|
|
|
// 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.
|
|
|
|
|
Fix a bug for SeekForPrev with partitioned filter and prefix (#8137)
Summary:
According to https://github.com/facebook/rocksdb/issues/5907, each filter partition "should include the bloom of the prefix of the last
key in the previous partition" so that SeekForPrev() in prefix mode can return correct result.
The prefix of the last key in the previous partition does not necessarily have the same prefix
as the first key in the current partition. Regardless of the first key in current partition, the
prefix of the last key in the previous partition should be added. The existing code, however,
does not follow this. Furthermore, there is another issue: when finishing current filter partition,
`FullFilterBlockBuilder::AddPrefix()` is called for the first key in next filter partition, which effectively
overwrites `last_prefix_str_` prematurely. Consequently, when the filter block builder proceeds
to the next partition, `last_prefix_str_` will be the prefix of its first key, leaving no way of adding
the bloom of the prefix of the last key of the previous partition.
Prefix extractor is FixedLength.2.
```
[ filter part 1 ] [ filter part 2 ]
abc d
```
When SeekForPrev("abcd"), checking the filter partition will land on filter part 2 because "abcd" > "abc"
but smaller than "d".
If the filter in filter part 2 happens to return false for the test for "ab", then SeekForPrev("abcd") will build
incorrect iterator tree in non-total-order mode.
Also fix a unit test which starts to fail following this PR. `InDomain` should not fail due to assertion
error when checking on an arbitrary key.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8137
Test Plan:
```
make check
```
Without this fix, the following command will fail pretty soon.
```
./db_stress --acquire_snapshot_one_in=10000 --avoid_flush_during_recovery=0 \
--avoid_unnecessary_blocking_io=0 --backup_max_size=104857600 --backup_one_in=0 \
--batch_protection_bytes_per_key=0 --block_size=16384 --bloom_bits=17 \
--bottommost_compression_type=disable --cache_index_and_filter_blocks=1 --cache_size=1048576 \
--checkpoint_one_in=0 --checksum_type=kxxHash64 --clear_column_family_one_in=0 \
--compact_files_one_in=1000000 --compact_range_one_in=1000000 --compaction_ttl=0 \
--compression_max_dict_buffer_bytes=0 --compression_max_dict_bytes=0 \
--compression_parallel_threads=1 --compression_type=zstd --compression_zstd_max_train_bytes=0 \
--continuous_verification_interval=0 --db=/dev/shm/rocksdb/rocksdb_crashtest_whitebox \
--db_write_buffer_size=8388608 --delpercent=5 --delrangepercent=0 --destroy_db_initially=0 --enable_blob_files=0 \
--enable_compaction_filter=0 --enable_pipelined_write=1 --file_checksum_impl=big --flush_one_in=1000000 \
--format_version=5 --get_current_wal_file_one_in=0 --get_live_files_one_in=1000000 --get_property_one_in=1000000 \
--get_sorted_wal_files_one_in=0 --index_block_restart_interval=4 --index_type=2 --ingest_external_file_one_in=0 \
--iterpercent=10 --key_len_percent_dist=1,30,69 --level_compaction_dynamic_level_bytes=True \
--log2_keys_per_lock=10 --long_running_snapshots=1 --mark_for_compaction_one_file_in=0 \
--max_background_compactions=20 --max_bytes_for_level_base=10485760 --max_key=100000000 --max_key_len=3 \
--max_manifest_file_size=1073741824 --max_write_batch_group_size_bytes=16777216 --max_write_buffer_number=3 \
--max_write_buffer_size_to_maintain=8388608 --memtablerep=skip_list --mmap_read=1 --mock_direct_io=False \
--nooverwritepercent=0 --open_files=500000 --ops_per_thread=20000000 --optimize_filters_for_memory=0 --paranoid_file_checks=1 --partition_filters=1 --partition_pinning=0 --pause_background_one_in=1000000 \
--periodic_compaction_seconds=0 --prefixpercent=5 --progress_reports=0 --read_fault_one_in=0 --read_only=0 \
--readpercent=45 --recycle_log_file_num=0 --reopen=20 --secondary_catch_up_one_in=0 \
--snapshot_hold_ops=100000 --sst_file_manager_bytes_per_sec=104857600 \
--sst_file_manager_bytes_per_truncate=0 --subcompactions=2 --sync=0 --sync_fault_injection=False \
--target_file_size_base=2097152 --target_file_size_multiplier=2 --test_batches_snapshots=0 --test_cf_consistency=0 \
--top_level_index_pinning=0 --unpartitioned_pinning=1 --use_blob_db=0 --use_block_based_filter=0 \
--use_direct_io_for_flush_and_compaction=0 --use_direct_reads=0 --use_full_merge_v1=0 --use_merge=0 \
--use_multiget=0 --use_ribbon_filter=0 --use_txn=0 --user_timestamp_size=8 --verify_checksum=1 \
--verify_checksum_one_in=1000000 --verify_db_one_in=100000 --write_buffer_size=4194304 \
--write_dbid_to_manifest=1 --writepercent=35
```
Reviewed By: pdillinger
Differential Revision: D27553054
Pulled By: riversand963
fbshipit-source-id: 60e391e4a2d8d98a9a3172ec5d6176b90ec3de98
4 years ago
|
|
|
#include <iomanip>
|
|
|
|
#include <sstream>
|
|
|
|
|
|
|
|
#include "db/db_test_util.h"
|
|
|
|
#include "options/options_helper.h"
|
|
|
|
#include "port/stack_trace.h"
|
|
|
|
#include "rocksdb/perf_context.h"
|
|
|
|
#include "table/block_based/filter_policy_internal.h"
|
|
|
|
#include "util/string_util.h"
|
|
|
|
|
|
|
|
namespace ROCKSDB_NAMESPACE {
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
using BFP = BloomFilterPolicy;
|
|
|
|
} // namespace
|
|
|
|
|
|
|
|
// DB tests related to bloom filter.
|
|
|
|
|
|
|
|
class DBBloomFilterTest : public DBTestBase {
|
|
|
|
public:
|
|
|
|
DBBloomFilterTest()
|
|
|
|
: DBTestBase("db_bloom_filter_test", /*env_do_fsync=*/true) {}
|
|
|
|
};
|
|
|
|
|
|
|
|
class DBBloomFilterTestWithParam : public DBTestBase,
|
|
|
|
public testing::WithParamInterface<
|
New Bloom filter implementation for full and partitioned filters (#6007)
Summary:
Adds an improved, replacement Bloom filter implementation (FastLocalBloom) for full and partitioned filters in the block-based table. This replacement is faster and more accurate, especially for high bits per key or millions of keys in a single filter.
Speed
The improved speed, at least on recent x86_64, comes from
* Using fastrange instead of modulo (%)
* Using our new hash function (XXH3 preview, added in a previous commit), which is much faster for large keys and only *slightly* slower on keys around 12 bytes if hashing the same size many thousands of times in a row.
* Optimizing the Bloom filter queries with AVX2 SIMD operations. (Added AVX2 to the USE_SSE=1 build.) Careful design was required to support (a) SIMD-optimized queries, (b) compatible non-SIMD code that's simple and efficient, (c) flexible choice of number of probes, and (d) essentially maximized accuracy for a cache-local Bloom filter. Probes are made eight at a time, so any number of probes up to 8 is the same speed, then up to 16, etc.
* Prefetching cache lines when building the filter. Although this optimization could be applied to the old structure as well, it seems to balance out the small added cost of accumulating 64 bit hashes for adding to the filter rather than 32 bit hashes.
Here's nominal speed data from filter_bench (200MB in filters, about 10k keys each, 10 bits filter data / key, 6 probes, avg key size 24 bytes, includes hashing time) on Skylake DE (relatively low clock speed):
$ ./filter_bench -quick -impl=2 -net_includes_hashing # New Bloom filter
Build avg ns/key: 47.7135
Mixed inside/outside queries...
Single filter net ns/op: 26.2825
Random filter net ns/op: 150.459
Average FP rate %: 0.954651
$ ./filter_bench -quick -impl=0 -net_includes_hashing # Old Bloom filter
Build avg ns/key: 47.2245
Mixed inside/outside queries...
Single filter net ns/op: 63.2978
Random filter net ns/op: 188.038
Average FP rate %: 1.13823
Similar build time but dramatically faster query times on hot data (63 ns to 26 ns), and somewhat faster on stale data (188 ns to 150 ns). Performance differences on batched and skewed query loads are between these extremes as expected.
The only other interesting thing about speed is "inside" (query key was added to filter) vs. "outside" (query key was not added to filter) query times. The non-SIMD implementations are substantially slower when most queries are "outside" vs. "inside". This goes against what one might expect or would have observed years ago, as "outside" queries only need about two probes on average, due to short-circuiting, while "inside" always have num_probes (say 6). The problem is probably the nastily unpredictable branch. The SIMD implementation has few branches (very predictable) and has pretty consistent running time regardless of query outcome.
Accuracy
The generally improved accuracy (re: Issue https://github.com/facebook/rocksdb/issues/5857) comes from a better design for probing indices
within a cache line (re: Issue https://github.com/facebook/rocksdb/issues/4120) and improved accuracy for millions of keys in a single filter from using a 64-bit hash function (XXH3p). Design details in code comments.
Accuracy data (generalizes, except old impl gets worse with millions of keys):
Memory bits per key: FP rate percent old impl -> FP rate percent new impl
6: 5.70953 -> 5.69888
8: 2.45766 -> 2.29709
10: 1.13977 -> 0.959254
12: 0.662498 -> 0.411593
16: 0.353023 -> 0.0873754
24: 0.261552 -> 0.0060971
50: 0.225453 -> ~0.00003 (less than 1 in a million queries are FP)
Fixes https://github.com/facebook/rocksdb/issues/5857
Fixes https://github.com/facebook/rocksdb/issues/4120
Unlike the old implementation, this implementation has a fixed cache line size (64 bytes). At 10 bits per key, the accuracy of this new implementation is very close to the old implementation with 128-byte cache line size. If there's sufficient demand, this implementation could be generalized.
Compatibility
Although old releases would see the new structure as corrupt filter data and read the table as if there's no filter, we've decided only to enable the new Bloom filter with new format_version=5. This provides a smooth path for automatic adoption over time, with an option for early opt-in.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6007
Test Plan: filter_bench has been used thoroughly to validate speed, accuracy, and correctness. Unit tests have been carefully updated to exercise new and old implementations, as well as the logic to select an implementation based on context (format_version).
Differential Revision: D18294749
Pulled By: pdillinger
fbshipit-source-id: d44c9db3696e4d0a17caaec47075b7755c262c5f
5 years ago
|
|
|
std::tuple<BFP::Mode, bool, uint32_t>> {
|
|
|
|
// public testing::WithParamInterface<bool> {
|
|
|
|
protected:
|
New Bloom filter implementation for full and partitioned filters (#6007)
Summary:
Adds an improved, replacement Bloom filter implementation (FastLocalBloom) for full and partitioned filters in the block-based table. This replacement is faster and more accurate, especially for high bits per key or millions of keys in a single filter.
Speed
The improved speed, at least on recent x86_64, comes from
* Using fastrange instead of modulo (%)
* Using our new hash function (XXH3 preview, added in a previous commit), which is much faster for large keys and only *slightly* slower on keys around 12 bytes if hashing the same size many thousands of times in a row.
* Optimizing the Bloom filter queries with AVX2 SIMD operations. (Added AVX2 to the USE_SSE=1 build.) Careful design was required to support (a) SIMD-optimized queries, (b) compatible non-SIMD code that's simple and efficient, (c) flexible choice of number of probes, and (d) essentially maximized accuracy for a cache-local Bloom filter. Probes are made eight at a time, so any number of probes up to 8 is the same speed, then up to 16, etc.
* Prefetching cache lines when building the filter. Although this optimization could be applied to the old structure as well, it seems to balance out the small added cost of accumulating 64 bit hashes for adding to the filter rather than 32 bit hashes.
Here's nominal speed data from filter_bench (200MB in filters, about 10k keys each, 10 bits filter data / key, 6 probes, avg key size 24 bytes, includes hashing time) on Skylake DE (relatively low clock speed):
$ ./filter_bench -quick -impl=2 -net_includes_hashing # New Bloom filter
Build avg ns/key: 47.7135
Mixed inside/outside queries...
Single filter net ns/op: 26.2825
Random filter net ns/op: 150.459
Average FP rate %: 0.954651
$ ./filter_bench -quick -impl=0 -net_includes_hashing # Old Bloom filter
Build avg ns/key: 47.2245
Mixed inside/outside queries...
Single filter net ns/op: 63.2978
Random filter net ns/op: 188.038
Average FP rate %: 1.13823
Similar build time but dramatically faster query times on hot data (63 ns to 26 ns), and somewhat faster on stale data (188 ns to 150 ns). Performance differences on batched and skewed query loads are between these extremes as expected.
The only other interesting thing about speed is "inside" (query key was added to filter) vs. "outside" (query key was not added to filter) query times. The non-SIMD implementations are substantially slower when most queries are "outside" vs. "inside". This goes against what one might expect or would have observed years ago, as "outside" queries only need about two probes on average, due to short-circuiting, while "inside" always have num_probes (say 6). The problem is probably the nastily unpredictable branch. The SIMD implementation has few branches (very predictable) and has pretty consistent running time regardless of query outcome.
Accuracy
The generally improved accuracy (re: Issue https://github.com/facebook/rocksdb/issues/5857) comes from a better design for probing indices
within a cache line (re: Issue https://github.com/facebook/rocksdb/issues/4120) and improved accuracy for millions of keys in a single filter from using a 64-bit hash function (XXH3p). Design details in code comments.
Accuracy data (generalizes, except old impl gets worse with millions of keys):
Memory bits per key: FP rate percent old impl -> FP rate percent new impl
6: 5.70953 -> 5.69888
8: 2.45766 -> 2.29709
10: 1.13977 -> 0.959254
12: 0.662498 -> 0.411593
16: 0.353023 -> 0.0873754
24: 0.261552 -> 0.0060971
50: 0.225453 -> ~0.00003 (less than 1 in a million queries are FP)
Fixes https://github.com/facebook/rocksdb/issues/5857
Fixes https://github.com/facebook/rocksdb/issues/4120
Unlike the old implementation, this implementation has a fixed cache line size (64 bytes). At 10 bits per key, the accuracy of this new implementation is very close to the old implementation with 128-byte cache line size. If there's sufficient demand, this implementation could be generalized.
Compatibility
Although old releases would see the new structure as corrupt filter data and read the table as if there's no filter, we've decided only to enable the new Bloom filter with new format_version=5. This provides a smooth path for automatic adoption over time, with an option for early opt-in.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6007
Test Plan: filter_bench has been used thoroughly to validate speed, accuracy, and correctness. Unit tests have been carefully updated to exercise new and old implementations, as well as the logic to select an implementation based on context (format_version).
Differential Revision: D18294749
Pulled By: pdillinger
fbshipit-source-id: d44c9db3696e4d0a17caaec47075b7755c262c5f
5 years ago
|
|
|
BFP::Mode bfp_impl_;
|
|
|
|
bool partition_filters_;
|
|
|
|
uint32_t format_version_;
|
|
|
|
|
|
|
|
public:
|
|
|
|
DBBloomFilterTestWithParam()
|
|
|
|
: DBTestBase("db_bloom_filter_tests", /*env_do_fsync=*/true) {}
|
|
|
|
|
|
|
|
~DBBloomFilterTestWithParam() override {}
|
|
|
|
|
|
|
|
void SetUp() override {
|
|
|
|
bfp_impl_ = std::get<0>(GetParam());
|
|
|
|
partition_filters_ = std::get<1>(GetParam());
|
|
|
|
format_version_ = std::get<2>(GetParam());
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
class DBBloomFilterTestDefFormatVersion : public DBBloomFilterTestWithParam {};
|
|
|
|
|
|
|
|
class SliceTransformLimitedDomainGeneric : public SliceTransform {
|
|
|
|
const char* Name() const override {
|
|
|
|
return "SliceTransformLimitedDomainGeneric";
|
|
|
|
}
|
|
|
|
|
|
|
|
Slice Transform(const Slice& src) const override {
|
|
|
|
return Slice(src.data(), 5);
|
|
|
|
}
|
|
|
|
|
|
|
|
bool InDomain(const Slice& src) const override {
|
|
|
|
// prefix will be x????
|
|
|
|
return src.size() >= 5;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool InRange(const Slice& dst) const override {
|
|
|
|
// prefix will be x????
|
|
|
|
return dst.size() == 5;
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
// KeyMayExist can lead to a few false positives, but not false negatives.
|
|
|
|
// To make test deterministic, use a much larger number of bits per key-20 than
|
|
|
|
// bits in the key, so that false positives are eliminated
|
|
|
|
TEST_P(DBBloomFilterTestDefFormatVersion, KeyMayExist) {
|
|
|
|
do {
|
|
|
|
ReadOptions ropts;
|
|
|
|
std::string value;
|
|
|
|
anon::OptionsOverride options_override;
|
|
|
|
options_override.filter_policy.reset(new BFP(20, bfp_impl_));
|
|
|
|
options_override.partition_filters = partition_filters_;
|
|
|
|
options_override.metadata_block_size = 32;
|
|
|
|
Options options = CurrentOptions(options_override);
|
|
|
|
if (partition_filters_) {
|
|
|
|
auto* table_options =
|
|
|
|
options.table_factory->GetOptions<BlockBasedTableOptions>();
|
|
|
|
if (table_options != nullptr &&
|
|
|
|
table_options->index_type !=
|
|
|
|
BlockBasedTableOptions::kTwoLevelIndexSearch) {
|
|
|
|
// In the current implementation partitioned filters depend on
|
|
|
|
// partitioned indexes
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
|
|
|
|
CreateAndReopenWithCF({"pikachu"}, options);
|
|
|
|
|
|
|
|
ASSERT_TRUE(!db_->KeyMayExist(ropts, handles_[1], "a", &value));
|
|
|
|
|
|
|
|
ASSERT_OK(Put(1, "a", "b"));
|
|
|
|
bool value_found = false;
|
|
|
|
ASSERT_TRUE(
|
|
|
|
db_->KeyMayExist(ropts, handles_[1], "a", &value, &value_found));
|
|
|
|
ASSERT_TRUE(value_found);
|
|
|
|
ASSERT_EQ("b", value);
|
|
|
|
|
|
|
|
ASSERT_OK(Flush(1));
|
|
|
|
value.clear();
|
|
|
|
|
|
|
|
uint64_t numopen = TestGetTickerCount(options, NO_FILE_OPENS);
|
|
|
|
uint64_t cache_added = TestGetTickerCount(options, BLOCK_CACHE_ADD);
|
|
|
|
ASSERT_TRUE(
|
|
|
|
db_->KeyMayExist(ropts, handles_[1], "a", &value, &value_found));
|
|
|
|
ASSERT_TRUE(!value_found);
|
|
|
|
// assert that no new files were opened and no new blocks were
|
|
|
|
// read into block cache.
|
|
|
|
ASSERT_EQ(numopen, TestGetTickerCount(options, NO_FILE_OPENS));
|
|
|
|
ASSERT_EQ(cache_added, TestGetTickerCount(options, BLOCK_CACHE_ADD));
|
|
|
|
|
|
|
|
ASSERT_OK(Delete(1, "a"));
|
|
|
|
|
|
|
|
numopen = TestGetTickerCount(options, NO_FILE_OPENS);
|
|
|
|
cache_added = TestGetTickerCount(options, BLOCK_CACHE_ADD);
|
|
|
|
ASSERT_TRUE(!db_->KeyMayExist(ropts, handles_[1], "a", &value));
|
|
|
|
ASSERT_EQ(numopen, TestGetTickerCount(options, NO_FILE_OPENS));
|
|
|
|
ASSERT_EQ(cache_added, TestGetTickerCount(options, BLOCK_CACHE_ADD));
|
|
|
|
|
|
|
|
ASSERT_OK(Flush(1));
|
|
|
|
ASSERT_OK(dbfull()->TEST_CompactRange(0, nullptr, nullptr, handles_[1],
|
|
|
|
true /* disallow trivial move */));
|
|
|
|
|
|
|
|
numopen = TestGetTickerCount(options, NO_FILE_OPENS);
|
|
|
|
cache_added = TestGetTickerCount(options, BLOCK_CACHE_ADD);
|
|
|
|
ASSERT_TRUE(!db_->KeyMayExist(ropts, handles_[1], "a", &value));
|
|
|
|
ASSERT_EQ(numopen, TestGetTickerCount(options, NO_FILE_OPENS));
|
|
|
|
ASSERT_EQ(cache_added, TestGetTickerCount(options, BLOCK_CACHE_ADD));
|
|
|
|
|
|
|
|
ASSERT_OK(Delete(1, "c"));
|
|
|
|
|
|
|
|
numopen = TestGetTickerCount(options, NO_FILE_OPENS);
|
|
|
|
cache_added = TestGetTickerCount(options, BLOCK_CACHE_ADD);
|
|
|
|
ASSERT_TRUE(!db_->KeyMayExist(ropts, handles_[1], "c", &value));
|
|
|
|
ASSERT_EQ(numopen, TestGetTickerCount(options, NO_FILE_OPENS));
|
|
|
|
ASSERT_EQ(cache_added, TestGetTickerCount(options, BLOCK_CACHE_ADD));
|
|
|
|
|
|
|
|
// KeyMayExist function only checks data in block caches, which is not used
|
|
|
|
// by plain table format.
|
|
|
|
} while (
|
|
|
|
ChangeOptions(kSkipPlainTable | kSkipHashIndex | kSkipFIFOCompaction));
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(DBBloomFilterTest, GetFilterByPrefixBloomCustomPrefixExtractor) {
|
|
|
|
for (bool partition_filters : {true, false}) {
|
|
|
|
Options options = last_options_;
|
|
|
|
options.prefix_extractor =
|
|
|
|
std::make_shared<SliceTransformLimitedDomainGeneric>();
|
|
|
|
options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
|
|
|
|
get_perf_context()->EnablePerLevelPerfContext();
|
|
|
|
BlockBasedTableOptions bbto;
|
|
|
|
bbto.filter_policy.reset(NewBloomFilterPolicy(10, false));
|
|
|
|
if (partition_filters) {
|
|
|
|
bbto.partition_filters = true;
|
|
|
|
bbto.index_type = BlockBasedTableOptions::IndexType::kTwoLevelIndexSearch;
|
|
|
|
}
|
|
|
|
bbto.whole_key_filtering = false;
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
|
|
|
|
DestroyAndReopen(options);
|
|
|
|
|
|
|
|
WriteOptions wo;
|
|
|
|
ReadOptions ro;
|
|
|
|
FlushOptions fo;
|
|
|
|
fo.wait = true;
|
|
|
|
std::string value;
|
|
|
|
|
|
|
|
ASSERT_OK(dbfull()->Put(wo, "barbarbar", "foo"));
|
|
|
|
ASSERT_OK(dbfull()->Put(wo, "barbarbar2", "foo2"));
|
|
|
|
ASSERT_OK(dbfull()->Put(wo, "foofoofoo", "bar"));
|
|
|
|
|
|
|
|
ASSERT_OK(dbfull()->Flush(fo));
|
|
|
|
|
|
|
|
ASSERT_EQ("foo", Get("barbarbar"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 0);
|
|
|
|
ASSERT_EQ(
|
|
|
|
0,
|
|
|
|
(*(get_perf_context()->level_to_perf_context))[0].bloom_filter_useful);
|
|
|
|
ASSERT_EQ("foo2", Get("barbarbar2"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 0);
|
|
|
|
ASSERT_EQ(
|
|
|
|
0,
|
|
|
|
(*(get_perf_context()->level_to_perf_context))[0].bloom_filter_useful);
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get("barbarbar3"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 0);
|
|
|
|
ASSERT_EQ(
|
|
|
|
0,
|
|
|
|
(*(get_perf_context()->level_to_perf_context))[0].bloom_filter_useful);
|
|
|
|
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get("barfoofoo"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 1);
|
|
|
|
ASSERT_EQ(
|
|
|
|
1,
|
|
|
|
(*(get_perf_context()->level_to_perf_context))[0].bloom_filter_useful);
|
|
|
|
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get("foobarbar"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 2);
|
|
|
|
ASSERT_EQ(
|
|
|
|
2,
|
|
|
|
(*(get_perf_context()->level_to_perf_context))[0].bloom_filter_useful);
|
|
|
|
|
|
|
|
ro.total_order_seek = true;
|
|
|
|
ASSERT_TRUE(db_->Get(ro, "foobarbar", &value).IsNotFound());
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 2);
|
|
|
|
ASSERT_EQ(
|
|
|
|
2,
|
|
|
|
(*(get_perf_context()->level_to_perf_context))[0].bloom_filter_useful);
|
|
|
|
get_perf_context()->Reset();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(DBBloomFilterTest, GetFilterByPrefixBloom) {
|
|
|
|
for (bool partition_filters : {true, false}) {
|
|
|
|
Options options = last_options_;
|
|
|
|
options.prefix_extractor.reset(NewFixedPrefixTransform(8));
|
|
|
|
options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
|
|
|
|
get_perf_context()->EnablePerLevelPerfContext();
|
|
|
|
BlockBasedTableOptions bbto;
|
|
|
|
bbto.filter_policy.reset(NewBloomFilterPolicy(10, false));
|
|
|
|
if (partition_filters) {
|
|
|
|
bbto.partition_filters = true;
|
|
|
|
bbto.index_type = BlockBasedTableOptions::IndexType::kTwoLevelIndexSearch;
|
|
|
|
}
|
|
|
|
bbto.whole_key_filtering = false;
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
|
|
|
|
DestroyAndReopen(options);
|
|
|
|
|
|
|
|
WriteOptions wo;
|
|
|
|
ReadOptions ro;
|
|
|
|
FlushOptions fo;
|
|
|
|
fo.wait = true;
|
|
|
|
std::string value;
|
|
|
|
|
|
|
|
ASSERT_OK(dbfull()->Put(wo, "barbarbar", "foo"));
|
|
|
|
ASSERT_OK(dbfull()->Put(wo, "barbarbar2", "foo2"));
|
|
|
|
ASSERT_OK(dbfull()->Put(wo, "foofoofoo", "bar"));
|
|
|
|
|
|
|
|
ASSERT_OK(dbfull()->Flush(fo));
|
|
|
|
|
|
|
|
ASSERT_EQ("foo", Get("barbarbar"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 0);
|
|
|
|
ASSERT_EQ("foo2", Get("barbarbar2"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 0);
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get("barbarbar3"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 0);
|
|
|
|
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get("barfoofoo"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 1);
|
|
|
|
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get("foobarbar"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 2);
|
|
|
|
|
|
|
|
ro.total_order_seek = true;
|
|
|
|
ASSERT_TRUE(db_->Get(ro, "foobarbar", &value).IsNotFound());
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 2);
|
|
|
|
ASSERT_EQ(
|
|
|
|
2,
|
|
|
|
(*(get_perf_context()->level_to_perf_context))[0].bloom_filter_useful);
|
|
|
|
get_perf_context()->Reset();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(DBBloomFilterTest, WholeKeyFilterProp) {
|
|
|
|
for (bool partition_filters : {true, false}) {
|
|
|
|
Options options = last_options_;
|
|
|
|
options.prefix_extractor.reset(NewFixedPrefixTransform(3));
|
|
|
|
options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
|
|
|
|
get_perf_context()->EnablePerLevelPerfContext();
|
|
|
|
|
|
|
|
BlockBasedTableOptions bbto;
|
|
|
|
bbto.filter_policy.reset(NewBloomFilterPolicy(10, false));
|
|
|
|
bbto.whole_key_filtering = false;
|
|
|
|
if (partition_filters) {
|
|
|
|
bbto.partition_filters = true;
|
|
|
|
bbto.index_type = BlockBasedTableOptions::IndexType::kTwoLevelIndexSearch;
|
|
|
|
}
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
|
|
|
|
DestroyAndReopen(options);
|
|
|
|
|
|
|
|
WriteOptions wo;
|
|
|
|
ReadOptions ro;
|
|
|
|
FlushOptions fo;
|
|
|
|
fo.wait = true;
|
|
|
|
std::string value;
|
|
|
|
|
|
|
|
ASSERT_OK(dbfull()->Put(wo, "foobar", "foo"));
|
|
|
|
// Needs insert some keys to make sure files are not filtered out by key
|
|
|
|
// ranges.
