andreacavalli/rocksdb
0
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity
78a6e07c83
rocksdb/table/block_based_table_factory.cc
Mike Kolupaev df38c1ce66 Add BlockBasedTableOptions::index_shortening (#5174) 
Summary:
Introduce BlockBasedTableOptions::index_shortening to give users control on which key shortening techniques to be used in building index blocks. Before this patch, both separators and successor keys where shortened in indexes. With this patch, the default is set to kShortenSeparators to only shorten the separators. Since each index block has many separators and only one successor (last key), the change should not have negative impact on index block size. However it should prevent many unnecessary block loads where due to approximation introduced by shorted successor, seek would land us to the previous block and then fix it by moving to the next one.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5174

Differential Revision: D14884185

Pulled By: al13n321

fbshipit-source-id: 1b08bc8c03edcf09b6b8c16e9a7eea08ad4dd534
2019-04-22 08:20:35 -07:00

632 lines
25 KiB
C++
Raw Blame History

// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
// This source code is licensed under both the GPLv2 (found in the
// COPYING file in the root directory) and Apache 2.0 License
// (found in the LICENSE.Apache file in the root directory).
//
// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file. See the AUTHORS file for names of contributors.
#include "table/block_based_table_factory.h"
#ifndef __STDC_FORMAT_MACROS
#define __STDC_FORMAT_MACROS
#endif
#include <inttypes.h>
#include <stdint.h>
#include <memory>
#include <string>
#include "options/options_helper.h"
#include "port/port.h"
#include "rocksdb/cache.h"
#include "rocksdb/convenience.h"
#include "rocksdb/flush_block_policy.h"
#include "table/block_based_table_builder.h"
#include "table/block_based_table_reader.h"
#include "table/format.h"
#include "util/mutexlock.h"
#include "util/string_util.h"
namespace rocksdb {
void TailPrefetchStats::RecordEffectiveSize(size_t len) {
MutexLock l(&mutex_);
if (num_records_ < kNumTracked) {
num_records_++;
}
records_[next_++] = len;
if (next_ == kNumTracked) {
next_ = 0;
}
}
size_t TailPrefetchStats::GetSuggestedPrefetchSize() {
std::vector<size_t> sorted;
{
MutexLock l(&mutex_);
if (num_records_ == 0) {
return 0;
}
sorted.assign(records_, records_ + num_records_);
}
// Of the historic size, we find the maximum one that satisifis the condtiion
// that if prefetching all, less than 1/8 will be wasted.
std::sort(sorted.begin(), sorted.end());
// Assuming we have 5 data points, and after sorting it looks like this:
//
// +---+
// +---+ | |
// | | | |
// | | | |
// | | | |
// | | | |
// +---+ | | | |
// | | | | | |
// +---+ | | | | | |
// | | | | | | | |
// +---+ | | | | | | | |
// | | | | | | | | | |
// | | | | | | | | | |
// | | | | | | | | | |
// | | | | | | | | | |
// | | | | | | | | | |
// +---+ +---+ +---+ +---+ +---+
//
// and we use every of the value as a candidate, and estimate how much we
// wasted, compared to read. For example, when we use the 3rd record
// as candiate. This area is what we read:
// +---+
// +---+ | |
// | | | |
// | | | |
// | | | |
// | | | |
// *** *** *** ***+ *** *** *** *** **
// * | | | | | |
// +---+ | | | | | *
// * | | | | | | | |
// +---+ | | | | | | | *
// * | | | | X | | | | |
// | | | | | | | | | *
// * | | | | | | | | |
// | | | | | | | | | *
// * | | | | | | | | |
// *** *** ***-*** ***--*** ***--*** +****
// which is (size of the record) X (number of records).
//
// While wasted is this area:
// +---+
// +---+ | |
// | | | |
// | | | |
// | | | |
// | | | |
// *** *** *** ****---+ | | | |
// * * | | | | |
// * *-*** *** | | | | |
// * * | | | | | | |
// *--** *** | | | | | | |
// | | | | | X | | | | |
// | | | | | | | | | |
// | | | | | | | | | |
// | | | | | | | | | |
// | | | | | | | | | |
// +---+ +---+ +---+ +---+ +---+
//
// Which can be calculated iteratively.
