue>::SeekForPrev(
|
|
|
|
const Slice& target) {
|
|
|
|
is_out_of_bound_ = false;
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
is_at_first_key_from_index_ = false;
|
|
|
|
if (!CheckPrefixMayMatch(target)) {
|
|
|
|
ResetDataIter();
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
SavePrevIndexValue();
|
|
|
|
|
|
|
|
// Call Seek() rather than SeekForPrev() in the index block, because the
|
|
|
|
// target data block will likely to contain the position for `target`, the
|
|
|
|
// same as Seek(), rather than than before.
|
|
|
|
// For example, if we have three data blocks, each containing two keys:
|
|
|
|
// [2, 4] [6, 8] [10, 12]
|
|
|
|
// (the keys in the index block would be [4, 8, 12])
|
|
|
|
// and the user calls SeekForPrev(7), we need to go to the second block,
|
|
|
|
// just like if they call Seek(7).
|
|
|
|
// The only case where the block is difference is when they seek to a position
|
|
|
|
// in the boundary. For example, if they SeekForPrev(5), we should go to the
|
|
|
|
// first block, rather than the second. However, we don't have the information
|
|
|
|
// to distinguish the two unless we read the second block. In this case, we'll
|
|
|
|
// end up with reading two blocks.
|
|
|
|
index_iter_->Seek(target);
|
|
|
|
|
|
|
|
if (!index_iter_->Valid()) {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
if (!index_iter_->status().ok()) {
|
|
|
|
ResetDataIter();
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
index_iter_->SeekToLast();
|
|
|
|
if (!index_iter_->Valid()) {
|
|
|
|
ResetDataIter();
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
InitDataBlock();
|
|
|
|
|
|
|
|
block_iter_.SeekForPrev(target);
|
|
|
|
|
|
|
|
FindKeyBackward();
|
|
|
|
CheckDataBlockWithinUpperBound();
|
|
|
|
assert(!block_iter_.Valid() ||
|
|
|
|
icomp_.Compare(target, block_iter_.key()) >= 0);
|
|
|
|
}
|
|
|
|
|
|
|
|
template <class TBlockIter, typename TValue>
|
|
|
|
void BlockBasedTableIterator<TBlockIter, TValue>::SeekToLast() {
|
|
|
|
is_out_of_bound_ = false;
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
is_at_first_key_from_index_ = false;
|
|
|
|
SavePrevIndexValue();
|
|
|
|
index_iter_->SeekToLast();
|
|
|
|
if (!index_iter_->Valid()) {
|
|
|
|
ResetDataIter();
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
InitDataBlock();
|
|
|
|
block_iter_.SeekToLast();
|
|
|
|
FindKeyBackward();
|
|
|
|
CheckDataBlockWithinUpperBound();
|
|
|
|
}
|
|
|
|
|
|
|
|
template <class TBlockIter, typename TValue>
|
|
|
|
void BlockBasedTableIterator<TBlockIter, TValue>::Next() {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
if (is_at_first_key_from_index_ && !MaterializeCurrentBlock()) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
assert(block_iter_points_to_real_block_);
|
|
|
|
block_iter_.Next();
|
|
|
|
FindKeyForward();
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
CheckOutOfBound();
|
|
|
|
}
|
|
|
|
|
|
|
|
template <class TBlockIter, typename TValue>
|
|
|
|
bool BlockBasedTableIterator<TBlockIter, TValue>::NextAndGetResult(
|
|
|
|
IterateResult* result) {
|
|
|
|
Next();
|
|
|
|
bool is_valid = Valid();
|
|
|
|
if (is_valid) {
|
|
|
|
result->key = key();
|
|
|
|
result->may_be_out_of_upper_bound = MayBeOutOfUpperBound();
|
|
|
|
}
|
|
|
|
return is_valid;
|
|
|
|
}
|
|
|
|
|
|
|
|
template <class TBlockIter, typename TValue>
|
|
|
|
void BlockBasedTableIterator<TBlockIter, TValue>::Prev() {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
if (is_at_first_key_from_index_) {
|
|
|
|
is_at_first_key_from_index_ = false;
|
|
|
|
|
|
|
|
index_iter_->Prev();
|
|
|
|
if (!index_iter_->Valid()) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
InitDataBlock();
|
|
|
|
block_iter_.SeekToLast();
|
|
|
|
} else {
|
|
|
|
assert(block_iter_points_to_real_block_);
|
|
|
|
block_iter_.Prev();
|
|
|
|
}
|
|
|
|
|
|
|
|
FindKeyBackward();
|
|
|
|
}
|
|
|
|
|
|
|
|
template <class TBlockIter, typename TValue>
|
|
|
|
void BlockBasedTableIterator<TBlockIter, TValue>::InitDataBlock() {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
BlockHandle data_block_handle = index_iter_->value().handle;
|
|
|
|
if (!block_iter_points_to_real_block_ ||
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
data_block_handle.offset() != prev_block_offset_ ||
|
Change and clarify the relationship between Valid(), status() and Seek*() for all iterators. Also fix some bugs
Summary:
Before this PR, Iterator/InternalIterator may simultaneously have non-ok status() and Valid() = true. That state means that the last operation failed, but the iterator is nevertheless positioned on some unspecified record. Likely intended uses of that are:
* If some sst files are corrupted, a normal iterator can be used to read the data from files that are not corrupted.
* When using read_tier = kBlockCacheTier, read the data that's in block cache, skipping over the data that is not.
However, this behavior wasn't documented well (and until recently the wiki on github had misleading incorrect information). In the code there's a lot of confusion about the relationship between status() and Valid(), and about whether Seek()/SeekToLast()/etc reset the status or not. There were a number of bugs caused by this confusion, both inside rocksdb and in the code that uses rocksdb (including ours).
This PR changes the convention to:
* If status() is not ok, Valid() always returns false.
* Any seek operation resets status. (Before the PR, it depended on iterator type and on particular error.)
This does sacrifice the two use cases listed above, but siying said it's ok.
Overview of the changes:
* A commit that adds missing status checks in MergingIterator. This fixes a bug that actually affects us, and we need it fixed. `DBIteratorTest.NonBlockingIterationBugRepro` explains the scenario.
* Changes to lots of iterator types to make all of them conform to the new convention. Some bug fixes along the way. By far the biggest changes are in DBIter, which is a big messy piece of code; I tried to make it less big and messy but mostly failed.
* A stress-test for DBIter, to gain some confidence that I didn't break it. It does a few million random operations on the iterator, while occasionally modifying the underlying data (like ForwardIterator does) and occasionally returning non-ok status from internal iterator.
To find the iterator types that needed changes I searched for "public .*Iterator" in the code. Here's an overview of all 27 iterator types:
Iterators that didn't need changes:
* status() is always ok(), or Valid() is always false: MemTableIterator, ModelIter, TestIterator, KVIter (2 classes with this name anonymous namespaces), LoggingForwardVectorIterator, VectorIterator, MockTableIterator, EmptyIterator, EmptyInternalIterator.
* Thin wrappers that always pass through Valid() and status(): ArenaWrappedDBIter, TtlIterator, InternalIteratorFromIterator.
Iterators with changes (see inline comments for details):
* DBIter - an overhaul:
- It used to silently skip corrupted keys (`FindParseableKey()`), which seems dangerous. This PR makes it just stop immediately after encountering a corrupted key, just like it would for other kinds of corruption. Let me know if there was actually some deeper meaning in this behavior and I should put it back.
- It had a few code paths silently discarding subiterator's status. The stress test caught a few.
- The backwards iteration code path was expecting the internal iterator's set of keys to be immutable. It's probably always true in practice at the moment, since ForwardIterator doesn't support backwards iteration, but this PR fixes it anyway. See added DBIteratorTest.ReverseToForwardBug for an example.
- Some parts of backwards iteration code path even did things like `assert(iter_->Valid())` after a seek, which is never a safe assumption.
- It used to not reset status on seek for some types of errors.
- Some simplifications and better comments.
- Some things got more complicated from the added error handling. I'm open to ideas for how to make it nicer.
* MergingIterator - check status after every operation on every subiterator, and in some places assert that valid subiterators have ok status.
* ForwardIterator - changed to the new convention, also slightly simplified.
* ForwardLevelIterator - fixed some bugs and simplified.
* LevelIterator - simplified.
* TwoLevelIterator - changed to the new convention. Also fixed a bug that would make SeekForPrev() sometimes silently ignore errors from first_level_iter_.
* BlockBasedTableIterator - minor changes.
* BlockIter - replaced `SetStatus()` with `Invalidate()` to make sure non-ok BlockIter is always invalid.
* PlainTableIterator - some seeks used to not reset status.
* CuckooTableIterator - tiny code cleanup.
* ManagedIterator - fixed some bugs.
* BaseDeltaIterator - changed to the new convention and fixed a bug.
* BlobDBIterator - seeks used to not reset status.
* KeyConvertingIterator - some small change.
Closes https://github.com/facebook/rocksdb/pull/3810
Differential Revision: D7888019
Pulled By: al13n321
fbshipit-source-id: 4aaf6d3421c545d16722a815b2fa2e7912bc851d
7 years ago
|
|
|
// if previous attempt of reading the block missed cache, try again
|
|
|
|
block_iter_.status().IsIncomplete()) {
|
|
|
|
if (block_iter_points_to_real_block_) {
|
|
|
|
ResetDataIter();
|
|
|
|
}
|
|
|
|
auto* rep = table_->get_rep();
|
|
|
|
|
|
|
|
// Prefetch additional data for range scans (iterators). Enabled only for
|
|
|
|
// user reads.
|
|
|
|
// Implicit auto readahead:
|
|
|
|
// Enabled after 2 sequential IOs when ReadOptions.readahead_size == 0.
|
|
|
|
// Explicit user requested readahead:
|
|
|
|
// Enabled from the very first IO when ReadOptions.readahead_size is set.
|
|
|
|
if (lookup_context_.caller != TableReaderCaller::kCompaction) {
|
|
|
|
if (read_options_.readahead_size == 0) {
|
|
|
|
// Implicit auto readahead
|
|
|
|
num_file_reads_++;
|
|
|
|
if (num_file_reads_ >
|
|
|
|
BlockBasedTable::kMinNumFileReadsToStartAutoReadahead) {
|
|
|
|
if (!rep->file->use_direct_io() &&
|
|
|
|
(data_block_handle.offset() +
|
|
|
|
static_cast<size_t>(data_block_handle.size()) +
|
|
|
|
kBlockTrailerSize >
|
|
|
|
readahead_limit_)) {
|
|
|
|
// Buffered I/O
|
|
|
|
// Discarding the return status of Prefetch calls intentionally, as
|
|
|
|
// we can fallback to reading from disk if Prefetch fails.
|
|
|
|
rep->file->Prefetch(data_block_handle.offset(), readahead_size_);
|
|
|
|
readahead_limit_ = static_cast<size_t>(data_block_handle.offset() +
|
|
|
|
readahead_size_);
|
|
|
|
// Keep exponentially increasing readahead size until
|
|
|
|
// kMaxAutoReadaheadSize.
|
|
|
|
readahead_size_ = std::min(BlockBasedTable::kMaxAutoReadaheadSize,
|
|
|
|
readahead_size_ * 2);
|
|
|
|
} else if (rep->file->use_direct_io() && !prefetch_buffer_) {
|
|
|
|
// Direct I/O
|
|
|
|
// Let FilePrefetchBuffer take care of the readahead.
|
|
|
|
prefetch_buffer_.reset(new FilePrefetchBuffer(
|
|
|
|
rep->file.get(), BlockBasedTable::kInitAutoReadaheadSize,
|
|
|
|
BlockBasedTable::kMaxAutoReadaheadSize));
|
|
|
|
}
|
Improve direct IO range scan performance with readahead (#3884)
Summary:
This PR extends the improvements in #3282 to also work when using Direct IO.
We see **4.5X performance improvement** in seekrandom benchmark doing long range scans, when using direct reads, on flash.
**Description:**
This change improves the performance of iterators doing long range scans (e.g. big/full index or table scans in MyRocks) by using readahead and prefetching additional data on each disk IO, and storing in a local buffer. This prefetching is automatically enabled on noticing more than 2 IOs for the same table file during iteration. The readahead size starts with 8KB and is exponentially increased on each additional sequential IO, up to a max of 256 KB. This helps in cutting down the number of IOs needed to complete the range scan.