|
|
|
|
ASSERT_OK(dbfull()->Put(wo, "aaa", ""));
|
|
|
|
ASSERT_OK(dbfull()->Put(wo, "zzz", ""));
|
|
|
|
ASSERT_OK(dbfull()->Flush(fo));
|
|
|
|
|
|
|
|
Reopen(options);
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get("foo"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 0);
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get("bar"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 1);
|
|
|
|
ASSERT_EQ("foo", Get("foobar"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 1);
|
|
|
|
|
|
|
|
// Reopen with whole key filtering enabled and prefix extractor
|
|
|
|
// NULL. Bloom filter should be off for both of whole key and
|
|
|
|
// prefix bloom.
|
|
|
|
bbto.whole_key_filtering = true;
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
|
|
|
|
options.prefix_extractor.reset();
|
|
|
|
Reopen(options);
|
|
|
|
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 1);
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get("foo"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 1);
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get("bar"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 1);
|
|
|
|
ASSERT_EQ("foo", Get("foobar"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 1);
|
|
|
|
// Write DB with only full key filtering.
|
|
|
|
ASSERT_OK(dbfull()->Put(wo, "foobar", "foo"));
|
|
|
|
// Needs insert some keys to make sure files are not filtered out by key
|
|
|
|
// ranges.
|
|
|
|
ASSERT_OK(dbfull()->Put(wo, "aaa", ""));
|
|
|
|
ASSERT_OK(dbfull()->Put(wo, "zzz", ""));
|
|
|
|
ASSERT_OK(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
|
|
|
|
|
|
|
|
// Reopen with both of whole key off and prefix extractor enabled.
|
|
|
|
// Still no bloom filter should be used.
|
|
|
|
options.prefix_extractor.reset(NewFixedPrefixTransform(3));
|
|
|
|
bbto.whole_key_filtering = false;
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
|
|
|
|
Reopen(options);
|
|
|
|
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 1);
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get("foo"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 1);
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get("bar"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 1);
|
|
|
|
ASSERT_EQ("foo", Get("foobar"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 1);
|
|
|
|
|
|
|
|
// Try to create a DB with mixed files:
|
|
|
|
ASSERT_OK(dbfull()->Put(wo, "foobar", "foo"));
|
|
|
|
// Needs insert some keys to make sure files are not filtered out by key
|
|
|
|
// ranges.
|
|
|
|
ASSERT_OK(dbfull()->Put(wo, "aaa", ""));
|
|
|
|
ASSERT_OK(dbfull()->Put(wo, "zzz", ""));
|
|
|
|
ASSERT_OK(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
|
|
|
|
|
|
|
|
options.prefix_extractor.reset();
|
|
|
|
bbto.whole_key_filtering = true;
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
|
|
|
|
Reopen(options);
|
|
|
|
|
|
|
|
// Try to create a DB with mixed files.
|
|
|
|
ASSERT_OK(dbfull()->Put(wo, "barfoo", "bar"));
|
|
|
|
// In this case needs insert some keys to make sure files are
|
|
|
|
// not filtered out by key ranges.
|
|
|
|
ASSERT_OK(dbfull()->Put(wo, "aaa", ""));
|
|
|
|
ASSERT_OK(dbfull()->Put(wo, "zzz", ""));
|
|
|
|
ASSERT_OK(Flush());
|
|
|
|
|
|
|
|
// Now we have two files:
|
|
|
|
// File 1: An older file with prefix bloom.
|
|
|
|
// File 2: A newer file with whole bloom filter.
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 1);
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get("foo"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 2);
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get("bar"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 3);
|
|
|
|
ASSERT_EQ("foo", Get("foobar"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 4);
|
|
|
|
ASSERT_EQ("bar", Get("barfoo"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 4);
|
|
|
|
|
|
|
|
// Reopen with the same setting: only whole key is used
|
|
|
|
Reopen(options);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 4);
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get("foo"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 5);
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get("bar"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 6);
|
|
|
|
ASSERT_EQ("foo", Get("foobar"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 7);
|
|
|
|
ASSERT_EQ("bar", Get("barfoo"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 7);
|
|
|
|
|
|
|
|
// Restart with both filters are allowed
|
|
|
|
options.prefix_extractor.reset(NewFixedPrefixTransform(3));
|
|
|
|
bbto.whole_key_filtering = true;
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
|
|
|
|
Reopen(options);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 7);
|
|
|
|
// File 1 will has it filtered out.
|
|
|
|
// File 2 will not, as prefix `foo` exists in the file.
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get("foo"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 8);
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get("bar"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 10);
|
|
|
|
ASSERT_EQ("foo", Get("foobar"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 11);
|
|
|
|
ASSERT_EQ("bar", Get("barfoo"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 11);
|
|
|
|
|
|
|
|
// Restart with only prefix bloom is allowed.
|
|
|
|
options.prefix_extractor.reset(NewFixedPrefixTransform(3));
|
|
|
|
bbto.whole_key_filtering = false;
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
|
|
|
|
Reopen(options);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 11);
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get("foo"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 11);
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get("bar"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 12);
|
|
|
|
ASSERT_EQ("foo", Get("foobar"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 12);
|
|
|
|
ASSERT_EQ("bar", Get("barfoo"));
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 12);
|
|
|
|
uint64_t bloom_filter_useful_all_levels = 0;
|
|
|
|
for (auto& kv : (*(get_perf_context()->level_to_perf_context))) {
|
|
|
|
if (kv.second.bloom_filter_useful > 0) {
|
|
|
|
bloom_filter_useful_all_levels += kv.second.bloom_filter_useful;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
ASSERT_EQ(12, bloom_filter_useful_all_levels);
|
|
|
|
get_perf_context()->Reset();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_P(DBBloomFilterTestWithParam, BloomFilter) {
|
|
|
|
do {
|
|
|
|
Options options = CurrentOptions();
|
|
|
|
env_->count_random_reads_ = true;
|
|
|
|
options.env = env_;
|
|
|
|
// ChangeCompactOptions() only changes compaction style, which does not
|
|
|
|
// trigger reset of table_factory
|
|
|
|
BlockBasedTableOptions table_options;
|
|
|
|
table_options.no_block_cache = true;
|
|
|
|
table_options.filter_policy.reset(new BFP(10, bfp_impl_));
|
|
|
|
table_options.partition_filters = partition_filters_;
|
|
|
|
if (partition_filters_) {
|
|
|
|
table_options.index_type =
|
|
|
|
BlockBasedTableOptions::IndexType::kTwoLevelIndexSearch;
|
|
|
|
}
|
|
|
|
table_options.format_version = format_version_;
|
|
|
|
if (format_version_ >= 4) {
|
|
|
|
// value delta encoding challenged more with index interval > 1
|
|
|
|
table_options.index_block_restart_interval = 8;
|
|
|
|
}
|
|
|
|
table_options.metadata_block_size = 32;
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
|
|
|
|
|
|
|
|
CreateAndReopenWithCF({"pikachu"}, options);
|
|
|
|
|
|
|
|
// Populate multiple layers
|
|
|
|
const int N = 10000;
|
|
|
|
for (int i = 0; i < N; i++) {
|
|
|
|
ASSERT_OK(Put(1, Key(i), Key(i)));
|
|
|
|
}
|
|
|
|
Compact(1, "a", "z");
|
|
|
|
for (int i = 0; i < N; i += 100) {
|
|
|
|
ASSERT_OK(Put(1, Key(i), Key(i)));
|
|
|
|
}
|
|
|
|
ASSERT_OK(Flush(1));
|
|
|
|
|
|
|
|
// Prevent auto compactions triggered by seeks
|
|
|
|
env_->delay_sstable_sync_.store(true, std::memory_order_release);
|
|
|
|
|
|
|
|
// Lookup present keys. Should rarely read from small sstable.
|
|
|
|
env_->random_read_counter_.Reset();
|
|
|
|
for (int i = 0; i < N; i++) {
|
|
|
|
ASSERT_EQ(Key(i), Get(1, Key(i)));
|
|
|
|
}
|
|
|
|
int reads = env_->random_read_counter_.Read();
|
|
|
|
fprintf(stderr, "%d present => %d reads\n", N, reads);
|
|
|
|
ASSERT_GE(reads, N);
|
|
|
|
if (partition_filters_) {
|
|
|
|
// Without block cache, we read an extra partition filter per each
|
|
|
|
// level*read and a partition index per each read
|
|
|
|
ASSERT_LE(reads, 4 * N + 2 * N / 100);
|
|
|
|
} else {
|
|
|
|
ASSERT_LE(reads, N + 2 * N / 100);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Lookup present keys. Should rarely read from either sstable.
|
|
|
|
env_->random_read_counter_.Reset();
|
|
|
|
for (int i = 0; i < N; i++) {
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get(1, Key(i) + ".missing"));
|
|
|
|
}
|
|
|
|
reads = env_->random_read_counter_.Read();
|
|
|
|
fprintf(stderr, "%d missing => %d reads\n", N, reads);
|
|
|
|
if (partition_filters_) {
|
|
|
|
// With partitioned filter we read one extra filter per level per each
|
|
|
|
// missed read.
|
|
|
|
ASSERT_LE(reads, 2 * N + 3 * N / 100);
|
|
|
|
} else {
|
|
|
|
ASSERT_LE(reads, 3 * N / 100);
|
|
|
|
}
|
|
|
|
|
|
|
|
#ifndef ROCKSDB_LITE
|
|
|
|
// Sanity check some table properties
|
|
|
|
std::map<std::string, std::string> props;
|
|
|
|
ASSERT_TRUE(db_->GetMapProperty(
|
|
|
|
handles_[1], DB::Properties::kAggregatedTableProperties, &props));
|
|
|
|
uint64_t nkeys = N + N / 100;
|
|
|
|
uint64_t filter_size = ParseUint64(props["filter_size"]);
|
|
|
|
EXPECT_LE(filter_size,
|
|
|
|
(partition_filters_ ? 12 : 11) * nkeys / /*bits / byte*/ 8);
|
|
|
|
EXPECT_GE(filter_size, 10 * nkeys / /*bits / byte*/ 8);
|
|
|
|
|
|
|
|
uint64_t num_filter_entries = ParseUint64(props["num_filter_entries"]);
|
|
|
|
EXPECT_EQ(num_filter_entries, nkeys);
|
|
|
|
#endif // ROCKSDB_LITE
|
|
|
|
|
|
|
|
env_->delay_sstable_sync_.store(false, std::memory_order_release);
|
|
|
|
Close();
|
|
|
|
} while (ChangeCompactOptions());
|
|
|
|
}
|
|
|
|
|
|
|
|
#if !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
|
|
|
|
INSTANTIATE_TEST_CASE_P(
|
|
|
|
FormatDef, DBBloomFilterTestDefFormatVersion,
|
|
|
|
::testing::Values(
|
New Bloom filter implementation for full and partitioned filters (#6007)
Summary:
Adds an improved, replacement Bloom filter implementation (FastLocalBloom) for full and partitioned filters in the block-based table. This replacement is faster and more accurate, especially for high bits per key or millions of keys in a single filter.
Speed
The improved speed, at least on recent x86_64, comes from
* Using fastrange instead of modulo (%)
* Using our new hash function (XXH3 preview, added in a previous commit), which is much faster for large keys and only *slightly* slower on keys around 12 bytes if hashing the same size many thousands of times in a row.
* Optimizing the Bloom filter queries with AVX2 SIMD operations. (Added AVX2 to the USE_SSE=1 build.) Careful design was required to support (a) SIMD-optimized queries, (b) compatible non-SIMD code that's simple and efficient, (c) flexible choice of number of probes, and (d) essentially maximized accuracy for a cache-local Bloom filter. Probes are made eight at a time, so any number of probes up to 8 is the same speed, then up to 16, etc.
* Prefetching cache lines when building the filter. Although this optimization could be applied to the old structure as well, it seems to balance out the small added cost of accumulating 64 bit hashes for adding to the filter rather than 32 bit hashes.
Here's nominal speed data from filter_bench (200MB in filters, about 10k keys each, 10 bits filter data / key, 6 probes, avg key size 24 bytes, includes hashing time) on Skylake DE (relatively low clock speed):
$ ./filter_bench -quick -impl=2 -net_includes_hashing # New Bloom filter
Build avg ns/key: 47.7135
Mixed inside/outside queries...
Single filter net ns/op: 26.2825
Random filter net ns/op: 150.459
Average FP rate %: 0.954651
$ ./filter_bench -quick -impl=0 -net_includes_hashing # Old Bloom filter
Build avg ns/key: 47.2245
Mixed inside/outside queries...
Single filter net ns/op: 63.2978
Random filter net ns/op: 188.038
Average FP rate %: 1.13823
Similar build time but dramatically faster query times on hot data (63 ns to 26 ns), and somewhat faster on stale data (188 ns to 150 ns). Performance differences on batched and skewed query loads are between these extremes as expected.
The only other interesting thing about speed is "inside" (query key was added to filter) vs. "outside" (query key was not added to filter) query times. The non-SIMD implementations are substantially slower when most queries are "outside" vs. "inside". This goes against what one might expect or would have observed years ago, as "outside" queries only need about two probes on average, due to short-circuiting, while "inside" always have num_probes (say 6). The problem is probably the nastily unpredictable branch. The SIMD implementation has few branches (very predictable) and has pretty consistent running time regardless of query outcome.
Accuracy
The generally improved accuracy (re: Issue https://github.com/facebook/rocksdb/issues/5857) comes from a better design for probing indices
within a cache line (re: Issue https://github.com/facebook/rocksdb/issues/4120) and improved accuracy for millions of keys in a single filter from using a 64-bit hash function (XXH3p). Design details in code comments.
Accuracy data (generalizes, except old impl gets worse with millions of keys):
Memory bits per key: FP rate percent old impl -> FP rate percent new impl
6: 5.70953 -> 5.69888
8: 2.45766 -> 2.29709
10: 1.13977 -> 0.959254
12: 0.662498 -> 0.411593
16: 0.353023 -> 0.0873754
24: 0.261552 -> 0.0060971
50: 0.225453 -> ~0.00003 (less than 1 in a million queries are FP)
Fixes https://github.com/facebook/rocksdb/issues/5857
Fixes https://github.com/facebook/rocksdb/issues/4120
Unlike the old implementation, this implementation has a fixed cache line size (64 bytes). At 10 bits per key, the accuracy of this new implementation is very close to the old implementation with 128-byte cache line size. If there's sufficient demand, this implementation could be generalized.
Compatibility
Although old releases would see the new structure as corrupt filter data and read the table as if there's no filter, we've decided only to enable the new Bloom filter with new format_version=5. This provides a smooth path for automatic adoption over time, with an option for early opt-in.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6007
Test Plan: filter_bench has been used thoroughly to validate speed, accuracy, and correctness. Unit tests have been carefully updated to exercise new and old implementations, as well as the logic to select an implementation based on context (format_version).
Differential Revision: D18294749
Pulled By: pdillinger
fbshipit-source-id: d44c9db3696e4d0a17caaec47075b7755c262c5f
5 years ago
|
|
|
std::make_tuple(BFP::kDeprecatedBlock, false,
|
|
|
|
test::kDefaultFormatVersion),
|
Experimental (production candidate) SST schema for Ribbon filter (#7658)
Summary:
Added experimental public API for Ribbon filter:
NewExperimentalRibbonFilterPolicy(). This experimental API will
take a "Bloom equivalent" bits per key, and configure the Ribbon
filter for the same FP rate as Bloom would have but ~30% space
savings. (Note: optimize_filters_for_memory is not yet implemented
for Ribbon filter. That can be added with no effect on schema.)
Internally, the Ribbon filter is configured using a "one_in_fp_rate"
value, which is 1 over desired FP rate. For example, use 100 for 1%
FP rate. I'm expecting this will be used in the future for configuring
Bloom-like filters, as I expect people to more commonly hold constant
the filter accuracy and change the space vs. time trade-off, rather than
hold constant the space (per key) and change the accuracy vs. time
trade-off, though we might make that available.
### Benchmarking
```
$ ./filter_bench -impl=2 -quick -m_keys_total_max=200 -average_keys_per_filter=100000 -net_includes_hashing
Building...
Build avg ns/key: 34.1341
Number of filters: 1993
Total size (MB): 238.488
Reported total allocated memory (MB): 262.875
Reported internal fragmentation: 10.2255%
Bits/key stored: 10.0029
----------------------------
Mixed inside/outside queries...
Single filter net ns/op: 18.7508
Random filter net ns/op: 258.246
Average FP rate %: 0.968672
----------------------------
Done. (For more info, run with -legend or -help.)
$ ./filter_bench -impl=3 -quick -m_keys_total_max=200 -average_keys_per_filter=100000 -net_includes_hashing
Building...
Build avg ns/key: 130.851
Number of filters: 1993
Total size (MB): 168.166
Reported total allocated memory (MB): 183.211
Reported internal fragmentation: 8.94626%
Bits/key stored: 7.05341
----------------------------
Mixed inside/outside queries...
Single filter net ns/op: 58.4523
Random filter net ns/op: 363.717
Average FP rate %: 0.952978
----------------------------
Done. (For more info, run with -legend or -help.)
```
168.166 / 238.488 = 0.705 -> 29.5% space reduction
130.851 / 34.1341 = 3.83x construction time for this Ribbon filter vs. lastest Bloom filter (could make that as little as about 2.5x for less space reduction)
### Working around a hashing "flaw"
bloom_test discovered a flaw in the simple hashing applied in
StandardHasher when num_starts == 1 (num_slots == 128), showing an
excessively high FP rate. The problem is that when many entries, on the
order of number of hash bits or kCoeffBits, are associated with the same
start location, the correlation between the CoeffRow and ResultRow (for
efficiency) can lead to a solution that is "universal," or nearly so, for
entries mapping to that start location. (Normally, variance in start
location breaks the effective association between CoeffRow and
ResultRow; the same value for CoeffRow is effectively different if start
locations are different.) Without kUseSmash and with num_starts > 1 (thus
num_starts ~= num_slots), this flaw should be completely irrelevant. Even
with 10M slots, the chances of a single slot having just 16 (or more)
entries map to it--not enough to cause an FP problem, which would be local
to that slot if it happened--is 1 in millions. This spreadsheet formula
shows that: =1/(10000000*(1 - POISSON(15, 1, TRUE)))
As kUseSmash==false (the setting for Standard128RibbonBitsBuilder) is
intended for CPU efficiency of filters with many more entries/slots than
kCoeffBits, a very reasonable work-around is to disallow num_starts==1
when !kUseSmash, by making the minimum non-zero number of slots
2*kCoeffBits. This is the work-around I've applied. This also means that
the new Ribbon filter schema (Standard128RibbonBitsBuilder) is not
space-efficient for less than a few hundred entries. Because of this, I
have made it fall back on constructing a Bloom filter, under existing
schema, when that is more space efficient for small filters. (We can
change this in the future if we want.)
TODO: better unit tests for this case in ribbon_test, and probably
update StandardHasher for kUseSmash case so that it can scale nicely to
small filters.
### Other related changes
* Add Ribbon filter to stress/crash test
* Add Ribbon filter to filter_bench as -impl=3
* Add option string support, as in "filter_policy=experimental_ribbon:5.678;"
where 5.678 is the Bloom equivalent bits per key.
* Rename internal mode BloomFilterPolicy::kAuto to kAutoBloom
* Add a general BuiltinFilterBitsBuilder::CalculateNumEntry based on
binary searching CalculateSpace (inefficient), so that subclasses
(especially experimental ones) don't have to provide an efficient
implementation inverting CalculateSpace.
* Minor refactor FastLocalBloomBitsBuilder for new base class
XXH3pFilterBitsBuilder shared with new Standard128RibbonBitsBuilder,
which allows the latter to fall back on Bloom construction in some
extreme cases.
* Mostly updated bloom_test for Ribbon filter, though a test like
FullBloomTest::Schema is a next TODO to ensure schema stability
(in case this becomes production-ready schema as it is).
* Add some APIs to ribbon_impl.h for configuring Ribbon filters.
Although these are reasonably covered by bloom_test, TODO more unit
tests in ribbon_test
* Added a "tool" FindOccupancyForSuccessRate to ribbon_test to get data
for constructing the linear approximations in GetNumSlotsFor95PctSuccess.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7658
Test Plan:
Some unit tests updated but other testing is left TODO. This
is considered experimental but laying down schema compatibility as early
as possible in case it proves production-quality. Also tested in
stress/crash test.
Reviewed By: jay-zhuang
Differential Revision: D24899349
Pulled By: pdillinger
fbshipit-source-id: 9715f3e6371c959d923aea8077c9423c7a9f82b8
4 years ago
|
|
|
std::make_tuple(BFP::kAutoBloom, true, test::kDefaultFormatVersion),
|
|
|
|
std::make_tuple(BFP::kAutoBloom, false, test::kDefaultFormatVersion)));
|
|
|
|
|
|
|
|
INSTANTIATE_TEST_CASE_P(
|
|
|
|
FormatDef, DBBloomFilterTestWithParam,
|
|
|
|
::testing::Values(
|
New Bloom filter implementation for full and partitioned filters (#6007)
Summary:
Adds an improved, replacement Bloom filter implementation (FastLocalBloom) for full and partitioned filters in the block-based table. This replacement is faster and more accurate, especially for high bits per key or millions of keys in a single filter.
Speed
The improved speed, at least on recent x86_64, comes from
* Using fastrange instead of modulo (%)
* Using our new hash function (XXH3 preview, added in a previous commit), which is much faster for large keys and only *slightly* slower on keys around 12 bytes if hashing the same size many thousands of times in a row.
* Optimizing the Bloom filter queries with AVX2 SIMD operations. (Added AVX2 to the USE_SSE=1 build.) Careful design was required to support (a) SIMD-optimized queries, (b) compatible non-SIMD code that's simple and efficient, (c) flexible choice of number of probes, and (d) essentially maximized accuracy for a cache-local Bloom filter. Probes are made eight at a time, so any number of probes up to 8 is the same speed, then up to 16, etc.
* Prefetching cache lines when building the filter. Although this optimization could be applied to the old structure as well, it seems to balance out the small added cost of accumulating 64 bit hashes for adding to the filter rather than 32 bit hashes.
Here's nominal speed data from filter_bench (200MB in filters, about 10k keys each, 10 bits filter data / key, 6 probes, avg key size 24 bytes, includes hashing time) on Skylake DE (relatively low clock speed):
$ ./filter_bench -quick -impl=2 -net_includes_hashing # New Bloom filter
Build avg ns/key: 47.7135
Mixed inside/outside queries...
Single filter net ns/op: 26.2825
Random filter net ns/op: 150.459
Average FP rate %: 0.954651
$ ./filter_bench -quick -impl=0 -net_includes_hashing # Old Bloom filter
Build avg ns/key: 47.2245
Mixed inside/outside queries...
Single filter net ns/op: 63.2978
Random filter net ns/op: 188.038
Average FP rate %: 1.13823
Similar build time but dramatically faster query times on hot data (63 ns to 26 ns), and somewhat faster on stale data (188 ns to 150 ns). Performance differences on batched and skewed query loads are between these extremes as expected.
The only other interesting thing about speed is "inside" (query key was added to filter) vs. "outside" (query key was not added to filter) query times. The non-SIMD implementations are substantially slower when most queries are "outside" vs. "inside". This goes against what one might expect or would have observed years ago, as "outside" queries only need about two probes on average, due to short-circuiting, while "inside" always have num_probes (say 6). The problem is probably the nastily unpredictable branch. The SIMD implementation has few branches (very predictable) and has pretty consistent running time regardless of query outcome.
Accuracy
The generally improved accuracy (re: Issue https://github.com/facebook/rocksdb/issues/5857) comes from a better design for probing indices
within a cache line (re: Issue https://github.com/facebook/rocksdb/issues/4120) and improved accuracy for millions of keys in a single filter from using a 64-bit hash function (XXH3p). Design details in code comments.
Accuracy data (generalizes, except old impl gets worse with millions of keys):
Memory bits per key: FP rate percent old impl -> FP rate percent new impl
6: 5.70953 -> 5.69888
8: 2.45766 -> 2.29709
10: 1.13977 -> 0.959254
12: 0.662498 -> 0.411593
16: 0.353023 -> 0.0873754
24: 0.261552 -> 0.0060971
50: 0.225453 -> ~0.00003 (less than 1 in a million queries are FP)
Fixes https://github.com/facebook/rocksdb/issues/5857
Fixes https://github.com/facebook/rocksdb/issues/4120
Unlike the old implementation, this implementation has a fixed cache line size (64 bytes). At 10 bits per key, the accuracy of this new implementation is very close to the old implementation with 128-byte cache line size. If there's sufficient demand, this implementation could be generalized.
Compatibility
Although old releases would see the new structure as corrupt filter data and read the table as if there's no filter, we've decided only to enable the new Bloom filter with new format_version=5. This provides a smooth path for automatic adoption over time, with an option for early opt-in.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6007
Test Plan: filter_bench has been used thoroughly to validate speed, accuracy, and correctness. Unit tests have been carefully updated to exercise new and old implementations, as well as the logic to select an implementation based on context (format_version).
Differential Revision: D18294749
Pulled By: pdillinger
fbshipit-source-id: d44c9db3696e4d0a17caaec47075b7755c262c5f
5 years ago
|
|
|
std::make_tuple(BFP::kDeprecatedBlock, false,
|
|
|
|
test::kDefaultFormatVersion),
|
Experimental (production candidate) SST schema for Ribbon filter (#7658)
Summary:
Added experimental public API for Ribbon filter:
NewExperimentalRibbonFilterPolicy(). This experimental API will
take a "Bloom equivalent" bits per key, and configure the Ribbon
filter for the same FP rate as Bloom would have but ~30% space
savings. (Note: optimize_filters_for_memory is not yet implemented
for Ribbon filter. That can be added with no effect on schema.)