// The difference between wasted using 4st and 3rd record, will
// be following area:
// +---+
// +--+ +-+ ++ +-+ +-+ +---+ | |
// + xxxxxxxxxxxxxxxxxxxxxxxx | | | |
// xxxxxxxxxxxxxxxxxxxxxxxx | | | |
// + xxxxxxxxxxxxxxxxxxxxxxxx | | | |
// | xxxxxxxxxxxxxxxxxxxxxxxx | | | |
// +-+ +-+ +-+ ++ +---+ +--+ | | |
// | | | | | | |
// +---+ ++ | | | | | |
// | | | | | | X | | |
// +---+ ++ | | | | | | | |
// | | | | | | | | | |
// | | | | | | | | | |
// | | | | | | | | | |
// | | | | | | | | | |
// | | | | | | | | | |
// +---+ +---+ +---+ +---+ +---+
//
// which will be the size difference between 4st and 3rd record,
// times 3, which is number of records before the 4st.
// Here we assume that all data within the prefetch range will be useful. In
// reality, it may not be the case when a partial block is inside the range,
// or there are data in the middle that is not read. We ignore those cases
// for simplicity.
assert(!sorted.empty());
size_t prev_size = sorted[0];
size_t max_qualified_size = sorted[0];
size_t wasted = 0;
for (size_t i = 1; i < sorted.size(); i++) {
size_t read = sorted[i] * sorted.size();
wasted += (sorted[i] - prev_size) * i;
if (wasted <= read / 8) {
max_qualified_size = sorted[i];
}
prev_size = sorted[i];
}
const size_t kMaxPrefetchSize = 512 * 1024; // Never exceed 512KB
return std::min(kMaxPrefetchSize, max_qualified_size);
}
BlockBasedTableFactory::BlockBasedTableFactory(
const BlockBasedTableOptions& _table_options)
: table_options_(_table_options) {
if (table_options_.flush_block_policy_factory == nullptr) {
table_options_.flush_block_policy_factory.reset(
new FlushBlockBySizePolicyFactory());
}
if (table_options_.no_block_cache) {
table_options_.block_cache.reset();
} else if (table_options_.block_cache == nullptr) {
table_options_.block_cache = NewLRUCache(8 << 20);
}
if (table_options_.block_size_deviation < 0 ||
table_options_.block_size_deviation > 100) {
table_options_.block_size_deviation = 0;
}
if (table_options_.block_restart_interval < 1) {
table_options_.block_restart_interval = 1;
}
if (table_options_.index_block_restart_interval < 1) {
table_options_.index_block_restart_interval = 1;
}
if (table_options_.partition_filters &&
table_options_.index_type !=
BlockBasedTableOptions::kTwoLevelIndexSearch) {
// We do not support partitioned filters without partitioning indexes
table_options_.partition_filters = false;
}
}
Status BlockBasedTableFactory::NewTableReader(
const TableReaderOptions& table_reader_options,
std::unique_ptr<RandomAccessFileReader>&& file, uint64_t file_size,
std::unique_ptr<TableReader>* table_reader,
bool prefetch_index_and_filter_in_cache) const {
return BlockBasedTable::Open(
table_reader_options.ioptions, table_reader_options.env_options,
table_options_, table_reader_options.internal_comparator, std::move(file),
file_size, table_reader, table_reader_options.prefix_extractor,
prefetch_index_and_filter_in_cache, table_reader_options.skip_filters,
table_reader_options.level, table_reader_options.immortal,
table_reader_options.largest_seqno, &tail_prefetch_stats_);
}
TableBuilder* BlockBasedTableFactory::NewTableBuilder(
const TableBuilderOptions& table_builder_options, uint32_t column_family_id,
WritableFileWriter* file) const {
auto table_builder = new BlockBasedTableBuilder(
table_builder_options.ioptions, table_builder_options.moptions,
table_options_, table_builder_options.internal_comparator,
table_builder_options.int_tbl_prop_collector_factories, column_family_id,
file, table_builder_options.compression_type,
table_builder_options.sample_for_compression,
table_builder_options.compression_opts,
table_builder_options.skip_filters,
table_builder_options.column_family_name,
table_builder_options.creation_time,
table_builder_options.oldest_key_time,
table_builder_options.target_file_size,