**Implementation Details:**
- Used `FilePrefetchBuffer` as the underlying buffer to store the readahead data. `FilePrefetchBuffer` can now take file_reader, readahead_size and max_readahead_size as input to the constructor, and automatically do readahead.
- `FilePrefetchBuffer::TryReadFromCache` can now call `FilePrefetchBuffer::Prefetch` if readahead is enabled.
- `AlignedBuffer` (which is the underlying store for `FilePrefetchBuffer`) now takes a few additional args in `AlignedBuffer::AllocateNewBuffer` to allow copying data from the old buffer.
- Made sure not to re-read partial chunks of data that were already available in the buffer, from device again.
- Fixed a couple of cases where `AlignedBuffer::cursize_` was not being properly kept up-to-date.
**Constraints:**
- Similar to #3282, this gets currently enabled only when ReadOptions.readahead_size = 0 (which is the default value).
- Since the prefetched data is stored in a temporary buffer allocated on heap, this could increase the memory usage if you have many iterators doing long range scans simultaneously.
- Enabled only for user reads, and disabled for compactions. Compaction reads are controlled by the options `use_direct_io_for_flush_and_compaction` and `compaction_readahead_size`, and the current feature takes precautions not to mess with them.
**Benchmarks:**
I used the same benchmark as used in #3282.
Data fill:
```
TEST_TMPDIR=/data/users/$USER/benchmarks/iter ./db_bench -benchmarks=fillrandom -num=1000000000 -compression_type="none" -level_compaction_dynamic_level_bytes
```
Do a long range scan: Seekrandom with large number of nexts
```
TEST_TMPDIR=/data/users/$USER/benchmarks/iter ./db_bench -benchmarks=seekrandom -use_direct_reads -duration=60 -num=1000000000 -use_existing_db -seek_nexts=10000 -statistics -histogram
```
```
Before:
seekrandom : 37939.906 micros/op 26 ops/sec; 29.2 MB/s (1636 of 1999 found)
With this change:
seekrandom : 8527.720 micros/op 117 ops/sec; 129.7 MB/s (6530 of 7999 found)
```
~4.5X perf improvement. Taken on an average of 3 runs.
Closes https://github.com/facebook/rocksdb/pull/3884
Differential Revision: D8082143
Pulled By: sagar0
fbshipit-source-id: 4d7a8561cbac03478663713df4d31ad2620253bb
7 years ago
|
|
|
}
|
|
|
|
} else if (!prefetch_buffer_) {
|
|
|
|
// Explicit user requested readahead
|
|
|
|
// The actual condition is:
|
|
|
|
// if (read_options_.readahead_size != 0 && !prefetch_buffer_)
|
|
|
|
prefetch_buffer_.reset(new FilePrefetchBuffer(
|
|
|
|
rep->file.get(), read_options_.readahead_size,
|
|
|
|
read_options_.readahead_size));
|
|
|
|
}
|
|
|
|
} else if (!prefetch_buffer_) {
|
|
|
|
prefetch_buffer_.reset(
|
|
|
|
new FilePrefetchBuffer(rep->file.get(), compaction_readahead_size_,
|
|
|
|
compaction_readahead_size_));
|
|
|
|
}
|
|
|
|
|
|
|
|
Status s;
|
|
|
|
table_->NewDataBlockIterator<TBlockIter>(
|
|
|
|
read_options_, data_block_handle, &block_iter_, block_type_,
|
|
|
|
/*get_context=*/nullptr, &lookup_context_, s, prefetch_buffer_.get(),
|
|
|
|
/*for_compaction=*/lookup_context_.caller ==
|
|
|
|
TableReaderCaller::kCompaction);
|
|
|
|
block_iter_points_to_real_block_ = true;
|
|
|
|
CheckDataBlockWithinUpperBound();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
template <class TBlockIter, typename TValue>
|
|
|
|
bool BlockBasedTableIterator<TBlockIter, TValue>::MaterializeCurrentBlock() {
|
|
|
|
assert(is_at_first_key_from_index_);
|
|
|
|
assert(!block_iter_points_to_real_block_);
|
|
|
|
assert(index_iter_->Valid());
|
|
|
|
|
|
|
|
is_at_first_key_from_index_ = false;
|
|
|
|
InitDataBlock();
|
|
|
|
assert(block_iter_points_to_real_block_);
|
|
|
|
block_iter_.SeekToFirst();
|
|
|
|
|
|
|
|
if (!block_iter_.Valid() ||
|
|
|
|
icomp_.Compare(block_iter_.key(),
|
|
|
|
index_iter_->value().first_internal_key) != 0) {
|
|
|
|
// Uh oh.
|
|
|
|
block_iter_.Invalidate(Status::Corruption(
|
|
|
|
"first key in index doesn't match first key in block"));
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
template <class TBlockIter, typename TValue>
|
|
|
|
void BlockBasedTableIterator<TBlockIter, TValue>::FindKeyForward() {
|
|
|
|
// This method's code is kept short to make it likely to be inlined.
|
|
|
|
|
|
|
|
assert(!is_out_of_bound_);
|
|
|
|
assert(block_iter_points_to_real_block_);
|
|
|
|
|
|
|
|
if (!block_iter_.Valid()) {
|
|
|
|
// This is the only call site of FindBlockForward(), but it's extracted into
|
|
|
|
// a separate method to keep FindKeyForward() short and likely to be
|
|
|
|
// inlined. When transitioning to a different block, we call
|
|
|
|
// FindBlockForward(), which is much longer and is probably not inlined.
|
|
|
|
FindBlockForward();
|
|
|
|
} else {
|
|
|
|
// This is the fast path that avoids a function call.
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
template <class TBlockIter, typename TValue>
|
|
|
|
void BlockBasedTableIterator<TBlockIter, TValue>::FindBlockForward() {
|
|
|
|
// TODO the while loop inherits from two-level-iterator. We don't know
|
|
|
|
// whether a block can be empty so it can be replaced by an "if".
|
|
|
|
do {
|
|
|
|
if (!block_iter_.status().ok()) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
// Whether next data block is out of upper bound, if there is one.
|
|
|
|
const bool next_block_is_out_of_bound =
|
|
|
|
read_options_.iterate_upper_bound != nullptr &&
|
|
|
|
block_iter_points_to_real_block_ && !data_block_within_upper_bound_;
|
|
|
|
assert(!next_block_is_out_of_bound ||
|
|
|
|
user_comparator_.Compare(*read_options_.iterate_upper_bound,
|
|
|
|
index_iter_->user_key()) <= 0);
|
|
|
|
ResetDataIter();
|
|
|
|
index_iter_->Next();
|
|
|
|
if (next_block_is_out_of_bound) {
|
|
|
|
// The next block is out of bound. No need to read it.
|
|
|
|
TEST_SYNC_POINT_CALLBACK("BlockBasedTableIterator:out_of_bound", nullptr);
|
|
|
|
// We need to make sure this is not the last data block before setting
|
|
|
|
// is_out_of_bound_, since the index key for the last data block can be
|
|
|
|
// larger than smallest key of the next file on the same level.
|
|
|
|
if (index_iter_->Valid()) {
|
|
|
|
is_out_of_bound_ = true;
|
|
|
|
}
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
if (!index_iter_->Valid()) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
IndexValue v = index_iter_->value();
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
// TODO(kolmike): Remove the != kBlockCacheTier condition.
|
|
|
|
if (!v.first_internal_key.empty() &&
|
|
|
|
read_options_.read_tier != kBlockCacheTier) {
|
|
|
|
// Index contains the first key of the block. Defer reading the block.
|
|
|
|
is_at_first_key_from_index_ = true;
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
InitDataBlock();
|
|
|
|
block_iter_.SeekToFirst();
|
|
|
|
} while (!block_iter_.Valid());
|
|
|
|
}
|
|
|
|
|
|
|
|
template <class TBlockIter, typename TValue>
|
|
|
|
void BlockBasedTableIterator<TBlockIter, TValue>::FindKeyBackward() {
|
|
|
|
while (!block_iter_.Valid()) {
|
|
|
|
if (!block_iter_.status().ok()) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
ResetDataIter();
|
|
|
|
index_iter_->Prev();
|
|
|
|
|
|
|
|
if (index_iter_->Valid()) {
|
|
|
|
InitDataBlock();
|
|
|
|
block_iter_.SeekToLast();
|
|
|
|
} else {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// We could have check lower bound here too, but we opt not to do it for
|
|
|
|
// code simplicity.
|
|
|
|
}
|
|
|
|
|
|
|
|
template <class TBlockIter, typename TValue>
|
|
|
|
void BlockBasedTableIterator<TBlockIter, TValue>::CheckOutOfBound() {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
if (read_options_.iterate_upper_bound != nullptr && Valid()) {
|
|
|
|
is_out_of_bound_ = user_comparator_.Compare(
|
|
|
|
*read_options_.iterate_upper_bound, user_key()) <= 0;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
template <class TBlockIter, typename TValue>
|
|
|
|
void BlockBasedTableIterator<TBlockIter,
|
|
|
|
TValue>::CheckDataBlockWithinUpperBound() {
|
|
|
|
if (read_options_.iterate_upper_bound != nullptr &&
|
|
|
|
block_iter_points_to_real_block_) {
|
|
|
|
data_block_within_upper_bound_ =
|
|
|
|
(user_comparator_.Compare(*read_options_.iterate_upper_bound,
|
|
|
|
index_iter_->user_key()) > 0);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
InternalIterator* BlockBasedTable::NewIterator(
|
|
|
|
const ReadOptions& read_options, const SliceTransform* prefix_extractor,
|
|
|
|
Arena* arena, bool skip_filters, TableReaderCaller caller,
|
|
|
|
size_t compaction_readahead_size) {
|
|
|
|
BlockCacheLookupContext lookup_context{caller};
|
|
|
|
bool need_upper_bound_check =
|
|
|
|
PrefixExtractorChanged(rep_->table_properties.get(), prefix_extractor);
|
|
|
|
if (arena == nullptr) {
|
|
|
|
return new BlockBasedTableIterator<DataBlockIter>(
|
|
|
|
this, read_options, rep_->internal_comparator,
|
|
|
|
NewIndexIterator(
|
|
|
|
read_options,
|
|
|
|
need_upper_bound_check &&
|
|
|
|
rep_->index_type == BlockBasedTableOptions::kHashSearch,
|
|
|
|
/*input_iter=*/nullptr, /*get_context=*/nullptr, &lookup_context),
|
|
|
|
!skip_filters && !read_options.total_order_seek &&
|
|
|
|
prefix_extractor != nullptr,
|
|
|
|
need_upper_bound_check, prefix_extractor, BlockType::kData, caller,
|
|
|
|
compaction_readahead_size);
|
|
|
|
} else {
|
|
|
|
auto* mem =
|
|
|
|
arena->AllocateAligned(sizeof(BlockBasedTableIterator<DataBlockIter>));
|
|
|
|
return new (mem) BlockBasedTableIterator<DataBlockIter>(
|
|
|
|
this, read_options, rep_->internal_comparator,
|
|
|
|
NewIndexIterator(
|
|
|
|
read_options,
|
|
|
|
need_upper_bound_check &&
|
|
|
|
rep_->index_type == BlockBasedTableOptions::kHashSearch,
|
|
|
|
/*input_iter=*/nullptr, /*get_context=*/nullptr, &lookup_context),
|
|
|
|
!skip_filters && !read_options.total_order_seek &&
|
|
|
|
prefix_extractor != nullptr,
|
|
|
|
need_upper_bound_check, prefix_extractor, BlockType::kData, caller,
|
|
|
|
compaction_readahead_size);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
FragmentedRangeTombstoneIterator* BlockBasedTable::NewRangeTombstoneIterator(
|
|
|
|
const ReadOptions& read_options) {
|
Cache fragmented range tombstones in BlockBasedTableReader (#4493)
Summary:
This allows tombstone fragmenting to only be performed when the table is opened, and cached for subsequent accesses.