Internally, the Ribbon filter is configured using a "one_in_fp_rate"
value, which is 1 over desired FP rate. For example, use 100 for 1%
FP rate. I'm expecting this will be used in the future for configuring
Bloom-like filters, as I expect people to more commonly hold constant
the filter accuracy and change the space vs. time trade-off, rather than
hold constant the space (per key) and change the accuracy vs. time
trade-off, though we might make that available.
### Benchmarking
```
$ ./filter_bench -impl=2 -quick -m_keys_total_max=200 -average_keys_per_filter=100000 -net_includes_hashing
Building...
Build avg ns/key: 34.1341
Number of filters: 1993
Total size (MB): 238.488
Reported total allocated memory (MB): 262.875
Reported internal fragmentation: 10.2255%
Bits/key stored: 10.0029
----------------------------
Mixed inside/outside queries...
Single filter net ns/op: 18.7508
Random filter net ns/op: 258.246
Average FP rate %: 0.968672
----------------------------
Done. (For more info, run with -legend or -help.)
$ ./filter_bench -impl=3 -quick -m_keys_total_max=200 -average_keys_per_filter=100000 -net_includes_hashing
Building...
Build avg ns/key: 130.851
Number of filters: 1993
Total size (MB): 168.166
Reported total allocated memory (MB): 183.211
Reported internal fragmentation: 8.94626%
Bits/key stored: 7.05341
----------------------------
Mixed inside/outside queries...
Single filter net ns/op: 58.4523
Random filter net ns/op: 363.717
Average FP rate %: 0.952978
----------------------------
Done. (For more info, run with -legend or -help.)
```
168.166 / 238.488 = 0.705 -> 29.5% space reduction
130.851 / 34.1341 = 3.83x construction time for this Ribbon filter vs. lastest Bloom filter (could make that as little as about 2.5x for less space reduction)
### Working around a hashing "flaw"
bloom_test discovered a flaw in the simple hashing applied in
StandardHasher when num_starts == 1 (num_slots == 128), showing an
excessively high FP rate. The problem is that when many entries, on the
order of number of hash bits or kCoeffBits, are associated with the same
start location, the correlation between the CoeffRow and ResultRow (for
efficiency) can lead to a solution that is "universal," or nearly so, for
entries mapping to that start location. (Normally, variance in start
location breaks the effective association between CoeffRow and
ResultRow; the same value for CoeffRow is effectively different if start
locations are different.) Without kUseSmash and with num_starts > 1 (thus
num_starts ~= num_slots), this flaw should be completely irrelevant. Even
with 10M slots, the chances of a single slot having just 16 (or more)
entries map to it--not enough to cause an FP problem, which would be local
to that slot if it happened--is 1 in millions. This spreadsheet formula
shows that: =1/(10000000*(1 - POISSON(15, 1, TRUE)))
As kUseSmash==false (the setting for Standard128RibbonBitsBuilder) is
intended for CPU efficiency of filters with many more entries/slots than
kCoeffBits, a very reasonable work-around is to disallow num_starts==1
when !kUseSmash, by making the minimum non-zero number of slots
2*kCoeffBits. This is the work-around I've applied. This also means that
the new Ribbon filter schema (Standard128RibbonBitsBuilder) is not
space-efficient for less than a few hundred entries. Because of this, I
have made it fall back on constructing a Bloom filter, under existing
schema, when that is more space efficient for small filters. (We can
change this in the future if we want.)
TODO: better unit tests for this case in ribbon_test, and probably
update StandardHasher for kUseSmash case so that it can scale nicely to
small filters.
### Other related changes
* Add Ribbon filter to stress/crash test
* Add Ribbon filter to filter_bench as -impl=3
* Add option string support, as in "filter_policy=experimental_ribbon:5.678;"
where 5.678 is the Bloom equivalent bits per key.
* Rename internal mode BloomFilterPolicy::kAuto to kAutoBloom
* Add a general BuiltinFilterBitsBuilder::CalculateNumEntry based on
binary searching CalculateSpace (inefficient), so that subclasses
(especially experimental ones) don't have to provide an efficient
implementation inverting CalculateSpace.
* Minor refactor FastLocalBloomBitsBuilder for new base class
XXH3pFilterBitsBuilder shared with new Standard128RibbonBitsBuilder,
which allows the latter to fall back on Bloom construction in some
extreme cases.
* Mostly updated bloom_test for Ribbon filter, though a test like
FullBloomTest::Schema is a next TODO to ensure schema stability
(in case this becomes production-ready schema as it is).
* Add some APIs to ribbon_impl.h for configuring Ribbon filters.
Although these are reasonably covered by bloom_test, TODO more unit
tests in ribbon_test
* Added a "tool" FindOccupancyForSuccessRate to ribbon_test to get data
for constructing the linear approximations in GetNumSlotsFor95PctSuccess.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7658
Test Plan:
Some unit tests updated but other testing is left TODO. This
is considered experimental but laying down schema compatibility as early
as possible in case it proves production-quality. Also tested in
stress/crash test.
Reviewed By: jay-zhuang
Differential Revision: D24899349
Pulled By: pdillinger
fbshipit-source-id: 9715f3e6371c959d923aea8077c9423c7a9f82b8
4 years ago
|
|
|
std::make_tuple(BFP::kAutoBloom, true, test::kDefaultFormatVersion),
|
|
|
|
std::make_tuple(BFP::kAutoBloom, false, test::kDefaultFormatVersion)));
|
|
|
|
|
|
|
|
INSTANTIATE_TEST_CASE_P(
|
|
|
|
FormatLatest, DBBloomFilterTestWithParam,
|
|
|
|
::testing::Values(
|
New Bloom filter implementation for full and partitioned filters (#6007)
Summary:
Adds an improved, replacement Bloom filter implementation (FastLocalBloom) for full and partitioned filters in the block-based table. This replacement is faster and more accurate, especially for high bits per key or millions of keys in a single filter.
Speed
The improved speed, at least on recent x86_64, comes from
* Using fastrange instead of modulo (%)
* Using our new hash function (XXH3 preview, added in a previous commit), which is much faster for large keys and only *slightly* slower on keys around 12 bytes if hashing the same size many thousands of times in a row.
* Optimizing the Bloom filter queries with AVX2 SIMD operations. (Added AVX2 to the USE_SSE=1 build.) Careful design was required to support (a) SIMD-optimized queries, (b) compatible non-SIMD code that's simple and efficient, (c) flexible choice of number of probes, and (d) essentially maximized accuracy for a cache-local Bloom filter. Probes are made eight at a time, so any number of probes up to 8 is the same speed, then up to 16, etc.
* Prefetching cache lines when building the filter. Although this optimization could be applied to the old structure as well, it seems to balance out the small added cost of accumulating 64 bit hashes for adding to the filter rather than 32 bit hashes.
Here's nominal speed data from filter_bench (200MB in filters, about 10k keys each, 10 bits filter data / key, 6 probes, avg key size 24 bytes, includes hashing time) on Skylake DE (relatively low clock speed):
$ ./filter_bench -quick -impl=2 -net_includes_hashing # New Bloom filter
Build avg ns/key: 47.7135
Mixed inside/outside queries...
Single filter net ns/op: 26.2825
Random filter net ns/op: 150.459
Average FP rate %: 0.954651
$ ./filter_bench -quick -impl=0 -net_includes_hashing # Old Bloom filter
Build avg ns/key: 47.2245
Mixed inside/outside queries...
Single filter net ns/op: 63.2978
Random filter net ns/op: 188.038
Average FP rate %: 1.13823
Similar build time but dramatically faster query times on hot data (63 ns to 26 ns), and somewhat faster on stale data (188 ns to 150 ns). Performance differences on batched and skewed query loads are between these extremes as expected.
The only other interesting thing about speed is "inside" (query key was added to filter) vs. "outside" (query key was not added to filter) query times. The non-SIMD implementations are substantially slower when most queries are "outside" vs. "inside". This goes against what one might expect or would have observed years ago, as "outside" queries only need about two probes on average, due to short-circuiting, while "inside" always have num_probes (say 6). The problem is probably the nastily unpredictable branch. The SIMD implementation has few branches (very predictable) and has pretty consistent running time regardless of query outcome.
Accuracy
The generally improved accuracy (re: Issue https://github.com/facebook/rocksdb/issues/5857) comes from a better design for probing indices
within a cache line (re: Issue https://github.com/facebook/rocksdb/issues/4120) and improved accuracy for millions of keys in a single filter from using a 64-bit hash function (XXH3p). Design details in code comments.
Accuracy data (generalizes, except old impl gets worse with millions of keys):
Memory bits per key: FP rate percent old impl -> FP rate percent new impl
6: 5.70953 -> 5.69888
8: 2.45766 -> 2.29709
10: 1.13977 -> 0.959254
12: 0.662498 -> 0.411593
16: 0.353023 -> 0.0873754
24: 0.261552 -> 0.0060971
50: 0.225453 -> ~0.00003 (less than 1 in a million queries are FP)
Fixes https://github.com/facebook/rocksdb/issues/5857
Fixes https://github.com/facebook/rocksdb/issues/4120
Unlike the old implementation, this implementation has a fixed cache line size (64 bytes). At 10 bits per key, the accuracy of this new implementation is very close to the old implementation with 128-byte cache line size. If there's sufficient demand, this implementation could be generalized.
Compatibility
Although old releases would see the new structure as corrupt filter data and read the table as if there's no filter, we've decided only to enable the new Bloom filter with new format_version=5. This provides a smooth path for automatic adoption over time, with an option for early opt-in.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6007
Test Plan: filter_bench has been used thoroughly to validate speed, accuracy, and correctness. Unit tests have been carefully updated to exercise new and old implementations, as well as the logic to select an implementation based on context (format_version).
Differential Revision: D18294749
Pulled By: pdillinger
fbshipit-source-id: d44c9db3696e4d0a17caaec47075b7755c262c5f
5 years ago
|
|
|
std::make_tuple(BFP::kDeprecatedBlock, false,
|
|
|
|
test::kLatestFormatVersion),
|
Experimental (production candidate) SST schema for Ribbon filter (#7658)
Summary:
Added experimental public API for Ribbon filter:
NewExperimentalRibbonFilterPolicy(). This experimental API will
take a "Bloom equivalent" bits per key, and configure the Ribbon
filter for the same FP rate as Bloom would have but ~30% space
savings. (Note: optimize_filters_for_memory is not yet implemented
for Ribbon filter. That can be added with no effect on schema.)
Internally, the Ribbon filter is configured using a "one_in_fp_rate"
value, which is 1 over desired FP rate. For example, use 100 for 1%
FP rate. I'm expecting this will be used in the future for configuring
Bloom-like filters, as I expect people to more commonly hold constant
the filter accuracy and change the space vs. time trade-off, rather than
hold constant the space (per key) and change the accuracy vs. time
trade-off, though we might make that available.
### Benchmarking
```
$ ./filter_bench -impl=2 -quick -m_keys_total_max=200 -average_keys_per_filter=100000 -net_includes_hashing
Building...
Build avg ns/key: 34.1341
Number of filters: 1993
Total size (MB): 238.488
Reported total allocated memory (MB): 262.875
Reported internal fragmentation: 10.2255%
Bits/key stored: 10.0029
----------------------------
Mixed inside/outside queries...
Single filter net ns/op: 18.7508
Random filter net ns/op: 258.246
Average FP rate %: 0.968672
----------------------------
Done. (For more info, run with -legend or -help.)
$ ./filter_bench -impl=3 -quick -m_keys_total_max=200 -average_keys_per_filter=100000 -net_includes_hashing
Building...
Build avg ns/key: 130.851
Number of filters: 1993
Total size (MB): 168.166
Reported total allocated memory (MB): 183.211
Reported internal fragmentation: 8.94626%
Bits/key stored: 7.05341
----------------------------
Mixed inside/outside queries...
Single filter net ns/op: 58.4523
Random filter net ns/op: 363.717
Average FP rate %: 0.952978
----------------------------
Done. (For more info, run with -legend or -help.)
```
168.166 / 238.488 = 0.705 -> 29.5% space reduction
130.851 / 34.1341 = 3.83x construction time for this Ribbon filter vs. lastest Bloom filter (could make that as little as about 2.5x for less space reduction)
### Working around a hashing "flaw"
bloom_test discovered a flaw in the simple hashing applied in
StandardHasher when num_starts == 1 (num_slots == 128), showing an
excessively high FP rate. The problem is that when many entries, on the
order of number of hash bits or kCoeffBits, are associated with the same
start location, the correlation between the CoeffRow and ResultRow (for
efficiency) can lead to a solution that is "universal," or nearly so, for
entries mapping to that start location. (Normally, variance in start
location breaks the effective association between CoeffRow and
ResultRow; the same value for CoeffRow is effectively different if start
locations are different.) Without kUseSmash and with num_starts > 1 (thus
num_starts ~= num_slots), this flaw should be completely irrelevant. Even
with 10M slots, the chances of a single slot having just 16 (or more)
entries map to it--not enough to cause an FP problem, which would be local
to that slot if it happened--is 1 in millions. This spreadsheet formula
shows that: =1/(10000000*(1 - POISSON(15, 1, TRUE)))
As kUseSmash==false (the setting for Standard128RibbonBitsBuilder) is
intended for CPU efficiency of filters with many more entries/slots than
kCoeffBits, a very reasonable work-around is to disallow num_starts==1
when !kUseSmash, by making the minimum non-zero number of slots
2*kCoeffBits. This is the work-around I've applied. This also means that
the new Ribbon filter schema (Standard128RibbonBitsBuilder) is not
space-efficient for less than a few hundred entries. Because of this, I
have made it fall back on constructing a Bloom filter, under existing
schema, when that is more space efficient for small filters. (We can
change this in the future if we want.)
TODO: better unit tests for this case in ribbon_test, and probably
update StandardHasher for kUseSmash case so that it can scale nicely to
small filters.
### Other related changes
* Add Ribbon filter to stress/crash test
* Add Ribbon filter to filter_bench as -impl=3
* Add option string support, as in "filter_policy=experimental_ribbon:5.678;"
where 5.678 is the Bloom equivalent bits per key.
* Rename internal mode BloomFilterPolicy::kAuto to kAutoBloom
* Add a general BuiltinFilterBitsBuilder::CalculateNumEntry based on
binary searching CalculateSpace (inefficient), so that subclasses
(especially experimental ones) don't have to provide an efficient
implementation inverting CalculateSpace.
* Minor refactor FastLocalBloomBitsBuilder for new base class
XXH3pFilterBitsBuilder shared with new Standard128RibbonBitsBuilder,
which allows the latter to fall back on Bloom construction in some
extreme cases.
* Mostly updated bloom_test for Ribbon filter, though a test like
FullBloomTest::Schema is a next TODO to ensure schema stability
(in case this becomes production-ready schema as it is).
* Add some APIs to ribbon_impl.h for configuring Ribbon filters.
Although these are reasonably covered by bloom_test, TODO more unit
tests in ribbon_test
* Added a "tool" FindOccupancyForSuccessRate to ribbon_test to get data
for constructing the linear approximations in GetNumSlotsFor95PctSuccess.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7658
Test Plan:
Some unit tests updated but other testing is left TODO. This
is considered experimental but laying down schema compatibility as early
as possible in case it proves production-quality. Also tested in
stress/crash test.
Reviewed By: jay-zhuang
Differential Revision: D24899349
Pulled By: pdillinger
fbshipit-source-id: 9715f3e6371c959d923aea8077c9423c7a9f82b8
4 years ago
|
|
|
std::make_tuple(BFP::kAutoBloom, true, test::kLatestFormatVersion),
|
|
|
|
std::make_tuple(BFP::kAutoBloom, false, test::kLatestFormatVersion)));
|
|
|
|
#endif // !defined(ROCKSDB_VALGRIND_RUN) || defined(ROCKSDB_FULL_VALGRIND_RUN)
|
|
|
|
|
|
|
|
TEST_F(DBBloomFilterTest, BloomFilterRate) {
|
|
|
|
while (ChangeFilterOptions()) {
|
|
|
|
Options options = CurrentOptions();
|
|
|
|
options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
|
|
|
|
get_perf_context()->EnablePerLevelPerfContext();
|
|
|
|
CreateAndReopenWithCF({"pikachu"}, options);
|
|
|
|
|
|
|
|
const int maxKey = 10000;
|
|
|
|
for (int i = 0; i < maxKey; i++) {
|
|
|
|
ASSERT_OK(Put(1, Key(i), Key(i)));
|
|
|
|
}
|
|
|
|
// Add a large key to make the file contain wide range
|
|
|
|
ASSERT_OK(Put(1, Key(maxKey + 55555), Key(maxKey + 55555)));
|
|
|
|
Flush(1);
|
|
|
|
|
|
|
|
// Check if they can be found
|
|
|
|
for (int i = 0; i < maxKey; i++) {
|
|
|
|
ASSERT_EQ(Key(i), Get(1, Key(i)));
|
|
|
|
}
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 0);
|
|
|
|
|
|
|
|
// Check if filter is useful
|
|
|
|
for (int i = 0; i < maxKey; i++) {
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get(1, Key(i + 33333)));
|
|
|
|
}
|
|
|
|
ASSERT_GE(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), maxKey * 0.98);
|
|
|
|
ASSERT_GE(
|
|
|
|
(*(get_perf_context()->level_to_perf_context))[0].bloom_filter_useful,
|
|
|
|
maxKey * 0.98);
|
|
|
|
get_perf_context()->Reset();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(DBBloomFilterTest, BloomFilterCompatibility) {
|
|
|
|
Options options = CurrentOptions();
|
|
|
|
options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
|
|
|
|
BlockBasedTableOptions table_options;
|
|
|
|
table_options.filter_policy.reset(NewBloomFilterPolicy(10, true));
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
|
|
|
|
|
|
|
|
// Create with block based filter
|
|
|
|
CreateAndReopenWithCF({"pikachu"}, options);
|
|
|
|
|
|
|
|
const int maxKey = 10000;
|
|
|
|
for (int i = 0; i < maxKey; i++) {
|
|
|
|
ASSERT_OK(Put(1, Key(i), Key(i)));
|
|
|
|
}
|
|
|
|
ASSERT_OK(Put(1, Key(maxKey + 55555), Key(maxKey + 55555)));
|
|
|
|
Flush(1);
|
|
|
|
|
|
|
|
// Check db with full filter
|
|
|
|
table_options.filter_policy.reset(NewBloomFilterPolicy(10, false));
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
|
|
|
|
ReopenWithColumnFamilies({"default", "pikachu"}, options);
|
|
|
|
|
|
|
|
// Check if they can be found
|
|
|
|
for (int i = 0; i < maxKey; i++) {
|
|
|
|
ASSERT_EQ(Key(i), Get(1, Key(i)));
|
|
|
|
}
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 0);
|
|
|
|
|
|
|
|
// Check db with partitioned full filter
|
|
|
|
table_options.partition_filters = true;
|
|
|
|
table_options.index_type =
|
|
|
|
BlockBasedTableOptions::IndexType::kTwoLevelIndexSearch;
|
|
|
|
table_options.filter_policy.reset(NewBloomFilterPolicy(10, false));
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
|
|
|
|
ReopenWithColumnFamilies({"default", "pikachu"}, options);
|
|
|
|
|
|
|
|
// Check if they can be found
|
|
|
|
for (int i = 0; i < maxKey; i++) {
|
|
|
|
ASSERT_EQ(Key(i), Get(1, Key(i)));
|
|
|
|
}
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 0);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(DBBloomFilterTest, BloomFilterReverseCompatibility) {
|
|
|
|
for (bool partition_filters : {true, false}) {
|
|
|
|
Options options = CurrentOptions();
|
|
|
|
options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
|
|
|
|
BlockBasedTableOptions table_options;
|
|
|
|
if (partition_filters) {
|
|
|
|
table_options.partition_filters = true;
|
|
|
|
table_options.index_type =
|
|
|
|
BlockBasedTableOptions::IndexType::kTwoLevelIndexSearch;
|
|
|
|
}
|
|
|
|
table_options.filter_policy.reset(NewBloomFilterPolicy(10, false));
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
|
|
|
|
DestroyAndReopen(options);
|
|
|
|
|
|
|
|
// Create with full filter
|
|
|
|
CreateAndReopenWithCF({"pikachu"}, options);
|
|
|
|
|
|
|
|
const int maxKey = 10000;
|
|
|
|
for (int i = 0; i < maxKey; i++) {
|
|
|
|
ASSERT_OK(Put(1, Key(i), Key(i)));
|
|
|
|
}
|
|
|
|
ASSERT_OK(Put(1, Key(maxKey + 55555), Key(maxKey + 55555)));
|
|
|
|
Flush(1);
|
|
|
|
|
|
|
|
// Check db with block_based filter
|
|
|
|
table_options.filter_policy.reset(NewBloomFilterPolicy(10, true));
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
|
|
|
|
ReopenWithColumnFamilies({"default", "pikachu"}, options);
|
|
|
|
|
|
|
|
// Check if they can be found
|
|
|
|
for (int i = 0; i < maxKey; i++) {
|
|
|
|
ASSERT_EQ(Key(i), Get(1, Key(i)));
|
|
|
|
}
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 0);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
// A wrapped bloom over block-based FilterPolicy
|
|
|
|
class TestingWrappedBlockBasedFilterPolicy : public FilterPolicy {
|
|
|
|
public:
|
|
|
|
explicit TestingWrappedBlockBasedFilterPolicy(int bits_per_key)
|
|
|
|
: filter_(NewBloomFilterPolicy(bits_per_key, true)), counter_(0) {}
|
|
|
|
|
|
|
|
~TestingWrappedBlockBasedFilterPolicy() override { delete filter_; }
|
|
|
|
|
|
|
|
const char* Name() const override {
|
|
|
|
return "TestingWrappedBlockBasedFilterPolicy";
|
|
|
|
}
|
|
|
|
|
|
|
|
void CreateFilter(const ROCKSDB_NAMESPACE::Slice* keys, int n,
|
|
|
|
std::string* dst) const override {
|
|
|
|
std::unique_ptr<ROCKSDB_NAMESPACE::Slice[]> user_keys(
|
|
|
|
new ROCKSDB_NAMESPACE::Slice[n]);
|
|
|
|
for (int i = 0; i < n; ++i) {
|
|
|
|
user_keys[i] = convertKey(keys[i]);
|
|
|
|
}
|
|
|
|
return filter_->CreateFilter(user_keys.get(), n, dst);
|
|
|
|
}
|
|
|
|
|
|
|
|
bool KeyMayMatch(const ROCKSDB_NAMESPACE::Slice& key,
|
|
|
|
const ROCKSDB_NAMESPACE::Slice& filter) const override {
|
|
|
|
counter_++;
|
|
|
|
return filter_->KeyMayMatch(convertKey(key), filter);
|
|
|
|
}
|
|
|
|
|
|
|
|
uint32_t GetCounter() { return counter_; }
|
|
|
|
|
|
|
|
private:
|
|
|
|
const FilterPolicy* filter_;
|
|
|
|
mutable uint32_t counter_;
|
|
|
|
|
|
|
|
ROCKSDB_NAMESPACE::Slice convertKey(
|
|
|
|
const ROCKSDB_NAMESPACE::Slice& key) const {
|
|
|
|
return key;
|
|
|
|
}
|
|
|
|
};
|
|
|
|
} // namespace
|
|
|
|
|
|
|
|
TEST_F(DBBloomFilterTest, WrappedBlockBasedFilterPolicy) {
|
|
|
|
Options options = CurrentOptions();
|
|
|
|
options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
|
|
|
|
|
|
|
|
BlockBasedTableOptions table_options;
|
|
|
|
TestingWrappedBlockBasedFilterPolicy* policy =
|
|
|
|
new TestingWrappedBlockBasedFilterPolicy(10);
|
|
|
|
table_options.filter_policy.reset(policy);
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
|
|
|
|
|
|
|
|
CreateAndReopenWithCF({"pikachu"}, options);
|
|
|
|
|
|
|
|
const int maxKey = 10000;
|
|
|
|
for (int i = 0; i < maxKey; i++) {
|
|
|
|
ASSERT_OK(Put(1, Key(i), Key(i)));
|
|
|
|
}
|
|
|
|
// Add a large key to make the file contain wide range
|
|
|
|
ASSERT_OK(Put(1, Key(maxKey + 55555), Key(maxKey + 55555)));
|
|
|
|
ASSERT_EQ(0U, policy->GetCounter());
|
|
|
|
Flush(1);
|
|
|
|
|
|
|
|
// Check if they can be found
|
|
|
|
for (int i = 0; i < maxKey; i++) {
|
|
|
|
ASSERT_EQ(Key(i), Get(1, Key(i)));
|
|
|
|
}
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 0);
|
|
|
|
ASSERT_EQ(1U * maxKey, policy->GetCounter());
|
|
|
|
|
|
|
|
// Check if filter is useful
|
|
|
|
for (int i = 0; i < maxKey; i++) {
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get(1, Key(i + 33333)));
|
|
|
|
}
|
|
|
|
ASSERT_GE(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), maxKey * 0.98);
|
|
|
|
ASSERT_EQ(2U * maxKey, policy->GetCounter());
|
|
|
|
}
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
// NOTE: This class is referenced by HISTORY.md as a model for a wrapper
|
|
|
|
// FilterPolicy selecting among configurations based on context.
|
|
|
|
class LevelAndStyleCustomFilterPolicy : public FilterPolicy {
|
|
|
|
public:
|
|
|
|
explicit LevelAndStyleCustomFilterPolicy(int bpk_fifo, int bpk_l0_other,
|
|
|
|
int bpk_otherwise)
|
|
|
|
: policy_fifo_(NewBloomFilterPolicy(bpk_fifo)),
|
|
|
|
policy_l0_other_(NewBloomFilterPolicy(bpk_l0_other)),
|
|
|
|
policy_otherwise_(NewBloomFilterPolicy(bpk_otherwise)) {}
|
|
|
|
|
|
|
|
// OK to use built-in policy name because we are deferring to a
|
|
|
|
// built-in builder. We aren't changing the serialized format.
|
|
|
|
const char* Name() const override { return policy_fifo_->Name(); }
|
|
|
|
|
|
|
|
FilterBitsBuilder* GetBuilderWithContext(
|
|
|
|
const FilterBuildingContext& context) const override {
|
|
|
|
if (context.compaction_style == kCompactionStyleFIFO) {
|
|
|
|
return policy_fifo_->GetBuilderWithContext(context);
|
|
|
|
} else if (context.level_at_creation == 0) {
|
|
|
|
return policy_l0_other_->GetBuilderWithContext(context);
|
|
|
|
} else {
|
|
|
|
return policy_otherwise_->GetBuilderWithContext(context);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
FilterBitsReader* GetFilterBitsReader(const Slice& contents) const override {
|
|
|
|
// OK to defer to any of them; they all can parse built-in filters
|
|
|
|
// from any settings.
|
|
|
|
return policy_fifo_->GetFilterBitsReader(contents);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Defer just in case configuration uses block-based filter
|
|
|
|
void CreateFilter(const Slice* keys, int n, std::string* dst) const override {
|
|
|
|
policy_otherwise_->CreateFilter(keys, n, dst);
|
|
|
|
}
|
|
|
|
bool KeyMayMatch(const Slice& key, const Slice& filter) const override {
|
|
|
|
return policy_otherwise_->KeyMayMatch(key, filter);
|
|
|
|
}
|
|
|
|
|
|
|
|
private:
|
|
|
|
const std::unique_ptr<const FilterPolicy> policy_fifo_;
|
|
|
|
const std::unique_ptr<const FilterPolicy> policy_l0_other_;
|
|
|
|
const std::unique_ptr<const FilterPolicy> policy_otherwise_;
|
|
|
|
};
|
|
|
|
|
Add more LSM info to FilterBuildingContext (#8246)
Summary:
Add `num_levels`, `is_bottommost`, and table file creation
`reason` to `FilterBuildingContext`, in anticipation of more powerful
Bloom-like filter support.
