2019-05-30 23:47:29 +02:00
|
|
|
|
2017-05-10 23:54:35 +02:00
|
|
|
// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
|
2017-07-16 01:03:42 +02:00
|
|
|
// This source code is licensed under both the GPLv2 (found in the
|
|
|
|
// COPYING file in the root directory) and Apache 2.0 License
|
|
|
|
// (found in the LICENSE.Apache file in the root directory).
|
2017-05-10 23:54:35 +02:00
|
|
|
#ifndef ROCKSDB_LITE
|
|
|
|
|
|
|
|
#include "utilities/blob_db/blob_db_impl.h"
|
|
|
|
#include <algorithm>
|
|
|
|
#include <cinttypes>
|
|
|
|
#include <iomanip>
|
|
|
|
#include <memory>
|
|
|
|
|
2019-10-14 21:52:15 +02:00
|
|
|
#include "db/blob_index.h"
|
2019-05-31 20:52:59 +02:00
|
|
|
#include "db/db_impl/db_impl.h"
|
2017-05-10 23:54:35 +02:00
|
|
|
#include "db/write_batch_internal.h"
|
2019-05-30 05:44:08 +02:00
|
|
|
#include "file/file_util.h"
|
|
|
|
#include "file/filename.h"
|
2019-09-16 19:31:27 +02:00
|
|
|
#include "file/random_access_file_reader.h"
|
2019-05-30 05:44:08 +02:00
|
|
|
#include "file/sst_file_manager_impl.h"
|
2019-09-16 19:31:27 +02:00
|
|
|
#include "file/writable_file_writer.h"
|
2019-06-01 02:19:43 +02:00
|
|
|
#include "logging/logging.h"
|
2017-05-10 23:54:35 +02:00
|
|
|
#include "monitoring/instrumented_mutex.h"
|
2017-11-28 20:42:28 +01:00
|
|
|
#include "monitoring/statistics.h"
|
2017-05-10 23:54:35 +02:00
|
|
|
#include "rocksdb/convenience.h"
|
|
|
|
#include "rocksdb/env.h"
|
|
|
|
#include "rocksdb/iterator.h"
|
|
|
|
#include "rocksdb/utilities/stackable_db.h"
|
|
|
|
#include "rocksdb/utilities/transaction.h"
|
2019-05-30 23:47:29 +02:00
|
|
|
#include "table/block_based/block.h"
|
|
|
|
#include "table/block_based/block_based_table_builder.h"
|
|
|
|
#include "table/block_based/block_builder.h"
|
2017-05-10 23:54:35 +02:00
|
|
|
#include "table/meta_blocks.h"
|
2019-05-31 02:39:43 +02:00
|
|
|
#include "test_util/sync_point.h"
|
2017-07-29 01:23:50 +02:00
|
|
|
#include "util/cast_util.h"
|
2017-05-10 23:54:35 +02:00
|
|
|
#include "util/crc32c.h"
|
2017-08-04 00:07:01 +02:00
|
|
|
#include "util/mutexlock.h"
|
2017-05-10 23:54:35 +02:00
|
|
|
#include "util/random.h"
|
2017-11-28 20:42:28 +01:00
|
|
|
#include "util/stop_watch.h"
|
2017-05-10 23:54:35 +02:00
|
|
|
#include "util/timer_queue.h"
|
2017-12-11 21:01:22 +01:00
|
|
|
#include "utilities/blob_db/blob_compaction_filter.h"
|
2017-10-18 02:24:25 +02:00
|
|
|
#include "utilities/blob_db/blob_db_iterator.h"
|
2018-03-06 20:46:20 +01:00
|
|
|
#include "utilities/blob_db/blob_db_listener.h"
|
2017-05-10 23:54:35 +02:00
|
|
|
|
|
|
|
namespace {
|
|
|
|
int kBlockBasedTableVersionFormat = 2;
|
|
|
|
} // end namespace
|
|
|
|
|
|
|
|
namespace rocksdb {
|
|
|
|
namespace blob_db {
|
|
|
|
|
2018-06-26 07:32:29 +02:00
|
|
|
bool BlobFileComparator::operator()(
|
|
|
|
const std::shared_ptr<BlobFile>& lhs,
|
|
|
|
const std::shared_ptr<BlobFile>& rhs) const {
|
|
|
|
return lhs->BlobFileNumber() > rhs->BlobFileNumber();
|
|
|
|
}
|
|
|
|
|
|
|
|
bool BlobFileComparatorTTL::operator()(
|
|
|
|
const std::shared_ptr<BlobFile>& lhs,
|
|
|
|
const std::shared_ptr<BlobFile>& rhs) const {
|
2018-03-06 20:46:20 +01:00
|
|
|
assert(lhs->HasTTL() && rhs->HasTTL());
|
2017-10-27 22:14:34 +02:00
|
|
|
if (lhs->expiration_range_.first < rhs->expiration_range_.first) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
if (lhs->expiration_range_.first > rhs->expiration_range_.first) {
|
|
|
|
return false;
|
|
|
|
}
|
2017-11-02 20:02:42 +01:00
|
|
|
return lhs->BlobFileNumber() < rhs->BlobFileNumber();
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
BlobDBImpl::BlobDBImpl(const std::string& dbname,
|
|
|
|
const BlobDBOptions& blob_db_options,
|
2017-12-11 21:01:22 +01:00
|
|
|
const DBOptions& db_options,
|
|
|
|
const ColumnFamilyOptions& cf_options)
|
|
|
|
: BlobDB(),
|
|
|
|
dbname_(dbname),
|
2017-05-10 23:54:35 +02:00
|
|
|
db_impl_(nullptr),
|
2017-07-28 08:16:18 +02:00
|
|
|
env_(db_options.env),
|
2017-05-10 23:54:35 +02:00
|
|
|
bdb_options_(blob_db_options),
|
|
|
|
db_options_(db_options),
|
2017-12-11 21:01:22 +01:00
|
|
|
cf_options_(cf_options),
|
2017-05-10 23:54:35 +02:00
|
|
|
env_options_(db_options),
|
2017-11-28 20:42:28 +01:00
|
|
|
statistics_(db_options_.statistics.get()),
|
2017-05-10 23:54:35 +02:00
|
|
|
next_file_number_(1),
|
|
|
|
epoch_of_(0),
|
2018-03-06 20:46:20 +01:00
|
|
|
closed_(true),
|
2017-05-10 23:54:35 +02:00
|
|
|
open_file_count_(0),
|
2018-03-06 20:46:20 +01:00
|
|
|
total_blob_size_(0),
|
|
|
|
live_sst_size_(0),
|
|
|
|
fifo_eviction_seq_(0),
|
|
|
|
evict_expiration_up_to_(0),
|
|
|
|
debug_level_(0) {
|
2017-05-10 23:54:35 +02:00
|
|
|
blob_dir_ = (bdb_options_.path_relative)
|
|
|
|
? dbname + "/" + bdb_options_.blob_dir
|
|
|
|
: bdb_options_.blob_dir;
|
2017-12-20 01:34:25 +01:00
|
|
|
env_options_.bytes_per_sync = blob_db_options.bytes_per_sync;
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2017-12-11 21:01:22 +01:00
|
|
|
BlobDBImpl::~BlobDBImpl() {
|
2019-02-22 02:23:05 +01:00
|
|
|
tqueue_.shutdown();
|
2017-12-11 21:01:22 +01:00
|
|
|
// CancelAllBackgroundWork(db_, true);
|
2018-03-06 20:46:20 +01:00
|
|
|
Status s __attribute__((__unused__)) = Close();
|
|
|
|
assert(s.ok());
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlobDBImpl::Close() {
|
|
|
|
if (closed_) {
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
closed_ = true;
|
2017-12-11 21:01:22 +01:00
|
|
|
|
2018-03-06 20:46:20 +01:00
|
|
|
// Close base DB before BlobDBImpl destructs to stop event listener and
|
|
|
|
// compaction filter call.