On the same DB used in #4449, running `readrandom` results in the following:
```
readrandom : 0.983 micros/op 1017076 ops/sec; 78.3 MB/s (63103 of 100000 found)
```
Now that Get performance in the presence of range tombstones is reasonable, I also compared the performance between a DB with range tombstones, "expanded" range tombstones (several point tombstones that cover the same keys the equivalent range tombstone would cover, a common workaround for DeleteRange), and no range tombstones. The created DBs had 5 million keys each, and DeleteRange was called at regular intervals (depending on the total number of range tombstones being written) after 4.5 million Puts. The table below summarizes the results of a `readwhilewriting` benchmark (in order to provide somewhat more realistic results):
```
Tombstones? | avg micros/op | stddev micros/op | avg ops/s | stddev ops/s
----------------- | ------------- | ---------------- | ------------ | ------------
None | 0.6186 | 0.04637 | 1,625,252.90 | 124,679.41
500 Expanded | 0.6019 | 0.03628 | 1,666,670.40 | 101,142.65
500 Unexpanded | 0.6435 | 0.03994 | 1,559,979.40 | 104,090.52
1k Expanded | 0.6034 | 0.04349 | 1,665,128.10 | 125,144.57
1k Unexpanded | 0.6261 | 0.03093 | 1,600,457.50 | 79,024.94
5k Expanded | 0.6163 | 0.05926 | 1,636,668.80 | 154,888.85
5k Unexpanded | 0.6402 | 0.04002 | 1,567,804.70 | 100,965.55
10k Expanded | 0.6036 | 0.05105 | 1,667,237.70 | 142,830.36
10k Unexpanded | 0.6128 | 0.02598 | 1,634,633.40 | 72,161.82
25k Expanded | 0.6198 | 0.04542 | 1,620,980.50 | 116,662.93
25k Unexpanded | 0.5478 | 0.0362 | 1,833,059.10 | 121,233.81
50k Expanded | 0.5104 | 0.04347 | 1,973,107.90 | 184,073.49
50k Unexpanded | 0.4528 | 0.03387 | 2,219,034.50 | 170,984.32
```
After a large enough quantity of range tombstones are written, range tombstone Gets can become faster than reading from an equivalent DB with several point tombstones.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4493
Differential Revision: D10842844
Pulled By: abhimadan
fbshipit-source-id: a7d44534f8120e6aabb65779d26c6b9df954c509
6 years ago
|
|
|
if (rep_->fragmented_range_dels == nullptr) {
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
SequenceNumber snapshot = kMaxSequenceNumber;
|
|
|
|
if (read_options.snapshot != nullptr) {
|
|
|
|
snapshot = read_options.snapshot->GetSequenceNumber();
|
|
|
|
}
|
|
|
|
return new FragmentedRangeTombstoneIterator(
|
|
|
|
rep_->fragmented_range_dels, rep_->internal_comparator, snapshot);
|
Cache fragmented range tombstones in BlockBasedTableReader (#4493)
Summary:
This allows tombstone fragmenting to only be performed when the table is opened, and cached for subsequent accesses.
On the same DB used in #4449, running `readrandom` results in the following:
```
readrandom : 0.983 micros/op 1017076 ops/sec; 78.3 MB/s (63103 of 100000 found)
```
Now that Get performance in the presence of range tombstones is reasonable, I also compared the performance between a DB with range tombstones, "expanded" range tombstones (several point tombstones that cover the same keys the equivalent range tombstone would cover, a common workaround for DeleteRange), and no range tombstones. The created DBs had 5 million keys each, and DeleteRange was called at regular intervals (depending on the total number of range tombstones being written) after 4.5 million Puts. The table below summarizes the results of a `readwhilewriting` benchmark (in order to provide somewhat more realistic results):
```
Tombstones? | avg micros/op | stddev micros/op | avg ops/s | stddev ops/s
----------------- | ------------- | ---------------- | ------------ | ------------
None | 0.6186 | 0.04637 | 1,625,252.90 | 124,679.41
500 Expanded | 0.6019 | 0.03628 | 1,666,670.40 | 101,142.65
500 Unexpanded | 0.6435 | 0.03994 | 1,559,979.40 | 104,090.52
1k Expanded | 0.6034 | 0.04349 | 1,665,128.10 | 125,144.57
1k Unexpanded | 0.6261 | 0.03093 | 1,600,457.50 | 79,024.94
5k Expanded | 0.6163 | 0.05926 | 1,636,668.80 | 154,888.85
5k Unexpanded | 0.6402 | 0.04002 | 1,567,804.70 | 100,965.55
10k Expanded | 0.6036 | 0.05105 | 1,667,237.70 | 142,830.36
10k Unexpanded | 0.6128 | 0.02598 | 1,634,633.40 | 72,161.82
25k Expanded | 0.6198 | 0.04542 | 1,620,980.50 | 116,662.93
25k Unexpanded | 0.5478 | 0.0362 | 1,833,059.10 | 121,233.81
50k Expanded | 0.5104 | 0.04347 | 1,973,107.90 | 184,073.49
50k Unexpanded | 0.4528 | 0.03387 | 2,219,034.50 | 170,984.32
```
After a large enough quantity of range tombstones are written, range tombstone Gets can become faster than reading from an equivalent DB with several point tombstones.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4493
Differential Revision: D10842844
Pulled By: abhimadan
fbshipit-source-id: a7d44534f8120e6aabb65779d26c6b9df954c509
6 years ago
|
|
|
}
|
|
|
|
|
|
|
|
bool BlockBasedTable::FullFilterKeyMayMatch(
|
|
|
|
const ReadOptions& read_options, FilterBlockReader* filter,
|
|
|
|
const Slice& internal_key, const bool no_io,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
const SliceTransform* prefix_extractor, GetContext* get_context,
|
|
|
|
BlockCacheLookupContext* lookup_context) const {
|
|
|
|
if (filter == nullptr || filter->IsBlockBased()) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
Slice user_key = ExtractUserKey(internal_key);
|
|
|
|
const Slice* const const_ikey_ptr = &internal_key;
|
|
|
|
bool may_match = true;
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
if (rep_->whole_key_filtering) {
|
|
|
|
size_t ts_sz =
|
|
|
|
rep_->internal_comparator.user_comparator()->timestamp_size();
|
|
|
|
Slice user_key_without_ts = StripTimestampFromUserKey(user_key, ts_sz);
|
|
|
|
may_match =
|
|
|
|
filter->KeyMayMatch(user_key_without_ts, prefix_extractor, kNotValid,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
no_io, const_ikey_ptr, get_context, lookup_context);
|
|
|
|
} else if (!read_options.total_order_seek && prefix_extractor &&
|
|
|
|
rep_->table_properties->prefix_extractor_name.compare(
|
|
|
|
prefix_extractor->Name()) == 0 &&
|
|
|
|
prefix_extractor->InDomain(user_key) &&
|
|
|
|
!filter->PrefixMayMatch(prefix_extractor->Transform(user_key),
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
prefix_extractor, kNotValid, no_io,
|
|
|
|
const_ikey_ptr, get_context,
|
|
|
|
lookup_context)) {
|
|
|
|
may_match = false;
|
|
|
|
}
|
|
|
|
if (may_match) {
|
|
|
|
RecordTick(rep_->ioptions.statistics, BLOOM_FILTER_FULL_POSITIVE);
|
|
|
|
PERF_COUNTER_BY_LEVEL_ADD(bloom_filter_full_positive, 1, rep_->level);
|
|
|
|
}
|
|
|
|
return may_match;
|
|
|
|
}
|
|
|
|
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
void BlockBasedTable::FullFilterKeysMayMatch(
|
|
|
|
const ReadOptions& read_options, FilterBlockReader* filter,
|
|
|
|
MultiGetRange* range, const bool no_io,
|
|
|
|
const SliceTransform* prefix_extractor,
|
|
|
|
BlockCacheLookupContext* lookup_context) const {
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
if (filter == nullptr || filter->IsBlockBased()) {
|
|
|
|
return;
|
|
|
|
}
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
if (rep_->whole_key_filtering) {
|
|
|
|
filter->KeysMayMatch(range, prefix_extractor, kNotValid, no_io,
|
|
|
|
lookup_context);
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
} else if (!read_options.total_order_seek && prefix_extractor &&
|
|
|
|
rep_->table_properties->prefix_extractor_name.compare(
|
|
|
|
prefix_extractor->Name()) == 0) {
|
|
|
|
filter->PrefixesMayMatch(range, prefix_extractor, kNotValid, false,
|
|
|
|
lookup_context);
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Skip bottom-level filter block caching when hit-optimized
Summary:
When Get() or NewIterator() trigger file loads, skip caching the filter block if
(1) optimize_filters_for_hits is set and (2) the file is on the bottommost
level. Also skip checking filters under the same conditions, which means that
for a preloaded file or a file that was trivially-moved to the bottom level, its
filter block will eventually expire from the cache.
- added parameters/instance variables in various places in order to propagate the config ("skip_filters") from version_set to block_based_table_reader
- in BlockBasedTable::Rep, this optimization prevents filter from being loaded when the file is opened simply by setting filter_policy = nullptr
- in BlockBasedTable::Get/BlockBasedTable::NewIterator, this optimization prevents filter from being used (even if it was loaded already) by setting filter = nullptr
Test Plan:
updated unit test:
$ ./db_test --gtest_filter=DBTest.OptimizeFiltersForHits
will also run 'make check'
Reviewers: sdong, igor, paultuckfield, anthony, rven, kradhakrishnan, IslamAbdelRahman, yhchiang
Reviewed By: yhchiang
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D51633
9 years ago
|
|
|
Status BlockBasedTable::Get(const ReadOptions& read_options, const Slice& key,
|
|
|
|
GetContext* get_context,
|
|
|
|
const SliceTransform* prefix_extractor,
|
|
|
|
bool skip_filters) {
|
|
|
|
assert(key.size() >= 8); // key must be internal key
|
|
|
|
assert(get_context != nullptr);
|
|
|
|
Status s;
|
|
|
|
const bool no_io = read_options.read_tier == kBlockCacheTier;
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
|
|
|
|
FilterBlockReader* const filter =
|
|
|
|
!skip_filters ? rep_->filter.get() : nullptr;
|
|
|
|
|
|
|
|
// First check the full filter
|
|
|
|
// If full filter not useful, Then go into each block
|
|
|
|
uint64_t tracing_get_id = get_context->get_tracing_get_id();
|
|
|
|
BlockCacheLookupContext lookup_context{
|
|
|
|
TableReaderCaller::kUserGet, tracing_get_id,
|
|
|
|
/*get_from_user_specified_snapshot=*/read_options.snapshot != nullptr};
|
|
|
|
if (block_cache_tracer_ && block_cache_tracer_->is_tracing_enabled()) {
|
|
|
|
// Trace the key since it contains both user key and sequence number.
|
|
|
|
lookup_context.referenced_key = key.ToString();
|
|
|
|
lookup_context.get_from_user_specified_snapshot =
|
|
|
|
read_options.snapshot != nullptr;
|
|
|
|
}
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
const bool may_match =
|
|
|
|
FullFilterKeyMayMatch(read_options, filter, key, no_io, prefix_extractor,
|
|
|
|
get_context, &lookup_context);
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
if (!may_match) {
|
|
|
|
RecordTick(rep_->ioptions.statistics, BLOOM_FILTER_USEFUL);
|
|
|
|
PERF_COUNTER_BY_LEVEL_ADD(bloom_filter_useful, 1, rep_->level);
|
|
|
|
} else {
|
|
|
|
IndexBlockIter iiter_on_stack;
|
|
|
|
// if prefix_extractor found in block differs from options, disable
|
|
|
|
// BlockPrefixIndex. Only do this check when index_type is kHashSearch.