To support this, added `is_bottommost` and `reason` to
`TableBuilderOptions`, which allowed removing `reason` parameter from
`rocksdb::BuildTable`.
I attempted to remove `skip_filters` from `TableBuilderOptions`, because
filter construction decisions should arise from options, not one-off
parameters. I could not completely remove it because the public API for
SstFileWriter takes a `skip_filters` parameter, and translating this
into an option change would mean awkwardly replacing the table_factory
if it is BlockBasedTableFactory with new filter_policy=nullptr option.
I marked this public skip_filters option as deprecated because of this
oddity. (skip_filters on the read side probably makes sense.)
At least `skip_filters` is now largely hidden for users of
`TableBuilderOptions` and is no longer used for implementing the
optimize_filters_for_hits option. Bringing the logic for that option
closer to handling of FilterBuildingContext makes it more obvious that
hese two are using the same notion of "bottommost." (Planned:
configuration options for Bloom-like filters that generalize
`optimize_filters_for_hits`)
Recommended follow-up: Try to get away from "bottommost level" naming of
things, which is inaccurate (see
VersionStorageInfo::RangeMightExistAfterSortedRun), and move to
"bottommost run" or just "bottommost."
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8246
Test Plan:
extended an existing unit test to exercise and check various
filter building contexts. Also, existing tests for
optimize_filters_for_hits validate some of the "bottommost" handling,
which is now closely connected to FilterBuildingContext::is_bottommost
through TableBuilderOptions::is_bottommost
Reviewed By: mrambacher
Differential Revision: D28099346
Pulled By: pdillinger
fbshipit-source-id: 2c1072e29c24d4ac404c761a7b7663292372600a
3 years ago
|
|
|
static std::map<TableFileCreationReason, std::string>
|
|
|
|
table_file_creation_reason_to_string{
|
|
|
|
{TableFileCreationReason::kCompaction, "kCompaction"},
|
|
|
|
{TableFileCreationReason::kFlush, "kFlush"},
|
|
|
|
{TableFileCreationReason::kMisc, "kMisc"},
|
|
|
|
{TableFileCreationReason::kRecovery, "kRecovery"},
|
|
|
|
};
|
|
|
|
|
|
|
|
class TestingContextCustomFilterPolicy
|
|
|
|
: public LevelAndStyleCustomFilterPolicy {
|
|
|
|
public:
|
|
|
|
explicit TestingContextCustomFilterPolicy(int bpk_fifo, int bpk_l0_other,
|
|
|
|
int bpk_otherwise)
|
|
|
|
: LevelAndStyleCustomFilterPolicy(bpk_fifo, bpk_l0_other, bpk_otherwise) {
|
|
|
|
}
|
|
|
|
|
|
|
|
FilterBitsBuilder* GetBuilderWithContext(
|
|
|
|
const FilterBuildingContext& context) const override {
|
|
|
|
test_report_ += "cf=";
|
|
|
|
test_report_ += context.column_family_name;
|
Add more LSM info to FilterBuildingContext (#8246)
Summary:
Add `num_levels`, `is_bottommost`, and table file creation
`reason` to `FilterBuildingContext`, in anticipation of more powerful
Bloom-like filter support.
To support this, added `is_bottommost` and `reason` to
`TableBuilderOptions`, which allowed removing `reason` parameter from
`rocksdb::BuildTable`.
I attempted to remove `skip_filters` from `TableBuilderOptions`, because
filter construction decisions should arise from options, not one-off
parameters. I could not completely remove it because the public API for
SstFileWriter takes a `skip_filters` parameter, and translating this
into an option change would mean awkwardly replacing the table_factory
if it is BlockBasedTableFactory with new filter_policy=nullptr option.
I marked this public skip_filters option as deprecated because of this
oddity. (skip_filters on the read side probably makes sense.)
At least `skip_filters` is now largely hidden for users of
`TableBuilderOptions` and is no longer used for implementing the
optimize_filters_for_hits option. Bringing the logic for that option
closer to handling of FilterBuildingContext makes it more obvious that
hese two are using the same notion of "bottommost." (Planned:
configuration options for Bloom-like filters that generalize
`optimize_filters_for_hits`)
Recommended follow-up: Try to get away from "bottommost level" naming of
things, which is inaccurate (see
VersionStorageInfo::RangeMightExistAfterSortedRun), and move to
"bottommost run" or just "bottommost."
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8246
Test Plan:
extended an existing unit test to exercise and check various
filter building contexts. Also, existing tests for
optimize_filters_for_hits validate some of the "bottommost" handling,
which is now closely connected to FilterBuildingContext::is_bottommost
through TableBuilderOptions::is_bottommost
Reviewed By: mrambacher
Differential Revision: D28099346
Pulled By: pdillinger
fbshipit-source-id: 2c1072e29c24d4ac404c761a7b7663292372600a
3 years ago
|
|
|
test_report_ += ",s=";
|
|
|
|
test_report_ +=
|
|
|
|
OptionsHelper::compaction_style_to_string[context.compaction_style];
|
Add more LSM info to FilterBuildingContext (#8246)
Summary:
Add `num_levels`, `is_bottommost`, and table file creation
`reason` to `FilterBuildingContext`, in anticipation of more powerful
Bloom-like filter support.
To support this, added `is_bottommost` and `reason` to
`TableBuilderOptions`, which allowed removing `reason` parameter from
`rocksdb::BuildTable`.
I attempted to remove `skip_filters` from `TableBuilderOptions`, because
filter construction decisions should arise from options, not one-off
parameters. I could not completely remove it because the public API for
SstFileWriter takes a `skip_filters` parameter, and translating this
into an option change would mean awkwardly replacing the table_factory
if it is BlockBasedTableFactory with new filter_policy=nullptr option.
I marked this public skip_filters option as deprecated because of this
oddity. (skip_filters on the read side probably makes sense.)
At least `skip_filters` is now largely hidden for users of
`TableBuilderOptions` and is no longer used for implementing the
optimize_filters_for_hits option. Bringing the logic for that option
closer to handling of FilterBuildingContext makes it more obvious that
hese two are using the same notion of "bottommost." (Planned:
configuration options for Bloom-like filters that generalize
`optimize_filters_for_hits`)
Recommended follow-up: Try to get away from "bottommost level" naming of
things, which is inaccurate (see
VersionStorageInfo::RangeMightExistAfterSortedRun), and move to
"bottommost run" or just "bottommost."
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8246
Test Plan:
extended an existing unit test to exercise and check various
filter building contexts. Also, existing tests for
optimize_filters_for_hits validate some of the "bottommost" handling,
which is now closely connected to FilterBuildingContext::is_bottommost
through TableBuilderOptions::is_bottommost
Reviewed By: mrambacher
Differential Revision: D28099346
Pulled By: pdillinger
fbshipit-source-id: 2c1072e29c24d4ac404c761a7b7663292372600a
3 years ago
|
|
|
test_report_ += ",n=";
|
|
|
|
test_report_ += ToString(context.num_levels);
|
|
|
|
test_report_ += ",l=";
|
|
|
|
test_report_ += ToString(context.level_at_creation);
|
|
|
|
test_report_ += ",b=";
|
|
|
|
test_report_ += ToString(int{context.is_bottommost});
|
|
|
|
test_report_ += ",r=";
|
|
|
|
test_report_ += table_file_creation_reason_to_string[context.reason];
|
|
|
|
test_report_ += "\n";
|
|
|
|
|
|
|
|
return LevelAndStyleCustomFilterPolicy::GetBuilderWithContext(context);
|
|
|
|
}
|
|
|
|
|
|
|
|
std::string DumpTestReport() {
|
|
|
|
std::string rv;
|
|
|
|
std::swap(rv, test_report_);
|
|
|
|
return rv;
|
|
|
|
}
|
|
|
|
|
|
|
|
private:
|
|
|
|
mutable std::string test_report_;
|
|
|
|
};
|
|
|
|
} // namespace
|
|
|
|
|
|
|
|
TEST_F(DBBloomFilterTest, ContextCustomFilterPolicy) {
|
Add more LSM info to FilterBuildingContext (#8246)
Summary:
Add `num_levels`, `is_bottommost`, and table file creation
`reason` to `FilterBuildingContext`, in anticipation of more powerful
Bloom-like filter support.
To support this, added `is_bottommost` and `reason` to
`TableBuilderOptions`, which allowed removing `reason` parameter from
`rocksdb::BuildTable`.
I attempted to remove `skip_filters` from `TableBuilderOptions`, because
filter construction decisions should arise from options, not one-off
parameters. I could not completely remove it because the public API for
SstFileWriter takes a `skip_filters` parameter, and translating this
into an option change would mean awkwardly replacing the table_factory
if it is BlockBasedTableFactory with new filter_policy=nullptr option.
I marked this public skip_filters option as deprecated because of this
oddity. (skip_filters on the read side probably makes sense.)
At least `skip_filters` is now largely hidden for users of
`TableBuilderOptions` and is no longer used for implementing the
optimize_filters_for_hits option. Bringing the logic for that option
closer to handling of FilterBuildingContext makes it more obvious that
hese two are using the same notion of "bottommost." (Planned:
configuration options for Bloom-like filters that generalize
`optimize_filters_for_hits`)
Recommended follow-up: Try to get away from "bottommost level" naming of
things, which is inaccurate (see
VersionStorageInfo::RangeMightExistAfterSortedRun), and move to
"bottommost run" or just "bottommost."
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8246
Test Plan:
extended an existing unit test to exercise and check various
filter building contexts. Also, existing tests for
optimize_filters_for_hits validate some of the "bottommost" handling,
which is now closely connected to FilterBuildingContext::is_bottommost
through TableBuilderOptions::is_bottommost
Reviewed By: mrambacher
Differential Revision: D28099346
Pulled By: pdillinger
fbshipit-source-id: 2c1072e29c24d4ac404c761a7b7663292372600a
3 years ago
|
|
|
auto policy = std::make_shared<TestingContextCustomFilterPolicy>(15, 8, 5);
|
|
|
|
Options options;
|
|
|
|
for (bool fifo : {true, false}) {
|
Add more LSM info to FilterBuildingContext (#8246)
Summary:
Add `num_levels`, `is_bottommost`, and table file creation
`reason` to `FilterBuildingContext`, in anticipation of more powerful
Bloom-like filter support.
To support this, added `is_bottommost` and `reason` to
`TableBuilderOptions`, which allowed removing `reason` parameter from
`rocksdb::BuildTable`.
I attempted to remove `skip_filters` from `TableBuilderOptions`, because
filter construction decisions should arise from options, not one-off
parameters. I could not completely remove it because the public API for
SstFileWriter takes a `skip_filters` parameter, and translating this
into an option change would mean awkwardly replacing the table_factory
if it is BlockBasedTableFactory with new filter_policy=nullptr option.
I marked this public skip_filters option as deprecated because of this
oddity. (skip_filters on the read side probably makes sense.)
At least `skip_filters` is now largely hidden for users of
`TableBuilderOptions` and is no longer used for implementing the
optimize_filters_for_hits option. Bringing the logic for that option
closer to handling of FilterBuildingContext makes it more obvious that
hese two are using the same notion of "bottommost." (Planned:
configuration options for Bloom-like filters that generalize
`optimize_filters_for_hits`)
Recommended follow-up: Try to get away from "bottommost level" naming of
things, which is inaccurate (see
VersionStorageInfo::RangeMightExistAfterSortedRun), and move to
"bottommost run" or just "bottommost."
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8246
Test Plan:
extended an existing unit test to exercise and check various
filter building contexts. Also, existing tests for
optimize_filters_for_hits validate some of the "bottommost" handling,
which is now closely connected to FilterBuildingContext::is_bottommost
through TableBuilderOptions::is_bottommost
Reviewed By: mrambacher
Differential Revision: D28099346
Pulled By: pdillinger
fbshipit-source-id: 2c1072e29c24d4ac404c761a7b7663292372600a
3 years ago
|
|
|
options = CurrentOptions();
|
|
|
|
options.max_open_files = fifo ? -1 : options.max_open_files;
|
|
|
|
options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
|
|
|
|
options.compaction_style =
|
|
|
|
fifo ? kCompactionStyleFIFO : kCompactionStyleLevel;
|
|
|
|
|
|
|
|
BlockBasedTableOptions table_options;
|
|
|
|
table_options.filter_policy = policy;
|
|
|
|
table_options.format_version = 5;
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
|
|
|
|
|
|
|
|
TryReopen(options);
|
|
|
|
CreateAndReopenWithCF({fifo ? "abe" : "bob"}, options);
|
|
|
|
|
|
|
|
const int maxKey = 10000;
|
|
|
|
for (int i = 0; i < maxKey / 2; i++) {
|
|
|
|
ASSERT_OK(Put(1, Key(i), Key(i)));
|
|
|
|
}
|
|
|
|
// Add a large key to make the file contain wide range
|
|
|
|
ASSERT_OK(Put(1, Key(maxKey + 55555), Key(maxKey + 55555)));
|
|
|
|
Flush(1);
|
|
|
|
EXPECT_EQ(policy->DumpTestReport(),
|
Add more LSM info to FilterBuildingContext (#8246)
Summary:
Add `num_levels`, `is_bottommost`, and table file creation
`reason` to `FilterBuildingContext`, in anticipation of more powerful
Bloom-like filter support.
To support this, added `is_bottommost` and `reason` to
`TableBuilderOptions`, which allowed removing `reason` parameter from
`rocksdb::BuildTable`.
I attempted to remove `skip_filters` from `TableBuilderOptions`, because
filter construction decisions should arise from options, not one-off
parameters. I could not completely remove it because the public API for
SstFileWriter takes a `skip_filters` parameter, and translating this
into an option change would mean awkwardly replacing the table_factory
if it is BlockBasedTableFactory with new filter_policy=nullptr option.
I marked this public skip_filters option as deprecated because of this
oddity. (skip_filters on the read side probably makes sense.)
At least `skip_filters` is now largely hidden for users of
`TableBuilderOptions` and is no longer used for implementing the
optimize_filters_for_hits option. Bringing the logic for that option
closer to handling of FilterBuildingContext makes it more obvious that
hese two are using the same notion of "bottommost." (Planned:
configuration options for Bloom-like filters that generalize
`optimize_filters_for_hits`)
Recommended follow-up: Try to get away from "bottommost level" naming of
things, which is inaccurate (see
VersionStorageInfo::RangeMightExistAfterSortedRun), and move to
"bottommost run" or just "bottommost."
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8246
Test Plan:
extended an existing unit test to exercise and check various
filter building contexts. Also, existing tests for
optimize_filters_for_hits validate some of the "bottommost" handling,
which is now closely connected to FilterBuildingContext::is_bottommost
through TableBuilderOptions::is_bottommost
Reviewed By: mrambacher
Differential Revision: D28099346
Pulled By: pdillinger
fbshipit-source-id: 2c1072e29c24d4ac404c761a7b7663292372600a
3 years ago
|
|
|
fifo ? "cf=abe,s=kCompactionStyleFIFO,n=1,l=0,b=0,r=kFlush\n"
|
|
|
|
: "cf=bob,s=kCompactionStyleLevel,n=7,l=0,b=0,r=kFlush\n");
|
|
|
|
|
|
|
|
for (int i = maxKey / 2; i < maxKey; i++) {
|
|
|
|
ASSERT_OK(Put(1, Key(i), Key(i)));
|
|
|
|
}
|
|
|
|
Flush(1);
|
|
|
|
EXPECT_EQ(policy->DumpTestReport(),
|
Add more LSM info to FilterBuildingContext (#8246)
Summary:
Add `num_levels`, `is_bottommost`, and table file creation
`reason` to `FilterBuildingContext`, in anticipation of more powerful
Bloom-like filter support.
To support this, added `is_bottommost` and `reason` to
`TableBuilderOptions`, which allowed removing `reason` parameter from
`rocksdb::BuildTable`.
I attempted to remove `skip_filters` from `TableBuilderOptions`, because
filter construction decisions should arise from options, not one-off
parameters. I could not completely remove it because the public API for
SstFileWriter takes a `skip_filters` parameter, and translating this
into an option change would mean awkwardly replacing the table_factory
if it is BlockBasedTableFactory with new filter_policy=nullptr option.
I marked this public skip_filters option as deprecated because of this
oddity. (skip_filters on the read side probably makes sense.)
At least `skip_filters` is now largely hidden for users of
`TableBuilderOptions` and is no longer used for implementing the
optimize_filters_for_hits option. Bringing the logic for that option
closer to handling of FilterBuildingContext makes it more obvious that
hese two are using the same notion of "bottommost." (Planned:
configuration options for Bloom-like filters that generalize
`optimize_filters_for_hits`)
Recommended follow-up: Try to get away from "bottommost level" naming of
things, which is inaccurate (see
VersionStorageInfo::RangeMightExistAfterSortedRun), and move to
"bottommost run" or just "bottommost."
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8246
Test Plan:
extended an existing unit test to exercise and check various
filter building contexts. Also, existing tests for
optimize_filters_for_hits validate some of the "bottommost" handling,
which is now closely connected to FilterBuildingContext::is_bottommost
through TableBuilderOptions::is_bottommost
Reviewed By: mrambacher
Differential Revision: D28099346
Pulled By: pdillinger
fbshipit-source-id: 2c1072e29c24d4ac404c761a7b7663292372600a
3 years ago
|
|
|
fifo ? "cf=abe,s=kCompactionStyleFIFO,n=1,l=0,b=0,r=kFlush\n"
|
|
|
|
: "cf=bob,s=kCompactionStyleLevel,n=7,l=0,b=0,r=kFlush\n");
|
|
|
|
|
|
|
|
// Check that they can be found
|
|
|
|
for (int i = 0; i < maxKey; i++) {
|
|
|
|
ASSERT_EQ(Key(i), Get(1, Key(i)));
|
|
|
|
}
|
|
|
|
// Since we have two tables / two filters, we might have Bloom checks on
|
|
|
|
// our queries, but no more than one "useful" per query on a found key.
|
|
|
|
EXPECT_LE(TestGetAndResetTickerCount(options, BLOOM_FILTER_USEFUL), maxKey);
|
|
|
|
|
|
|
|
// Check that we have two filters, each about
|
|
|
|
// fifo: 0.12% FP rate (15 bits per key)
|
|
|
|
// level: 2.3% FP rate (8 bits per key)
|
|
|
|
for (int i = 0; i < maxKey; i++) {
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get(1, Key(i + 33333)));
|
|
|
|
}
|
|
|
|
{
|
|
|
|
auto useful_count =
|
|
|
|
TestGetAndResetTickerCount(options, BLOOM_FILTER_USEFUL);
|
|
|
|
EXPECT_GE(useful_count, maxKey * 2 * (fifo ? 0.9980 : 0.975));
|
|
|
|
EXPECT_LE(useful_count, maxKey * 2 * (fifo ? 0.9995 : 0.98));
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!fifo) { // FIFO only has L0
|
|
|
|
// Full compaction
|
|
|
|
ASSERT_OK(db_->CompactRange(CompactRangeOptions(), handles_[1], nullptr,
|
|
|
|
nullptr));
|
|
|
|
EXPECT_EQ(policy->DumpTestReport(),
|
Add more LSM info to FilterBuildingContext (#8246)
Summary:
Add `num_levels`, `is_bottommost`, and table file creation
`reason` to `FilterBuildingContext`, in anticipation of more powerful
Bloom-like filter support.
To support this, added `is_bottommost` and `reason` to
`TableBuilderOptions`, which allowed removing `reason` parameter from
`rocksdb::BuildTable`.
I attempted to remove `skip_filters` from `TableBuilderOptions`, because
filter construction decisions should arise from options, not one-off
parameters. I could not completely remove it because the public API for
SstFileWriter takes a `skip_filters` parameter, and translating this
into an option change would mean awkwardly replacing the table_factory
if it is BlockBasedTableFactory with new filter_policy=nullptr option.
I marked this public skip_filters option as deprecated because of this
oddity. (skip_filters on the read side probably makes sense.)
At least `skip_filters` is now largely hidden for users of
`TableBuilderOptions` and is no longer used for implementing the
optimize_filters_for_hits option. Bringing the logic for that option
closer to handling of FilterBuildingContext makes it more obvious that
hese two are using the same notion of "bottommost." (Planned:
configuration options for Bloom-like filters that generalize
`optimize_filters_for_hits`)
Recommended follow-up: Try to get away from "bottommost level" naming of
things, which is inaccurate (see
VersionStorageInfo::RangeMightExistAfterSortedRun), and move to
"bottommost run" or just "bottommost."
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8246
Test Plan:
extended an existing unit test to exercise and check various
filter building contexts. Also, existing tests for
optimize_filters_for_hits validate some of the "bottommost" handling,
which is now closely connected to FilterBuildingContext::is_bottommost
through TableBuilderOptions::is_bottommost
Reviewed By: mrambacher
Differential Revision: D28099346
Pulled By: pdillinger
fbshipit-source-id: 2c1072e29c24d4ac404c761a7b7663292372600a
3 years ago
|
|
|
"cf=bob,s=kCompactionStyleLevel,n=7,l=1,b=1,r=kCompaction\n");
|
|
|
|
|
|
|
|
// Check that we now have one filter, about 9.2% FP rate (5 bits per key)
|
|
|
|
for (int i = 0; i < maxKey; i++) {
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get(1, Key(i + 33333)));
|
|
|
|
}
|
|
|
|
{
|
|
|
|
auto useful_count =
|
|
|
|
TestGetAndResetTickerCount(options, BLOOM_FILTER_USEFUL);
|
|
|
|
EXPECT_GE(useful_count, maxKey * 0.90);
|
|
|
|
EXPECT_LE(useful_count, maxKey * 0.91);
|
|
|
|
}
|
Add more LSM info to FilterBuildingContext (#8246)
Summary:
Add `num_levels`, `is_bottommost`, and table file creation
`reason` to `FilterBuildingContext`, in anticipation of more powerful
Bloom-like filter support.
To support this, added `is_bottommost` and `reason` to
`TableBuilderOptions`, which allowed removing `reason` parameter from
`rocksdb::BuildTable`.
I attempted to remove `skip_filters` from `TableBuilderOptions`, because
filter construction decisions should arise from options, not one-off
parameters. I could not completely remove it because the public API for
SstFileWriter takes a `skip_filters` parameter, and translating this
into an option change would mean awkwardly replacing the table_factory
if it is BlockBasedTableFactory with new filter_policy=nullptr option.
I marked this public skip_filters option as deprecated because of this
oddity. (skip_filters on the read side probably makes sense.)
At least `skip_filters` is now largely hidden for users of
`TableBuilderOptions` and is no longer used for implementing the
optimize_filters_for_hits option. Bringing the logic for that option
closer to handling of FilterBuildingContext makes it more obvious that
hese two are using the same notion of "bottommost." (Planned:
configuration options for Bloom-like filters that generalize
`optimize_filters_for_hits`)
Recommended follow-up: Try to get away from "bottommost level" naming of
things, which is inaccurate (see
VersionStorageInfo::RangeMightExistAfterSortedRun), and move to
"bottommost run" or just "bottommost."