|
|
|
|
Status s = db_->Close();
|
|
|
|
// delete db_ anyway even if close failed.
|
|
|
|
delete db_;
|
|
|
|
// Reset pointers to avoid StackableDB delete the pointer again.
|
|
|
|
db_ = nullptr;
|
|
|
|
db_impl_ = nullptr;
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
s = SyncBlobFiles();
|
|
|
|
return s;
|
2017-12-11 21:01:22 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
BlobDBOptions BlobDBImpl::GetBlobDBOptions() const { return bdb_options_; }
|
|
|
|
|
|
|
|
Status BlobDBImpl::Open(std::vector<ColumnFamilyHandle*>* handles) {
|
|
|
|
assert(handles != nullptr);
|
2017-05-10 23:54:35 +02:00
|
|
|
assert(db_ == nullptr);
|
2017-12-11 21:01:22 +01:00
|
|
|
if (blob_dir_.empty()) {
|
|
|
|
return Status::NotSupported("No blob directory in options");
|
|
|
|
}
|
|
|
|
if (cf_options_.compaction_filter != nullptr ||
|
|
|
|
cf_options_.compaction_filter_factory != nullptr) {
|
|
|
|
return Status::NotSupported("Blob DB doesn't support compaction filter.");
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2017-12-11 21:01:22 +01:00
|
|
|
Status s;
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2017-12-11 21:01:22 +01:00
|
|
|
// Create info log.
|
|
|
|
if (db_options_.info_log == nullptr) {
|
|
|
|
s = CreateLoggerFromOptions(dbname_, db_options_, &db_options_.info_log);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2017-12-11 21:01:22 +01:00
|
|
|
ROCKS_LOG_INFO(db_options_.info_log, "Opening BlobDB...");
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2017-12-11 21:01:22 +01:00
|
|
|
// Open blob directory.
|
|
|
|
s = env_->CreateDirIfMissing(blob_dir_);
|
2017-05-10 23:54:35 +02:00
|
|
|
if (!s.ok()) {
|
2017-12-11 21:01:22 +01:00
|
|
|
ROCKS_LOG_ERROR(db_options_.info_log,
|
|
|
|
"Failed to create blob_dir %s, status: %s",
|
|
|
|
blob_dir_.c_str(), s.ToString().c_str());
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
2017-07-28 08:16:18 +02:00
|
|
|
s = env_->NewDirectory(blob_dir_, &dir_ent_);
|
2017-05-10 23:54:35 +02:00
|
|
|
if (!s.ok()) {
|
2017-12-11 21:01:22 +01:00
|
|
|
ROCKS_LOG_ERROR(db_options_.info_log,
|
|
|
|
"Failed to open blob_dir %s, status: %s", blob_dir_.c_str(),
|
|
|
|
s.ToString().c_str());
|
|
|
|
return s;
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2017-12-11 21:01:22 +01:00
|
|
|
// Open blob files.
|
|
|
|
s = OpenAllBlobFiles();
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
2017-06-14 22:08:54 +02:00
|
|
|
}
|
|
|
|
|
2017-12-11 21:01:22 +01:00
|
|
|
// Update options
|
2018-03-06 20:46:20 +01:00
|
|
|
db_options_.listeners.push_back(std::make_shared<BlobDBListener>(this));
|
2017-12-11 21:01:22 +01:00
|
|
|
cf_options_.compaction_filter_factory.reset(
|
2018-03-06 20:46:20 +01:00
|
|
|
new BlobIndexCompactionFilterFactory(this, env_, statistics_));
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2017-12-11 21:01:22 +01:00
|
|
|
// Open base db.
|
|
|
|
ColumnFamilyDescriptor cf_descriptor(kDefaultColumnFamilyName, cf_options_);
|
|
|
|
s = DB::Open(db_options_, dbname_, {cf_descriptor}, handles, &db_);
|
2017-05-10 23:54:35 +02:00
|
|
|
if (!s.ok()) {
|
2017-12-11 21:01:22 +01:00
|
|
|
return s;
|
|
|
|
}
|
|
|
|
db_impl_ = static_cast_with_check<DBImpl, DB>(db_->GetRootDB());
|
2019-03-26 18:43:22 +01:00
|
|
|
|
|
|
|
// Add trash files in blob dir to file delete scheduler.
|
|
|
|
SstFileManagerImpl* sfm = static_cast<SstFileManagerImpl*>(
|
|
|
|
db_impl_->immutable_db_options().sst_file_manager.get());
|
|
|
|
DeleteScheduler::CleanupDirectory(env_, sfm, blob_dir_);
|
|
|
|
|
2018-03-06 20:46:20 +01:00
|
|
|
UpdateLiveSSTSize();
|
2017-12-11 21:01:22 +01:00
|
|
|
|
|
|
|
// Start background jobs.
|
|
|
|
if (!bdb_options_.disable_background_tasks) {
|
|
|
|
StartBackgroundTasks();
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2017-12-11 21:01:22 +01:00
|
|
|
ROCKS_LOG_INFO(db_options_.info_log, "BlobDB pointer %p", this);
|
2017-12-20 01:50:59 +01:00
|
|
|
bdb_options_.Dump(db_options_.info_log.get());
|
2018-03-06 20:46:20 +01:00
|
|
|
closed_ = false;
|
2017-05-10 23:54:35 +02:00
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
void BlobDBImpl::StartBackgroundTasks() {
|
|
|
|
// store a call to a member function and object
|
|
|
|
tqueue_.add(
|
2017-08-01 21:48:22 +02:00
|
|
|
kReclaimOpenFilesPeriodMillisecs,
|
2017-05-10 23:54:35 +02:00
|
|
|
std::bind(&BlobDBImpl::ReclaimOpenFiles, this, std::placeholders::_1));
|
|
|
|
tqueue_.add(
|
2017-08-21 03:12:38 +02:00
|
|
|
kDeleteObsoleteFilesPeriodMillisecs,
|
|
|
|
std::bind(&BlobDBImpl::DeleteObsoleteFiles, this, std::placeholders::_1));
|
2017-08-01 21:48:22 +02:00
|
|
|
tqueue_.add(kSanityCheckPeriodMillisecs,
|
2017-05-10 23:54:35 +02:00
|
|
|
std::bind(&BlobDBImpl::SanityCheck, this, std::placeholders::_1));
|
|
|
|
tqueue_.add(
|
2018-08-21 07:38:34 +02:00
|
|
|
kEvictExpiredFilesPeriodMillisecs,
|
|
|
|
std::bind(&BlobDBImpl::EvictExpiredFiles, this, std::placeholders::_1));
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2017-12-11 21:01:22 +01:00
|
|
|
Status BlobDBImpl::GetAllBlobFiles(std::set<uint64_t>* file_numbers) {
|
|
|
|
assert(file_numbers != nullptr);
|
2017-05-10 23:54:35 +02:00
|
|
|
std::vector<std::string> all_files;
|
2017-12-11 21:01:22 +01:00
|
|
|
Status s = env_->GetChildren(blob_dir_, &all_files);
|
|
|
|
if (!s.ok()) {
|
|
|
|
ROCKS_LOG_ERROR(db_options_.info_log,
|
|
|
|
"Failed to get list of blob files, status: %s",
|
|
|
|
s.ToString().c_str());
|
|
|
|
return s;
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2017-12-11 21:01:22 +01:00
|
|
|
for (const auto& file_name : all_files) {
|
|
|
|
uint64_t file_number;
|
2017-05-10 23:54:35 +02:00
|
|
|
FileType type;
|
2017-12-11 21:01:22 +01:00
|
|
|
bool success = ParseFileName(file_name, &file_number, &type);
|
|
|
|
if (success && type == kBlobFile) {
|
|
|
|
file_numbers->insert(file_number);
|
2017-05-10 23:54:35 +02:00
|
|
|
} else {
|
2017-07-13 19:03:39 +02:00
|
|
|
ROCKS_LOG_WARN(db_options_.info_log,
|