|
|
|
|
bool need_upper_bound_check = false;
|
|
|
|
if (rep_->index_type == BlockBasedTableOptions::kHashSearch) {
|
|
|
|
need_upper_bound_check = PrefixExtractorChanged(
|
|
|
|
rep_->table_properties.get(), prefix_extractor);
|
|
|
|
}
|
|
|
|
auto iiter =
|
|
|
|
NewIndexIterator(read_options, need_upper_bound_check, &iiter_on_stack,
|
|
|
|
get_context, &lookup_context);
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
std::unique_ptr<InternalIteratorBase<IndexValue>> iiter_unique_ptr;
|
|
|
|
if (iiter != &iiter_on_stack) {
|
|
|
|
iiter_unique_ptr.reset(iiter);
|
|
|
|
}
|
|
|
|
|
|
|
|
size_t ts_sz =
|
|
|
|
rep_->internal_comparator.user_comparator()->timestamp_size();
|
|
|
|
bool matched = false; // if such user key mathced a key in SST
|
|
|
|
bool done = false;
|
|
|
|
for (iiter->Seek(key); iiter->Valid() && !done; iiter->Next()) {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
IndexValue v = iiter->value();
|
|
|
|
|
|
|
|
bool not_exist_in_filter =
|
|
|
|
filter != nullptr && filter->IsBlockBased() == true &&
|
|
|
|
!filter->KeyMayMatch(ExtractUserKeyAndStripTimestamp(key, ts_sz),
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
prefix_extractor, v.handle.offset(), no_io,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
/*const_ikey_ptr=*/nullptr, get_context,
|
|
|
|
&lookup_context);
|
|
|
|
|
|
|
|
if (not_exist_in_filter) {
|
|
|
|
// Not found
|
|
|
|
// TODO: think about interaction with Merge. If a user key cannot
|
|
|
|
// cross one data block, we should be fine.
|
|
|
|
RecordTick(rep_->ioptions.statistics, BLOOM_FILTER_USEFUL);
|
|
|
|
PERF_COUNTER_BY_LEVEL_ADD(bloom_filter_useful, 1, rep_->level);
|
|
|
|
break;
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
}
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
if (!v.first_internal_key.empty() && !skip_filters &&
|
|
|
|
UserComparatorWrapper(rep_->internal_comparator.user_comparator())
|
|
|
|
.Compare(ExtractUserKey(key),
|
|
|
|
ExtractUserKey(v.first_internal_key)) < 0) {
|
|
|
|
// The requested key falls between highest key in previous block and
|
|
|
|
// lowest key in current block.
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
BlockCacheLookupContext lookup_data_block_context{
|
|
|
|
TableReaderCaller::kUserGet, tracing_get_id,
|
|
|
|
/*get_from_user_specified_snapshot=*/read_options.snapshot !=
|
|
|
|
nullptr};
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
bool does_referenced_key_exist = false;
|
|
|
|
DataBlockIter biter;
|
|
|
|
uint64_t referenced_data_size = 0;
|
|
|
|
NewDataBlockIterator<DataBlockIter>(
|
|
|
|
read_options, v.handle, &biter, BlockType::kData, get_context,
|
|
|
|
&lookup_data_block_context,
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
/*s=*/Status(), /*prefetch_buffer*/ nullptr);
|
|
|
|
|
|
|
|
if (no_io && biter.status().IsIncomplete()) {
|
|
|
|
// couldn't get block from block_cache
|
|
|
|
// Update Saver.state to Found because we are only looking for
|
|
|
|
// whether we can guarantee the key is not there when "no_io" is set
|
|
|
|
get_context->MarkKeyMayExist();
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
if (!biter.status().ok()) {
|
|
|
|
s = biter.status();
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
bool may_exist = biter.SeekForGet(key);
|
|
|
|
// If user-specified timestamp is supported, we cannot end the search
|
|
|
|
// just because hash index lookup indicates the key+ts does not exist.
|
|
|
|
if (!may_exist && ts_sz == 0) {
|
|
|
|
// HashSeek cannot find the key this block and the the iter is not
|
|
|
|
// the end of the block, i.e. cannot be in the following blocks
|
|
|
|
// either. In this case, the seek_key cannot be found, so we break
|
|
|
|
// from the top level for-loop.
|
|
|
|
done = true;
|
|
|
|
} else {
|
|
|
|
// Call the *saver function on each entry/block until it returns false
|
|
|
|
for (; biter.Valid(); biter.Next()) {
|
|
|
|
ParsedInternalKey parsed_key;
|
|
|
|
if (!ParseInternalKey(biter.key(), &parsed_key)) {
|
|
|
|
s = Status::Corruption(Slice());
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!get_context->SaveValue(
|
|
|
|
parsed_key, biter.value(), &matched,
|
|
|
|
biter.IsValuePinned() ? &biter : nullptr)) {
|
|
|
|
if (get_context->State() == GetContext::GetState::kFound) {
|
|
|
|
does_referenced_key_exist = true;
|
|
|
|
referenced_data_size = biter.key().size() + biter.value().size();
|
|
|
|
}
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
done = true;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
s = biter.status();
|
|
|
|
}
|
|
|
|
// Write the block cache access record.
|
|
|
|
if (block_cache_tracer_ && block_cache_tracer_->is_tracing_enabled()) {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
// Avoid making copy of block_key, cf_name, and referenced_key when
|
|
|
|
// constructing the access record.
|
|
|
|
Slice referenced_key;
|
|
|
|
if (does_referenced_key_exist) {
|
|
|
|
referenced_key = biter.key();
|
|
|
|
} else {
|
|
|
|
referenced_key = key;
|
|
|
|
}
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
BlockCacheTraceRecord access_record(
|
|
|
|
rep_->ioptions.env->NowMicros(),
|
|
|
|
/*block_key=*/"", lookup_data_block_context.block_type,
|
|
|
|
lookup_data_block_context.block_size, rep_->cf_id_for_tracing(),
|
|
|
|
/*cf_name=*/"", rep_->level_for_tracing(),
|
|
|
|
rep_->sst_number_for_tracing(), lookup_data_block_context.caller,
|
|
|
|
lookup_data_block_context.is_cache_hit,
|
|
|
|
lookup_data_block_context.no_insert,
|
|
|
|
lookup_data_block_context.get_id,
|
|
|
|
lookup_data_block_context.get_from_user_specified_snapshot,
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
/*referenced_key=*/"", referenced_data_size,
|
|
|
|
lookup_data_block_context.num_keys_in_block,
|
|
|
|
does_referenced_key_exist);
|
|
|
|
block_cache_tracer_->WriteBlockAccess(
|
|
|
|
access_record, lookup_data_block_context.block_key,
|
|
|
|
rep_->cf_name_for_tracing(), referenced_key);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (done) {
|
|
|
|
// Avoid the extra Next which is expensive in two-level indexes
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (matched && filter != nullptr && !filter->IsBlockBased()) {
|
|
|
|
RecordTick(rep_->ioptions.statistics, BLOOM_FILTER_FULL_TRUE_POSITIVE);
|
|
|
|
PERF_COUNTER_BY_LEVEL_ADD(bloom_filter_full_true_positive, 1,
|
|
|
|
rep_->level);
|
|
|
|
}
|
|
|
|
if (s.ok()) {
|
|
|
|
s = iiter->status();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
using MultiGetRange = MultiGetContext::Range;
|
|
|
|
void BlockBasedTable::MultiGet(const ReadOptions& read_options,
|
|
|
|
const MultiGetRange* mget_range,
|
|
|
|
const SliceTransform* prefix_extractor,
|
|
|
|
bool skip_filters) {
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
FilterBlockReader* const filter =
|
|
|
|
!skip_filters ? rep_->filter.get() : nullptr;
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
MultiGetRange sst_file_range(*mget_range, mget_range->begin(),
|
|
|
|
mget_range->end());
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
|
|
|
|
// First check the full filter
|
|
|
|
// If full filter not useful, Then go into each block
|
|
|
|
const bool no_io = read_options.read_tier == kBlockCacheTier;
|
|
|
|
uint64_t tracing_mget_id = BlockCacheTraceHelper::kReservedGetId;
|
|
|
|
if (!sst_file_range.empty() && sst_file_range.begin()->get_context) {
|
|
|
|
tracing_mget_id = sst_file_range.begin()->get_context->get_tracing_get_id();
|
|
|
|
}
|
|
|
|
BlockCacheLookupContext lookup_context{
|
|
|
|
TableReaderCaller::kUserMultiGet, tracing_mget_id,
|
|
|
|
/*get_from_user_specified_snapshot=*/read_options.snapshot != nullptr};
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
FullFilterKeysMayMatch(read_options, filter, &sst_file_range, no_io,
|
|
|
|
prefix_extractor, &lookup_context);
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
|
|
|
|
if (skip_filters || !sst_file_range.empty()) {
|
|
|
|
IndexBlockIter iiter_on_stack;
|
|
|
|
// if prefix_extractor found in block differs from options, disable
|
|
|
|
// BlockPrefixIndex. Only do this check when index_type is kHashSearch.
|
|
|
|
bool need_upper_bound_check = false;
|
|
|
|
if (rep_->index_type == BlockBasedTableOptions::kHashSearch) {
|
|
|
|
need_upper_bound_check = PrefixExtractorChanged(
|
|
|
|
rep_->table_properties.get(), prefix_extractor);
|
|
|
|
}
|
|
|
|
auto iiter =
|
|
|
|
NewIndexIterator(read_options, need_upper_bound_check, &iiter_on_stack,
|
|
|
|
sst_file_range.begin()->get_context, &lookup_context);
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
std::unique_ptr<InternalIteratorBase<IndexValue>> iiter_unique_ptr;
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
if (iiter != &iiter_on_stack) {
|
|
|
|
iiter_unique_ptr.reset(iiter);
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t offset = std::numeric_limits<uint64_t>::max();
|
|
|
|
autovector<BlockHandle, MultiGetContext::MAX_BATCH_SIZE> block_handles;
|
|
|
|
autovector<CachableEntry<Block>, MultiGetContext::MAX_BATCH_SIZE> results;
|
|
|
|
autovector<Status, MultiGetContext::MAX_BATCH_SIZE> statuses;
|
|
|
|
static const size_t kMultiGetReadStackBufSize = 8192;
|
|
|
|
char stack_buf[kMultiGetReadStackBufSize];
|
|
|
|
std::unique_ptr<char[]> block_buf;
|
|
|
|
{
|
|
|
|
MultiGetRange data_block_range(sst_file_range, sst_file_range.begin(),
|
|
|
|
sst_file_range.end());
|
|
|
|
|
|
|
|
CachableEntry<UncompressionDict> uncompression_dict;
|
|
|
|
Status uncompression_dict_status;
|
|
|
|
if (rep_->uncompression_dict_reader) {
|
|
|
|
uncompression_dict_status =
|
|
|
|
rep_->uncompression_dict_reader->GetOrReadUncompressionDictionary(
|
|
|
|
nullptr /* prefetch_buffer */, no_io,
|
|
|
|
sst_file_range.begin()->get_context, &lookup_context,
|
|
|
|
&uncompression_dict);
|
|
|
|
}
|
|
|
|
|
|
|
|
const UncompressionDict& dict = uncompression_dict.GetValue()
|
|
|
|
? *uncompression_dict.GetValue()
|
|
|
|
: UncompressionDict::GetEmptyDict();
|
|
|
|
|
|
|
|
size_t total_len = 0;
|
|
|
|
ReadOptions ro = read_options;
|
|
|
|
ro.read_tier = kBlockCacheTier;
|
|
|
|
|
|
|
|
for (auto miter = data_block_range.begin();
|
|
|
|
miter != data_block_range.end(); ++miter) {
|
|
|
|
const Slice& key = miter->ikey;
|
|
|
|
iiter->Seek(miter->ikey);
|
|
|
|
|
|
|
|
IndexValue v;
|
|
|
|
if (iiter->Valid()) {
|
|
|
|
v = iiter->value();
|
|
|
|
}
|
|
|
|
if (!iiter->Valid() ||
|
|
|
|
(!v.first_internal_key.empty() && !skip_filters &&
|
|
|
|
UserComparatorWrapper(rep_->internal_comparator.user_comparator())
|
|
|
|
.Compare(ExtractUserKey(key),
|
|
|
|
ExtractUserKey(v.first_internal_key)) < 0)) {
|
|
|
|
// The requested key falls between highest key in previous block and
|
|
|
|
// lowest key in current block.