table_builder_options.file_creation_time);
return table_builder;
}
Status BlockBasedTableFactory::SanitizeOptions(
const DBOptions& /*db_opts*/, const ColumnFamilyOptions& cf_opts) const {
if (table_options_.index_type == BlockBasedTableOptions::kHashSearch &&
cf_opts.prefix_extractor == nullptr) {
return Status::InvalidArgument(
"Hash index is specified for block-based "
"table, but prefix_extractor is not given");
}
if (table_options_.cache_index_and_filter_blocks &&
table_options_.no_block_cache) {
return Status::InvalidArgument(
"Enable cache_index_and_filter_blocks, "
", but block cache is disabled");
}
if (table_options_.pin_l0_filter_and_index_blocks_in_cache &&
table_options_.no_block_cache) {
return Status::InvalidArgument(
"Enable pin_l0_filter_and_index_blocks_in_cache, "
", but block cache is disabled");
}
if (!BlockBasedTableSupportedVersion(table_options_.format_version)) {
return Status::InvalidArgument(
"Unsupported BlockBasedTable format_version. Please check "
"include/rocksdb/table.h for more info");
}
if (table_options_.block_align && (cf_opts.compression != kNoCompression)) {
return Status::InvalidArgument(
"Enable block_align, but compression "
"enabled");
}
if (table_options_.block_align &&
(table_options_.block_size & (table_options_.block_size - 1))) {
return Status::InvalidArgument(
"Block alignment requested but block size is not a power of 2");
}
if (table_options_.data_block_index_type ==
BlockBasedTableOptions::kDataBlockBinaryAndHash &&
table_options_.data_block_hash_table_util_ratio <= 0) {
return Status::InvalidArgument(
"data_block_hash_table_util_ratio should be greater than 0 when "
"data_block_index_type is set to kDataBlockBinaryAndHash");
}
return Status::OK();
}
std::string BlockBasedTableFactory::GetPrintableTableOptions() const {
std::string ret;
ret.reserve(20000);
const int kBufferSize = 200;
char buffer[kBufferSize];
snprintf(buffer, kBufferSize, " flush_block_policy_factory: %s (%p)\n",
table_options_.flush_block_policy_factory->Name(),
static_cast<void*>(table_options_.flush_block_policy_factory.get()));
ret.append(buffer);
snprintf(buffer, kBufferSize, " cache_index_and_filter_blocks: %d\n",
table_options_.cache_index_and_filter_blocks);
ret.append(buffer);
snprintf(buffer, kBufferSize,
" cache_index_and_filter_blocks_with_high_priority: %d\n",
table_options_.cache_index_and_filter_blocks_with_high_priority);
ret.append(buffer);
snprintf(buffer, kBufferSize,
" pin_l0_filter_and_index_blocks_in_cache: %d\n",
table_options_.pin_l0_filter_and_index_blocks_in_cache);
ret.append(buffer);
snprintf(buffer, kBufferSize, " pin_top_level_index_and_filter: %d\n",
table_options_.pin_top_level_index_and_filter);
ret.append(buffer);
snprintf(buffer, kBufferSize, " index_type: %d\n",
table_options_.index_type);
ret.append(buffer);
snprintf(buffer, kBufferSize, " data_block_index_type: %d\n",
table_options_.data_block_index_type);
ret.append(buffer);
snprintf(buffer, kBufferSize, " index_shortening: %d\n",
static_cast<int>(table_options_.index_shortening));
ret.append(buffer);
snprintf(buffer, kBufferSize, " data_block_hash_table_util_ratio: %lf\n",
table_options_.data_block_hash_table_util_ratio);
ret.append(buffer);
snprintf(buffer, kBufferSize, " hash_index_allow_collision: %d\n",
table_options_.hash_index_allow_collision);
ret.append(buffer);
snprintf(buffer, kBufferSize, " checksum: %d\n", table_options_.checksum);
ret.append(buffer);
snprintf(buffer, kBufferSize, " no_block_cache: %d\n",
table_options_.no_block_cache);
ret.append(buffer);
snprintf(buffer, kBufferSize, " block_cache: %p\n",
static_cast<void*>(table_options_.block_cache.get()));
ret.append(buffer);
if (table_options_.block_cache) {
const char* block_cache_name = table_options_.block_cache->Name();
if (block_cache_name != nullptr) {
snprintf(buffer, kBufferSize, " block_cache_name: %s\n",