2017-12-11 21:01:22 +01:00
|
|
|
"Skipping file in blob directory: %s", file_name.c_str());
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-12-11 21:01:22 +01:00
|
|
|
return s;
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2017-12-11 21:01:22 +01:00
|
|
|
Status BlobDBImpl::OpenAllBlobFiles() {
|
|
|
|
std::set<uint64_t> file_numbers;
|
|
|
|
Status s = GetAllBlobFiles(&file_numbers);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2017-12-11 21:01:22 +01:00
|
|
|
if (!file_numbers.empty()) {
|
|
|
|
next_file_number_.store(*file_numbers.rbegin() + 1);
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2017-12-11 21:01:22 +01:00
|
|
|
std::string blob_file_list;
|
|
|
|
std::string obsolete_file_list;
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2017-12-11 21:01:22 +01:00
|
|
|
for (auto& file_number : file_numbers) {
|
|
|
|
std::shared_ptr<BlobFile> blob_file = std::make_shared<BlobFile>(
|
|
|
|
this, blob_dir_, file_number, db_options_.info_log.get());
|
|
|
|
blob_file->MarkImmutable();
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2017-12-11 21:01:22 +01:00
|
|
|
// Read file header and footer
|
|
|
|
Status read_metadata_status = blob_file->ReadMetadata(env_, env_options_);
|
|
|
|
if (read_metadata_status.IsCorruption()) {
|
|
|
|
// Remove incomplete file.
|
2018-03-06 20:46:20 +01:00
|
|
|
ObsoleteBlobFile(blob_file, 0 /*obsolete_seq*/, false /*update_size*/);
|
2017-12-11 21:01:22 +01:00
|
|
|
if (!obsolete_file_list.empty()) {
|
|
|
|
obsolete_file_list.append(", ");
|
|
|
|
}
|
|
|
|
obsolete_file_list.append(ToString(file_number));
|
|
|
|
continue;
|
|
|
|
} else if (!read_metadata_status.ok()) {
|
|
|
|
ROCKS_LOG_ERROR(db_options_.info_log,
|
2019-04-04 21:05:42 +02:00
|
|
|
"Unable to read metadata of blob file %" PRIu64
|
2017-12-11 21:01:22 +01:00
|
|
|
", status: '%s'",
|
|
|
|
file_number, read_metadata_status.ToString().c_str());
|
|
|
|
return read_metadata_status;
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2018-03-06 20:46:20 +01:00
|
|
|
total_blob_size_ += blob_file->GetFileSize();
|
|
|
|
|
2017-12-11 21:01:22 +01:00
|
|
|
blob_files_[file_number] = blob_file;
|
|
|
|
if (!blob_file_list.empty()) {
|
|
|
|
blob_file_list.append(", ");
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
2017-12-11 21:01:22 +01:00
|
|
|
blob_file_list.append(ToString(file_number));
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2017-12-11 21:01:22 +01:00
|
|
|
ROCKS_LOG_INFO(db_options_.info_log,
|
|
|
|
"Found %" ROCKSDB_PRIszt " blob files: %s", blob_files_.size(),
|
|
|
|
blob_file_list.c_str());
|
|
|
|
ROCKS_LOG_INFO(db_options_.info_log,
|
|
|
|
"Found %" ROCKSDB_PRIszt
|
|
|
|
" incomplete or corrupted blob files: %s",
|
|
|
|
obsolete_files_.size(), obsolete_file_list.c_str());
|
|
|
|
return s;
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
void BlobDBImpl::CloseRandomAccessLocked(
|
|
|
|
const std::shared_ptr<BlobFile>& bfile) {
|
|
|
|
bfile->CloseRandomAccessLocked();
|
|
|
|
open_file_count_--;
|
|
|
|
}
|
|
|
|
|
2018-09-21 01:50:07 +02:00
|
|
|
Status BlobDBImpl::GetBlobFileReader(
|
|
|
|
const std::shared_ptr<BlobFile>& blob_file,
|
|
|
|
std::shared_ptr<RandomAccessFileReader>* reader) {
|
|
|
|
assert(reader != nullptr);
|
2017-05-10 23:54:35 +02:00
|
|
|
bool fresh_open = false;
|
2018-09-21 01:50:07 +02:00
|
|
|
Status s = blob_file->GetReader(env_, env_options_, reader, &fresh_open);
|
|
|
|
if (s.ok() && fresh_open) {
|
|
|
|
assert(*reader != nullptr);
|
|
|
|
open_file_count_++;
|
|
|
|
}
|
|
|
|
return s;
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
std::shared_ptr<BlobFile> BlobDBImpl::NewBlobFile(const std::string& reason) {
|
|
|
|
uint64_t file_num = next_file_number_++;
|
2017-12-11 21:01:22 +01:00
|
|
|
auto bfile = std::make_shared<BlobFile>(this, blob_dir_, file_num,
|
|
|
|
db_options_.info_log.get());
|
2017-07-13 19:03:39 +02:00
|
|
|
ROCKS_LOG_DEBUG(db_options_.info_log, "New blob file created: %s reason='%s'",
|
|
|
|
bfile->PathName().c_str(), reason.c_str());
|
2017-05-10 23:54:35 +02:00
|
|
|
LogFlush(db_options_.info_log);
|
|
|
|
return bfile;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlobDBImpl::CreateWriterLocked(const std::shared_ptr<BlobFile>& bfile) {
|
|
|
|
std::string fpath(bfile->PathName());
|
|
|
|
std::unique_ptr<WritableFile> wfile;
|
|
|
|
|
2017-08-04 22:09:56 +02:00
|
|
|
Status s = env_->ReopenWritableFile(fpath, &wfile, env_options_);
|
2017-05-10 23:54:35 +02:00
|
|
|
if (!s.ok()) {
|
2017-07-13 19:03:39 +02:00
|
|
|
ROCKS_LOG_ERROR(db_options_.info_log,
|
|
|
|
"Failed to open blob file for write: %s status: '%s'"
|
|
|
|
" exists: '%s'",
|
|
|
|
fpath.c_str(), s.ToString().c_str(),
|
2017-07-28 08:16:18 +02:00
|
|
|
env_->FileExists(fpath).ToString().c_str());
|
2017-05-10 23:54:35 +02:00
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
std::unique_ptr<WritableFileWriter> fwriter;
|
2018-08-23 19:04:10 +02:00
|
|
|
fwriter.reset(new WritableFileWriter(std::move(wfile), fpath, env_options_));
|
2017-05-10 23:54:35 +02:00
|
|
|
|
|
|
|
uint64_t boffset = bfile->GetFileSize();
|
|
|
|
if (debug_level_ >= 2 && boffset) {
|
2019-04-04 21:05:42 +02:00
|
|
|
ROCKS_LOG_DEBUG(db_options_.info_log,
|
|
|
|
"Open blob file: %s with offset: %" PRIu64, fpath.c_str(),
|
|
|
|
boffset);
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
Writer::ElemType et = Writer::kEtNone;
|
2017-10-27 22:14:34 +02:00
|
|
|
if (bfile->file_size_ == BlobLogHeader::kSize) {
|
2017-05-10 23:54:35 +02:00
|
|
|
et = Writer::kEtFileHdr;
|
2017-10-27 22:14:34 +02:00
|
|
|
} else if (bfile->file_size_ > BlobLogHeader::kSize) {
|
2017-10-17 21:11:52 +02:00
|
|
|
et = Writer::kEtRecord;
|
|
|
|
} else if (bfile->file_size_) {
|
2017-07-13 19:03:39 +02:00
|
|
|
ROCKS_LOG_WARN(db_options_.info_log,
|
2019-04-04 21:05:42 +02:00
|
|
|
"Open blob file: %s with wrong size: %" PRIu64,
|
|
|
|
fpath.c_str(), boffset);
|
2017-05-10 23:54:35 +02:00
|
|
|