|
|
|
|
*(miter->s) = iiter->status();
|
|
|
|
data_block_range.SkipKey(miter);
|
|
|
|
sst_file_range.SkipKey(miter);
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!uncompression_dict_status.ok()) {
|
|
|
|
*(miter->s) = uncompression_dict_status;
|
|
|
|
data_block_range.SkipKey(miter);
|
|
|
|
sst_file_range.SkipKey(miter);
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
statuses.emplace_back();
|
|
|
|
results.emplace_back();
|
|
|
|
if (v.handle.offset() == offset) {
|
|
|
|
// We're going to reuse the block for this key later on. No need to
|
|
|
|
// look it up now. Place a null handle
|
|
|
|
block_handles.emplace_back(BlockHandle::NullBlockHandle());
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
// Lookup the cache for the given data block referenced by an index
|
|
|
|
// iterator value (i.e BlockHandle). If it exists in the cache,
|
|
|
|
// initialize block to the contents of the data block.
|
|
|
|
offset = v.handle.offset();
|
|
|
|
BlockHandle handle = v.handle;
|
|
|
|
BlockCacheLookupContext lookup_data_block_context(
|
|
|
|
TableReaderCaller::kUserMultiGet);
|
|
|
|
Status s = RetrieveBlock(
|
|
|
|
nullptr, ro, handle, dict, &(results.back()), BlockType::kData,
|
|
|
|
miter->get_context, &lookup_data_block_context,
|
|
|
|
/* for_compaction */ false, /* use_cache */ true);
|
|
|
|
if (s.IsIncomplete()) {
|
|
|
|
s = Status::OK();
|
|
|
|
}
|
|
|
|
if (s.ok() && !results.back().IsEmpty()) {
|
|
|
|
// Found it in the cache. Add NULL handle to indicate there is
|
|
|
|
// nothing to read from disk
|
|
|
|
block_handles.emplace_back(BlockHandle::NullBlockHandle());
|
|
|
|
} else {
|
|
|
|
block_handles.emplace_back(handle);
|
|
|
|
total_len += handle.size();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (total_len) {
|
|
|
|
char* scratch = nullptr;
|
|
|
|
// If the blocks need to be uncompressed and we don't need the
|
|
|
|
// compressed blocks, then we can use a contiguous block of
|
|
|
|
// memory to read in all the blocks as it will be temporary
|
|
|
|
// storage
|
|
|
|
// 1. If blocks are compressed and compressed block cache is there,
|
|
|
|
// alloc heap bufs
|
|
|
|
// 2. If blocks are uncompressed, alloc heap bufs
|
|
|
|
// 3. If blocks are compressed and no compressed block cache, use
|
|
|
|
// stack buf
|
|
|
|
if (rep_->table_options.block_cache_compressed == nullptr &&
|
|
|
|
rep_->blocks_maybe_compressed) {
|
|
|
|
if (total_len <= kMultiGetReadStackBufSize) {
|
|
|
|
scratch = stack_buf;
|
|
|
|
} else {
|
|
|
|
scratch = new char[total_len];
|
|
|
|
block_buf.reset(scratch);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
RetrieveMultipleBlocks(read_options, &data_block_range, &block_handles,
|
|
|
|
&statuses, &results, scratch, dict);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
DataBlockIter first_biter;
|
|
|
|
DataBlockIter next_biter;
|
|
|
|
size_t idx_in_batch = 0;
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
for (auto miter = sst_file_range.begin(); miter != sst_file_range.end();
|
|
|
|
++miter) {
|
|
|
|
Status s;
|
|
|
|
GetContext* get_context = miter->get_context;
|
|
|
|
const Slice& key = miter->ikey;
|
|
|
|
bool matched = false; // if such user key matched a key in SST
|
|
|
|
bool done = false;
|
|
|
|
bool first_block = true;
|
|
|
|
do {
|
|
|
|
DataBlockIter* biter = nullptr;
|
|
|
|
bool reusing_block = true;
|
|
|
|
uint64_t referenced_data_size = 0;
|
|
|
|
bool does_referenced_key_exist = false;
|
|
|
|
BlockCacheLookupContext lookup_data_block_context(
|
|
|
|
TableReaderCaller::kUserMultiGet, tracing_mget_id,
|
|
|
|
/*get_from_user_specified_snapshot=*/read_options.snapshot !=
|
|
|
|
nullptr);
|
|
|
|
if (first_block) {
|
|
|
|
if (!block_handles[idx_in_batch].IsNull() ||
|
|
|
|
!results[idx_in_batch].IsEmpty()) {
|
|
|
|
first_biter.Invalidate(Status::OK());
|
|
|
|
NewDataBlockIterator<DataBlockIter>(
|
|
|
|
read_options, results[idx_in_batch], &first_biter,
|
|
|
|
statuses[idx_in_batch]);
|
|
|
|
reusing_block = false;
|
|
|
|
}
|
|
|
|
biter = &first_biter;
|
|
|
|
idx_in_batch++;
|
|
|
|
} else {
|
|
|
|
IndexValue v = iiter->value();
|
|
|
|
if (!v.first_internal_key.empty() && !skip_filters &&
|
|
|
|
UserComparatorWrapper(rep_->internal_comparator.user_comparator())
|
|
|
|
.Compare(ExtractUserKey(key),
|
|
|
|
ExtractUserKey(v.first_internal_key)) < 0) {
|
|
|
|
// The requested key falls between highest key in previous block and
|
|
|
|
// lowest key in current block.
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
next_biter.Invalidate(Status::OK());
|
|
|
|
NewDataBlockIterator<DataBlockIter>(
|
|
|
|
read_options, iiter->value().handle, &next_biter,
|
|
|
|
BlockType::kData, get_context, &lookup_data_block_context,
|
|
|
|
Status(), nullptr);
|
|
|
|
biter = &next_biter;
|
|
|
|
reusing_block = false;
|
|
|
|
}
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
if (read_options.read_tier == kBlockCacheTier &&
|
|
|
|
biter->status().IsIncomplete()) {
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
// couldn't get block from block_cache
|
|
|
|
// Update Saver.state to Found because we are only looking for
|
|
|
|
// whether we can guarantee the key is not there when "no_io" is set
|
|
|
|
get_context->MarkKeyMayExist();
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
if (!biter->status().ok()) {
|
|
|
|
s = biter->status();
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool may_exist = biter->SeekForGet(key);
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
if (!may_exist) {
|
|
|
|
// HashSeek cannot find the key this block and the the iter is not
|
|
|
|
// the end of the block, i.e. cannot be in the following blocks
|
|
|
|
// either. In this case, the seek_key cannot be found, so we break
|
|
|
|
// from the top level for-loop.
|
|
|
|
break;
|
|
|
|
}
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
|
|
|
|
// Call the *saver function on each entry/block until it returns false
|
|
|
|
for (; biter->Valid(); biter->Next()) {
|
|
|
|
ParsedInternalKey parsed_key;
|
|
|
|
Cleanable dummy;
|
|
|
|
Cleanable* value_pinner = nullptr;
|
|
|
|
if (!ParseInternalKey(biter->key(), &parsed_key)) {
|
|
|
|
s = Status::Corruption(Slice());
|
|
|
|
}
|
|
|
|
if (biter->IsValuePinned()) {
|
|
|
|
if (reusing_block) {
|
|
|
|
Cache* block_cache = rep_->table_options.block_cache.get();
|
|
|
|
assert(biter->cache_handle() != nullptr);
|
|
|
|
block_cache->Ref(biter->cache_handle());
|
|
|
|
dummy.RegisterCleanup(&ReleaseCachedEntry, block_cache,
|
|
|
|
biter->cache_handle());
|
|
|
|
value_pinner = &dummy;
|
|
|
|
} else {
|
|
|
|
value_pinner = biter;
|
|
|
|
}
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
}
|
|
|
|
if (!get_context->SaveValue(parsed_key, biter->value(), &matched,
|
|
|
|
value_pinner)) {
|
|
|
|
if (get_context->State() == GetContext::GetState::kFound) {
|
|
|
|
does_referenced_key_exist = true;
|
|
|
|
referenced_data_size =
|
|
|
|
biter->key().size() + biter->value().size();
|
|
|
|
}
|
|
|
|
done = true;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
s = biter->status();
|
|
|
|
}
|
|
|
|
// Write the block cache access.
|
|
|
|
if (block_cache_tracer_ && block_cache_tracer_->is_tracing_enabled()) {
|
|
|
|
// Avoid making copy of block_key, cf_name, and referenced_key when
|
|
|
|
// constructing the access record.
|
|
|
|
Slice referenced_key;
|
|
|
|
if (does_referenced_key_exist) {
|
|
|
|
referenced_key = biter->key();
|
|
|
|
} else {
|
|
|
|
referenced_key = key;
|
|
|
|
}
|
|
|
|
BlockCacheTraceRecord access_record(
|
|
|
|
rep_->ioptions.env->NowMicros(),
|
|
|
|
/*block_key=*/"", lookup_data_block_context.block_type,
|
|
|
|
lookup_data_block_context.block_size, rep_->cf_id_for_tracing(),
|
|
|
|
/*cf_name=*/"", rep_->level_for_tracing(),
|
|
|
|
rep_->sst_number_for_tracing(), lookup_data_block_context.caller,
|
|
|
|
lookup_data_block_context.is_cache_hit,
|
|
|
|
lookup_data_block_context.no_insert,
|
|
|
|
lookup_data_block_context.get_id,
|
|
|
|
lookup_data_block_context.get_from_user_specified_snapshot,
|
|
|
|
/*referenced_key=*/"", referenced_data_size,
|
|
|
|
lookup_data_block_context.num_keys_in_block,
|
|
|
|
does_referenced_key_exist);
|
|
|
|
block_cache_tracer_->WriteBlockAccess(
|
|
|
|
access_record, lookup_data_block_context.block_key,
|
|
|
|
rep_->cf_name_for_tracing(), referenced_key);
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
}
|
|
|
|
s = biter->status();
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
if (done) {
|
|
|
|
// Avoid the extra Next which is expensive in two-level indexes
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
if (first_block) {
|
|
|
|
iiter->Seek(key);
|
|
|
|
}
|
|
|
|
first_block = false;
|
|
|
|
iiter->Next();
|
|
|
|
} while (iiter->Valid());
|
|
|
|
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
if (matched && filter != nullptr && !filter->IsBlockBased()) {
|
|
|
|
RecordTick(rep_->ioptions.statistics, BLOOM_FILTER_FULL_TRUE_POSITIVE);
|
|
|
|
PERF_COUNTER_BY_LEVEL_ADD(bloom_filter_full_true_positive, 1,
|
|
|
|
rep_->level);
|
|
|
|
}
|
|
|
|
if (s.ok()) {
|
|
|
|
s = iiter->status();
|
|
|
|
}
|
|
|
|
*(miter->s) = s;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::Prefetch(const Slice* const begin,
|
|
|
|
const Slice* const end) {
|
|
|
|
auto& comparator = rep_->internal_comparator;
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
UserComparatorWrapper user_comparator(comparator.user_comparator());
|
|
|
|
// pre-condition
|
|
|
|
if (begin && end && comparator.Compare(*begin, *end) > 0) {
|
|
|
|
return Status::InvalidArgument(*begin, *end);
|
|
|
|
}
|
|
|
|
BlockCacheLookupContext lookup_context{TableReaderCaller::kPrefetch};
|
|
|
|
IndexBlockIter iiter_on_stack;
|
|
|
|
auto iiter = NewIndexIterator(ReadOptions(), /*need_upper_bound_check=*/false,
|
|
|
|
&iiter_on_stack, /*get_context=*/nullptr,
|
|
|
|
&lookup_context);
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
std::unique_ptr<InternalIteratorBase<IndexValue>> iiter_unique_ptr;
|
|
|
|
if (iiter != &iiter_on_stack) {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
iiter_unique_ptr = std::unique_ptr<InternalIteratorBase<IndexValue>>(iiter);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!iiter->status().ok()) {
|
|
|
|
// error opening index iterator
|
|
|
|
return iiter->status();
|
|
|
|
}
|
|
|
|
|
|
|
|
// indicates if we are on the last page that need to be pre-fetched
|
|
|
|
bool prefetching_boundary_page = false;
|
|
|
|
|
|
|
|
for (begin ? iiter->Seek(*begin) : iiter->SeekToFirst(); iiter->Valid();
|
|
|
|
iiter->Next()) {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
BlockHandle block_handle = iiter->value().handle;
|
|
|
|
const bool is_user_key = !rep_->index_key_includes_seq;
|
|
|
|
if (end &&
|
|
|
|
((!is_user_key && comparator.Compare(iiter->key(), *end) >= 0) ||
|
|
|
|
(is_user_key &&
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
user_comparator.Compare(iiter->key(), ExtractUserKey(*end)) >= 0))) {
|
|
|
|
if (prefetching_boundary_page) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
// The index entry represents the last key in the data block.