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8246
Test Plan:
extended an existing unit test to exercise and check various
filter building contexts. Also, existing tests for
optimize_filters_for_hits validate some of the "bottommost" handling,
which is now closely connected to FilterBuildingContext::is_bottommost
through TableBuilderOptions::is_bottommost
Reviewed By: mrambacher
Differential Revision: D28099346
Pulled By: pdillinger
fbshipit-source-id: 2c1072e29c24d4ac404c761a7b7663292372600a
3 years ago
|
|
|
} else {
|
|
|
|
#ifndef ROCKSDB_LITE
|
|
|
|
// Also try external SST file
|
|
|
|
{
|
|
|
|
std::string file_path = dbname_ + "/external.sst";
|
|
|
|
SstFileWriter sst_file_writer(EnvOptions(), options, handles_[1]);
|
|
|
|
ASSERT_OK(sst_file_writer.Open(file_path));
|
|
|
|
ASSERT_OK(sst_file_writer.Put("key", "value"));
|
|
|
|
ASSERT_OK(sst_file_writer.Finish());
|
|
|
|
}
|
|
|
|
// Note: kCompactionStyleLevel is default, ignored if num_levels == -1
|
|
|
|
EXPECT_EQ(policy->DumpTestReport(),
|
|
|
|
"cf=abe,s=kCompactionStyleLevel,n=-1,l=-1,b=0,r=kMisc\n");
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
|
|
|
|
// Destroy
|
|
|
|
ASSERT_OK(dbfull()->DropColumnFamily(handles_[1]));
|
|
|
|
ASSERT_OK(dbfull()->DestroyColumnFamilyHandle(handles_[1]));
|
|
|
|
handles_[1] = nullptr;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
class SliceTransformLimitedDomain : public SliceTransform {
|
|
|
|
const char* Name() const override { return "SliceTransformLimitedDomain"; }
|
|
|
|
|
|
|
|
Slice Transform(const Slice& src) const override {
|
|
|
|
return Slice(src.data(), 5);
|
|
|
|
}
|
|
|
|
|
|
|
|
bool InDomain(const Slice& src) const override {
|
|
|
|
// prefix will be x????
|
|
|
|
return src.size() >= 5 && src[0] == 'x';
|
|
|
|
}
|
|
|
|
|
|
|
|
bool InRange(const Slice& dst) const override {
|
|
|
|
// prefix will be x????
|
|
|
|
return dst.size() == 5 && dst[0] == 'x';
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
TEST_F(DBBloomFilterTest, PrefixExtractorFullFilter) {
|
|
|
|
BlockBasedTableOptions bbto;
|
|
|
|
// Full Filter Block
|
|
|
|
bbto.filter_policy.reset(ROCKSDB_NAMESPACE::NewBloomFilterPolicy(10, false));
|
|
|
|
bbto.whole_key_filtering = false;
|
|
|
|
|
|
|
|
Options options = CurrentOptions();
|
|
|
|
options.prefix_extractor = std::make_shared<SliceTransformLimitedDomain>();
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
|
|
|
|
|
|
|
|
DestroyAndReopen(options);
|
|
|
|
|
|
|
|
ASSERT_OK(Put("x1111_AAAA", "val1"));
|
|
|
|
ASSERT_OK(Put("x1112_AAAA", "val2"));
|
|
|
|
ASSERT_OK(Put("x1113_AAAA", "val3"));
|
|
|
|
ASSERT_OK(Put("x1114_AAAA", "val4"));
|
|
|
|
// Not in domain, wont be added to filter
|
|
|
|
ASSERT_OK(Put("zzzzz_AAAA", "val5"));
|
|
|
|
|
|
|
|
ASSERT_OK(Flush());
|
|
|
|
|
|
|
|
ASSERT_EQ(Get("x1111_AAAA"), "val1");
|
|
|
|
ASSERT_EQ(Get("x1112_AAAA"), "val2");
|
|
|
|
ASSERT_EQ(Get("x1113_AAAA"), "val3");
|
|
|
|
ASSERT_EQ(Get("x1114_AAAA"), "val4");
|
|
|
|
// Was not added to filter but rocksdb will try to read it from the filter
|
|
|
|
ASSERT_EQ(Get("zzzzz_AAAA"), "val5");
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(DBBloomFilterTest, PrefixExtractorBlockFilter) {
|
|
|
|
BlockBasedTableOptions bbto;
|
|
|
|
// Block Filter Block
|
|
|
|
bbto.filter_policy.reset(ROCKSDB_NAMESPACE::NewBloomFilterPolicy(10, true));
|
|
|
|
|
|
|
|
Options options = CurrentOptions();
|
|
|
|
options.prefix_extractor = std::make_shared<SliceTransformLimitedDomain>();
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
|
|
|
|
|
|
|
|
DestroyAndReopen(options);
|
|
|
|
|
|
|
|
ASSERT_OK(Put("x1113_AAAA", "val3"));
|
|
|
|
ASSERT_OK(Put("x1114_AAAA", "val4"));
|
|
|
|
// Not in domain, wont be added to filter
|
|
|
|
ASSERT_OK(Put("zzzzz_AAAA", "val1"));
|
|
|
|
ASSERT_OK(Put("zzzzz_AAAB", "val2"));
|
|
|
|
ASSERT_OK(Put("zzzzz_AAAC", "val3"));
|
|
|
|
ASSERT_OK(Put("zzzzz_AAAD", "val4"));
|
|
|
|
|
|
|
|
ASSERT_OK(Flush());
|
|
|
|
|
|
|
|
std::vector<std::string> iter_res;
|
|
|
|
auto iter = db_->NewIterator(ReadOptions());
|
|
|
|
// Seek to a key that was not in Domain
|
|
|
|
for (iter->Seek("zzzzz_AAAA"); iter->Valid(); iter->Next()) {
|
|
|
|
iter_res.emplace_back(iter->value().ToString());
|
|
|
|
}
|
|
|
|
|
|
|
|
std::vector<std::string> expected_res = {"val1", "val2", "val3", "val4"};
|
|
|
|
ASSERT_EQ(iter_res, expected_res);
|
|
|
|
delete iter;
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(DBBloomFilterTest, MemtableWholeKeyBloomFilter) {
|
|
|
|
// regression test for #2743. the range delete tombstones in memtable should
|
|
|
|
// be added even when Get() skips searching due to its prefix bloom filter
|
|
|
|
const int kMemtableSize = 1 << 20; // 1MB
|
|
|
|
const int kMemtablePrefixFilterSize = 1 << 13; // 8KB
|
|
|
|
const int kPrefixLen = 4;
|
|
|
|
Options options = CurrentOptions();
|
|
|
|
options.memtable_prefix_bloom_size_ratio =
|
|
|
|
static_cast<double>(kMemtablePrefixFilterSize) / kMemtableSize;
|
|
|
|
options.prefix_extractor.reset(
|
|
|
|
ROCKSDB_NAMESPACE::NewFixedPrefixTransform(kPrefixLen));
|
|
|
|
options.write_buffer_size = kMemtableSize;
|
|
|
|
options.memtable_whole_key_filtering = false;
|
|
|
|
Reopen(options);
|
|
|
|
std::string key1("AAAABBBB");
|
|
|
|
std::string key2("AAAACCCC"); // not in DB
|
|
|
|
std::string key3("AAAADDDD");
|
|
|
|
std::string key4("AAAAEEEE");
|
|
|
|
std::string value1("Value1");
|
|
|
|
std::string value3("Value3");
|
|
|
|
std::string value4("Value4");
|
|
|
|
|
|
|
|
ASSERT_OK(Put(key1, value1, WriteOptions()));
|
|
|
|
|
|
|
|
// check memtable bloom stats
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get(key2));
|
|
|
|
ASSERT_EQ(0, get_perf_context()->bloom_memtable_miss_count);
|
|
|
|
// same prefix, bloom filter false positive
|
|
|
|
ASSERT_EQ(1, get_perf_context()->bloom_memtable_hit_count);
|
|
|
|
|
|
|
|
// enable whole key bloom filter
|
|
|
|
options.memtable_whole_key_filtering = true;
|
|
|
|
Reopen(options);
|
|
|
|
// check memtable bloom stats
|
|
|
|
ASSERT_OK(Put(key3, value3, WriteOptions()));
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get(key2));
|
|
|
|
// whole key bloom filter kicks in and determines it's a miss
|
|
|
|
ASSERT_EQ(1, get_perf_context()->bloom_memtable_miss_count);
|
|
|
|
ASSERT_EQ(1, get_perf_context()->bloom_memtable_hit_count);
|
|
|
|
|
|
|
|
// verify whole key filtering does not depend on prefix_extractor
|
|
|
|
options.prefix_extractor.reset();
|
|
|
|
Reopen(options);
|
|
|
|
// check memtable bloom stats
|
|
|
|
ASSERT_OK(Put(key4, value4, WriteOptions()));
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get(key2));
|
|
|
|
// whole key bloom filter kicks in and determines it's a miss
|
|
|
|
ASSERT_EQ(2, get_perf_context()->bloom_memtable_miss_count);
|
|
|
|
ASSERT_EQ(1, get_perf_context()->bloom_memtable_hit_count);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(DBBloomFilterTest, MemtablePrefixBloomOutOfDomain) {
|
|
|
|
constexpr size_t kPrefixSize = 8;
|
|
|
|
const std::string kKey = "key";
|
|
|
|
assert(kKey.size() < kPrefixSize);
|
|
|
|
Options options = CurrentOptions();
|
|
|
|
options.prefix_extractor.reset(NewFixedPrefixTransform(kPrefixSize));
|
|
|
|
options.memtable_prefix_bloom_size_ratio = 0.25;
|
|
|
|
Reopen(options);
|
|
|
|
ASSERT_OK(Put(kKey, "v"));
|
|
|
|
ASSERT_EQ("v", Get(kKey));
|
|
|
|
std::unique_ptr<Iterator> iter(dbfull()->NewIterator(ReadOptions()));
|
|
|
|
iter->Seek(kKey);
|
|
|
|
ASSERT_TRUE(iter->Valid());
|
|
|
|
ASSERT_EQ(kKey, iter->key());
|
|
|
|
iter->SeekForPrev(kKey);
|
|
|
|
ASSERT_TRUE(iter->Valid());
|
|
|
|
ASSERT_EQ(kKey, iter->key());
|
|
|
|
}
|
|
|
|
|
|
|
|
class DBBloomFilterTestVaryPrefixAndFormatVer
|
|
|
|
: public DBTestBase,
|
|
|
|
public testing::WithParamInterface<std::tuple<bool, uint32_t>> {
|
|
|
|
protected:
|
|
|
|
bool use_prefix_;
|
|
|
|
uint32_t format_version_;
|
|
|
|
|
|
|
|
public:
|
|
|
|
DBBloomFilterTestVaryPrefixAndFormatVer()
|
|
|
|
: DBTestBase("db_bloom_filter_tests", /*env_do_fsync=*/true) {}
|
|
|
|
|
|
|
|
~DBBloomFilterTestVaryPrefixAndFormatVer() override {}
|
|
|
|
|
|
|
|
void SetUp() override {
|
|
|
|
use_prefix_ = std::get<0>(GetParam());
|
|
|
|
format_version_ = std::get<1>(GetParam());
|
|
|
|
}
|
|
|
|
|
|
|
|
static std::string UKey(uint32_t i) { return Key(static_cast<int>(i)); }
|
|
|
|
};
|
|
|
|
|
|
|
|
TEST_P(DBBloomFilterTestVaryPrefixAndFormatVer, PartitionedMultiGet) {
|
|
|
|
Options options = CurrentOptions();
|
|
|
|
if (use_prefix_) {
|
|
|
|
// Entire key from UKey()
|
|
|
|
options.prefix_extractor.reset(NewCappedPrefixTransform(9));
|
|
|
|
}
|
|
|
|
options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
|
|
|
|
BlockBasedTableOptions bbto;
|
|
|
|
bbto.filter_policy.reset(NewBloomFilterPolicy(20));
|
|
|
|
bbto.partition_filters = true;
|
|
|
|
bbto.index_type = BlockBasedTableOptions::IndexType::kTwoLevelIndexSearch;
|
|
|
|
bbto.whole_key_filtering = !use_prefix_;
|
|
|
|
if (use_prefix_) { // (not related to prefix, just alternating between)
|
|
|
|
// Make sure code appropriately deals with metadata block size setting
|
|
|
|
// that is "too small" (smaller than minimum size for filter builder)
|
|
|
|
bbto.metadata_block_size = 63;
|
|
|
|
} else {
|
|
|
|
// Make sure the test will work even on platforms with large minimum
|
|
|
|
// filter size, due to large cache line size.
|
|
|
|
// (Largest cache line size + 10+% overhead.)
|
|
|
|
bbto.metadata_block_size = 290;
|
|
|
|
}
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
|
|
|
|
DestroyAndReopen(options);
|
|
|
|
ReadOptions ropts;
|
|
|
|
|
|
|
|
constexpr uint32_t N = 12000;
|
|
|
|
// Add N/2 evens
|
|
|
|
for (uint32_t i = 0; i < N; i += 2) {
|
|
|
|
ASSERT_OK(Put(UKey(i), UKey(i)));
|
|
|
|
}
|
|
|
|
ASSERT_OK(Flush());
|
|
|
|
#ifndef ROCKSDB_LITE
|
|
|
|
ASSERT_EQ(TotalTableFiles(), 1);
|
|
|
|
#endif
|
|
|
|
|
|
|
|
constexpr uint32_t Q = 29;
|
|
|
|
// MultiGet In
|
|
|
|
std::array<std::string, Q> keys;
|
|
|
|
std::array<Slice, Q> key_slices;
|
|
|
|
std::array<ColumnFamilyHandle*, Q> column_families;
|
|
|
|
// MultiGet Out
|
|
|
|
std::array<Status, Q> statuses;
|
|
|
|
std::array<PinnableSlice, Q> values;
|
|
|
|
|
|
|
|
TestGetAndResetTickerCount(options, BLOCK_CACHE_FILTER_HIT);
|
|
|
|
TestGetAndResetTickerCount(options, BLOCK_CACHE_FILTER_MISS);
|
|
|
|
TestGetAndResetTickerCount(options, BLOOM_FILTER_PREFIX_USEFUL);
|
|
|
|
TestGetAndResetTickerCount(options, BLOOM_FILTER_USEFUL);
|
|
|
|
TestGetAndResetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED);
|
|
|
|
TestGetAndResetTickerCount(options, BLOOM_FILTER_FULL_POSITIVE);
|
|
|
|
TestGetAndResetTickerCount(options, BLOOM_FILTER_FULL_TRUE_POSITIVE);
|
|
|
|
|
|
|
|
// Check that initial clump of keys only loads one partition filter from
|
|
|
|
// block cache.
|
|
|
|
// And that spread out keys load many partition filters.
|
|
|
|
// In both cases, mix present vs. not present keys.
|
|
|
|
for (uint32_t stride : {uint32_t{1}, (N / Q) | 1}) {
|
|
|
|
for (uint32_t i = 0; i < Q; ++i) {
|
|
|
|
keys[i] = UKey(i * stride);
|
|
|
|
key_slices[i] = Slice(keys[i]);
|
|
|
|
column_families[i] = db_->DefaultColumnFamily();
|
|
|
|
statuses[i] = Status();
|
|
|
|
values[i] = PinnableSlice();
|
|
|
|
}
|
|
|
|
|
|
|
|
db_->MultiGet(ropts, Q, &column_families[0], &key_slices[0], &values[0],
|
|
|
|
/*timestamps=*/nullptr, &statuses[0], true);
|
|
|
|
|
|
|
|
// Confirm correct status results
|
|
|
|
uint32_t number_not_found = 0;
|
|
|
|
for (uint32_t i = 0; i < Q; ++i) {
|
|
|
|
if ((i * stride % 2) == 0) {
|
|
|
|
ASSERT_OK(statuses[i]);
|
|
|
|
} else {
|
|
|
|
ASSERT_TRUE(statuses[i].IsNotFound());
|
|
|
|
++number_not_found;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Confirm correct Bloom stats (no FPs)
|
|
|
|
uint64_t filter_useful = TestGetAndResetTickerCount(
|
|
|
|
options,
|
|
|
|
use_prefix_ ? BLOOM_FILTER_PREFIX_USEFUL : BLOOM_FILTER_USEFUL);
|
|
|
|
uint64_t filter_checked =
|
|
|
|
TestGetAndResetTickerCount(options, use_prefix_
|
|
|
|
? BLOOM_FILTER_PREFIX_CHECKED
|
|
|
|
: BLOOM_FILTER_FULL_POSITIVE) +
|
|
|
|
(use_prefix_ ? 0 : filter_useful);
|
|
|
|
EXPECT_EQ(filter_useful, number_not_found);
|
|
|
|
EXPECT_EQ(filter_checked, Q);
|
|
|
|
if (!use_prefix_) {
|
|
|
|
EXPECT_EQ(
|
|
|
|
TestGetAndResetTickerCount(options, BLOOM_FILTER_FULL_TRUE_POSITIVE),
|
|
|
|
Q - number_not_found);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Confirm no duplicate loading same filter partition
|
|
|
|
uint64_t filter_accesses =
|
|
|
|
TestGetAndResetTickerCount(options, BLOCK_CACHE_FILTER_HIT) +
|
|
|
|
TestGetAndResetTickerCount(options, BLOCK_CACHE_FILTER_MISS);
|
|
|
|
if (stride == 1) {
|
|
|
|
EXPECT_EQ(filter_accesses, 1);
|
|
|
|
} else {
|
|
|
|
// for large stride
|
|
|
|
EXPECT_GE(filter_accesses, Q / 2 + 1);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Check that a clump of keys (present and not) works when spanning
|
|
|
|
// two partitions
|
|
|
|
int found_spanning = 0;
|
|
|
|
for (uint32_t start = 0; start < N / 2;) {
|
|
|
|
for (uint32_t i = 0; i < Q; ++i) {
|
|
|
|
keys[i] = UKey(start + i);
|
|
|
|
key_slices[i] = Slice(keys[i]);
|
|
|
|
column_families[i] = db_->DefaultColumnFamily();
|
|
|
|
statuses[i] = Status();
|
|
|
|
values[i] = PinnableSlice();
|
|
|
|
}
|
|
|
|
|
|
|
|
db_->MultiGet(ropts, Q, &column_families[0], &key_slices[0], &values[0],
|
|
|
|
/*timestamps=*/nullptr, &statuses[0], true);
|
|
|
|
|
|
|
|
// Confirm correct status results
|
|
|
|
uint32_t number_not_found = 0;
|
|
|
|
for (uint32_t i = 0; i < Q; ++i) {
|
|
|
|
if (((start + i) % 2) == 0) {
|
|
|
|
ASSERT_OK(statuses[i]);
|
|
|
|
} else {
|
|
|
|
ASSERT_TRUE(statuses[i].IsNotFound());
|
|
|
|
++number_not_found;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Confirm correct Bloom stats (might see some FPs)
|
|
|
|
uint64_t filter_useful = TestGetAndResetTickerCount(
|
|
|
|
options,
|
|
|
|
use_prefix_ ? BLOOM_FILTER_PREFIX_USEFUL : BLOOM_FILTER_USEFUL);
|
|
|
|
uint64_t filter_checked =
|
|
|
|
TestGetAndResetTickerCount(options, use_prefix_
|
|
|
|
? BLOOM_FILTER_PREFIX_CHECKED
|
|
|
|
: BLOOM_FILTER_FULL_POSITIVE) +
|
|
|
|
(use_prefix_ ? 0 : filter_useful);
|
|
|
|
EXPECT_GE(filter_useful, number_not_found - 2); // possible FP
|
|
|
|
EXPECT_EQ(filter_checked, Q);
|
|
|
|
if (!use_prefix_) {
|
|
|
|
EXPECT_EQ(
|
|
|
|
TestGetAndResetTickerCount(options, BLOOM_FILTER_FULL_TRUE_POSITIVE),
|
|
|
|
Q - number_not_found);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Confirm no duplicate loading of same filter partition
|
|
|
|
uint64_t filter_accesses =
|
|
|
|
TestGetAndResetTickerCount(options, BLOCK_CACHE_FILTER_HIT) +
|
|
|
|
TestGetAndResetTickerCount(options, BLOCK_CACHE_FILTER_MISS);
|
|
|
|
if (filter_accesses == 2) {
|
|
|
|
// Spanned across partitions.
|
|
|
|
++found_spanning;
|
|
|
|
if (found_spanning >= 2) {
|
|
|
|
break;
|
|
|
|
} else {
|
|
|
|
// Ensure that at least once we have at least one present and
|
|
|
|
// one non-present key on both sides of partition boundary.
|
|
|
|
start += 2;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
EXPECT_EQ(filter_accesses, 1);
|
|
|
|
// See explanation at "start += 2"
|
|
|
|
start += Q - 4;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
EXPECT_TRUE(found_spanning >= 2);
|
|
|
|
}
|
|
|
|
|
|
|
|
INSTANTIATE_TEST_CASE_P(DBBloomFilterTestVaryPrefixAndFormatVer,
|
|
|
|
DBBloomFilterTestVaryPrefixAndFormatVer,
|
|
|
|
::testing::Values(
|
|
|
|
// (use_prefix, format_version)
|
|
|
|
std::make_tuple(false, 2),
|
|
|
|
std::make_tuple(false, 3),
|
|
|
|
std::make_tuple(false, 4),
|
|
|
|
std::make_tuple(false, 5),
|
|
|
|
std::make_tuple(true, 2),
|
|
|
|
std::make_tuple(true, 3),
|
|
|
|
std::make_tuple(true, 4),
|
|
|
|
std::make_tuple(true, 5)));
|
|
|
|
|
|
|
|
#ifndef ROCKSDB_LITE
|
|
|
|
namespace {
|
|
|
|
namespace BFP2 {
|
|
|
|
// Extends BFP::Mode with option to use Plain table
|
|
|
|
using PseudoMode = int;
|
|
|
|
static constexpr PseudoMode kPlainTable = -1;
|
|
|
|
} // namespace BFP2
|
|
|
|
} // namespace
|
|
|
|
|
|
|
|
class BloomStatsTestWithParam
|
|
|
|
: public DBBloomFilterTest,
|
|
|
|
public testing::WithParamInterface<std::tuple<BFP2::PseudoMode, bool>> {
|
|
|
|
public:
|
|
|
|
BloomStatsTestWithParam() {
|
New Bloom filter implementation for full and partitioned filters (#6007)
Summary:
Adds an improved, replacement Bloom filter implementation (FastLocalBloom) for full and partitioned filters in the block-based table. This replacement is faster and more accurate, especially for high bits per key or millions of keys in a single filter.
Speed
The improved speed, at least on recent x86_64, comes from
* Using fastrange instead of modulo (%)
* Using our new hash function (XXH3 preview, added in a previous commit), which is much faster for large keys and only *slightly* slower on keys around 12 bytes if hashing the same size many thousands of times in a row.
* Optimizing the Bloom filter queries with AVX2 SIMD operations. (Added AVX2 to the USE_SSE=1 build.) Careful design was required to support (a) SIMD-optimized queries, (b) compatible non-SIMD code that's simple and efficient, (c) flexible choice of number of probes, and (d) essentially maximized accuracy for a cache-local Bloom filter. Probes are made eight at a time, so any number of probes up to 8 is the same speed, then up to 16, etc.
* Prefetching cache lines when building the filter. Although this optimization could be applied to the old structure as well, it seems to balance out the small added cost of accumulating 64 bit hashes for adding to the filter rather than 32 bit hashes.
Here's nominal speed data from filter_bench (200MB in filters, about 10k keys each, 10 bits filter data / key, 6 probes, avg key size 24 bytes, includes hashing time) on Skylake DE (relatively low clock speed):
$ ./filter_bench -quick -impl=2 -net_includes_hashing # New Bloom filter
Build avg ns/key: 47.7135
Mixed inside/outside queries...
Single filter net ns/op: 26.2825
Random filter net ns/op: 150.459
Average FP rate %: 0.954651
$ ./filter_bench -quick -impl=0 -net_includes_hashing # Old Bloom filter
Build avg ns/key: 47.2245
Mixed inside/outside queries...
Single filter net ns/op: 63.2978
Random filter net ns/op: 188.038
Average FP rate %: 1.13823
Similar build time but dramatically faster query times on hot data (63 ns to 26 ns), and somewhat faster on stale data (188 ns to 150 ns). Performance differences on batched and skewed query loads are between these extremes as expected.
The only other interesting thing about speed is "inside" (query key was added to filter) vs. "outside" (query key was not added to filter) query times. The non-SIMD implementations are substantially slower when most queries are "outside" vs. "inside". This goes against what one might expect or would have observed years ago, as "outside" queries only need about two probes on average, due to short-circuiting, while "inside" always have num_probes (say 6). The problem is probably the nastily unpredictable branch. The SIMD implementation has few branches (very predictable) and has pretty consistent running time regardless of query outcome.
Accuracy
The generally improved accuracy (re: Issue https://github.com/facebook/rocksdb/issues/5857) comes from a better design for probing indices
within a cache line (re: Issue https://github.com/facebook/rocksdb/issues/4120) and improved accuracy for millions of keys in a single filter from using a 64-bit hash function (XXH3p). Design details in code comments.
Accuracy data (generalizes, except old impl gets worse with millions of keys):
Memory bits per key: FP rate percent old impl -> FP rate percent new impl
6: 5.70953 -> 5.69888
8: 2.45766 -> 2.29709
10: 1.13977 -> 0.959254
12: 0.662498 -> 0.411593
16: 0.353023 -> 0.0873754
24: 0.261552 -> 0.0060971
50: 0.225453 -> ~0.00003 (less than 1 in a million queries are FP)
Fixes https://github.com/facebook/rocksdb/issues/5857
Fixes https://github.com/facebook/rocksdb/issues/4120
Unlike the old implementation, this implementation has a fixed cache line size (64 bytes). At 10 bits per key, the accuracy of this new implementation is very close to the old implementation with 128-byte cache line size. If there's sufficient demand, this implementation could be generalized.
Compatibility
Although old releases would see the new structure as corrupt filter data and read the table as if there's no filter, we've decided only to enable the new Bloom filter with new format_version=5. This provides a smooth path for automatic adoption over time, with an option for early opt-in.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6007
Test Plan: filter_bench has been used thoroughly to validate speed, accuracy, and correctness. Unit tests have been carefully updated to exercise new and old implementations, as well as the logic to select an implementation based on context (format_version).