return Status::Corruption("Invalid blob file size");
|
|
|
|
}
|
|
|
|
|
|
|
|
bfile->log_writer_ = std::make_shared<Writer>(
|
2017-11-28 20:42:28 +01:00
|
|
|
std::move(fwriter), env_, statistics_, bfile->file_number_,
|
|
|
|
bdb_options_.bytes_per_sync, db_options_.use_fsync, boffset);
|
2017-05-10 23:54:35 +02:00
|
|
|
bfile->log_writer_->last_elem_type_ = et;
|
|
|
|
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
std::shared_ptr<BlobFile> BlobDBImpl::FindBlobFileLocked(
|
2017-08-04 02:46:00 +02:00
|
|
|
uint64_t expiration) const {
|
2018-03-06 20:46:20 +01:00
|
|
|
if (open_ttl_files_.empty()) {
|
|
|
|
return nullptr;
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
|
|
|
std::shared_ptr<BlobFile> tmp = std::make_shared<BlobFile>();
|
2018-03-06 20:46:20 +01:00
|
|
|
tmp->SetHasTTL(true);
|
2017-10-27 22:14:34 +02:00
|
|
|
tmp->expiration_range_ = std::make_pair(expiration, 0);
|
2018-03-06 20:46:20 +01:00
|
|
|
tmp->file_number_ = std::numeric_limits<uint64_t>::max();
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2017-11-01 00:33:55 +01:00
|
|
|
auto citr = open_ttl_files_.equal_range(tmp);
|
|
|
|
if (citr.first == open_ttl_files_.end()) {
|
|
|
|
assert(citr.second == open_ttl_files_.end());
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2017-11-01 00:33:55 +01:00
|
|
|
std::shared_ptr<BlobFile> check = *(open_ttl_files_.rbegin());
|
2018-03-06 20:46:20 +01:00
|
|
|
return (check->expiration_range_.second <= expiration) ? nullptr : check;
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2018-03-06 20:46:20 +01:00
|
|
|
if (citr.first != citr.second) {
|
|
|
|
return *(citr.first);
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
|
|
|
auto finditr = citr.second;
|
2018-03-06 20:46:20 +01:00
|
|
|
if (finditr != open_ttl_files_.begin()) {
|
|
|
|
--finditr;
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2018-03-06 20:46:20 +01:00
|
|
|
bool b2 = (*finditr)->expiration_range_.second <= expiration;
|
2017-10-27 22:14:34 +02:00
|
|
|
bool b1 = (*finditr)->expiration_range_.first > expiration;
|
2017-05-10 23:54:35 +02:00
|
|
|
|
|
|
|
return (b1 || b2) ? nullptr : (*finditr);
|
|
|
|
}
|
|
|
|
|
2018-10-24 00:01:36 +02:00
|
|
|
Status BlobDBImpl::CheckOrCreateWriterLocked(
|
|
|
|
const std::shared_ptr<BlobFile>& blob_file,
|
|
|
|
std::shared_ptr<Writer>* writer) {
|
|
|
|
assert(writer != nullptr);
|
|
|
|
*writer = blob_file->GetWriter();
|
|
|
|
if (*writer != nullptr) {
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
Status s = CreateWriterLocked(blob_file);
|
|
|
|
if (s.ok()) {
|
|
|
|
*writer = blob_file->GetWriter();
|
|
|
|
}
|
|
|
|
return s;
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2018-10-24 00:01:36 +02:00
|
|
|
Status BlobDBImpl::SelectBlobFile(std::shared_ptr<BlobFile>* blob_file) {
|
|
|
|
assert(blob_file != nullptr);
|
2017-05-10 23:54:35 +02:00
|
|
|
{
|
|
|
|
ReadLock rl(&mutex_);
|
2017-11-01 00:33:55 +01:00
|
|
|
if (open_non_ttl_file_ != nullptr) {
|
2018-10-24 00:01:36 +02:00
|
|
|
*blob_file = open_non_ttl_file_;
|
|
|
|
return Status::OK();
|
2017-11-01 00:33:55 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// CHECK again
|
|
|
|
WriteLock wl(&mutex_);
|
|
|
|
if (open_non_ttl_file_ != nullptr) {
|
2018-10-24 00:01:36 +02:00
|
|
|
*blob_file = open_non_ttl_file_;
|
|
|
|
return Status::OK();
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2018-10-24 00:01:36 +02:00
|
|
|
*blob_file = NewBlobFile("SelectBlobFile");
|
|
|
|
assert(*blob_file != nullptr);
|
2017-05-10 23:54:35 +02:00
|
|
|
|
|
|
|
// file not visible, hence no lock
|
2018-10-24 00:01:36 +02:00
|
|
|
std::shared_ptr<Writer> writer;
|
|
|
|
Status s = CheckOrCreateWriterLocked(*blob_file, &writer);
|
|
|
|
if (!s.ok()) {
|
2017-07-13 19:03:39 +02:00
|
|
|
ROCKS_LOG_ERROR(db_options_.info_log,
|
2018-10-24 00:01:36 +02:00
|
|
|
"Failed to get writer from blob file: %s, error: %s",
|
|
|
|
(*blob_file)->PathName().c_str(), s.ToString().c_str());
|
|
|
|
return s;
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2018-10-24 00:01:36 +02:00
|
|
|
(*blob_file)->file_size_ = BlobLogHeader::kSize;
|
|
|
|
(*blob_file)->header_.compression = bdb_options_.compression;
|
|
|
|
(*blob_file)->header_.has_ttl = false;
|
|
|
|
(*blob_file)->header_.column_family_id =
|
2017-11-01 00:27:22 +01:00
|
|
|
reinterpret_cast<ColumnFamilyHandleImpl*>(DefaultColumnFamily())->GetID();
|
2018-10-24 00:01:36 +02:00
|
|
|
(*blob_file)->header_valid_ = true;
|
|
|
|
(*blob_file)->SetColumnFamilyId((*blob_file)->header_.column_family_id);
|
|
|
|
(*blob_file)->SetHasTTL(false);
|
|
|
|
(*blob_file)->SetCompression(bdb_options_.compression);
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2018-10-24 00:01:36 +02:00
|
|
|
s = writer->WriteHeader((*blob_file)->header_);
|
2017-05-10 23:54:35 +02:00
|
|
|
if (!s.ok()) {
|
2017-07-13 19:03:39 +02:00
|
|
|
ROCKS_LOG_ERROR(db_options_.info_log,
|
|
|
|
"Failed to write header to new blob file: %s"
|
|
|
|
" status: '%s'",
|
2018-10-24 00:01:36 +02:00
|
|
|
(*blob_file)->PathName().c_str(), s.ToString().c_str());
|
|
|
|
return s;
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2018-10-24 00:01:36 +02:00
|
|
|
blob_files_.insert(
|
|
|
|
std::make_pair((*blob_file)->BlobFileNumber(), *blob_file));
|
|
|
|
open_non_ttl_file_ = *blob_file;
|
2018-03-06 20:46:20 +01:00
|
|
|
total_blob_size_ += BlobLogHeader::kSize;
|
2018-10-24 00:01:36 +02:00
|
|
|
return s;
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2018-10-24 00:01:36 +02:00
|
|
|
Status BlobDBImpl::SelectBlobFileTTL(uint64_t expiration,
|
|
|
|
std::shared_ptr<BlobFile>* blob_file) {
|
|
|
|
assert(blob_file != nullptr);
|
2017-08-04 02:46:00 +02:00
|
|
|
assert(expiration != kNoExpiration);
|
2017-05-10 23:54:35 +02:00
|
|
|
uint64_t epoch_read = 0;
|
|
|
|
{
|
|
|
|
ReadLock rl(&mutex_);
|
2018-10-24 00:01:36 +02:00
|
|
|
*blob_file = FindBlobFileLocked(expiration);
|
2017-05-10 23:54:35 +02:00
|
|
|
epoch_read = epoch_of_.load();
|
|
|
|
}
|