|
|
|
|
// We should load this page into memory as well, but no more
|
|
|
|
prefetching_boundary_page = true;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Load the block specified by the block_handle into the block cache
|
|
|
|
DataBlockIter biter;
|
|
|
|
|
|
|
|
NewDataBlockIterator<DataBlockIter>(
|
|
|
|
ReadOptions(), block_handle, &biter, /*type=*/BlockType::kData,
|
|
|
|
/*get_context=*/nullptr, &lookup_context, Status(),
|
|
|
|
/*prefetch_buffer=*/nullptr);
|
|
|
|
|
|
|
|
if (!biter.status().ok()) {
|
|
|
|
// there was an unexpected error while pre-fetching
|
|
|
|
return biter.status();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::VerifyChecksum(const ReadOptions& read_options,
|
|
|
|
TableReaderCaller caller) {
|
|
|
|
Status s;
|
|
|
|
// Check Meta blocks
|
|
|
|
std::unique_ptr<Block> meta;
|
|
|
|
std::unique_ptr<InternalIterator> meta_iter;
|
|
|
|
s = ReadMetaBlock(nullptr /* prefetch buffer */, &meta, &meta_iter);
|
|
|
|
if (s.ok()) {
|
|
|
|
s = VerifyChecksumInMetaBlocks(meta_iter.get());
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
// Check Data blocks
|
|
|
|
IndexBlockIter iiter_on_stack;
|
|
|
|
BlockCacheLookupContext context{caller};
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
InternalIteratorBase<IndexValue>* iiter = NewIndexIterator(
|
|
|
|
read_options, /*disable_prefix_seek=*/false, &iiter_on_stack,
|
|
|
|
/*get_context=*/nullptr, &context);
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
std::unique_ptr<InternalIteratorBase<IndexValue>> iiter_unique_ptr;
|
|
|
|
if (iiter != &iiter_on_stack) {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
iiter_unique_ptr = std::unique_ptr<InternalIteratorBase<IndexValue>>(iiter);
|
|
|
|
}
|
|
|
|
if (!iiter->status().ok()) {
|
|
|
|
// error opening index iterator
|
|
|
|
return iiter->status();
|
|
|
|
}
|
|
|
|
s = VerifyChecksumInBlocks(read_options, iiter);
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::VerifyChecksumInBlocks(
|
|
|
|
const ReadOptions& read_options,
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
InternalIteratorBase<IndexValue>* index_iter) {
|
|
|
|
Status s;
|
|
|
|
// We are scanning the whole file, so no need to do exponential
|
|
|
|
// increasing of the buffer size.
|
|
|
|
size_t readahead_size = (read_options.readahead_size != 0)
|
|
|
|
? read_options.readahead_size
|
|
|
|
: kMaxAutoReadaheadSize;
|
|
|
|
// FilePrefetchBuffer doesn't work in mmap mode and readahead is not
|
|
|
|
// needed there.
|
|
|
|
FilePrefetchBuffer prefetch_buffer(
|
|
|
|
rep_->file.get(), readahead_size /* readadhead_size */,
|
|
|
|
readahead_size /* max_readahead_size */,
|
|
|
|
!rep_->ioptions.allow_mmap_reads /* enable */);
|
|
|
|
|
|
|
|
for (index_iter->SeekToFirst(); index_iter->Valid(); index_iter->Next()) {
|
|
|
|
s = index_iter->status();
|
|
|
|
if (!s.ok()) {
|
|
|
|
break;
|
|
|
|
}
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
BlockHandle handle = index_iter->value().handle;
|
|
|
|
BlockContents contents;
|
|
|
|
BlockFetcher block_fetcher(
|
|
|
|
rep_->file.get(), &prefetch_buffer, rep_->footer, ReadOptions(), handle,
|
|
|
|
&contents, rep_->ioptions, false /* decompress */,
|
|
|
|
false /*maybe_compressed*/, BlockType::kData,
|
|
|
|
UncompressionDict::GetEmptyDict(), rep_->persistent_cache_options);
|
|
|
|
s = block_fetcher.ReadBlockContents();
|
|
|
|
if (!s.ok()) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
BlockType BlockBasedTable::GetBlockTypeForMetaBlockByName(
|
|
|
|
const Slice& meta_block_name) {
|
|
|
|
if (meta_block_name.starts_with(kFilterBlockPrefix) ||
|
|
|
|
meta_block_name.starts_with(kFullFilterBlockPrefix) ||
|
|
|
|
meta_block_name.starts_with(kPartitionedFilterBlockPrefix)) {
|
|
|
|
return BlockType::kFilter;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (meta_block_name == kPropertiesBlock) {
|
|
|
|
return BlockType::kProperties;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (meta_block_name == kCompressionDictBlock) {
|
|
|
|
return BlockType::kCompressionDictionary;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (meta_block_name == kRangeDelBlock) {
|
|
|
|
return BlockType::kRangeDeletion;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (meta_block_name == kHashIndexPrefixesBlock) {
|
|
|
|
return BlockType::kHashIndexPrefixes;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (meta_block_name == kHashIndexPrefixesMetadataBlock) {
|
|
|
|
return BlockType::kHashIndexMetadata;
|
|
|
|
}
|
|
|
|
|
|
|
|
assert(false);
|
|
|
|
return BlockType::kInvalid;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::VerifyChecksumInMetaBlocks(
|
|
|
|
InternalIteratorBase<Slice>* index_iter) {
|
|
|
|
Status s;
|
|
|
|
for (index_iter->SeekToFirst(); index_iter->Valid(); index_iter->Next()) {
|
|
|
|
s = index_iter->status();
|
|
|
|
if (!s.ok()) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
BlockHandle handle;
|
|
|
|
Slice input = index_iter->value();
|
|
|
|
s = handle.DecodeFrom(&input);
|
|
|
|
BlockContents contents;
|
|
|
|
const Slice meta_block_name = index_iter->key();
|
|
|
|
BlockFetcher block_fetcher(
|
|
|
|
rep_->file.get(), nullptr /* prefetch buffer */, rep_->footer,
|
|
|
|
ReadOptions(), handle, &contents, rep_->ioptions,
|
|
|
|
false /* decompress */, false /*maybe_compressed*/,
|
|
|
|
GetBlockTypeForMetaBlockByName(meta_block_name),
|
|
|
|
UncompressionDict::GetEmptyDict(), rep_->persistent_cache_options);
|
|
|
|
s = block_fetcher.ReadBlockContents();
|
|
|
|
if (s.IsCorruption() && meta_block_name == kPropertiesBlock) {
|
|
|
|
TableProperties* table_properties;
|
|
|
|
s = TryReadPropertiesWithGlobalSeqno(nullptr /* prefetch_buffer */,
|
|
|
|
index_iter->value(),
|
|
|
|
&table_properties);
|
|
|
|
delete table_properties;
|
|
|
|
}
|
|
|
|
if (!s.ok()) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
bool BlockBasedTable::TEST_BlockInCache(const BlockHandle& handle) const {
|
|
|
|
assert(rep_ != nullptr);
|
|
|
|
|
|
|
|
Cache* const cache = rep_->table_options.block_cache.get();
|
|
|
|
if (cache == nullptr) {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
char cache_key_storage[kMaxCacheKeyPrefixSize + kMaxVarint64Length];
|
|
|
|
Slice cache_key =
|
|
|
|
GetCacheKey(rep_->cache_key_prefix, rep_->cache_key_prefix_size, handle,
|
|
|
|
cache_key_storage);
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
|
|
|
|
Cache::Handle* const cache_handle = cache->Lookup(cache_key);
|
|
|
|
if (cache_handle == nullptr) {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
cache->Release(cache_handle);
|
|
|
|
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool BlockBasedTable::TEST_KeyInCache(const ReadOptions& options,
|
|
|
|
const Slice& key) {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
std::unique_ptr<InternalIteratorBase<IndexValue>> iiter(NewIndexIterator(
|
|
|
|
options, /*need_upper_bound_check=*/false, /*input_iter=*/nullptr,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
/*get_context=*/nullptr, /*lookup_context=*/nullptr));
|
|
|
|
iiter->Seek(key);
|
|
|
|
assert(iiter->Valid());
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
return TEST_BlockInCache(iiter->value().handle);
|
|
|
|
}
|
|
|
|
|
|
|
|
// REQUIRES: The following fields of rep_ should have already been populated:
|
|
|
|
// 1. file
|
|
|
|
// 2. index_handle,
|
|
|
|
// 3. options
|
|
|
|
// 4. internal_comparator
|
|
|
|
// 5. index_type
|
|
|
|
Status BlockBasedTable::CreateIndexReader(
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
FilePrefetchBuffer* prefetch_buffer,
|
|
|
|
InternalIterator* preloaded_meta_index_iter, bool use_cache, bool prefetch,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
bool pin, BlockCacheLookupContext* lookup_context,
|
|
|
|
std::unique_ptr<IndexReader>* index_reader) {
|
|
|
|
// kHashSearch requires non-empty prefix_extractor but bypass checking
|
|
|
|
// prefix_extractor here since we have no access to MutableCFOptions.
|
|
|
|
// Add need_upper_bound_check flag in BlockBasedTable::NewIndexIterator.
|
|
|
|
// If prefix_extractor does not match prefix_extractor_name from table
|
|
|
|
// properties, turn off Hash Index by setting total_order_seek to true
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
switch (rep_->index_type) {
|
|
|
|
case BlockBasedTableOptions::kTwoLevelIndexSearch: {
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
return PartitionIndexReader::Create(this, prefetch_buffer, use_cache,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
prefetch, pin, lookup_context,
|
|
|
|
index_reader);
|
|
|
|
}
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
case BlockBasedTableOptions::kBinarySearch:
|
|
|
|
case BlockBasedTableOptions::kBinarySearchWithFirstKey: {
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
return BinarySearchIndexReader::Create(this, prefetch_buffer, use_cache,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
prefetch, pin, lookup_context,
|
|
|
|
index_reader);
|
|
|
|
}
|
|
|
|
case BlockBasedTableOptions::kHashSearch: {
|
|
|
|
std::unique_ptr<Block> meta_guard;
|
|
|
|
std::unique_ptr<InternalIterator> meta_iter_guard;
|
|
|
|
auto meta_index_iter = preloaded_meta_index_iter;
|
|
|
|
if (meta_index_iter == nullptr) {
|
|
|
|
auto s = ReadMetaBlock(prefetch_buffer, &meta_guard, &meta_iter_guard);
|
|
|
|
if (!s.ok()) {
|
|
|
|
// we simply fall back to binary search in case there is any
|
|
|
|
// problem with prefix hash index loading.
|
|
|
|
ROCKS_LOG_WARN(rep_->ioptions.info_log,
|
|
|
|
"Unable to read the metaindex block."
|
|
|
|
" Fall back to binary search index.");
|
|
|
|
return BinarySearchIndexReader::Create(this, prefetch_buffer,
|
|
|
|
use_cache, prefetch, pin,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
lookup_context, index_reader);
|
|
|
|
}
|
|
|
|
meta_index_iter = meta_iter_guard.get();
|
|
|
|
}
|
|
|
|
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
return HashIndexReader::Create(this, prefetch_buffer, meta_index_iter,
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
use_cache, prefetch, pin, lookup_context,
|
|
|
|
index_reader);
|
|
|
|
}
|
|
|
|
default: {
|
|
|
|
std::string error_message =
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
"Unrecognized index type: " + ToString(rep_->index_type);
|
|
|
|
return Status::InvalidArgument(error_message.c_str());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t BlockBasedTable::ApproximateOffsetOf(
|
|
|
|
const InternalIteratorBase<IndexValue>& index_iter) const {
|
|
|
|
uint64_t result = 0;
|
|
|
|
if (index_iter.Valid()) {
|
|
|
|
BlockHandle handle = index_iter.value().handle;
|
|
|
|
result = handle.offset();
|
|
|
|
} else {
|
|
|
|
// The iterator is past the last key in the file. If table_properties is not
|
|
|
|
// available, approximate the offset by returning the offset of the
|
|
|
|
// metaindex block (which is right near the end of the file).
|
RocksDB 2.8 to be able to read files generated by 2.6
Summary:
From 2.6 to 2.7, property block name is renamed from rocksdb.stats to rocksdb.properties. Older properties were not able to be loaded. In 2.8, we seem to have added some logic that uses property block without checking null pointers, which create segment faults.