Differential Revision: D18294749
Pulled By: pdillinger
fbshipit-source-id: d44c9db3696e4d0a17caaec47075b7755c262c5f
5 years ago
|
|
|
bfp_impl_ = std::get<0>(GetParam());
|
|
|
|
partition_filters_ = std::get<1>(GetParam());
|
|
|
|
|
|
|
|
options_.create_if_missing = true;
|
|
|
|
options_.prefix_extractor.reset(
|
|
|
|
ROCKSDB_NAMESPACE::NewFixedPrefixTransform(4));
|
|
|
|
options_.memtable_prefix_bloom_size_ratio =
|
|
|
|
8.0 * 1024.0 / static_cast<double>(options_.write_buffer_size);
|
|
|
|
if (bfp_impl_ == BFP2::kPlainTable) {
|
|
|
|
assert(!partition_filters_); // not supported in plain table
|
|
|
|
PlainTableOptions table_options;
|
|
|
|
options_.table_factory.reset(NewPlainTableFactory(table_options));
|
|
|
|
} else {
|
|
|
|
BlockBasedTableOptions table_options;
|
|
|
|
table_options.hash_index_allow_collision = false;
|
|
|
|
if (partition_filters_) {
|
|
|
|
assert(bfp_impl_ != BFP::kDeprecatedBlock);
|
|
|
|
table_options.partition_filters = partition_filters_;
|
|
|
|
table_options.index_type =
|
|
|
|
BlockBasedTableOptions::IndexType::kTwoLevelIndexSearch;
|
|
|
|
}
|
|
|
|
table_options.filter_policy.reset(
|
|
|
|
new BFP(10, static_cast<BFP::Mode>(bfp_impl_)));
|
|
|
|
options_.table_factory.reset(NewBlockBasedTableFactory(table_options));
|
|
|
|
}
|
|
|
|
options_.env = env_;
|
|
|
|
|
|
|
|
get_perf_context()->Reset();
|
|
|
|
DestroyAndReopen(options_);
|
|
|
|
}
|
|
|
|
|
|
|
|
~BloomStatsTestWithParam() override {
|
|
|
|
get_perf_context()->Reset();
|
|
|
|
Destroy(options_);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Required if inheriting from testing::WithParamInterface<>
|
|
|
|
static void SetUpTestCase() {}
|
|
|
|
static void TearDownTestCase() {}
|
|
|
|
|
|
|
|
BFP2::PseudoMode bfp_impl_;
|
|
|
|
bool partition_filters_;
|
|
|
|
Options options_;
|
|
|
|
};
|
|
|
|
|
|
|
|
// 1 Insert 2 K-V pairs into DB
|
|
|
|
// 2 Call Get() for both keys - expext memtable bloom hit stat to be 2
|
|
|
|
// 3 Call Get() for nonexisting key - expect memtable bloom miss stat to be 1
|
|
|
|
// 4 Call Flush() to create SST
|
|
|
|
// 5 Call Get() for both keys - expext SST bloom hit stat to be 2
|
|
|
|
// 6 Call Get() for nonexisting key - expect SST bloom miss stat to be 1
|
|
|
|
// Test both: block and plain SST
|
|
|
|
TEST_P(BloomStatsTestWithParam, BloomStatsTest) {
|
|
|
|
std::string key1("AAAA");
|
|
|
|
std::string key2("RXDB"); // not in DB
|
|
|
|
std::string key3("ZBRA");
|
|
|
|
std::string value1("Value1");
|
|
|
|
std::string value3("Value3");
|
|
|
|
|
|
|
|
ASSERT_OK(Put(key1, value1, WriteOptions()));
|
|
|
|
ASSERT_OK(Put(key3, value3, WriteOptions()));
|
|
|
|
|
|
|
|
// check memtable bloom stats
|
|
|
|
ASSERT_EQ(value1, Get(key1));
|
|
|
|
ASSERT_EQ(1, get_perf_context()->bloom_memtable_hit_count);
|
|
|
|
ASSERT_EQ(value3, Get(key3));
|
|
|
|
ASSERT_EQ(2, get_perf_context()->bloom_memtable_hit_count);
|
|
|
|
ASSERT_EQ(0, get_perf_context()->bloom_memtable_miss_count);
|
|
|
|
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get(key2));
|
|
|
|
ASSERT_EQ(1, get_perf_context()->bloom_memtable_miss_count);
|
|
|
|
ASSERT_EQ(2, get_perf_context()->bloom_memtable_hit_count);
|
|
|
|
|
|
|
|
// sanity checks
|
|
|
|
ASSERT_EQ(0, get_perf_context()->bloom_sst_hit_count);
|
|
|
|
ASSERT_EQ(0, get_perf_context()->bloom_sst_miss_count);
|
|
|
|
|
|
|
|
Flush();
|
|
|
|
|
|
|
|
// sanity checks
|
|
|
|
ASSERT_EQ(0, get_perf_context()->bloom_sst_hit_count);
|
|
|
|
ASSERT_EQ(0, get_perf_context()->bloom_sst_miss_count);
|
|
|
|
|
|
|
|
// check SST bloom stats
|
|
|
|
ASSERT_EQ(value1, Get(key1));
|
|
|
|
ASSERT_EQ(1, get_perf_context()->bloom_sst_hit_count);
|
|
|
|
ASSERT_EQ(value3, Get(key3));
|
|
|
|
ASSERT_EQ(2, get_perf_context()->bloom_sst_hit_count);
|
|
|
|
|
|
|
|
ASSERT_EQ("NOT_FOUND", Get(key2));
|
|
|
|
ASSERT_EQ(1, get_perf_context()->bloom_sst_miss_count);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Same scenario as in BloomStatsTest but using an iterator
|
|
|
|
TEST_P(BloomStatsTestWithParam, BloomStatsTestWithIter) {
|
|
|
|
std::string key1("AAAA");
|
|
|
|
std::string key2("RXDB"); // not in DB
|
|
|
|
std::string key3("ZBRA");
|
|
|
|
std::string value1("Value1");
|
|
|
|
std::string value3("Value3");
|
|
|
|
|
|
|
|
ASSERT_OK(Put(key1, value1, WriteOptions()));
|
|
|
|
ASSERT_OK(Put(key3, value3, WriteOptions()));
|
|
|
|
|
|
|
|
std::unique_ptr<Iterator> iter(dbfull()->NewIterator(ReadOptions()));
|
|
|
|
|
|
|
|
// check memtable bloom stats
|
|
|
|
iter->Seek(key1);
|
|
|
|
ASSERT_OK(iter->status());
|
|
|
|
ASSERT_TRUE(iter->Valid());
|
|
|
|
ASSERT_EQ(value1, iter->value().ToString());
|
|
|
|
ASSERT_EQ(1, get_perf_context()->bloom_memtable_hit_count);
|
|
|
|
ASSERT_EQ(0, get_perf_context()->bloom_memtable_miss_count);
|
|
|
|
|
|
|
|
iter->Seek(key3);
|
|
|
|
ASSERT_OK(iter->status());
|
|
|
|
ASSERT_TRUE(iter->Valid());
|
|
|
|
ASSERT_EQ(value3, iter->value().ToString());
|
|
|
|
ASSERT_EQ(2, get_perf_context()->bloom_memtable_hit_count);
|
|
|
|
ASSERT_EQ(0, get_perf_context()->bloom_memtable_miss_count);
|
|
|
|
|
|
|
|
iter->Seek(key2);
|
|
|
|
ASSERT_OK(iter->status());
|
|
|
|
ASSERT_TRUE(!iter->Valid());
|
|
|
|
ASSERT_EQ(1, get_perf_context()->bloom_memtable_miss_count);
|
|
|
|
ASSERT_EQ(2, get_perf_context()->bloom_memtable_hit_count);
|
|
|
|
|
|
|
|
Flush();
|
|
|
|
|
|
|
|
iter.reset(dbfull()->NewIterator(ReadOptions()));
|
|
|
|
|
|
|
|
// Check SST bloom stats
|
|
|
|
iter->Seek(key1);
|
|
|
|
ASSERT_OK(iter->status());
|
|
|
|
ASSERT_TRUE(iter->Valid());
|
|
|
|
ASSERT_EQ(value1, iter->value().ToString());
|
|
|
|
ASSERT_EQ(1, get_perf_context()->bloom_sst_hit_count);
|
|
|
|
|
|
|
|
iter->Seek(key3);
|
|
|
|
ASSERT_OK(iter->status());
|
|
|
|
ASSERT_TRUE(iter->Valid());
|
|
|
|
ASSERT_EQ(value3, iter->value().ToString());
|
|
|
|
// The seek doesn't check block-based bloom filter because last index key
|
|
|
|
// starts with the same prefix we're seeking to.
|
New Bloom filter implementation for full and partitioned filters (#6007)
Summary:
Adds an improved, replacement Bloom filter implementation (FastLocalBloom) for full and partitioned filters in the block-based table. This replacement is faster and more accurate, especially for high bits per key or millions of keys in a single filter.
Speed
The improved speed, at least on recent x86_64, comes from
* Using fastrange instead of modulo (%)
* Using our new hash function (XXH3 preview, added in a previous commit), which is much faster for large keys and only *slightly* slower on keys around 12 bytes if hashing the same size many thousands of times in a row.
* Optimizing the Bloom filter queries with AVX2 SIMD operations. (Added AVX2 to the USE_SSE=1 build.) Careful design was required to support (a) SIMD-optimized queries, (b) compatible non-SIMD code that's simple and efficient, (c) flexible choice of number of probes, and (d) essentially maximized accuracy for a cache-local Bloom filter. Probes are made eight at a time, so any number of probes up to 8 is the same speed, then up to 16, etc.
* Prefetching cache lines when building the filter. Although this optimization could be applied to the old structure as well, it seems to balance out the small added cost of accumulating 64 bit hashes for adding to the filter rather than 32 bit hashes.
Here's nominal speed data from filter_bench (200MB in filters, about 10k keys each, 10 bits filter data / key, 6 probes, avg key size 24 bytes, includes hashing time) on Skylake DE (relatively low clock speed):
$ ./filter_bench -quick -impl=2 -net_includes_hashing # New Bloom filter
Build avg ns/key: 47.7135
Mixed inside/outside queries...
Single filter net ns/op: 26.2825
Random filter net ns/op: 150.459
Average FP rate %: 0.954651
$ ./filter_bench -quick -impl=0 -net_includes_hashing # Old Bloom filter
Build avg ns/key: 47.2245
Mixed inside/outside queries...
Single filter net ns/op: 63.2978
Random filter net ns/op: 188.038
Average FP rate %: 1.13823
Similar build time but dramatically faster query times on hot data (63 ns to 26 ns), and somewhat faster on stale data (188 ns to 150 ns). Performance differences on batched and skewed query loads are between these extremes as expected.
The only other interesting thing about speed is "inside" (query key was added to filter) vs. "outside" (query key was not added to filter) query times. The non-SIMD implementations are substantially slower when most queries are "outside" vs. "inside". This goes against what one might expect or would have observed years ago, as "outside" queries only need about two probes on average, due to short-circuiting, while "inside" always have num_probes (say 6). The problem is probably the nastily unpredictable branch. The SIMD implementation has few branches (very predictable) and has pretty consistent running time regardless of query outcome.
Accuracy
The generally improved accuracy (re: Issue https://github.com/facebook/rocksdb/issues/5857) comes from a better design for probing indices
within a cache line (re: Issue https://github.com/facebook/rocksdb/issues/4120) and improved accuracy for millions of keys in a single filter from using a 64-bit hash function (XXH3p). Design details in code comments.
Accuracy data (generalizes, except old impl gets worse with millions of keys):
Memory bits per key: FP rate percent old impl -> FP rate percent new impl
6: 5.70953 -> 5.69888
8: 2.45766 -> 2.29709
10: 1.13977 -> 0.959254
12: 0.662498 -> 0.411593
16: 0.353023 -> 0.0873754
24: 0.261552 -> 0.0060971
50: 0.225453 -> ~0.00003 (less than 1 in a million queries are FP)
Fixes https://github.com/facebook/rocksdb/issues/5857
Fixes https://github.com/facebook/rocksdb/issues/4120
Unlike the old implementation, this implementation has a fixed cache line size (64 bytes). At 10 bits per key, the accuracy of this new implementation is very close to the old implementation with 128-byte cache line size. If there's sufficient demand, this implementation could be generalized.
Compatibility
Although old releases would see the new structure as corrupt filter data and read the table as if there's no filter, we've decided only to enable the new Bloom filter with new format_version=5. This provides a smooth path for automatic adoption over time, with an option for early opt-in.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6007
Test Plan: filter_bench has been used thoroughly to validate speed, accuracy, and correctness. Unit tests have been carefully updated to exercise new and old implementations, as well as the logic to select an implementation based on context (format_version).
Differential Revision: D18294749
Pulled By: pdillinger
fbshipit-source-id: d44c9db3696e4d0a17caaec47075b7755c262c5f
5 years ago
|
|
|
uint64_t expected_hits = bfp_impl_ == BFP::kDeprecatedBlock ? 1 : 2;
|
|
|
|
ASSERT_EQ(expected_hits, get_perf_context()->bloom_sst_hit_count);
|
|
|
|
|
|
|
|
iter->Seek(key2);
|
|
|
|
ASSERT_OK(iter->status());
|
|
|
|
ASSERT_TRUE(!iter->Valid());
|
|
|
|
ASSERT_EQ(1, get_perf_context()->bloom_sst_miss_count);
|
|
|
|
ASSERT_EQ(expected_hits, get_perf_context()->bloom_sst_hit_count);
|
|
|
|
}
|
|
|
|
|
|
|
|
INSTANTIATE_TEST_CASE_P(
|
|
|
|
BloomStatsTestWithParam, BloomStatsTestWithParam,
|
New Bloom filter implementation for full and partitioned filters (#6007)
Summary:
Adds an improved, replacement Bloom filter implementation (FastLocalBloom) for full and partitioned filters in the block-based table. This replacement is faster and more accurate, especially for high bits per key or millions of keys in a single filter.
Speed
The improved speed, at least on recent x86_64, comes from
* Using fastrange instead of modulo (%)
* Using our new hash function (XXH3 preview, added in a previous commit), which is much faster for large keys and only *slightly* slower on keys around 12 bytes if hashing the same size many thousands of times in a row.
* Optimizing the Bloom filter queries with AVX2 SIMD operations. (Added AVX2 to the USE_SSE=1 build.) Careful design was required to support (a) SIMD-optimized queries, (b) compatible non-SIMD code that's simple and efficient, (c) flexible choice of number of probes, and (d) essentially maximized accuracy for a cache-local Bloom filter. Probes are made eight at a time, so any number of probes up to 8 is the same speed, then up to 16, etc.
* Prefetching cache lines when building the filter. Although this optimization could be applied to the old structure as well, it seems to balance out the small added cost of accumulating 64 bit hashes for adding to the filter rather than 32 bit hashes.
Here's nominal speed data from filter_bench (200MB in filters, about 10k keys each, 10 bits filter data / key, 6 probes, avg key size 24 bytes, includes hashing time) on Skylake DE (relatively low clock speed):
$ ./filter_bench -quick -impl=2 -net_includes_hashing # New Bloom filter
Build avg ns/key: 47.7135
Mixed inside/outside queries...
Single filter net ns/op: 26.2825
Random filter net ns/op: 150.459
Average FP rate %: 0.954651
$ ./filter_bench -quick -impl=0 -net_includes_hashing # Old Bloom filter
Build avg ns/key: 47.2245
Mixed inside/outside queries...
Single filter net ns/op: 63.2978
Random filter net ns/op: 188.038
Average FP rate %: 1.13823
Similar build time but dramatically faster query times on hot data (63 ns to 26 ns), and somewhat faster on stale data (188 ns to 150 ns). Performance differences on batched and skewed query loads are between these extremes as expected.
The only other interesting thing about speed is "inside" (query key was added to filter) vs. "outside" (query key was not added to filter) query times. The non-SIMD implementations are substantially slower when most queries are "outside" vs. "inside". This goes against what one might expect or would have observed years ago, as "outside" queries only need about two probes on average, due to short-circuiting, while "inside" always have num_probes (say 6). The problem is probably the nastily unpredictable branch. The SIMD implementation has few branches (very predictable) and has pretty consistent running time regardless of query outcome.
Accuracy
The generally improved accuracy (re: Issue https://github.com/facebook/rocksdb/issues/5857) comes from a better design for probing indices
within a cache line (re: Issue https://github.com/facebook/rocksdb/issues/4120) and improved accuracy for millions of keys in a single filter from using a 64-bit hash function (XXH3p). Design details in code comments.
Accuracy data (generalizes, except old impl gets worse with millions of keys):
Memory bits per key: FP rate percent old impl -> FP rate percent new impl
6: 5.70953 -> 5.69888
8: 2.45766 -> 2.29709
10: 1.13977 -> 0.959254
12: 0.662498 -> 0.411593
16: 0.353023 -> 0.0873754
24: 0.261552 -> 0.0060971
50: 0.225453 -> ~0.00003 (less than 1 in a million queries are FP)
Fixes https://github.com/facebook/rocksdb/issues/5857
Fixes https://github.com/facebook/rocksdb/issues/4120
Unlike the old implementation, this implementation has a fixed cache line size (64 bytes). At 10 bits per key, the accuracy of this new implementation is very close to the old implementation with 128-byte cache line size. If there's sufficient demand, this implementation could be generalized.
Compatibility
Although old releases would see the new structure as corrupt filter data and read the table as if there's no filter, we've decided only to enable the new Bloom filter with new format_version=5. This provides a smooth path for automatic adoption over time, with an option for early opt-in.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6007
Test Plan: filter_bench has been used thoroughly to validate speed, accuracy, and correctness. Unit tests have been carefully updated to exercise new and old implementations, as well as the logic to select an implementation based on context (format_version).
Differential Revision: D18294749
Pulled By: pdillinger
fbshipit-source-id: d44c9db3696e4d0a17caaec47075b7755c262c5f
5 years ago
|
|
|
::testing::Values(std::make_tuple(BFP::kDeprecatedBlock, false),
|
|
|
|
std::make_tuple(BFP::kLegacyBloom, false),
|
|
|
|
std::make_tuple(BFP::kLegacyBloom, true),
|
|
|
|
std::make_tuple(BFP::kFastLocalBloom, false),
|
|
|
|
std::make_tuple(BFP::kFastLocalBloom, true),
|
|
|
|
std::make_tuple(BFP2::kPlainTable, false)));
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
void PrefixScanInit(DBBloomFilterTest* dbtest) {
|
|
|
|
char buf[100];
|
|
|
|
std::string keystr;
|
|
|
|
const int small_range_sstfiles = 5;
|
|
|
|
const int big_range_sstfiles = 5;
|
|
|
|
|
|
|
|
// Generate 11 sst files with the following prefix ranges.
|
|
|
|
// GROUP 0: [0,10] (level 1)
|
|
|
|
// GROUP 1: [1,2], [2,3], [3,4], [4,5], [5, 6] (level 0)
|
|
|
|
// GROUP 2: [0,6], [0,7], [0,8], [0,9], [0,10] (level 0)
|
|
|
|
//
|
|
|
|
// A seek with the previous API would do 11 random I/Os (to all the
|
|
|
|
// files). With the new API and a prefix filter enabled, we should
|
|
|
|
// only do 2 random I/O, to the 2 files containing the key.
|
|
|
|
|
|
|
|
// GROUP 0
|
|
|
|
snprintf(buf, sizeof(buf), "%02d______:start", 0);
|
|
|
|
keystr = std::string(buf);
|
|
|
|
ASSERT_OK(dbtest->Put(keystr, keystr));
|
|
|
|
snprintf(buf, sizeof(buf), "%02d______:end", 10);
|
|
|
|
keystr = std::string(buf);
|
|
|
|
ASSERT_OK(dbtest->Put(keystr, keystr));
|
|
|
|
ASSERT_OK(dbtest->Flush());
|
|
|
|
ASSERT_OK(dbtest->dbfull()->CompactRange(CompactRangeOptions(), nullptr,
|
|
|
|
nullptr)); // move to level 1
|
|
|
|
|
|
|
|
// GROUP 1
|
|
|
|
for (int i = 1; i <= small_range_sstfiles; i++) {
|
|
|
|
snprintf(buf, sizeof(buf), "%02d______:start", i);
|
|
|
|
keystr = std::string(buf);
|
|
|
|
ASSERT_OK(dbtest->Put(keystr, keystr));
|
|
|
|
snprintf(buf, sizeof(buf), "%02d______:end", i + 1);
|
|
|
|
keystr = std::string(buf);
|
|
|
|
ASSERT_OK(dbtest->Put(keystr, keystr));
|
|
|
|
dbtest->Flush();
|
|
|
|
}
|
|
|
|
|
|
|
|
// GROUP 2
|
|
|
|
for (int i = 1; i <= big_range_sstfiles; i++) {
|
|
|
|
snprintf(buf, sizeof(buf), "%02d______:start", 0);
|
|
|
|
keystr = std::string(buf);
|
|
|
|
ASSERT_OK(dbtest->Put(keystr, keystr));
|
|
|
|
snprintf(buf, sizeof(buf), "%02d______:end", small_range_sstfiles + i + 1);
|
|
|
|
keystr = std::string(buf);
|
|
|
|
ASSERT_OK(dbtest->Put(keystr, keystr));
|
|
|
|
dbtest->Flush();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
} // namespace
|
|
|
|
|
|
|
|
TEST_F(DBBloomFilterTest, PrefixScan) {
|
|
|
|
while (ChangeFilterOptions()) {
|
|
|
|
int count;
|
|
|
|
Slice prefix;
|
|
|
|
Slice key;
|
|
|
|
char buf[100];
|
|
|
|
Iterator* iter;
|
|
|
|
snprintf(buf, sizeof(buf), "03______:");
|
|
|
|
prefix = Slice(buf, 8);
|
|
|
|
key = Slice(buf, 9);
|
|
|
|
ASSERT_EQ(key.difference_offset(prefix), 8);
|
|
|
|
ASSERT_EQ(prefix.difference_offset(key), 8);
|
|
|
|
// db configs
|
|
|
|
env_->count_random_reads_ = true;
|
|
|
|
Options options = CurrentOptions();
|
|
|
|
options.env = env_;
|
|
|
|
options.prefix_extractor.reset(NewFixedPrefixTransform(8));
|
|
|
|
options.disable_auto_compactions = true;
|
|
|
|
options.max_background_compactions = 2;
|
|
|
|
options.create_if_missing = true;
|
|
|
|
options.memtable_factory.reset(NewHashSkipListRepFactory(16));
|
|
|
|
assert(!options.unordered_write);
|
|
|
|
// It is incompatible with allow_concurrent_memtable_write=false
|
|
|
|
options.allow_concurrent_memtable_write = false;
|
|
|
|
|
|
|
|
BlockBasedTableOptions table_options;
|
|
|
|
table_options.no_block_cache = true;
|
|
|
|
table_options.filter_policy.reset(NewBloomFilterPolicy(10));
|
|
|
|
table_options.whole_key_filtering = false;
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
|
|
|
|
|
|
|
|
// 11 RAND I/Os
|
|
|
|
DestroyAndReopen(options);
|
|
|
|
PrefixScanInit(this);
|
|
|
|
count = 0;
|
|
|
|
env_->random_read_counter_.Reset();
|
|
|
|
iter = db_->NewIterator(ReadOptions());
|
|
|
|
for (iter->Seek(prefix); iter->Valid(); iter->Next()) {
|
|
|
|
if (!iter->key().starts_with(prefix)) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
count++;
|
|
|
|
}
|
|
|
|
ASSERT_OK(iter->status());
|
|
|
|
delete iter;
|
|
|
|
ASSERT_EQ(count, 2);
|
|
|
|
ASSERT_EQ(env_->random_read_counter_.Read(), 2);
|
|
|
|
Close();
|
|
|
|
} // end of while
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(DBBloomFilterTest, OptimizeFiltersForHits) {
|
|
|
|
Options options = CurrentOptions();
|
|
|
|
options.write_buffer_size = 64 * 1024;
|
|
|
|
options.arena_block_size = 4 * 1024;
|
|
|
|
options.target_file_size_base = 64 * 1024;
|
|
|
|
options.level0_file_num_compaction_trigger = 2;
|
|
|
|
options.level0_slowdown_writes_trigger = 2;
|
|
|
|
options.level0_stop_writes_trigger = 4;
|
|
|
|
options.max_bytes_for_level_base = 256 * 1024;
|
|
|
|
options.max_write_buffer_number = 2;
|
|
|
|
options.max_background_compactions = 8;
|
|
|
|
options.max_background_flushes = 8;
|
|
|
|
options.compression = kNoCompression;
|
|
|
|
options.compaction_style = kCompactionStyleLevel;
|
|
|
|
options.level_compaction_dynamic_level_bytes = true;
|
|
|
|
BlockBasedTableOptions bbto;
|
|
|
|
bbto.cache_index_and_filter_blocks = true;
|
|
|
|
bbto.filter_policy.reset(NewBloomFilterPolicy(10, true));
|
|
|
|
bbto.whole_key_filtering = true;
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
|
|
|
|
options.optimize_filters_for_hits = true;
|
|
|
|
options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
|
|
|
|
get_perf_context()->Reset();
|
|
|
|
get_perf_context()->EnablePerLevelPerfContext();
|
|
|
|
CreateAndReopenWithCF({"mypikachu"}, options);
|
|
|
|
|
|
|
|
int numkeys = 200000;
|
|
|
|
|
|
|
|
// Generate randomly shuffled keys, so the updates are almost
|
|
|
|
// random.
|
|
|
|
std::vector<int> keys;
|
|
|
|
keys.reserve(numkeys);
|
|
|
|
for (int i = 0; i < numkeys; i += 2) {
|
|
|
|
keys.push_back(i);
|
|
|
|
}
|
|
|
|
RandomShuffle(std::begin(keys), std::end(keys));
|
|
|
|
int num_inserted = 0;
|
|
|
|
for (int key : keys) {
|
|
|
|
ASSERT_OK(Put(1, Key(key), "val"));
|
|
|
|
if (++num_inserted % 1000 == 0) {
|
|
|
|
ASSERT_OK(dbfull()->TEST_WaitForFlushMemTable());
|
|
|
|
ASSERT_OK(dbfull()->TEST_WaitForCompact());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
ASSERT_OK(Put(1, Key(0), "val"));
|
|
|
|
ASSERT_OK(Put(1, Key(numkeys), "val"));
|
|
|
|
ASSERT_OK(Flush(1));
|
|
|
|
ASSERT_OK(dbfull()->TEST_WaitForCompact());
|
|
|
|
|
|
|
|
if (NumTableFilesAtLevel(0, 1) == 0) {
|
|
|
|
// No Level 0 file. Create one.
|
|
|
|
ASSERT_OK(Put(1, Key(0), "val"));
|
|
|
|
ASSERT_OK(Put(1, Key(numkeys), "val"));
|
|
|
|
ASSERT_OK(Flush(1));
|
|
|
|
ASSERT_OK(dbfull()->TEST_WaitForCompact());
|
|
|
|
}
|
|
|
|
|
|
|
|
for (int i = 1; i < numkeys; i += 2) {
|
|
|
|
ASSERT_EQ(Get(1, Key(i)), "NOT_FOUND");
|
|
|
|
}
|
|
|
|
|
|
|
|
ASSERT_EQ(0, TestGetTickerCount(options, GET_HIT_L0));
|
|
|
|
ASSERT_EQ(0, TestGetTickerCount(options, GET_HIT_L1));
|
|
|
|
ASSERT_EQ(0, TestGetTickerCount(options, GET_HIT_L2_AND_UP));
|
|
|
|
|
|
|
|
// Now we have three sorted run, L0, L5 and L6 with most files in L6 have
|
|
|
|
// no bloom filter. Most keys be checked bloom filters twice.