block_cache_name);
ret.append(buffer);
}
ret.append(" block_cache_options:\n");
ret.append(table_options_.block_cache->GetPrintableOptions());
}
snprintf(buffer, kBufferSize, " block_cache_compressed: %p\n",
static_cast<void*>(table_options_.block_cache_compressed.get()));
ret.append(buffer);
if (table_options_.block_cache_compressed) {
const char* block_cache_compressed_name =
table_options_.block_cache_compressed->Name();
if (block_cache_compressed_name != nullptr) {
snprintf(buffer, kBufferSize, " block_cache_name: %s\n",
block_cache_compressed_name);
ret.append(buffer);
}
ret.append(" block_cache_compressed_options:\n");
ret.append(table_options_.block_cache_compressed->GetPrintableOptions());
}
snprintf(buffer, kBufferSize, " persistent_cache: %p\n",
static_cast<void*>(table_options_.persistent_cache.get()));
ret.append(buffer);
if (table_options_.persistent_cache) {
snprintf(buffer, kBufferSize, " persistent_cache_options:\n");
ret.append(buffer);
ret.append(table_options_.persistent_cache->GetPrintableOptions());
}
snprintf(buffer, kBufferSize, " block_size: %" ROCKSDB_PRIszt "\n",
table_options_.block_size);
ret.append(buffer);
snprintf(buffer, kBufferSize, " block_size_deviation: %d\n",
table_options_.block_size_deviation);
ret.append(buffer);
snprintf(buffer, kBufferSize, " block_restart_interval: %d\n",
table_options_.block_restart_interval);
ret.append(buffer);
snprintf(buffer, kBufferSize, " index_block_restart_interval: %d\n",
table_options_.index_block_restart_interval);
ret.append(buffer);
snprintf(buffer, kBufferSize, " metadata_block_size: %" PRIu64 "\n",
table_options_.metadata_block_size);
ret.append(buffer);
snprintf(buffer, kBufferSize, " partition_filters: %d\n",
table_options_.partition_filters);
ret.append(buffer);
snprintf(buffer, kBufferSize, " use_delta_encoding: %d\n",
table_options_.use_delta_encoding);
ret.append(buffer);
snprintf(buffer, kBufferSize, " filter_policy: %s\n",
table_options_.filter_policy == nullptr
? "nullptr"
: table_options_.filter_policy->Name());
ret.append(buffer);
snprintf(buffer, kBufferSize, " whole_key_filtering: %d\n",
table_options_.whole_key_filtering);
ret.append(buffer);
snprintf(buffer, kBufferSize, " verify_compression: %d\n",
table_options_.verify_compression);
ret.append(buffer);
snprintf(buffer, kBufferSize, " read_amp_bytes_per_bit: %d\n",
table_options_.read_amp_bytes_per_bit);
ret.append(buffer);
snprintf(buffer, kBufferSize, " format_version: %d\n",
table_options_.format_version);
ret.append(buffer);
snprintf(buffer, kBufferSize, " enable_index_compression: %d\n",
table_options_.enable_index_compression);
ret.append(buffer);
snprintf(buffer, kBufferSize, " block_align: %d\n",
table_options_.block_align);
ret.append(buffer);
return ret;
}
#ifndef ROCKSDB_LITE
namespace {
bool SerializeSingleBlockBasedTableOption(
std::string* opt_string, const BlockBasedTableOptions& bbt_options,
const std::string& name, const std::string& delimiter) {
auto iter = block_based_table_type_info.find(name);
if (iter == block_based_table_type_info.end()) {
return false;
}
auto& opt_info = iter->second;
const char* opt_address =
reinterpret_cast<const char*>(&bbt_options) + opt_info.offset;
std::string value;
bool result = SerializeSingleOptionHelper(opt_address, opt_info.type, &value);
if (result) {
*opt_string = name + "=" + value + delimiter;
}
return result;
}
} // namespace
Status BlockBasedTableFactory::GetOptionString(
std::string* opt_string, const std::string& delimiter) const {
assert(opt_string);
opt_string->clear();
for (auto iter = block_based_table_type_info.begin();
iter != block_based_table_type_info.end(); ++iter) {
if (iter->second.verification == OptionVerificationType::kDeprecated) {
// If the option is no longer used in rocksdb and marked as deprecated,
// we skip it in the serialization.