In this patch, we fix it by:
(1) try rocksdb.stats if rocksdb.properties is not found
(2) add some null checking before consuming rep->table_properties
Test Plan: make sure a file generated in 2.7 couldn't be opened now can be opened.
Reviewers: haobo, igor, yhchiang
Reviewed By: igor
CC: ljin, xjin, dhruba, kailiu, leveldb
Differential Revision: https://reviews.facebook.net/D17961
11 years ago
|
|
|
if (rep_->table_properties) {
|
|
|
|
result = rep_->table_properties->data_size;
|
|
|
|
}
|
|
|
|
// table_properties is not present in the table.
|
|
|
|
if (result == 0) {
|
|
|
|
result = rep_->footer.metaindex_handle().offset();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t BlockBasedTable::ApproximateOffsetOf(const Slice& key,
|
|
|
|
TableReaderCaller caller) {
|
|
|
|
BlockCacheLookupContext context(caller);
|
|
|
|
IndexBlockIter iiter_on_stack;
|
|
|
|
auto index_iter =
|
|
|
|
NewIndexIterator(ReadOptions(), /*disable_prefix_seek=*/false,
|
|
|
|
/*input_iter=*/&iiter_on_stack, /*get_context=*/nullptr,
|
|
|
|
/*lookup_context=*/&context);
|
|
|
|
std::unique_ptr<InternalIteratorBase<IndexValue>> iiter_unique_ptr;
|
|
|
|
if (index_iter != &iiter_on_stack) {
|
|
|
|
iiter_unique_ptr.reset(index_iter);
|
|
|
|
}
|
|
|
|
|
|
|
|
index_iter->Seek(key);
|
|
|
|
return ApproximateOffsetOf(*index_iter);
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t BlockBasedTable::ApproximateSize(const Slice& start, const Slice& end,
|
|
|
|
TableReaderCaller caller) {
|
|
|
|
assert(rep_->internal_comparator.Compare(start, end) <= 0);
|
|
|
|
|
|
|
|
BlockCacheLookupContext context(caller);
|
|
|
|
IndexBlockIter iiter_on_stack;
|
|
|
|
auto index_iter =
|
|
|
|
NewIndexIterator(ReadOptions(), /*disable_prefix_seek=*/false,
|
|
|
|
/*input_iter=*/&iiter_on_stack, /*get_context=*/nullptr,
|
|
|
|
/*lookup_context=*/&context);
|
|
|
|
std::unique_ptr<InternalIteratorBase<IndexValue>> iiter_unique_ptr;
|
|
|
|
if (index_iter != &iiter_on_stack) {
|
|
|
|
iiter_unique_ptr.reset(index_iter);
|
|
|
|
}
|
|
|
|
|
|
|
|
index_iter->Seek(start);
|
|
|
|
uint64_t start_offset = ApproximateOffsetOf(*index_iter);
|
|
|
|
index_iter->Seek(end);
|
|
|
|
uint64_t end_offset = ApproximateOffsetOf(*index_iter);
|
|
|
|
|
|
|
|
assert(end_offset >= start_offset);
|
|
|
|
return end_offset - start_offset;
|
|
|
|
}
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
bool BlockBasedTable::TEST_FilterBlockInCache() const {
|
|
|
|
assert(rep_ != nullptr);
|
|
|
|
return TEST_BlockInCache(rep_->filter_handle);
|
|
|
|
}
|
|
|
|
|
Move the index readers out of the block cache (#5298)
Summary:
Currently, when the block cache is used for index blocks as well, it is
not really the index block that is stored in the cache but an
IndexReader object. Since this object is not pure data (it has, for
instance, pointers that might dangle), it's not really sharable. To
avoid the issues around this, the current code uses a dummy unique cache
key for each TableReader to store the IndexReader, and erases the
IndexReader entry when the TableReader is closed. Instead of doing this,
the new code moves the IndexReader out of the cache altogether. In
particular, instead of the TableReader owning, or caching/pinning the
IndexReader based on the customer's settings, the TableReader
unconditionally owns the IndexReader, which in turn owns/caches/pins
the index block (which is itself sharable and thus can be safely put in
the cache without any hacks).
Note: the change has two side effects:
1) Partitions of partitioned indexes no longer affect the read
amplification statistics.
2) Eviction statistics for index blocks are temporarily broken. We plan to fix
this in a separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5298
Differential Revision: D15303203
Pulled By: ltamasi
fbshipit-source-id: 935a69ba59d87d5e44f42e2310619b790c366e47
6 years ago
|
|
|
bool BlockBasedTable::TEST_IndexBlockInCache() const {
|
|
|
|
assert(rep_ != nullptr);
|
|
|
|
|
|
|
|
return TEST_BlockInCache(rep_->footer.index_handle());
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::GetKVPairsFromDataBlocks(
|
|
|
|
std::vector<KVPairBlock>* kv_pair_blocks) {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
std::unique_ptr<InternalIteratorBase<IndexValue>> blockhandles_iter(
|
|
|
|
NewIndexIterator(ReadOptions(), /*need_upper_bound_check=*/false,
|
|
|
|
/*input_iter=*/nullptr, /*get_context=*/nullptr,
|
|
|
|
/*lookup_contex=*/nullptr));
|
|
|
|
|
|
|
|
Status s = blockhandles_iter->status();
|
|
|
|
if (!s.ok()) {
|
|
|
|
// Cannot read Index Block
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
for (blockhandles_iter->SeekToFirst(); blockhandles_iter->Valid();
|
|
|
|
blockhandles_iter->Next()) {
|
|
|
|
s = blockhandles_iter->status();
|
|
|
|
|
|
|
|
if (!s.ok()) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
std::unique_ptr<InternalIterator> datablock_iter;
|
|
|
|
datablock_iter.reset(NewDataBlockIterator<DataBlockIter>(
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
ReadOptions(), blockhandles_iter->value().handle,
|
|
|
|
/*input_iter=*/nullptr, /*type=*/BlockType::kData,
|
|
|
|
/*get_context=*/nullptr, /*lookup_context=*/nullptr, Status(),
|
|
|
|
/*prefetch_buffer=*/nullptr));
|
|
|
|
s = datablock_iter->status();
|
|
|
|
|
|
|
|
if (!s.ok()) {
|
|
|
|
// Error reading the block - Skipped
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
KVPairBlock kv_pair_block;
|
|
|
|
for (datablock_iter->SeekToFirst(); datablock_iter->Valid();
|
|
|
|
datablock_iter->Next()) {
|
|
|
|
s = datablock_iter->status();
|
|
|
|
if (!s.ok()) {
|
|
|
|
// Error reading the block - Skipped
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
const Slice& key = datablock_iter->key();
|
|
|
|
const Slice& value = datablock_iter->value();
|
|
|
|
std::string key_copy = std::string(key.data(), key.size());
|
|
|
|
std::string value_copy = std::string(value.data(), value.size());
|
|
|
|
|
|
|
|
kv_pair_block.push_back(
|
|
|
|
std::make_pair(std::move(key_copy), std::move(value_copy)));
|
|
|
|
}
|
|
|
|
kv_pair_blocks->push_back(std::move(kv_pair_block));
|
|
|
|
}
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
Status BlockBasedTable::DumpTable(WritableFile* out_file) {
|
|
|
|
// Output Footer
|
|
|
|
out_file->Append(
|
|
|
|
"Footer Details:\n"
|
|
|
|
"--------------------------------------\n"
|
|
|
|
" ");
|
|
|
|
out_file->Append(rep_->footer.ToString().c_str());
|
|
|
|
out_file->Append("\n");
|
|
|
|
|
|
|
|
// Output MetaIndex
|
|
|
|
out_file->Append(
|
|
|
|
"Metaindex Details:\n"
|
|
|
|
"--------------------------------------\n");
|
|
|
|
std::unique_ptr<Block> meta;
|
|
|
|
std::unique_ptr<InternalIterator> meta_iter;
|
|
|
|
Status s = ReadMetaBlock(nullptr /* prefetch_buffer */, &meta, &meta_iter);
|
|
|
|
if (s.ok()) {
|
|
|
|
for (meta_iter->SeekToFirst(); meta_iter->Valid(); meta_iter->Next()) {
|
|
|
|
s = meta_iter->status();
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
if (meta_iter->key() == rocksdb::kPropertiesBlock) {
|
|
|
|
out_file->Append(" Properties block handle: ");
|
|
|
|
out_file->Append(meta_iter->value().ToString(true).c_str());
|
|
|
|
out_file->Append("\n");
|
Shared dictionary compression using reference block
Summary:
This adds a new metablock containing a shared dictionary that is used
to compress all data blocks in the SST file. The size of the shared dictionary
is configurable in CompressionOptions and defaults to 0. It's currently only
used for zlib/lz4/lz4hc, but the block will be stored in the SST regardless of
the compression type if the user chooses a nonzero dictionary size.
During compaction, computes the dictionary by randomly sampling the first
output file in each subcompaction. It pre-computes the intervals to sample
by assuming the output file will have the maximum allowable length. In case
the file is smaller, some of the pre-computed sampling intervals can be beyond
end-of-file, in which case we skip over those samples and the dictionary will
be a bit smaller. After the dictionary is generated using the first file in a
subcompaction, it is loaded into the compression library before writing each
block in each subsequent file of that subcompaction.
On the read path, gets the dictionary from the metablock, if it exists. Then,
loads that dictionary into the compression library before reading each block.
Test Plan: new unit test
Reviewers: yhchiang, IslamAbdelRahman, cyan, sdong
Reviewed By: sdong
Subscribers: andrewkr, yoshinorim, kradhakrishnan, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D52287
9 years ago
|
|
|
} else if (meta_iter->key() == rocksdb::kCompressionDictBlock) {
|
|
|
|
out_file->Append(" Compression dictionary block handle: ");
|
|
|
|
out_file->Append(meta_iter->value().ToString(true).c_str());
|
|
|
|
out_file->Append("\n");
|
|
|
|
} else if (strstr(meta_iter->key().ToString().c_str(),
|
|
|
|
"filter.rocksdb.") != nullptr) {
|
|
|
|
out_file->Append(" Filter block handle: ");
|
|
|
|
out_file->Append(meta_iter->value().ToString(true).c_str());
|
|
|
|
out_file->Append("\n");
|
|
|
|
} else if (meta_iter->key() == rocksdb::kRangeDelBlock) {
|
|
|
|
out_file->Append(" Range deletion block handle: ");
|
|
|
|
out_file->Append(meta_iter->value().ToString(true).c_str());
|
|
|
|
out_file->Append("\n");
|
|
|
|
}
|
|
|
|
}
|
|
|
|
out_file->Append("\n");
|
|
|
|
} else {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Output TableProperties
|
|
|
|
const rocksdb::TableProperties* table_properties;
|
|
|
|
table_properties = rep_->table_properties.get();
|
|
|
|
|
|
|
|
if (table_properties != nullptr) {
|
|
|
|
out_file->Append(
|
|
|
|
"Table Properties:\n"
|
|
|
|
"--------------------------------------\n"
|
|
|
|
" ");
|
|
|
|
out_file->Append(table_properties->ToString("\n ", ": ").c_str());
|
|
|
|
out_file->Append("\n");
|
|
|
|
}
|
Move the filter readers out of the block cache (#5504)
Summary:
Currently, when the block cache is used for the filter block, it is not
really the block itself that is stored in the cache but a FilterBlockReader
object. Since this object is not pure data (it has, for instance, pointers that
might dangle, including in one case a back pointer to the TableReader), it's not
really sharable. To avoid the issues around this, the current code erases the
cache entries when the TableReader is closed (which, BTW, is not sufficient
since a concurrent TableReader might have picked up the object in the meantime).