|
|
|
|
ASSERT_GT(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 65000 * 2);
|
|
|
|
ASSERT_LT(TestGetTickerCount(options, BLOOM_FILTER_USEFUL), 120000 * 2);
|
|
|
|
uint64_t bloom_filter_useful_all_levels = 0;
|
|
|
|
for (auto& kv : (*(get_perf_context()->level_to_perf_context))) {
|
|
|
|
if (kv.second.bloom_filter_useful > 0) {
|
|
|
|
bloom_filter_useful_all_levels += kv.second.bloom_filter_useful;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
ASSERT_GT(bloom_filter_useful_all_levels, 65000 * 2);
|
|
|
|
ASSERT_LT(bloom_filter_useful_all_levels, 120000 * 2);
|
|
|
|
|
|
|
|
for (int i = 0; i < numkeys; i += 2) {
|
|
|
|
ASSERT_EQ(Get(1, Key(i)), "val");
|
|
|
|
}
|
|
|
|
|
|
|
|
// Part 2 (read path): rewrite last level with blooms, then verify they get
|
|
|
|
// cached only if !optimize_filters_for_hits
|
|
|
|
options.disable_auto_compactions = true;
|
|
|
|
options.num_levels = 9;
|
|
|
|
options.optimize_filters_for_hits = false;
|
|
|
|
options.statistics = CreateDBStatistics();
|
|
|
|
bbto.block_cache.reset();
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
|
|
|
|
|
|
|
|
ReopenWithColumnFamilies({"default", "mypikachu"}, options);
|
|
|
|
MoveFilesToLevel(7 /* level */, 1 /* column family index */);
|
|
|
|
|
|
|
|
std::string value = Get(1, Key(0));
|
|
|
|
uint64_t prev_cache_filter_hits =
|
|
|
|
TestGetTickerCount(options, BLOCK_CACHE_FILTER_HIT);
|
|
|
|
value = Get(1, Key(0));
|
|
|
|
ASSERT_EQ(prev_cache_filter_hits + 1,
|
|
|
|
TestGetTickerCount(options, BLOCK_CACHE_FILTER_HIT));
|
|
|
|
|
|
|
|
// Now that we know the filter blocks exist in the last level files, see if
|
|
|
|
// filter caching is skipped for this optimization
|
|
|
|
options.optimize_filters_for_hits = true;
|
|
|
|
options.statistics = CreateDBStatistics();
|
|
|
|
bbto.block_cache.reset();
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
|
|
|
|
|
|
|
|
ReopenWithColumnFamilies({"default", "mypikachu"}, options);
|
|
|
|
|
|
|
|
value = Get(1, Key(0));
|
|
|
|
ASSERT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_FILTER_MISS));
|
|
|
|
ASSERT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_FILTER_HIT));
|
|
|
|
ASSERT_EQ(2 /* index and data block */,
|
|
|
|
TestGetTickerCount(options, BLOCK_CACHE_ADD));
|
|
|
|
|
|
|
|
// Check filter block ignored for files preloaded during DB::Open()
|
|
|
|
options.max_open_files = -1;
|
|
|
|
options.statistics = CreateDBStatistics();
|
|
|
|
bbto.block_cache.reset();
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
|
|
|
|
|
|
|
|
ReopenWithColumnFamilies({"default", "mypikachu"}, options);
|
|
|
|
|
|
|
|
uint64_t prev_cache_filter_misses =
|
|
|
|
TestGetTickerCount(options, BLOCK_CACHE_FILTER_MISS);
|
|
|
|
prev_cache_filter_hits = TestGetTickerCount(options, BLOCK_CACHE_FILTER_HIT);
|
|
|
|
Get(1, Key(0));
|
|
|
|
ASSERT_EQ(prev_cache_filter_misses,
|
|
|
|
TestGetTickerCount(options, BLOCK_CACHE_FILTER_MISS));
|
|
|
|
ASSERT_EQ(prev_cache_filter_hits,
|
|
|
|
TestGetTickerCount(options, BLOCK_CACHE_FILTER_HIT));
|
|
|
|
|
|
|
|
// Check filter block ignored for file trivially-moved to bottom level
|
|
|
|
bbto.block_cache.reset();
|
|
|
|
options.max_open_files = 100; // setting > -1 makes it not preload all files
|
|
|
|
options.statistics = CreateDBStatistics();
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
|
|
|
|
|
|
|
|
ReopenWithColumnFamilies({"default", "mypikachu"}, options);
|
|
|
|
|
|
|
|
ASSERT_OK(Put(1, Key(numkeys + 1), "val"));
|
|
|
|
ASSERT_OK(Flush(1));
|
|
|
|
|
|
|
|
int32_t trivial_move = 0;
|
|
|
|
int32_t non_trivial_move = 0;
|
|
|
|
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
|
|
|
|
"DBImpl::BackgroundCompaction:TrivialMove",
|
|
|
|
[&](void* /*arg*/) { trivial_move++; });
|
|
|
|
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
|
|
|
|
"DBImpl::BackgroundCompaction:NonTrivial",
|
|
|
|
[&](void* /*arg*/) { non_trivial_move++; });
|
|
|
|
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
|
|
|
|
|
|
|
|
CompactRangeOptions compact_options;
|
|
|
|
compact_options.bottommost_level_compaction =
|
|
|
|
BottommostLevelCompaction::kSkip;
|
|
|
|
compact_options.change_level = true;
|
|
|
|
compact_options.target_level = 7;
|
|
|
|
ASSERT_TRUE(db_->CompactRange(compact_options, handles_[1], nullptr, nullptr)
|
|
|
|
.IsNotSupported());
|
|
|
|
|
|
|
|
ASSERT_EQ(trivial_move, 1);
|
|
|
|
ASSERT_EQ(non_trivial_move, 0);
|
|
|
|
|
|
|
|
prev_cache_filter_hits = TestGetTickerCount(options, BLOCK_CACHE_FILTER_HIT);
|
|
|
|
prev_cache_filter_misses =
|
|
|
|
TestGetTickerCount(options, BLOCK_CACHE_FILTER_MISS);
|
|
|
|
value = Get(1, Key(numkeys + 1));
|
|
|
|
ASSERT_EQ(prev_cache_filter_hits,
|
|
|
|
TestGetTickerCount(options, BLOCK_CACHE_FILTER_HIT));
|
|
|
|
ASSERT_EQ(prev_cache_filter_misses,
|
|
|
|
TestGetTickerCount(options, BLOCK_CACHE_FILTER_MISS));
|
|
|
|
|
|
|
|
// Check filter block not cached for iterator
|
|
|
|
bbto.block_cache.reset();
|
|
|
|
options.statistics = CreateDBStatistics();
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
|
|
|
|
|
|
|
|
ReopenWithColumnFamilies({"default", "mypikachu"}, options);
|
|
|
|
|
|
|
|
std::unique_ptr<Iterator> iter(db_->NewIterator(ReadOptions(), handles_[1]));
|
|
|
|
iter->SeekToFirst();
|
|
|
|
ASSERT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_FILTER_MISS));
|
|
|
|
ASSERT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_FILTER_HIT));
|
|
|
|
ASSERT_EQ(2 /* index and data block */,
|
|
|
|
TestGetTickerCount(options, BLOCK_CACHE_ADD));
|
|
|
|
get_perf_context()->Reset();
|
|
|
|
}
|
|
|
|
|
|
|
|
int CountIter(std::unique_ptr<Iterator>& iter, const Slice& key) {
|
|
|
|
int count = 0;
|
|
|
|
for (iter->Seek(key); iter->Valid(); iter->Next()) {
|
|
|
|
count++;
|
|
|
|
}
|
|
|
|
EXPECT_OK(iter->status());
|
|
|
|
return count;
|
|
|
|
}
|
|
|
|
|
|
|
|
// use iterate_upper_bound to hint compatiability of existing bloom filters.
|
|
|
|
// The BF is considered compatible if 1) upper bound and seek key transform
|
|
|
|
// into the same string, or 2) the transformed seek key is of the same length
|
|
|
|
// as the upper bound and two keys are adjacent according to the comparator.
|
|
|
|
TEST_F(DBBloomFilterTest, DynamicBloomFilterUpperBound) {
|
New Bloom filter implementation for full and partitioned filters (#6007)
Summary:
Adds an improved, replacement Bloom filter implementation (FastLocalBloom) for full and partitioned filters in the block-based table. This replacement is faster and more accurate, especially for high bits per key or millions of keys in a single filter.
Speed
The improved speed, at least on recent x86_64, comes from
* Using fastrange instead of modulo (%)
* Using our new hash function (XXH3 preview, added in a previous commit), which is much faster for large keys and only *slightly* slower on keys around 12 bytes if hashing the same size many thousands of times in a row.
* Optimizing the Bloom filter queries with AVX2 SIMD operations. (Added AVX2 to the USE_SSE=1 build.) Careful design was required to support (a) SIMD-optimized queries, (b) compatible non-SIMD code that's simple and efficient, (c) flexible choice of number of probes, and (d) essentially maximized accuracy for a cache-local Bloom filter. Probes are made eight at a time, so any number of probes up to 8 is the same speed, then up to 16, etc.
* Prefetching cache lines when building the filter. Although this optimization could be applied to the old structure as well, it seems to balance out the small added cost of accumulating 64 bit hashes for adding to the filter rather than 32 bit hashes.
Here's nominal speed data from filter_bench (200MB in filters, about 10k keys each, 10 bits filter data / key, 6 probes, avg key size 24 bytes, includes hashing time) on Skylake DE (relatively low clock speed):
$ ./filter_bench -quick -impl=2 -net_includes_hashing # New Bloom filter
Build avg ns/key: 47.7135
Mixed inside/outside queries...
Single filter net ns/op: 26.2825
Random filter net ns/op: 150.459
Average FP rate %: 0.954651
$ ./filter_bench -quick -impl=0 -net_includes_hashing # Old Bloom filter
Build avg ns/key: 47.2245
Mixed inside/outside queries...
Single filter net ns/op: 63.2978
Random filter net ns/op: 188.038
Average FP rate %: 1.13823
Similar build time but dramatically faster query times on hot data (63 ns to 26 ns), and somewhat faster on stale data (188 ns to 150 ns). Performance differences on batched and skewed query loads are between these extremes as expected.
The only other interesting thing about speed is "inside" (query key was added to filter) vs. "outside" (query key was not added to filter) query times. The non-SIMD implementations are substantially slower when most queries are "outside" vs. "inside". This goes against what one might expect or would have observed years ago, as "outside" queries only need about two probes on average, due to short-circuiting, while "inside" always have num_probes (say 6). The problem is probably the nastily unpredictable branch. The SIMD implementation has few branches (very predictable) and has pretty consistent running time regardless of query outcome.
Accuracy
The generally improved accuracy (re: Issue https://github.com/facebook/rocksdb/issues/5857) comes from a better design for probing indices
within a cache line (re: Issue https://github.com/facebook/rocksdb/issues/4120) and improved accuracy for millions of keys in a single filter from using a 64-bit hash function (XXH3p). Design details in code comments.
Accuracy data (generalizes, except old impl gets worse with millions of keys):
Memory bits per key: FP rate percent old impl -> FP rate percent new impl
6: 5.70953 -> 5.69888
8: 2.45766 -> 2.29709
10: 1.13977 -> 0.959254
12: 0.662498 -> 0.411593
16: 0.353023 -> 0.0873754
24: 0.261552 -> 0.0060971
50: 0.225453 -> ~0.00003 (less than 1 in a million queries are FP)
Fixes https://github.com/facebook/rocksdb/issues/5857
Fixes https://github.com/facebook/rocksdb/issues/4120
Unlike the old implementation, this implementation has a fixed cache line size (64 bytes). At 10 bits per key, the accuracy of this new implementation is very close to the old implementation with 128-byte cache line size. If there's sufficient demand, this implementation could be generalized.
Compatibility
Although old releases would see the new structure as corrupt filter data and read the table as if there's no filter, we've decided only to enable the new Bloom filter with new format_version=5. This provides a smooth path for automatic adoption over time, with an option for early opt-in.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6007
Test Plan: filter_bench has been used thoroughly to validate speed, accuracy, and correctness. Unit tests have been carefully updated to exercise new and old implementations, as well as the logic to select an implementation based on context (format_version).
Differential Revision: D18294749
Pulled By: pdillinger
fbshipit-source-id: d44c9db3696e4d0a17caaec47075b7755c262c5f
5 years ago
|
|
|
for (auto bfp_impl : BFP::kAllFixedImpls) {
|
|
|
|
int using_full_builder = bfp_impl != BFP::kDeprecatedBlock;
|
|
|
|
Options options;
|
|
|
|
options.create_if_missing = true;
|
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566)
Summary:
This PR does a few things:
1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation.
2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed:
- The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated
- The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory).
3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10).
I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged.
Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently.
Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566
Reviewed By: zhichao-cao
Differential Revision: D24408980
Pulled By: jay-zhuang
fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
4 years ago
|
|
|
options.env = CurrentOptions().env;
|
|
|
|
options.prefix_extractor.reset(NewCappedPrefixTransform(4));
|
|
|
|
options.disable_auto_compactions = true;
|
|
|
|
options.statistics = CreateDBStatistics();
|
|
|
|
// Enable prefix bloom for SST files
|
|
|
|
BlockBasedTableOptions table_options;
|
|
|
|
table_options.cache_index_and_filter_blocks = true;
|
|
|
|
table_options.filter_policy.reset(new BFP(10, bfp_impl));
|
|
|
|
table_options.index_shortening = BlockBasedTableOptions::
|
|
|
|
IndexShorteningMode::kShortenSeparatorsAndSuccessor;
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
|
|
|
|
DestroyAndReopen(options);
|
|
|
|
|
|
|
|
ASSERT_OK(Put("abcdxxx0", "val1"));
|
|
|
|
ASSERT_OK(Put("abcdxxx1", "val2"));
|
|
|
|
ASSERT_OK(Put("abcdxxx2", "val3"));
|
|
|
|
ASSERT_OK(Put("abcdxxx3", "val4"));
|
|
|
|
ASSERT_OK(dbfull()->Flush(FlushOptions()));
|
|
|
|
{
|
|
|
|
// prefix_extractor has not changed, BF will always be read
|
|
|
|
Slice upper_bound("abce");
|
|
|
|
ReadOptions read_options;
|
|
|
|
read_options.prefix_same_as_start = true;
|
|
|
|
read_options.iterate_upper_bound = &upper_bound;
|
|
|
|
std::unique_ptr<Iterator> iter(db_->NewIterator(read_options));
|
|
|
|
ASSERT_EQ(CountIter(iter, "abcd0000"), 4);
|
|
|
|
}
|
|
|
|
{
|
|
|
|
Slice upper_bound("abcdzzzz");
|
|
|
|
ReadOptions read_options;
|
|
|
|
read_options.prefix_same_as_start = true;
|
|
|
|
read_options.iterate_upper_bound = &upper_bound;
|
|
|
|
std::unique_ptr<Iterator> iter(db_->NewIterator(read_options));
|
|
|
|
ASSERT_EQ(CountIter(iter, "abcd0000"), 4);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED), 2);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_USEFUL), 0);
|
|
|
|
}
|
|
|
|
ASSERT_OK(dbfull()->SetOptions({{"prefix_extractor", "fixed:5"}}));
|
|
|
|
ASSERT_EQ(0, strcmp(dbfull()->GetOptions().prefix_extractor->Name(),
|
|
|
|
"rocksdb.FixedPrefix.5"));
|
|
|
|
{
|
|
|
|
// BF changed, [abcdxx00, abce) is a valid bound, will trigger BF read
|
|
|
|
Slice upper_bound("abce");
|
|
|
|
ReadOptions read_options;
|
|
|
|
read_options.prefix_same_as_start = true;
|
|
|
|
read_options.iterate_upper_bound = &upper_bound;
|
|
|
|
std::unique_ptr<Iterator> iter(db_->NewIterator(read_options));
|
|
|
|
ASSERT_EQ(CountIter(iter, "abcdxx00"), 4);
|
|
|
|
// should check bloom filter since upper bound meets requirement
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED),
|
|
|
|
2 + using_full_builder);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_USEFUL), 0);
|
|
|
|
}
|
|
|
|
{
|
|
|
|
// [abcdxx01, abcey) is not valid bound since upper bound is too long for
|
|
|
|
// the BF in SST (capped:4)
|
|
|
|
Slice upper_bound("abcey");
|
|
|
|
ReadOptions read_options;
|
|
|
|
read_options.prefix_same_as_start = true;
|
|
|
|
read_options.iterate_upper_bound = &upper_bound;
|
|
|
|
std::unique_ptr<Iterator> iter(db_->NewIterator(read_options));
|
|
|
|
ASSERT_EQ(CountIter(iter, "abcdxx01"), 4);
|
|
|
|
// should skip bloom filter since upper bound is too long
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED),
|
|
|
|
2 + using_full_builder);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_USEFUL), 0);
|
|
|
|
}
|
|
|
|
{
|
|
|
|
// [abcdxx02, abcdy) is a valid bound since the prefix is the same
|
|
|
|
Slice upper_bound("abcdy");
|
|
|
|
ReadOptions read_options;
|
|
|
|
read_options.prefix_same_as_start = true;
|
|
|
|
read_options.iterate_upper_bound = &upper_bound;
|
|
|
|
std::unique_ptr<Iterator> iter(db_->NewIterator(read_options));
|
|
|
|
ASSERT_EQ(CountIter(iter, "abcdxx02"), 4);
|
|
|
|
// should check bloom filter since upper bound matches transformed seek
|
|
|
|
// key
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED),
|
|
|
|
2 + using_full_builder * 2);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_USEFUL), 0);
|
|
|
|
}
|
|
|
|
{
|
|
|
|
// [aaaaaaaa, abce) is not a valid bound since 1) they don't share the
|
|
|
|
// same prefix, 2) the prefixes are not consecutive
|
|
|
|
Slice upper_bound("abce");
|
|
|
|
ReadOptions read_options;
|
|
|
|
read_options.prefix_same_as_start = true;
|
|
|
|
read_options.iterate_upper_bound = &upper_bound;
|
|
|
|
std::unique_ptr<Iterator> iter(db_->NewIterator(read_options));
|
|
|
|
ASSERT_EQ(CountIter(iter, "aaaaaaaa"), 0);
|
|
|
|
// should skip bloom filter since mismatch is found
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED),
|
|
|
|
2 + using_full_builder * 2);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_USEFUL), 0);
|
|
|
|
}
|
|
|
|
ASSERT_OK(dbfull()->SetOptions({{"prefix_extractor", "fixed:3"}}));
|
|
|
|
{
|
|
|
|
// [abc, abd) is not a valid bound since the upper bound is too short
|
|
|
|
// for BF (capped:4)
|
|
|
|
Slice upper_bound("abd");
|
|
|
|
ReadOptions read_options;
|
|
|
|
read_options.prefix_same_as_start = true;
|
|
|
|
read_options.iterate_upper_bound = &upper_bound;
|
|
|
|
std::unique_ptr<Iterator> iter(db_->NewIterator(read_options));
|
|
|
|
ASSERT_EQ(CountIter(iter, "abc"), 4);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED),
|
|
|
|
2 + using_full_builder * 2);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_USEFUL), 0);
|
|
|
|
}
|
|
|
|
ASSERT_OK(dbfull()->SetOptions({{"prefix_extractor", "capped:4"}}));
|
|
|
|
{
|
|
|
|
// set back to capped:4 and verify BF is always read
|
|
|
|
Slice upper_bound("abd");
|
|
|
|
ReadOptions read_options;
|
|
|
|
read_options.prefix_same_as_start = true;
|
|
|
|
read_options.iterate_upper_bound = &upper_bound;
|
|
|
|
std::unique_ptr<Iterator> iter(db_->NewIterator(read_options));
|
|
|
|
ASSERT_EQ(CountIter(iter, "abc"), 0);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED),
|
|
|
|
3 + using_full_builder * 2);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_USEFUL), 1);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Create multiple SST files each with a different prefix_extractor config,
|
|
|
|
// verify iterators can read all SST files using the latest config.
|
|
|
|
TEST_F(DBBloomFilterTest, DynamicBloomFilterMultipleSST) {
|
New Bloom filter implementation for full and partitioned filters (#6007)
Summary:
Adds an improved, replacement Bloom filter implementation (FastLocalBloom) for full and partitioned filters in the block-based table. This replacement is faster and more accurate, especially for high bits per key or millions of keys in a single filter.
Speed
The improved speed, at least on recent x86_64, comes from
* Using fastrange instead of modulo (%)
* Using our new hash function (XXH3 preview, added in a previous commit), which is much faster for large keys and only *slightly* slower on keys around 12 bytes if hashing the same size many thousands of times in a row.
* Optimizing the Bloom filter queries with AVX2 SIMD operations. (Added AVX2 to the USE_SSE=1 build.) Careful design was required to support (a) SIMD-optimized queries, (b) compatible non-SIMD code that's simple and efficient, (c) flexible choice of number of probes, and (d) essentially maximized accuracy for a cache-local Bloom filter. Probes are made eight at a time, so any number of probes up to 8 is the same speed, then up to 16, etc.
* Prefetching cache lines when building the filter. Although this optimization could be applied to the old structure as well, it seems to balance out the small added cost of accumulating 64 bit hashes for adding to the filter rather than 32 bit hashes.
Here's nominal speed data from filter_bench (200MB in filters, about 10k keys each, 10 bits filter data / key, 6 probes, avg key size 24 bytes, includes hashing time) on Skylake DE (relatively low clock speed):
$ ./filter_bench -quick -impl=2 -net_includes_hashing # New Bloom filter
Build avg ns/key: 47.7135
Mixed inside/outside queries...
Single filter net ns/op: 26.2825
Random filter net ns/op: 150.459
Average FP rate %: 0.954651
$ ./filter_bench -quick -impl=0 -net_includes_hashing # Old Bloom filter
Build avg ns/key: 47.2245
Mixed inside/outside queries...
Single filter net ns/op: 63.2978
Random filter net ns/op: 188.038
Average FP rate %: 1.13823
Similar build time but dramatically faster query times on hot data (63 ns to 26 ns), and somewhat faster on stale data (188 ns to 150 ns). Performance differences on batched and skewed query loads are between these extremes as expected.
The only other interesting thing about speed is "inside" (query key was added to filter) vs. "outside" (query key was not added to filter) query times. The non-SIMD implementations are substantially slower when most queries are "outside" vs. "inside". This goes against what one might expect or would have observed years ago, as "outside" queries only need about two probes on average, due to short-circuiting, while "inside" always have num_probes (say 6). The problem is probably the nastily unpredictable branch. The SIMD implementation has few branches (very predictable) and has pretty consistent running time regardless of query outcome.
Accuracy
The generally improved accuracy (re: Issue https://github.com/facebook/rocksdb/issues/5857) comes from a better design for probing indices
within a cache line (re: Issue https://github.com/facebook/rocksdb/issues/4120) and improved accuracy for millions of keys in a single filter from using a 64-bit hash function (XXH3p). Design details in code comments.
Accuracy data (generalizes, except old impl gets worse with millions of keys):
Memory bits per key: FP rate percent old impl -> FP rate percent new impl
6: 5.70953 -> 5.69888
8: 2.45766 -> 2.29709
10: 1.13977 -> 0.959254
12: 0.662498 -> 0.411593
16: 0.353023 -> 0.0873754
24: 0.261552 -> 0.0060971
50: 0.225453 -> ~0.00003 (less than 1 in a million queries are FP)
Fixes https://github.com/facebook/rocksdb/issues/5857
Fixes https://github.com/facebook/rocksdb/issues/4120
Unlike the old implementation, this implementation has a fixed cache line size (64 bytes). At 10 bits per key, the accuracy of this new implementation is very close to the old implementation with 128-byte cache line size. If there's sufficient demand, this implementation could be generalized.
Compatibility
Although old releases would see the new structure as corrupt filter data and read the table as if there's no filter, we've decided only to enable the new Bloom filter with new format_version=5. This provides a smooth path for automatic adoption over time, with an option for early opt-in.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6007
Test Plan: filter_bench has been used thoroughly to validate speed, accuracy, and correctness. Unit tests have been carefully updated to exercise new and old implementations, as well as the logic to select an implementation based on context (format_version).
Differential Revision: D18294749
Pulled By: pdillinger
fbshipit-source-id: d44c9db3696e4d0a17caaec47075b7755c262c5f
5 years ago
|
|
|
for (auto bfp_impl : BFP::kAllFixedImpls) {
|
|
|
|
int using_full_builder = bfp_impl != BFP::kDeprecatedBlock;
|
|
|
|
Options options;
|
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566)
Summary:
This PR does a few things:
1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation.
2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed:
- The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated
- The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory).
3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10).
I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged.
Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently.
Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566
Reviewed By: zhichao-cao
Differential Revision: D24408980
Pulled By: jay-zhuang
fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
4 years ago
|
|
|
options.env = CurrentOptions().env;
|
|
|
|
options.create_if_missing = true;
|
|
|
|
options.prefix_extractor.reset(NewFixedPrefixTransform(1));
|
|
|
|
options.disable_auto_compactions = true;
|
|
|
|
options.statistics = CreateDBStatistics();
|
|
|
|
// Enable prefix bloom for SST files
|
|
|
|
BlockBasedTableOptions table_options;
|
|
|
|
table_options.filter_policy.reset(new BFP(10, bfp_impl));
|
|
|
|
table_options.cache_index_and_filter_blocks = true;
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
|
|
|
|
DestroyAndReopen(options);
|
|
|
|
|
|
|
|
Slice upper_bound("foz90000");
|
|
|
|
ReadOptions read_options;
|
|
|
|
read_options.prefix_same_as_start = true;
|
|
|
|
|
|
|
|
// first SST with fixed:1 BF
|
|
|
|
ASSERT_OK(Put("foo2", "bar2"));
|
|
|
|
ASSERT_OK(Put("foo", "bar"));
|
|
|
|
ASSERT_OK(Put("foq1", "bar1"));
|
|
|
|
ASSERT_OK(Put("fpa", "0"));
|
|
|
|
dbfull()->Flush(FlushOptions());
|
|
|
|
std::unique_ptr<Iterator> iter_old(db_->NewIterator(read_options));
|
|
|
|
ASSERT_EQ(CountIter(iter_old, "foo"), 4);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED), 1);
|
|
|
|
|
|
|
|
ASSERT_OK(dbfull()->SetOptions({{"prefix_extractor", "capped:3"}}));
|
|
|
|
ASSERT_EQ(0, strcmp(dbfull()->GetOptions().prefix_extractor->Name(),
|
|
|
|
"rocksdb.CappedPrefix.3"));
|
|
|
|
read_options.iterate_upper_bound = &upper_bound;
|
|
|
|
std::unique_ptr<Iterator> iter(db_->NewIterator(read_options));
|
|
|
|
ASSERT_EQ(CountIter(iter, "foo"), 2);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED),
|
|
|
|
1 + using_full_builder);
|
|
|
|
ASSERT_EQ(CountIter(iter, "gpk"), 0);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED),
|
|
|
|
1 + using_full_builder);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_USEFUL), 0);
|
|
|
|
|
|
|
|
// second SST with capped:3 BF
|
|
|
|
ASSERT_OK(Put("foo3", "bar3"));
|
|
|
|
ASSERT_OK(Put("foo4", "bar4"));
|
|
|
|
ASSERT_OK(Put("foq5", "bar5"));
|
|
|
|
ASSERT_OK(Put("fpb", "1"));
|
|
|
|
ASSERT_OK(dbfull()->Flush(FlushOptions()));
|
|
|
|
{
|
|
|
|
// BF is cappped:3 now
|
|
|
|
std::unique_ptr<Iterator> iter_tmp(db_->NewIterator(read_options));
|
|
|
|
ASSERT_EQ(CountIter(iter_tmp, "foo"), 4);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED),
|
|
|
|
2 + using_full_builder * 2);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_USEFUL), 0);
|
|
|
|
ASSERT_EQ(CountIter(iter_tmp, "gpk"), 0);
|
|
|
|
// both counters are incremented because BF is "not changed" for 1 of the
|
|
|
|
// 2 SST files, so filter is checked once and found no match.