|
|
|
|
2018-10-24 00:01:36 +02:00
|
|
|
if (*blob_file != nullptr) {
|
|
|
|
assert(!(*blob_file)->Immutable());
|
|
|
|
return Status::OK();
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2017-08-04 02:46:00 +02:00
|
|
|
uint64_t exp_low =
|
2017-05-10 23:54:35 +02:00
|
|
|
(expiration / bdb_options_.ttl_range_secs) * bdb_options_.ttl_range_secs;
|
2017-08-04 02:46:00 +02:00
|
|
|
uint64_t exp_high = exp_low + bdb_options_.ttl_range_secs;
|
2017-10-27 22:14:34 +02:00
|
|
|
ExpirationRange expiration_range = std::make_pair(exp_low, exp_high);
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2018-10-24 00:01:36 +02:00
|
|
|
*blob_file = NewBlobFile("SelectBlobFileTTL");
|
|
|
|
assert(*blob_file != nullptr);
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2019-04-04 21:05:42 +02:00
|
|
|
ROCKS_LOG_INFO(db_options_.info_log,
|
|
|
|
"New blob file TTL range: %s %" PRIu64 " %" PRIu64,
|
2018-10-24 00:01:36 +02:00
|
|
|
(*blob_file)->PathName().c_str(), exp_low, exp_high);
|
2017-05-10 23:54:35 +02:00
|
|
|
LogFlush(db_options_.info_log);
|
|
|
|
|
|
|
|
// we don't need to take lock as no other thread is seeing bfile yet
|
2018-10-24 00:01:36 +02:00
|
|
|
std::shared_ptr<Writer> writer;
|
|
|
|
Status s = CheckOrCreateWriterLocked(*blob_file, &writer);
|
|
|
|
if (!s.ok()) {
|
|
|
|
ROCKS_LOG_ERROR(
|
|
|
|
db_options_.info_log,
|
|
|
|
"Failed to get writer from blob file with TTL: %s, error: %s",
|
|
|
|
(*blob_file)->PathName().c_str(), s.ToString().c_str());
|
|
|
|
return s;
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2018-10-24 00:01:36 +02:00
|
|
|
(*blob_file)->header_.expiration_range = expiration_range;
|
|
|
|
(*blob_file)->header_.compression = bdb_options_.compression;
|
|
|
|
(*blob_file)->header_.has_ttl = true;
|
|
|
|
(*blob_file)->header_.column_family_id =
|
2017-11-01 00:27:22 +01:00
|
|
|
reinterpret_cast<ColumnFamilyHandleImpl*>(DefaultColumnFamily())->GetID();
|
2018-10-24 00:01:36 +02:00
|
|
|
(*blob_file)->header_valid_ = true;
|
|
|
|
(*blob_file)->SetColumnFamilyId((*blob_file)->header_.column_family_id);
|
|
|
|
(*blob_file)->SetHasTTL(true);
|
|
|
|
(*blob_file)->SetCompression(bdb_options_.compression);
|
|
|
|
(*blob_file)->file_size_ = BlobLogHeader::kSize;
|
2017-05-10 23:54:35 +02:00
|
|
|
|
|
|
|
// set the first value of the range, since that is
|
2017-11-01 00:33:55 +01:00
|
|
|
// concrete at this time. also necessary to add to open_ttl_files_
|
2018-10-24 00:01:36 +02:00
|
|
|
(*blob_file)->expiration_range_ = expiration_range;
|
2017-05-10 23:54:35 +02:00
|
|
|
|
|
|
|
WriteLock wl(&mutex_);
|
|
|
|
// in case the epoch has shifted in the interim, then check
|
|
|
|
// check condition again - should be rare.
|
|
|
|
if (epoch_of_.load() != epoch_read) {
|
2018-10-24 00:01:36 +02:00
|
|
|
std::shared_ptr<BlobFile> blob_file2 = FindBlobFileLocked(expiration);
|
|
|
|
if (blob_file2 != nullptr) {
|
|
|
|
*blob_file = std::move(blob_file2);
|
|
|
|
return Status::OK();
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2018-10-24 00:01:36 +02:00
|
|
|
s = writer->WriteHeader((*blob_file)->header_);
|
2017-05-10 23:54:35 +02:00
|
|
|
if (!s.ok()) {
|
2017-07-13 19:03:39 +02:00
|
|
|
ROCKS_LOG_ERROR(db_options_.info_log,
|
|
|
|
"Failed to write header to new blob file: %s"
|
|
|
|
" status: '%s'",
|
2018-10-24 00:01:36 +02:00
|
|
|
(*blob_file)->PathName().c_str(), s.ToString().c_str());
|
|
|
|
return s;
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2018-10-24 00:01:36 +02:00
|
|
|
blob_files_.insert(
|
|
|
|
std::make_pair((*blob_file)->BlobFileNumber(), *blob_file));
|
|
|
|
open_ttl_files_.insert(*blob_file);
|
2018-03-06 20:46:20 +01:00
|
|
|
total_blob_size_ += BlobLogHeader::kSize;
|
2017-05-10 23:54:35 +02:00
|
|
|
epoch_of_++;
|
|
|
|
|
2018-10-24 00:01:36 +02:00
|
|
|
return s;
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
class BlobDBImpl::BlobInserter : public WriteBatch::Handler {
|
|
|
|
private:
|
|
|
|
const WriteOptions& options_;
|
|
|
|
BlobDBImpl* blob_db_impl_;
|
|
|
|
uint32_t default_cf_id_;
|
|
|
|
WriteBatch batch_;
|
2017-05-10 23:54:35 +02:00
|
|
|
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
public:
|
|
|
|
BlobInserter(const WriteOptions& options, BlobDBImpl* blob_db_impl,
|
2017-12-15 22:18:32 +01:00
|
|
|
uint32_t default_cf_id)
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
: options_(options),
|
|
|
|
blob_db_impl_(blob_db_impl),
|
2017-12-15 22:18:32 +01:00
|
|
|
default_cf_id_(default_cf_id) {}
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
|
|
|
|
WriteBatch* batch() { return &batch_; }
|
|
|
|
|
2019-02-14 22:52:47 +01:00
|
|
|
Status PutCF(uint32_t column_family_id, const Slice& key,
|
|
|
|
const Slice& value) override {
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
if (column_family_id != default_cf_id_) {
|
|
|
|
return Status::NotSupported(
|
|
|
|
"Blob DB doesn't support non-default column family.");
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
2018-08-23 19:04:10 +02:00
|
|
|
Status s = blob_db_impl_->PutBlobValue(options_, key, value, kNoExpiration,
|
|
|
|
&batch_);
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
return s;
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2019-02-14 22:52:47 +01:00
|
|
|
Status DeleteCF(uint32_t column_family_id, const Slice& key) override {
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
if (column_family_id != default_cf_id_) {
|
|
|
|
return Status::NotSupported(
|
|
|
|
"Blob DB doesn't support non-default column family.");
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
Status s = WriteBatchInternal::Delete(&batch_, column_family_id, key);
|
|
|
|
return s;
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
virtual Status DeleteRange(uint32_t column_family_id, const Slice& begin_key,
|
|
|
|
const Slice& end_key) {
|
|
|
|
if (column_family_id != default_cf_id_) {
|
|
|
|
return Status::NotSupported(
|
|
|
|
"Blob DB doesn't support non-default column family.");