continue;
}
std::string single_output;
bool result = SerializeSingleBlockBasedTableOption(
&single_output, table_options_, iter->first, delimiter);
assert(result);
if (result) {
opt_string->append(single_output);
}
}
return Status::OK();
}
#else
Status BlockBasedTableFactory::GetOptionString(
std::string* /*opt_string*/, const std::string& /*delimiter*/) const {
return Status::OK();
}
#endif // !ROCKSDB_LITE
const BlockBasedTableOptions& BlockBasedTableFactory::table_options() const {
return table_options_;
}
#ifndef ROCKSDB_LITE
namespace {
std::string ParseBlockBasedTableOption(const std::string& name,
const std::string& org_value,
BlockBasedTableOptions* new_options,
bool input_strings_escaped = false,
bool ignore_unknown_options = false) {
const std::string& value =
input_strings_escaped ? UnescapeOptionString(org_value) : org_value;
if (!input_strings_escaped) {
// if the input string is not escaped, it means this function is
// invoked from SetOptions, which takes the old format.
if (name == "block_cache" || name == "block_cache_compressed") {
// cache options can be specified in the following format
// "block_cache={capacity=1M;num_shard_bits=4;
// strict_capacity_limit=true;high_pri_pool_ratio=0.5;}"
// To support backward compatibility, the following format
// is also supported.
// "block_cache=1M"
std::shared_ptr<Cache> cache;
// block_cache is specified in format block_cache=<cache_size>.
if (value.find('=') == std::string::npos) {
cache = NewLRUCache(ParseSizeT(value));
} else {
LRUCacheOptions cache_opts;
if (!ParseOptionHelper(reinterpret_cast<char*>(&cache_opts),
OptionType::kLRUCacheOptions, value)) {
return "Invalid cache options";
}
cache = NewLRUCache(cache_opts);
}
if (name == "block_cache") {
new_options->block_cache = cache;
} else {
new_options->block_cache_compressed = cache;
}
return "";
} else if (name == "filter_policy") {
// Expect the following format
// bloomfilter:int:bool
const std::string kName = "bloomfilter:";
if (value.compare(0, kName.size(), kName) != 0) {
return "Invalid filter policy name";
}
size_t pos = value.find(':', kName.size());
if (pos == std::string::npos) {
return "Invalid filter policy config, missing bits_per_key";
}
int bits_per_key =
ParseInt(trim(value.substr(kName.size(), pos - kName.size())));
bool use_block_based_builder =
ParseBoolean("use_block_based_builder", trim(value.substr(pos + 1)));
new_options->filter_policy.reset(
NewBloomFilterPolicy(bits_per_key, use_block_based_builder));
return "";
}
}
const auto iter = block_based_table_type_info.find(name);
if (iter == block_based_table_type_info.end()) {
if (ignore_unknown_options) {
return "";
} else {
return "Unrecognized option";
}
}
const auto& opt_info = iter->second;
if (opt_info.verification != OptionVerificationType::kDeprecated &&
!ParseOptionHelper(reinterpret_cast<char*>(new_options) + opt_info.offset,
opt_info.type, value)) {
return "Invalid value";
}
return "";
}
} // namespace
Status GetBlockBasedTableOptionsFromString(
const BlockBasedTableOptions& table_options, const std::string& opts_str,
BlockBasedTableOptions* new_table_options) {
std::unordered_map<std::string, std::string> opts_map;
Status s = StringToMap(opts_str, &opts_map);
if (!s.ok()) {
return s;
}
return GetBlockBasedTableOptionsFromMap(table_options, opts_map,