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED),
|
|
|
|
3 + using_full_builder * 2);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_USEFUL), 1);
|
|
|
|
}
|
|
|
|
|
|
|
|
ASSERT_OK(dbfull()->SetOptions({{"prefix_extractor", "fixed:2"}}));
|
|
|
|
ASSERT_EQ(0, strcmp(dbfull()->GetOptions().prefix_extractor->Name(),
|
|
|
|
"rocksdb.FixedPrefix.2"));
|
|
|
|
// third SST with fixed:2 BF
|
|
|
|
ASSERT_OK(Put("foo6", "bar6"));
|
|
|
|
ASSERT_OK(Put("foo7", "bar7"));
|
|
|
|
ASSERT_OK(Put("foq8", "bar8"));
|
|
|
|
ASSERT_OK(Put("fpc", "2"));
|
|
|
|
ASSERT_OK(dbfull()->Flush(FlushOptions()));
|
|
|
|
{
|
|
|
|
// BF is fixed:2 now
|
|
|
|
std::unique_ptr<Iterator> iter_tmp(db_->NewIterator(read_options));
|
|
|
|
ASSERT_EQ(CountIter(iter_tmp, "foo"), 9);
|
|
|
|
// the first and last BF are checked
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED),
|
|
|
|
4 + using_full_builder * 3);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_USEFUL), 1);
|
|
|
|
ASSERT_EQ(CountIter(iter_tmp, "gpk"), 0);
|
|
|
|
// only last BF is checked and not found
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED),
|
|
|
|
5 + using_full_builder * 3);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_USEFUL), 2);
|
|
|
|
}
|
|
|
|
|
|
|
|
// iter_old can only see the first SST, so checked plus 1
|
|
|
|
ASSERT_EQ(CountIter(iter_old, "foo"), 4);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED),
|
|
|
|
6 + using_full_builder * 3);
|
|
|
|
// iter was created after the first setoptions call so only full filter
|
|
|
|
// will check the filter
|
|
|
|
ASSERT_EQ(CountIter(iter, "foo"), 2);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED),
|
|
|
|
6 + using_full_builder * 4);
|
|
|
|
|
|
|
|
{
|
|
|
|
// keys in all three SSTs are visible to iterator
|
|
|
|
// The range of [foo, foz90000] is compatible with (fixed:1) and (fixed:2)
|
|
|
|
// so +2 for checked counter
|
|
|
|
std::unique_ptr<Iterator> iter_all(db_->NewIterator(read_options));
|
|
|
|
ASSERT_EQ(CountIter(iter_all, "foo"), 9);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED),
|
|
|
|
7 + using_full_builder * 5);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_USEFUL), 2);
|
|
|
|
ASSERT_EQ(CountIter(iter_all, "gpk"), 0);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED),
|
|
|
|
8 + using_full_builder * 5);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_USEFUL), 3);
|
|
|
|
}
|
|
|
|
ASSERT_OK(dbfull()->SetOptions({{"prefix_extractor", "capped:3"}}));
|
|
|
|
ASSERT_EQ(0, strcmp(dbfull()->GetOptions().prefix_extractor->Name(),
|
|
|
|
"rocksdb.CappedPrefix.3"));
|
|
|
|
{
|
|
|
|
std::unique_ptr<Iterator> iter_all(db_->NewIterator(read_options));
|
|
|
|
ASSERT_EQ(CountIter(iter_all, "foo"), 6);
|
|
|
|
// all three SST are checked because the current options has the same as
|
|
|
|
// the remaining SST (capped:3)
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED),
|
|
|
|
9 + using_full_builder * 7);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_USEFUL), 3);
|
|
|
|
ASSERT_EQ(CountIter(iter_all, "gpk"), 0);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED),
|
|
|
|
10 + using_full_builder * 7);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_USEFUL), 4);
|
|
|
|
}
|
|
|
|
// TODO(Zhongyi): Maybe also need to add Get calls to test point look up?
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Create a new column family in a running DB, change prefix_extractor
|
|
|
|
// dynamically, verify the iterator created on the new column family behaves
|
|
|
|
// as expected
|
|
|
|
TEST_F(DBBloomFilterTest, DynamicBloomFilterNewColumnFamily) {
|
|
|
|
int iteration = 0;
|
New Bloom filter implementation for full and partitioned filters (#6007)
Summary:
Adds an improved, replacement Bloom filter implementation (FastLocalBloom) for full and partitioned filters in the block-based table. This replacement is faster and more accurate, especially for high bits per key or millions of keys in a single filter.
Speed
The improved speed, at least on recent x86_64, comes from
* Using fastrange instead of modulo (%)
* Using our new hash function (XXH3 preview, added in a previous commit), which is much faster for large keys and only *slightly* slower on keys around 12 bytes if hashing the same size many thousands of times in a row.
* Optimizing the Bloom filter queries with AVX2 SIMD operations. (Added AVX2 to the USE_SSE=1 build.) Careful design was required to support (a) SIMD-optimized queries, (b) compatible non-SIMD code that's simple and efficient, (c) flexible choice of number of probes, and (d) essentially maximized accuracy for a cache-local Bloom filter. Probes are made eight at a time, so any number of probes up to 8 is the same speed, then up to 16, etc.
* Prefetching cache lines when building the filter. Although this optimization could be applied to the old structure as well, it seems to balance out the small added cost of accumulating 64 bit hashes for adding to the filter rather than 32 bit hashes.
Here's nominal speed data from filter_bench (200MB in filters, about 10k keys each, 10 bits filter data / key, 6 probes, avg key size 24 bytes, includes hashing time) on Skylake DE (relatively low clock speed):
$ ./filter_bench -quick -impl=2 -net_includes_hashing # New Bloom filter
Build avg ns/key: 47.7135
Mixed inside/outside queries...
Single filter net ns/op: 26.2825
Random filter net ns/op: 150.459
Average FP rate %: 0.954651
$ ./filter_bench -quick -impl=0 -net_includes_hashing # Old Bloom filter
Build avg ns/key: 47.2245
Mixed inside/outside queries...
Single filter net ns/op: 63.2978
Random filter net ns/op: 188.038
Average FP rate %: 1.13823
Similar build time but dramatically faster query times on hot data (63 ns to 26 ns), and somewhat faster on stale data (188 ns to 150 ns). Performance differences on batched and skewed query loads are between these extremes as expected.
The only other interesting thing about speed is "inside" (query key was added to filter) vs. "outside" (query key was not added to filter) query times. The non-SIMD implementations are substantially slower when most queries are "outside" vs. "inside". This goes against what one might expect or would have observed years ago, as "outside" queries only need about two probes on average, due to short-circuiting, while "inside" always have num_probes (say 6). The problem is probably the nastily unpredictable branch. The SIMD implementation has few branches (very predictable) and has pretty consistent running time regardless of query outcome.
Accuracy
The generally improved accuracy (re: Issue https://github.com/facebook/rocksdb/issues/5857) comes from a better design for probing indices
within a cache line (re: Issue https://github.com/facebook/rocksdb/issues/4120) and improved accuracy for millions of keys in a single filter from using a 64-bit hash function (XXH3p). Design details in code comments.
Accuracy data (generalizes, except old impl gets worse with millions of keys):
Memory bits per key: FP rate percent old impl -> FP rate percent new impl
6: 5.70953 -> 5.69888
8: 2.45766 -> 2.29709
10: 1.13977 -> 0.959254
12: 0.662498 -> 0.411593
16: 0.353023 -> 0.0873754
24: 0.261552 -> 0.0060971
50: 0.225453 -> ~0.00003 (less than 1 in a million queries are FP)
Fixes https://github.com/facebook/rocksdb/issues/5857
Fixes https://github.com/facebook/rocksdb/issues/4120
Unlike the old implementation, this implementation has a fixed cache line size (64 bytes). At 10 bits per key, the accuracy of this new implementation is very close to the old implementation with 128-byte cache line size. If there's sufficient demand, this implementation could be generalized.
Compatibility
Although old releases would see the new structure as corrupt filter data and read the table as if there's no filter, we've decided only to enable the new Bloom filter with new format_version=5. This provides a smooth path for automatic adoption over time, with an option for early opt-in.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6007
Test Plan: filter_bench has been used thoroughly to validate speed, accuracy, and correctness. Unit tests have been carefully updated to exercise new and old implementations, as well as the logic to select an implementation based on context (format_version).
Differential Revision: D18294749
Pulled By: pdillinger
fbshipit-source-id: d44c9db3696e4d0a17caaec47075b7755c262c5f
5 years ago
|
|
|
for (auto bfp_impl : BFP::kAllFixedImpls) {
|
|
|
|
Options options = CurrentOptions();
|
|
|
|
options.create_if_missing = true;
|
|
|
|
options.prefix_extractor.reset(NewFixedPrefixTransform(1));
|
|
|
|
options.disable_auto_compactions = true;
|
|
|
|
options.statistics = CreateDBStatistics();
|
|
|
|
// Enable prefix bloom for SST files
|
|
|
|
BlockBasedTableOptions table_options;
|
|
|
|
table_options.cache_index_and_filter_blocks = true;
|
|
|
|
table_options.filter_policy.reset(new BFP(10, bfp_impl));
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
|
|
|
|
CreateAndReopenWithCF({"pikachu" + std::to_string(iteration)}, options);
|
|
|
|
ReadOptions read_options;
|
|
|
|
read_options.prefix_same_as_start = true;
|
|
|
|
// create a new CF and set prefix_extractor dynamically
|
|
|
|
options.prefix_extractor.reset(NewCappedPrefixTransform(3));
|
|
|
|
CreateColumnFamilies({"ramen_dojo_" + std::to_string(iteration)}, options);
|
|
|
|
ASSERT_EQ(0,
|
|
|
|
strcmp(dbfull()->GetOptions(handles_[2]).prefix_extractor->Name(),
|
|
|
|
"rocksdb.CappedPrefix.3"));
|
|
|
|
ASSERT_OK(Put(2, "foo3", "bar3"));
|
|
|
|
ASSERT_OK(Put(2, "foo4", "bar4"));
|
|
|
|
ASSERT_OK(Put(2, "foo5", "bar5"));
|
|
|
|
ASSERT_OK(Put(2, "foq6", "bar6"));
|
|
|
|
ASSERT_OK(Put(2, "fpq7", "bar7"));
|
|
|
|
dbfull()->Flush(FlushOptions());
|
|
|
|
{
|
|
|
|
std::unique_ptr<Iterator> iter(
|
|
|
|
db_->NewIterator(read_options, handles_[2]));
|
|
|
|
ASSERT_EQ(CountIter(iter, "foo"), 3);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED), 0);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_USEFUL), 0);
|
|
|
|
}
|
|
|
|
ASSERT_OK(
|
|
|
|
dbfull()->SetOptions(handles_[2], {{"prefix_extractor", "fixed:2"}}));
|
|
|
|
ASSERT_EQ(0,
|
|
|
|
strcmp(dbfull()->GetOptions(handles_[2]).prefix_extractor->Name(),
|
|
|
|
"rocksdb.FixedPrefix.2"));
|
|
|
|
{
|
|
|
|
std::unique_ptr<Iterator> iter(
|
|
|
|
db_->NewIterator(read_options, handles_[2]));
|
|
|
|
ASSERT_EQ(CountIter(iter, "foo"), 4);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED), 0);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_USEFUL), 0);
|
|
|
|
}
|
|
|
|
ASSERT_OK(dbfull()->DropColumnFamily(handles_[2]));
|
|
|
|
ASSERT_OK(dbfull()->DestroyColumnFamilyHandle(handles_[2]));
|
|
|
|
handles_[2] = nullptr;
|
|
|
|
ASSERT_OK(dbfull()->DropColumnFamily(handles_[1]));
|
|
|
|
ASSERT_OK(dbfull()->DestroyColumnFamilyHandle(handles_[1]));
|
|
|
|
handles_[1] = nullptr;
|
|
|
|
iteration++;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Verify it's possible to change prefix_extractor at runtime and iterators
|
|
|
|
// behaves as expected
|
|
|
|
TEST_F(DBBloomFilterTest, DynamicBloomFilterOptions) {
|
New Bloom filter implementation for full and partitioned filters (#6007)
Summary:
Adds an improved, replacement Bloom filter implementation (FastLocalBloom) for full and partitioned filters in the block-based table. This replacement is faster and more accurate, especially for high bits per key or millions of keys in a single filter.
Speed
The improved speed, at least on recent x86_64, comes from
* Using fastrange instead of modulo (%)
* Using our new hash function (XXH3 preview, added in a previous commit), which is much faster for large keys and only *slightly* slower on keys around 12 bytes if hashing the same size many thousands of times in a row.
* Optimizing the Bloom filter queries with AVX2 SIMD operations. (Added AVX2 to the USE_SSE=1 build.) Careful design was required to support (a) SIMD-optimized queries, (b) compatible non-SIMD code that's simple and efficient, (c) flexible choice of number of probes, and (d) essentially maximized accuracy for a cache-local Bloom filter. Probes are made eight at a time, so any number of probes up to 8 is the same speed, then up to 16, etc.
* Prefetching cache lines when building the filter. Although this optimization could be applied to the old structure as well, it seems to balance out the small added cost of accumulating 64 bit hashes for adding to the filter rather than 32 bit hashes.
Here's nominal speed data from filter_bench (200MB in filters, about 10k keys each, 10 bits filter data / key, 6 probes, avg key size 24 bytes, includes hashing time) on Skylake DE (relatively low clock speed):
$ ./filter_bench -quick -impl=2 -net_includes_hashing # New Bloom filter
Build avg ns/key: 47.7135
Mixed inside/outside queries...
Single filter net ns/op: 26.2825
Random filter net ns/op: 150.459
Average FP rate %: 0.954651
$ ./filter_bench -quick -impl=0 -net_includes_hashing # Old Bloom filter
Build avg ns/key: 47.2245
Mixed inside/outside queries...
Single filter net ns/op: 63.2978
Random filter net ns/op: 188.038
Average FP rate %: 1.13823
Similar build time but dramatically faster query times on hot data (63 ns to 26 ns), and somewhat faster on stale data (188 ns to 150 ns). Performance differences on batched and skewed query loads are between these extremes as expected.
The only other interesting thing about speed is "inside" (query key was added to filter) vs. "outside" (query key was not added to filter) query times. The non-SIMD implementations are substantially slower when most queries are "outside" vs. "inside". This goes against what one might expect or would have observed years ago, as "outside" queries only need about two probes on average, due to short-circuiting, while "inside" always have num_probes (say 6). The problem is probably the nastily unpredictable branch. The SIMD implementation has few branches (very predictable) and has pretty consistent running time regardless of query outcome.
Accuracy
The generally improved accuracy (re: Issue https://github.com/facebook/rocksdb/issues/5857) comes from a better design for probing indices
within a cache line (re: Issue https://github.com/facebook/rocksdb/issues/4120) and improved accuracy for millions of keys in a single filter from using a 64-bit hash function (XXH3p). Design details in code comments.
Accuracy data (generalizes, except old impl gets worse with millions of keys):
Memory bits per key: FP rate percent old impl -> FP rate percent new impl
6: 5.70953 -> 5.69888
8: 2.45766 -> 2.29709
10: 1.13977 -> 0.959254
12: 0.662498 -> 0.411593
16: 0.353023 -> 0.0873754
24: 0.261552 -> 0.0060971
50: 0.225453 -> ~0.00003 (less than 1 in a million queries are FP)
Fixes https://github.com/facebook/rocksdb/issues/5857
Fixes https://github.com/facebook/rocksdb/issues/4120
Unlike the old implementation, this implementation has a fixed cache line size (64 bytes). At 10 bits per key, the accuracy of this new implementation is very close to the old implementation with 128-byte cache line size. If there's sufficient demand, this implementation could be generalized.
Compatibility
Although old releases would see the new structure as corrupt filter data and read the table as if there's no filter, we've decided only to enable the new Bloom filter with new format_version=5. This provides a smooth path for automatic adoption over time, with an option for early opt-in.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6007
Test Plan: filter_bench has been used thoroughly to validate speed, accuracy, and correctness. Unit tests have been carefully updated to exercise new and old implementations, as well as the logic to select an implementation based on context (format_version).
Differential Revision: D18294749
Pulled By: pdillinger
fbshipit-source-id: d44c9db3696e4d0a17caaec47075b7755c262c5f
5 years ago
|
|
|
for (auto bfp_impl : BFP::kAllFixedImpls) {
|
|
|
|
Options options;
|
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566)
Summary:
This PR does a few things:
1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation.
2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed:
- The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated
- The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory).
3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10).
I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged.
Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently.
Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566
Reviewed By: zhichao-cao
Differential Revision: D24408980
Pulled By: jay-zhuang
fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
4 years ago
|
|
|
options.env = CurrentOptions().env;
|
|
|
|
options.create_if_missing = true;
|
|
|
|
options.prefix_extractor.reset(NewFixedPrefixTransform(1));
|
|
|
|
options.disable_auto_compactions = true;
|
|
|
|
options.statistics = CreateDBStatistics();
|
|
|
|
// Enable prefix bloom for SST files
|
|
|
|
BlockBasedTableOptions table_options;
|
|
|
|
table_options.cache_index_and_filter_blocks = true;
|
|
|
|
table_options.filter_policy.reset(new BFP(10, bfp_impl));
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
|
|
|
|
DestroyAndReopen(options);
|
|
|
|
|
|
|
|
ASSERT_OK(Put("foo2", "bar2"));
|
|
|
|
ASSERT_OK(Put("foo", "bar"));
|
|
|
|
ASSERT_OK(Put("foo1", "bar1"));
|
|
|
|
ASSERT_OK(Put("fpa", "0"));
|
|
|
|
dbfull()->Flush(FlushOptions());
|
|
|
|
ASSERT_OK(Put("foo3", "bar3"));
|
|
|
|
ASSERT_OK(Put("foo4", "bar4"));
|
|
|
|
ASSERT_OK(Put("foo5", "bar5"));
|
|
|
|
ASSERT_OK(Put("fpb", "1"));
|
|
|
|
dbfull()->Flush(FlushOptions());
|
|
|
|
ASSERT_OK(Put("foo6", "bar6"));
|
|
|
|
ASSERT_OK(Put("foo7", "bar7"));
|
|
|
|
ASSERT_OK(Put("foo8", "bar8"));
|
|
|
|
ASSERT_OK(Put("fpc", "2"));
|
|
|
|
dbfull()->Flush(FlushOptions());
|
|
|
|
|
|
|
|
ReadOptions read_options;
|
|
|
|
read_options.prefix_same_as_start = true;
|
|
|
|
{
|
|
|
|
std::unique_ptr<Iterator> iter(db_->NewIterator(read_options));
|
|
|
|
ASSERT_EQ(CountIter(iter, "foo"), 12);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED), 3);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_USEFUL), 0);
|
|
|
|
}
|
|
|
|
std::unique_ptr<Iterator> iter_old(db_->NewIterator(read_options));
|
|
|
|
ASSERT_EQ(CountIter(iter_old, "foo"), 12);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED), 6);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_USEFUL), 0);
|
|
|
|
|
|
|
|
ASSERT_OK(dbfull()->SetOptions({{"prefix_extractor", "capped:3"}}));
|
|
|
|
ASSERT_EQ(0, strcmp(dbfull()->GetOptions().prefix_extractor->Name(),
|
|
|
|
"rocksdb.CappedPrefix.3"));
|
|
|
|
{
|
|
|
|
std::unique_ptr<Iterator> iter(db_->NewIterator(read_options));
|
|
|
|
// "fp*" should be skipped
|
|
|
|
ASSERT_EQ(CountIter(iter, "foo"), 9);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED), 6);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_USEFUL), 0);
|
|
|
|
}
|
|
|
|
|
|
|
|
// iterator created before should not be affected and see all keys
|
|
|
|
ASSERT_EQ(CountIter(iter_old, "foo"), 12);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED), 9);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_USEFUL), 0);
|
|
|
|
ASSERT_EQ(CountIter(iter_old, "abc"), 0);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED), 12);
|
|
|
|
ASSERT_EQ(TestGetTickerCount(options, BLOOM_FILTER_PREFIX_USEFUL), 3);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Fix a bug for SeekForPrev with partitioned filter and prefix (#8137)
Summary:
According to https://github.com/facebook/rocksdb/issues/5907, each filter partition "should include the bloom of the prefix of the last
key in the previous partition" so that SeekForPrev() in prefix mode can return correct result.
The prefix of the last key in the previous partition does not necessarily have the same prefix
as the first key in the current partition. Regardless of the first key in current partition, the
prefix of the last key in the previous partition should be added. The existing code, however,
does not follow this. Furthermore, there is another issue: when finishing current filter partition,
`FullFilterBlockBuilder::AddPrefix()` is called for the first key in next filter partition, which effectively
overwrites `last_prefix_str_` prematurely. Consequently, when the filter block builder proceeds
to the next partition, `last_prefix_str_` will be the prefix of its first key, leaving no way of adding
the bloom of the prefix of the last key of the previous partition.
Prefix extractor is FixedLength.2.
```
[ filter part 1 ] [ filter part 2 ]
abc d
```
When SeekForPrev("abcd"), checking the filter partition will land on filter part 2 because "abcd" > "abc"
but smaller than "d".
If the filter in filter part 2 happens to return false for the test for "ab", then SeekForPrev("abcd") will build
incorrect iterator tree in non-total-order mode.
Also fix a unit test which starts to fail following this PR. `InDomain` should not fail due to assertion
error when checking on an arbitrary key.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8137
Test Plan:
```
make check
```
Without this fix, the following command will fail pretty soon.
```
./db_stress --acquire_snapshot_one_in=10000 --avoid_flush_during_recovery=0 \
--avoid_unnecessary_blocking_io=0 --backup_max_size=104857600 --backup_one_in=0 \
--batch_protection_bytes_per_key=0 --block_size=16384 --bloom_bits=17 \
--bottommost_compression_type=disable --cache_index_and_filter_blocks=1 --cache_size=1048576 \
--checkpoint_one_in=0 --checksum_type=kxxHash64 --clear_column_family_one_in=0 \
--compact_files_one_in=1000000 --compact_range_one_in=1000000 --compaction_ttl=0 \
--compression_max_dict_buffer_bytes=0 --compression_max_dict_bytes=0 \
--compression_parallel_threads=1 --compression_type=zstd --compression_zstd_max_train_bytes=0 \
--continuous_verification_interval=0 --db=/dev/shm/rocksdb/rocksdb_crashtest_whitebox \
--db_write_buffer_size=8388608 --delpercent=5 --delrangepercent=0 --destroy_db_initially=0 --enable_blob_files=0 \
--enable_compaction_filter=0 --enable_pipelined_write=1 --file_checksum_impl=big --flush_one_in=1000000 \
--format_version=5 --get_current_wal_file_one_in=0 --get_live_files_one_in=1000000 --get_property_one_in=1000000 \
--get_sorted_wal_files_one_in=0 --index_block_restart_interval=4 --index_type=2 --ingest_external_file_one_in=0 \
--iterpercent=10 --key_len_percent_dist=1,30,69 --level_compaction_dynamic_level_bytes=True \
--log2_keys_per_lock=10 --long_running_snapshots=1 --mark_for_compaction_one_file_in=0 \
--max_background_compactions=20 --max_bytes_for_level_base=10485760 --max_key=100000000 --max_key_len=3 \
--max_manifest_file_size=1073741824 --max_write_batch_group_size_bytes=16777216 --max_write_buffer_number=3 \
--max_write_buffer_size_to_maintain=8388608 --memtablerep=skip_list --mmap_read=1 --mock_direct_io=False \
--nooverwritepercent=0 --open_files=500000 --ops_per_thread=20000000 --optimize_filters_for_memory=0 --paranoid_file_checks=1 --partition_filters=1 --partition_pinning=0 --pause_background_one_in=1000000 \
--periodic_compaction_seconds=0 --prefixpercent=5 --progress_reports=0 --read_fault_one_in=0 --read_only=0 \
--readpercent=45 --recycle_log_file_num=0 --reopen=20 --secondary_catch_up_one_in=0 \
--snapshot_hold_ops=100000 --sst_file_manager_bytes_per_sec=104857600 \
--sst_file_manager_bytes_per_truncate=0 --subcompactions=2 --sync=0 --sync_fault_injection=False \
--target_file_size_base=2097152 --target_file_size_multiplier=2 --test_batches_snapshots=0 --test_cf_consistency=0 \
--top_level_index_pinning=0 --unpartitioned_pinning=1 --use_blob_db=0 --use_block_based_filter=0 \
--use_direct_io_for_flush_and_compaction=0 --use_direct_reads=0 --use_full_merge_v1=0 --use_merge=0 \
--use_multiget=0 --use_ribbon_filter=0 --use_txn=0 --user_timestamp_size=8 --verify_checksum=1 \
--verify_checksum_one_in=1000000 --verify_db_one_in=100000 --write_buffer_size=4194304 \
--write_dbid_to_manifest=1 --writepercent=35
```
Reviewed By: pdillinger
Differential Revision: D27553054
Pulled By: riversand963
fbshipit-source-id: 60e391e4a2d8d98a9a3172ec5d6176b90ec3de98
4 years ago
|
|
|
TEST_F(DBBloomFilterTest, SeekForPrevWithPartitionedFilters) {
|
|
|
|
Options options = CurrentOptions();
|
|
|
|
constexpr size_t kNumKeys = 10000;
|
|
|
|
static_assert(kNumKeys <= 10000, "kNumKeys have to be <= 10000");
|
|
|
|
options.memtable_factory.reset(new SpecialSkipListFactory(kNumKeys + 10));
|
|
|
|
options.create_if_missing = true;
|
|
|
|
constexpr size_t kPrefixLength = 4;
|
|
|
|
options.prefix_extractor.reset(NewFixedPrefixTransform(kPrefixLength));
|
|
|
|
options.compression = kNoCompression;
|
|
|
|
BlockBasedTableOptions bbto;
|
|
|
|
bbto.filter_policy.reset(NewBloomFilterPolicy(50));
|
|
|
|
bbto.index_shortening =
|
|
|
|
BlockBasedTableOptions::IndexShorteningMode::kNoShortening;
|
|
|
|
bbto.block_size = 128;
|
|
|
|
bbto.metadata_block_size = 128;
|
|
|
|
bbto.partition_filters = true;
|
|
|
|
bbto.index_type = BlockBasedTableOptions::IndexType::kTwoLevelIndexSearch;
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
|
|
|
|
DestroyAndReopen(options);
|
|
|
|
|
|
|
|
const std::string value(64, '\0');
|
|
|
|
|
|
|
|
WriteOptions write_opts;
|
|
|
|
write_opts.disableWAL = true;
|
|
|
|
for (size_t i = 0; i < kNumKeys; ++i) {
|
|
|
|
std::ostringstream oss;
|
|
|
|
oss << std::setfill('0') << std::setw(4) << std::fixed << i;
|
|
|
|
ASSERT_OK(db_->Put(write_opts, oss.str(), value));
|
|
|
|
}
|
|
|
|
ASSERT_OK(Flush());
|
|
|
|
|
|
|
|
ReadOptions read_opts;
|
|
|
|
// Use legacy, implicit prefix seek
|
|
|
|
read_opts.total_order_seek = false;
|
|
|
|
read_opts.auto_prefix_mode = false;
|
|
|
|
std::unique_ptr<Iterator> it(db_->NewIterator(read_opts));
|
|
|
|
for (size_t i = 0; i < kNumKeys; ++i) {
|
|
|
|
// Seek with a key after each one added but with same prefix. One will
|
|
|
|
// surely cross a partition boundary.
|
|
|
|
std::ostringstream oss;
|
|
|
|
oss << std::setfill('0') << std::setw(4) << std::fixed << i << "a";
|
|
|
|
it->SeekForPrev(oss.str());
|
|
|
|
ASSERT_OK(it->status());
|
|
|
|
ASSERT_TRUE(it->Valid());
|
|
|
|
}
|
|
|
|
it.reset();
|
|
|
|
}
|
|
|
|
|
|
|
|
#endif // ROCKSDB_LITE
|
|
|
|
|
|
|
|
} // namespace ROCKSDB_NAMESPACE
|
|
|
|
|
|
|
|
int main(int argc, char** argv) {
|
|
|
|
ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
|
|
|
|
::testing::InitGoogleTest(&argc, argv);
|
|
|
|
return RUN_ALL_TESTS();
|
|
|
|
}
|