|
2017-09-08 19:57:12 +02:00
|
|
|
}
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
Status s = WriteBatchInternal::DeleteRange(&batch_, column_family_id,
|
|
|
|
begin_key, end_key);
|
|
|
|
return s;
|
|
|
|
}
|
2017-09-08 19:57:12 +02:00
|
|
|
|
2019-02-14 22:52:47 +01:00
|
|
|
Status SingleDeleteCF(uint32_t /*column_family_id*/,
|
|
|
|
const Slice& /*key*/) override {
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
return Status::NotSupported("Not supported operation in blob db.");
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2019-02-14 22:52:47 +01:00
|
|
|
Status MergeCF(uint32_t /*column_family_id*/, const Slice& /*key*/,
|
|
|
|
const Slice& /*value*/) override {
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
return Status::NotSupported("Not supported operation in blob db.");
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2019-02-14 22:52:47 +01:00
|
|
|
void LogData(const Slice& blob) override { batch_.PutLogData(blob); }
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
};
|
2017-05-10 23:54:35 +02:00
|
|
|
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
Status BlobDBImpl::Write(const WriteOptions& options, WriteBatch* updates) {
|
2017-11-28 20:42:28 +01:00
|
|
|
StopWatch write_sw(env_, statistics_, BLOB_DB_WRITE_MICROS);
|
|
|
|
RecordTick(statistics_, BLOB_DB_NUM_WRITE);
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
uint32_t default_cf_id =
|
|
|
|
reinterpret_cast<ColumnFamilyHandleImpl*>(DefaultColumnFamily())->GetID();
|
2017-11-09 04:33:12 +01:00
|
|
|
Status s;
|
2017-12-15 22:18:32 +01:00
|
|
|
BlobInserter blob_inserter(options, this, default_cf_id);
|
2017-11-09 04:33:12 +01:00
|
|
|
{
|
|
|
|
// Release write_mutex_ before DB write to avoid race condition with
|
|
|
|
// flush begin listener, which also require write_mutex_ to sync
|
|
|
|
// blob files.
|
|
|
|
MutexLock l(&write_mutex_);
|
|
|
|
s = updates->Iterate(&blob_inserter);
|
|
|
|
}
|
2017-10-17 21:11:52 +02:00
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
2018-03-02 21:54:24 +01:00
|
|
|
return db_->Write(options, blob_inserter.batch());
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
Status BlobDBImpl::Put(const WriteOptions& options, const Slice& key,
|
|
|
|
const Slice& value) {
|
2018-08-06 20:43:51 +02:00
|
|
|
return PutUntil(options, key, value, kNoExpiration);
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
}
|
|
|
|
|
2017-05-10 23:54:35 +02:00
|
|
|
Status BlobDBImpl::PutWithTTL(const WriteOptions& options,
|
|
|
|
const Slice& key, const Slice& value,
|
2017-08-04 02:46:00 +02:00
|
|
|
uint64_t ttl) {
|
|
|
|
uint64_t now = EpochNow();
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
uint64_t expiration = kNoExpiration - now > ttl ? now + ttl : kNoExpiration;
|
|
|
|
return PutUntil(options, key, value, expiration);
|
2017-06-14 22:44:36 +02:00
|
|
|
}
|
|
|
|
|
2017-09-08 19:57:12 +02:00
|
|
|
Status BlobDBImpl::PutUntil(const WriteOptions& options, const Slice& key,
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
const Slice& value, uint64_t expiration) {
|
2017-11-28 20:42:28 +01:00
|
|
|
StopWatch write_sw(env_, statistics_, BLOB_DB_WRITE_MICROS);
|
|
|
|
RecordTick(statistics_, BLOB_DB_NUM_PUT);
|
2017-11-02 21:20:28 +01:00
|
|
|
TEST_SYNC_POINT("BlobDBImpl::PutUntil:Start");
|
2017-11-09 04:33:12 +01:00
|
|
|
Status s;
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
WriteBatch batch;
|
2017-11-09 04:33:12 +01:00
|
|
|
{
|
|
|
|
// Release write_mutex_ before DB write to avoid race condition with
|
|
|
|
// flush begin listener, which also require write_mutex_ to sync
|
|
|
|
// blob files.
|
|
|
|
MutexLock l(&write_mutex_);
|
2017-12-15 22:18:32 +01:00
|
|
|
s = PutBlobValue(options, key, value, expiration, &batch);
|
2017-11-09 04:33:12 +01:00
|
|
|
}
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
if (s.ok()) {
|
|
|
|
s = db_->Write(options, &batch);
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
2017-11-02 21:20:28 +01:00
|
|
|
TEST_SYNC_POINT("BlobDBImpl::PutUntil:Finish");
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
return s;
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2018-03-05 22:08:17 +01:00
|
|
|
Status BlobDBImpl::PutBlobValue(const WriteOptions& /*options*/,
|
|
|
|
const Slice& key, const Slice& value,
|
|
|
|
uint64_t expiration, WriteBatch* batch) {
|
2018-03-06 20:46:20 +01:00
|
|
|
write_mutex_.AssertHeld();
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
Status s;
|
|
|
|
std::string index_entry;
|
2017-10-18 02:24:25 +02:00
|
|
|
uint32_t column_family_id =
|
|
|
|
reinterpret_cast<ColumnFamilyHandleImpl*>(DefaultColumnFamily())->GetID();
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
if (value.size() < bdb_options_.min_blob_size) {
|
|
|
|
if (expiration == kNoExpiration) {
|
|
|
|
// Put as normal value
|
|
|
|
s = batch->Put(key, value);
|
2017-11-28 20:42:28 +01:00
|
|
|
RecordTick(statistics_, BLOB_DB_WRITE_INLINED);
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
} else {
|
|
|
|
// Inlined with TTL
|
|
|
|
BlobIndex::EncodeInlinedTTL(&index_entry, expiration, value);
|
|
|
|
s = WriteBatchInternal::PutBlobIndex(batch, column_family_id, key,
|
|
|
|
index_entry);
|
2017-11-28 20:42:28 +01:00
|
|
|
RecordTick(statistics_, BLOB_DB_WRITE_INLINED_TTL);
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
}
|
|
|
|
} else {
|
|
|
|
std::string compression_output;
|
|
|
|
Slice value_compressed = GetCompressedSlice(value, &compression_output);
|
2017-05-10 23:54:35 +02:00
|
|
|
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
std::string headerbuf;
|
2017-10-27 22:14:34 +02:00
|
|
|
Writer::ConstructBlobHeader(&headerbuf, key, value_compressed, expiration);
|
2017-10-18 02:24:25 +02:00
|
|
|
|
2018-03-06 20:46:20 +01:00
|
|
|
// Check DB size limit before selecting blob file to
|
|
|
|
// Since CheckSizeAndEvictBlobFiles() can close blob files, it needs to be
|
|
|
|
// done before calling SelectBlobFile().