new_table_options);
}
Status GetBlockBasedTableOptionsFromMap(
const BlockBasedTableOptions& table_options,
const std::unordered_map<std::string, std::string>& opts_map,
BlockBasedTableOptions* new_table_options, bool input_strings_escaped,
bool ignore_unknown_options) {
assert(new_table_options);
*new_table_options = table_options;
for (const auto& o : opts_map) {
auto error_message = ParseBlockBasedTableOption(
o.first, o.second, new_table_options, input_strings_escaped,
ignore_unknown_options);
if (error_message != "") {
const auto iter = block_based_table_type_info.find(o.first);
if (iter == block_based_table_type_info.end() ||
!input_strings_escaped || // !input_strings_escaped indicates
// the old API, where everything is
// parsable.
(iter->second.verification != OptionVerificationType::kByName &&
iter->second.verification !=
OptionVerificationType::kByNameAllowNull &&
iter->second.verification !=
OptionVerificationType::kByNameAllowFromNull &&
iter->second.verification != OptionVerificationType::kDeprecated)) {
// Restore "new_options" to the default "base_options".
*new_table_options = table_options;
return Status::InvalidArgument("Can't parse BlockBasedTableOptions:",
o.first + " " + error_message);
}
}
}
return Status::OK();
}
Status VerifyBlockBasedTableFactory(
const BlockBasedTableFactory* base_tf,
const BlockBasedTableFactory* file_tf,
OptionsSanityCheckLevel sanity_check_level) {
if ((base_tf != nullptr) != (file_tf != nullptr) &&
sanity_check_level > kSanityLevelNone) {
return Status::Corruption(
"[RocksDBOptionsParser]: Inconsistent TableFactory class type");
}
if (base_tf == nullptr) {
return Status::OK();
}
assert(file_tf != nullptr);
const auto& base_opt = base_tf->table_options();
const auto& file_opt = file_tf->table_options();
for (auto& pair : block_based_table_type_info) {
if (pair.second.verification == OptionVerificationType::kDeprecated) {
// We skip checking deprecated variables as they might
// contain random values since they might not be initialized
continue;
}
if (BBTOptionSanityCheckLevel(pair.first) <= sanity_check_level) {
if (!AreEqualOptions(reinterpret_cast<const char*>(&base_opt),
reinterpret_cast<const char*>(&file_opt),
pair.second, pair.first, nullptr)) {
return Status::Corruption(
"[RocksDBOptionsParser]: "
"failed the verification on BlockBasedTableOptions::",
pair.first);
}
}
}
return Status::OK();
}
#endif // !ROCKSDB_LITE
TableFactory* NewBlockBasedTableFactory(
const BlockBasedTableOptions& _table_options) {
return new BlockBasedTableFactory(_table_options);
}
const std::string BlockBasedTableFactory::kName = "BlockBasedTable";
const std::string BlockBasedTablePropertyNames::kIndexType =
"rocksdb.block.based.table.index.type";
const std::string BlockBasedTablePropertyNames::kWholeKeyFiltering =
"rocksdb.block.based.table.whole.key.filtering";
const std::string BlockBasedTablePropertyNames::kPrefixFiltering =
"rocksdb.block.based.table.prefix.filtering";
const std::string kHashIndexPrefixesBlock = "rocksdb.hashindex.prefixes";
const std::string kHashIndexPrefixesMetadataBlock =
"rocksdb.hashindex.metadata";
const std::string kPropTrue = "1";
const std::string kPropFalse = "0";
} // namespace rocksdb
Reference in New Issue View Git Blame Copy Permalink