Instead of doing this, the patch moves the FilterBlockReader out of the cache
altogether, and decouples the filter reader object from the filter block.
In particular, instead of the TableReader owning, or caching/pinning the
FilterBlockReader (based on the customer's settings), with the change the
TableReader unconditionally owns the FilterBlockReader, which in turn
owns/caches/pins the filter block. This change also enables us to reuse the code
paths historically used for data blocks for filters as well.
Note:
Eviction statistics for filter blocks are temporarily broken. We plan to fix this in a
separate phase.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5504
Test Plan: make asan_check
Differential Revision: D16036974
Pulled By: ltamasi
fbshipit-source-id: 770f543c5fb4ed126fd1e04bfd3809cf4ff9c091
6 years ago
|
|
|
|
|
|
|
if (rep_->filter) {
|
|
|
|
out_file->Append(
|
|
|
|
"Filter Details:\n"
|
|
|
|
"--------------------------------------\n"
|
|
|
|
" ");
|
|
|
|
out_file->Append(rep_->filter->ToString().c_str());
|
|
|
|
out_file->Append("\n");
|
|
|
|
}
|
|
|
|
|
|
|
|
// Output Index block
|
|
|
|
s = DumpIndexBlock(out_file);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Output compression dictionary
|
|
|
|
if (rep_->uncompression_dict_reader) {
|
|
|
|
CachableEntry<UncompressionDict> uncompression_dict;
|
|
|
|
s = rep_->uncompression_dict_reader->GetOrReadUncompressionDictionary(
|
|
|
|
nullptr /* prefetch_buffer */, false /* no_io */,
|
|
|
|
nullptr /* get_context */, nullptr /* lookup_context */,
|
|
|
|
&uncompression_dict);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
assert(uncompression_dict.GetValue());
|
|
|
|
|
|
|
|
const Slice& raw_dict = uncompression_dict.GetValue()->GetRawDict();
|
|
|
|
out_file->Append(
|
|
|
|
"Compression Dictionary:\n"
|
|
|
|
"--------------------------------------\n");
|
|
|
|
out_file->Append(" size (bytes): ");
|
|
|
|
out_file->Append(rocksdb::ToString(raw_dict.size()));
|
|
|
|
out_file->Append("\n\n");
|
|
|
|
out_file->Append(" HEX ");
|
|
|
|
out_file->Append(raw_dict.ToString(true).c_str());
|
|
|
|
out_file->Append("\n\n");
|
|
|
|
}
|
|
|
|
|
|
|
|
// Output range deletions block
|
|
|
|
auto* range_del_iter = NewRangeTombstoneIterator(ReadOptions());
|
|
|
|
if (range_del_iter != nullptr) {
|
|
|
|
range_del_iter->SeekToFirst();
|
|
|
|
if (range_del_iter->Valid()) {
|
|
|
|
out_file->Append(
|
|
|
|
"Range deletions:\n"
|
|
|
|
"--------------------------------------\n"
|
|
|
|
" ");
|
|
|
|
for (; range_del_iter->Valid(); range_del_iter->Next()) {
|
|
|
|
DumpKeyValue(range_del_iter->key(), range_del_iter->value(), out_file);
|
|
|
|
}
|
|
|
|
out_file->Append("\n");
|
|
|
|
}
|
|
|
|
delete range_del_iter;
|
|
|
|
}
|
|
|
|
// Output Data blocks
|
|
|
|
s = DumpDataBlocks(out_file);
|
|
|
|
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::DumpIndexBlock(WritableFile* out_file) {
|
|
|
|
out_file->Append(
|
|
|
|
"Index Details:\n"
|
|
|
|
"--------------------------------------\n");
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
std::unique_ptr<InternalIteratorBase<IndexValue>> blockhandles_iter(
|
|
|
|
NewIndexIterator(ReadOptions(), /*need_upper_bound_check=*/false,
|
|
|
|
/*input_iter=*/nullptr, /*get_context=*/nullptr,
|
|
|
|
/*lookup_contex=*/nullptr));
|
|
|
|
Status s = blockhandles_iter->status();
|
|
|
|
if (!s.ok()) {
|
|
|
|
out_file->Append("Can not read Index Block \n\n");
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
out_file->Append(" Block key hex dump: Data block handle\n");
|
|
|
|
out_file->Append(" Block key ascii\n\n");
|
|
|
|
for (blockhandles_iter->SeekToFirst(); blockhandles_iter->Valid();
|
|
|
|
blockhandles_iter->Next()) {
|
|
|
|
s = blockhandles_iter->status();
|
|
|
|
if (!s.ok()) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
Slice key = blockhandles_iter->key();
|
|
|
|
Slice user_key;
|
|
|
|
InternalKey ikey;
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
if (!rep_->index_key_includes_seq) {
|
|
|
|
user_key = key;
|
|
|
|
} else {
|
|
|
|
ikey.DecodeFrom(key);
|
|
|
|
user_key = ikey.user_key();
|
|
|
|
}
|
|
|
|
|
|
|
|
out_file->Append(" HEX ");
|
|
|
|
out_file->Append(user_key.ToString(true).c_str());
|
|
|
|
out_file->Append(": ");
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
out_file->Append(blockhandles_iter->value()
|
|
|
|
.ToString(true, rep_->index_has_first_key)
|
|
|
|
.c_str());
|
|
|
|
out_file->Append("\n");
|
|
|
|
|
|
|
|
std::string str_key = user_key.ToString();
|
|
|
|
std::string res_key("");
|
|
|
|
char cspace = ' ';
|
|
|
|
for (size_t i = 0; i < str_key.size(); i++) {
|
|
|
|
res_key.append(&str_key[i], 1);
|
|
|
|
res_key.append(1, cspace);
|
|
|
|
}
|
|
|
|
out_file->Append(" ASCII ");
|
|
|
|
out_file->Append(res_key.c_str());
|
|
|
|
out_file->Append("\n ------\n");
|
|
|
|
}
|
|
|
|
out_file->Append("\n");
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlockBasedTable::DumpDataBlocks(WritableFile* out_file) {
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
std::unique_ptr<InternalIteratorBase<IndexValue>> blockhandles_iter(
|
|
|
|
NewIndexIterator(ReadOptions(), /*need_upper_bound_check=*/false,
|
|
|
|
/*input_iter=*/nullptr, /*get_context=*/nullptr,
|
|
|
|
/*lookup_contex=*/nullptr));
|
|
|
|
Status s = blockhandles_iter->status();
|
|
|
|
if (!s.ok()) {
|
|
|
|
out_file->Append("Can not read Index Block \n\n");
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t datablock_size_min = std::numeric_limits<uint64_t>::max();
|
|
|
|
uint64_t datablock_size_max = 0;
|
|
|
|
uint64_t datablock_size_sum = 0;
|
|
|
|
|
|
|
|
size_t block_id = 1;
|
|
|
|
for (blockhandles_iter->SeekToFirst(); blockhandles_iter->Valid();
|
|
|
|
block_id++, blockhandles_iter->Next()) {
|
|
|
|
s = blockhandles_iter->status();
|
|
|
|
if (!s.ok()) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
BlockHandle bh = blockhandles_iter->value().handle;
|
|
|
|
uint64_t datablock_size = bh.size();
|
|
|
|
datablock_size_min = std::min(datablock_size_min, datablock_size);
|
|
|
|
datablock_size_max = std::max(datablock_size_max, datablock_size);
|
|
|
|
datablock_size_sum += datablock_size;
|
|
|
|
|
|
|
|
out_file->Append("Data Block # ");
|
|
|
|
out_file->Append(rocksdb::ToString(block_id));
|
|
|
|
out_file->Append(" @ ");
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
out_file->Append(blockhandles_iter->value().handle.ToString(true).c_str());
|
|
|
|
out_file->Append("\n");
|
|
|
|
out_file->Append("--------------------------------------\n");
|
|
|
|
|
|
|
|
std::unique_ptr<InternalIterator> datablock_iter;
|
|
|
|
datablock_iter.reset(NewDataBlockIterator<DataBlockIter>(
|
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
|
|
|
ReadOptions(), blockhandles_iter->value().handle,
|
|
|
|
/*input_iter=*/nullptr, /*type=*/BlockType::kData,
|
|
|
|
/*get_context=*/nullptr, /*lookup_context=*/nullptr, Status(),
|
|
|
|
/*prefetch_buffer=*/nullptr));
|
|
|
|
s = datablock_iter->status();
|
|
|
|
|
|
|
|
if (!s.ok()) {
|
|
|
|
out_file->Append("Error reading the block - Skipped \n\n");
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
for (datablock_iter->SeekToFirst(); datablock_iter->Valid();
|
|
|
|
datablock_iter->Next()) {
|
|
|
|
s = datablock_iter->status();
|
|
|
|
if (!s.ok()) {
|
|
|
|
out_file->Append("Error reading the block - Skipped \n");
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
DumpKeyValue(datablock_iter->key(), datablock_iter->value(), out_file);
|
|
|
|
}
|
|
|
|
out_file->Append("\n");
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t num_datablocks = block_id - 1;
|
|
|
|
if (num_datablocks) {
|
|
|
|
double datablock_size_avg =
|
|
|
|
static_cast<double>(datablock_size_sum) / num_datablocks;
|
|
|
|
out_file->Append("Data Block Summary:\n");
|
|
|
|
out_file->Append("--------------------------------------");
|
|
|
|
out_file->Append("\n # data blocks: ");
|
|
|
|
out_file->Append(rocksdb::ToString(num_datablocks));
|
|
|
|
out_file->Append("\n min data block size: ");
|
|
|
|
out_file->Append(rocksdb::ToString(datablock_size_min));
|
|
|
|
out_file->Append("\n max data block size: ");
|
|
|
|
out_file->Append(rocksdb::ToString(datablock_size_max));
|
|
|
|
out_file->Append("\n avg data block size: ");
|
|
|
|
out_file->Append(rocksdb::ToString(datablock_size_avg));
|
|
|
|
out_file->Append("\n");
|
|
|
|
}
|
|
|
|
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
void BlockBasedTable::DumpKeyValue(const Slice& key, const Slice& value,
|
|
|
|
WritableFile* out_file) {
|
|
|
|
InternalKey ikey;
|
|
|
|
ikey.DecodeFrom(key);
|
|
|
|
|
|
|
|
out_file->Append(" HEX ");
|
|
|
|
out_file->Append(ikey.user_key().ToString(true).c_str());
|
|
|
|
out_file->Append(": ");
|
|
|
|
out_file->Append(value.ToString(true).c_str());
|
|
|
|
out_file->Append("\n");
|
|
|
|
|
|
|
|
std::string str_key = ikey.user_key().ToString();
|
|
|
|
std::string str_value = value.ToString();
|
|
|
|
std::string res_key(""), res_value("");
|
|
|
|
char cspace = ' ';
|
|
|
|
for (size_t i = 0; i < str_key.size(); i++) {
|
|
|
|
if (str_key[i] == '\0') {
|
|
|
|
res_key.append("\\0", 2);
|
|
|
|
} else {
|
|
|
|
res_key.append(&str_key[i], 1);
|
|
|
|
}
|
|
|
|
res_key.append(1, cspace);
|
|
|
|
}
|
|
|
|
for (size_t i = 0; i < str_value.size(); i++) {
|
|
|
|
if (str_value[i] == '\0') {
|
|
|
|
res_value.append("\\0", 2);
|
|
|
|
} else {
|
|
|
|
res_value.append(&str_value[i], 1);
|
|
|
|
}
|
|
|
|
res_value.append(1, cspace);
|
|
|
|
}
|
|
|
|
|
|
|
|
out_file->Append(" ASCII ");
|
|
|
|
out_file->Append(res_key.c_str());
|
|
|
|
out_file->Append(": ");
|
|
|
|
out_file->Append(res_value.c_str());
|
|
|
|
out_file->Append("\n ------\n");
|
|
|
|
}
|
|
|
|
|
|
|
|
} // namespace rocksdb
|