|
|
|
|
s = CheckSizeAndEvictBlobFiles(headerbuf.size() + key.size() +
|
|
|
|
value_compressed.size());
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2018-10-24 00:01:36 +02:00
|
|
|
std::shared_ptr<BlobFile> blob_file;
|
|
|
|
if (expiration != kNoExpiration) {
|
|
|
|
s = SelectBlobFileTTL(expiration, &blob_file);
|
2017-11-28 20:42:28 +01:00
|
|
|
} else {
|
2018-10-24 00:01:36 +02:00
|
|
|
s = SelectBlobFile(&blob_file);
|
|
|
|
}
|
|
|
|
if (s.ok()) {
|
|
|
|
assert(blob_file != nullptr);
|
|
|
|
assert(blob_file->compression() == bdb_options_.compression);
|
|
|
|
s = AppendBlob(blob_file, headerbuf, key, value_compressed, expiration,
|
|
|
|
&index_entry);
|
2017-11-28 20:42:28 +01:00
|
|
|
}
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
if (s.ok()) {
|
|
|
|
if (expiration != kNoExpiration) {
|
2019-08-15 01:07:03 +02:00
|
|
|
WriteLock file_lock(&blob_file->mutex_);
|
2018-10-24 00:01:36 +02:00
|
|
|
blob_file->ExtendExpirationRange(expiration);
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
}
|
2018-10-24 00:01:36 +02:00
|
|
|
s = CloseBlobFileIfNeeded(blob_file);
|
|
|
|
}
|
|
|
|
if (s.ok()) {
|
|
|
|
s = WriteBatchInternal::PutBlobIndex(batch, column_family_id, key,
|
|
|
|
index_entry);
|
|
|
|
}
|
|
|
|
if (s.ok()) {
|
|
|
|
if (expiration == kNoExpiration) {
|
|
|
|
RecordTick(statistics_, BLOB_DB_WRITE_BLOB);
|
|
|
|
} else {
|
|
|
|
RecordTick(statistics_, BLOB_DB_WRITE_BLOB_TTL);
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
}
|
|
|
|
} else {
|
2019-04-04 21:05:42 +02:00
|
|
|
ROCKS_LOG_ERROR(
|
|
|
|
db_options_.info_log,
|
|
|
|
"Failed to append blob to FILE: %s: KEY: %s VALSZ: %" ROCKSDB_PRIszt
|
|
|
|
" status: '%s' blob_file: '%s'",
|
|
|
|
blob_file->PathName().c_str(), key.ToString().c_str(), value.size(),
|
|
|
|
s.ToString().c_str(), blob_file->DumpState().c_str());
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
}
|
2017-08-25 19:40:25 +02:00
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2017-11-28 20:42:28 +01:00
|
|
|
RecordTick(statistics_, BLOB_DB_NUM_KEYS_WRITTEN);
|
|
|
|
RecordTick(statistics_, BLOB_DB_BYTES_WRITTEN, key.size() + value.size());
|
2019-02-28 19:14:19 +01:00
|
|
|
RecordInHistogram(statistics_, BLOB_DB_KEY_SIZE, key.size());
|
|
|
|
RecordInHistogram(statistics_, BLOB_DB_VALUE_SIZE, value.size());
|
2017-11-28 20:42:28 +01:00
|
|
|
|
2017-05-10 23:54:35 +02:00
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
Slice BlobDBImpl::GetCompressedSlice(const Slice& raw,
|
|
|
|
std::string* compression_output) const {
|
|
|
|
if (bdb_options_.compression == kNoCompression) {
|
|
|
|
return raw;
|
|
|
|
}
|
2017-11-28 20:42:28 +01:00
|
|
|
StopWatch compression_sw(env_, statistics_, BLOB_DB_COMPRESSION_MICROS);
|
2019-01-19 04:10:17 +01:00
|
|
|
CompressionType type = bdb_options_.compression;
|
|
|
|
CompressionOptions opts;
|
|
|
|
CompressionContext context(type);
|
2019-03-18 20:07:35 +01:00
|
|
|
CompressionInfo info(opts, context, CompressionDict::GetEmptyDict(), type,
|
|
|
|
0 /* sample_for_compression */);
|
|
|
|
CompressBlock(raw, info, &type, kBlockBasedTableVersionFormat, false,
|
|
|
|
compression_output, nullptr, nullptr);
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
return *compression_output;
|
|
|
|
}
|
|
|
|
|
2018-03-06 20:46:20 +01:00
|
|
|
void BlobDBImpl::GetCompactionContext(BlobCompactionContext* context) {
|
|
|
|
ReadLock l(&mutex_);
|
|
|
|
|
|
|
|
context->next_file_number = next_file_number_.load();
|
|
|
|
context->current_blob_files.clear();
|
|
|
|
for (auto& p : blob_files_) {
|
|
|
|
context->current_blob_files.insert(p.first);
|
2018-01-20 01:19:34 +01:00
|
|
|
}
|
2018-03-06 20:46:20 +01:00
|
|
|
context->fifo_eviction_seq = fifo_eviction_seq_;
|
|
|
|
context->evict_expiration_up_to = evict_expiration_up_to_;
|
2017-11-02 20:02:42 +01:00
|
|
|
}
|
|
|
|
|
2018-03-06 20:46:20 +01:00
|
|
|
void BlobDBImpl::UpdateLiveSSTSize() {
|
|
|
|
uint64_t live_sst_size = 0;
|
|
|
|
bool ok = GetIntProperty(DB::Properties::kLiveSstFilesSize, &live_sst_size);
|
|
|
|
if (ok) {
|
|
|
|
live_sst_size_.store(live_sst_size);
|
|
|
|
ROCKS_LOG_INFO(db_options_.info_log,
|
|
|
|
"Updated total SST file size: %" PRIu64 " bytes.",
|
|
|
|
live_sst_size);
|
|
|
|
} else {
|
|
|
|
ROCKS_LOG_ERROR(
|
|
|
|
db_options_.info_log,
|
|
|
|
"Failed to update total SST file size after flush or compaction.");
|
2017-11-02 20:02:42 +01:00
|
|
|
}
|
2018-03-06 20:46:20 +01:00
|
|
|
{
|
|
|
|
// Trigger FIFO eviction if needed.
|
|
|
|
MutexLock l(&write_mutex_);
|
|
|
|
Status s = CheckSizeAndEvictBlobFiles(0, true /*force*/);
|
|
|
|
if (s.IsNoSpace()) {
|
|
|
|
ROCKS_LOG_WARN(db_options_.info_log,
|
|
|
|
"DB grow out-of-space after SST size updated. Current live"
|
|
|
|
" SST size: %" PRIu64
|
|
|
|
" , current blob files size: %" PRIu64 ".",
|
|
|
|
live_sst_size_.load(), total_blob_size_.load());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2017-11-02 20:02:42 +01:00
|
|
|
|
2018-03-06 20:46:20 +01:00
|
|
|
Status BlobDBImpl::CheckSizeAndEvictBlobFiles(uint64_t blob_size,
|
|
|
|
bool force_evict) {
|
|
|
|
write_mutex_.AssertHeld();
|
|
|
|
|
|
|
|
uint64_t live_sst_size = live_sst_size_.load();
|
|
|
|
if (bdb_options_.max_db_size == 0 ||
|
|
|
|
live_sst_size + total_blob_size_.load() + blob_size <=
|
|
|
|
bdb_options_.max_db_size) {
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
if (bdb_options_.is_fifo == false ||
|
|
|
|
(!force_evict && live_sst_size + blob_size > bdb_options_.max_db_size)) {
|
|
|
|
// FIFO eviction is disabled, or no space to insert new blob even we evict
|
|
|
|
// all blob files.
|
|
|
|
return Status::NoSpace(
|
|
|
|
"Write failed, as writing it would exceed max_db_size limit.");
|
|
|
|
}
|
|
|
|
|
|
|
|
std::vector<std::shared_ptr<BlobFile>> candidate_files;
|
2018-06-26 07:32:29 +02:00
|
|
|
CopyBlobFiles(&candidate_files);
|
2018-03-06 20:46:20 +01:00
|
|
|
std::sort(candidate_files.begin(), candidate_files.end(),
|
2018-06-26 07:32:29 +02:00
|
|
|
BlobFileComparator());
|
2018-03-06 20:46:20 +01:00
|
|
|
fifo_eviction_seq_ = GetLatestSequenceNumber();
|
|
|
|
|
|
|
|
WriteLock l(&mutex_);
|
|
|
|
|
|
|
|
while (!candidate_files.empty() &&
|
|
|
|
live_sst_size + total_blob_size_.load() + blob_size >
|
|
|
|
bdb_options_.max_db_size) {
|
|
|
|
std::shared_ptr<BlobFile> blob_file = candidate_files.back();
|
|
|
|
candidate_files.pop_back();
|
|
|
|
WriteLock file_lock(&blob_file->mutex_);
|
|
|
|
if (blob_file->Obsolete()) {
|
|
|
|
// File already obsoleted by someone else.
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
// FIFO eviction can evict open blob files.
|
|
|
|
if (!blob_file->Immutable()) {
|
|
|
|
Status s = CloseBlobFile(blob_file, false /*need_lock*/);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
assert(blob_file->Immutable());
|
|
|
|
auto expiration_range = blob_file->GetExpirationRange();
|
2017-11-02 23:50:30 +01:00
|
|
|
ROCKS_LOG_INFO(db_options_.info_log,
|
|
|
|
"Evict oldest blob file since DB out of space. Current "
|
2018-03-06 20:46:20 +01:00
|
|
|
"live SST file size: %" PRIu64 ", total blob size: %" PRIu64
|
|
|
|
", max db size: %" PRIu64 ", evicted blob file #%" PRIu64
|
2018-06-26 07:32:29 +02:00
|
|
|
".",
|
2018-03-06 20:46:20 +01:00
|
|
|
live_sst_size, total_blob_size_.load(),
|
2018-06-26 07:32:29 +02:00
|
|
|
bdb_options_.max_db_size, blob_file->BlobFileNumber());
|
2018-03-06 20:46:20 +01:00
|
|
|
ObsoleteBlobFile(blob_file, fifo_eviction_seq_, true /*update_size*/);
|
|
|
|
evict_expiration_up_to_ = expiration_range.first;
|
2017-11-28 20:42:28 +01:00
|
|
|
RecordTick(statistics_, BLOB_DB_FIFO_NUM_FILES_EVICTED);
|
|
|
|
RecordTick(statistics_, BLOB_DB_FIFO_NUM_KEYS_EVICTED,
|
2018-03-06 20:46:20 +01:00
|
|
|
blob_file->BlobCount());
|
2017-11-28 20:42:28 +01:00
|
|
|
RecordTick(statistics_, BLOB_DB_FIFO_BYTES_EVICTED,
|
2018-03-06 20:46:20 +01:00
|
|
|
blob_file->GetFileSize());
|
2018-01-20 01:19:34 +01:00
|
|
|
TEST_SYNC_POINT("BlobDBImpl::EvictOldestBlobFile:Evicted");
|
2017-11-02 23:50:30 +01:00
|
|
|
}
|
2018-03-06 20:46:20 +01:00
|
|
|
if (live_sst_size + total_blob_size_.load() + blob_size >
|
|
|
|
bdb_options_.max_db_size) {
|
|
|
|
return Status::NoSpace(
|
|
|
|
"Write failed, as writing it would exceed max_db_size limit.");
|
2017-11-02 20:02:42 +01:00
|
|
|
}
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
2017-05-10 23:54:35 +02:00
|
|
|
Status BlobDBImpl::AppendBlob(const std::shared_ptr<BlobFile>& bfile,
|
|
|
|
const std::string& headerbuf, const Slice& key,
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
const Slice& value, uint64_t expiration,
|
|
|
|
std::string* index_entry) {
|
2018-03-06 20:46:20 +01:00
|
|
|
Status s;
|
2017-05-10 23:54:35 +02:00
|
|
|
uint64_t blob_offset = 0;
|
|
|
|
uint64_t key_offset = 0;
|
|
|
|
{
|
|
|
|
WriteLock lockbfile_w(&bfile->mutex_);
|
2018-10-24 00:01:36 +02:00
|
|
|
std::shared_ptr<Writer> writer;
|
|
|
|
s = CheckOrCreateWriterLocked(bfile, &writer);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
2018-03-06 20:46:20 +01:00
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
|
|
|
// write the blob to the blob log.
|
|
|
|
s = writer->EmitPhysicalRecord(headerbuf, key, value, &key_offset,
|
|
|
|
&blob_offset);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!s.ok()) {
|
2017-07-13 19:03:39 +02:00
|
|
|
ROCKS_LOG_ERROR(db_options_.info_log,
|
|
|
|
"Invalid status in AppendBlob: %s status: '%s'",
|
|
|
|
bfile->PathName().c_str(), s.ToString().c_str());
|
2017-05-10 23:54:35 +02:00
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
// increment blob count
|
|
|
|
bfile->blob_count_++;
|
|
|
|
|
2018-03-06 20:46:20 +01:00
|
|
|
uint64_t size_put = headerbuf.size() + key.size() + value.size();
|
2017-05-10 23:54:35 +02:00
|
|
|
bfile->file_size_ += size_put;
|
2018-03-06 20:46:20 +01:00
|
|
|
total_blob_size_ += size_put;
|
2017-05-10 23:54:35 +02:00
|
|
|
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
if (expiration == kNoExpiration) {
|
|
|
|
BlobIndex::EncodeBlob(index_entry, bfile->BlobFileNumber(), blob_offset,
|
|
|
|
value.size(), bdb_options_.compression);
|
|
|
|
} else {
|
|
|
|
BlobIndex::EncodeBlobTTL(index_entry, expiration, bfile->BlobFileNumber(),
|
|
|
|
blob_offset, value.size(),
|
|
|
|
bdb_options_.compression);
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
std::vector<Status> BlobDBImpl::MultiGet(
|
2017-08-21 03:12:38 +02:00
|
|
|
const ReadOptions& read_options,
|
2017-05-10 23:54:35 +02:00
|
|
|
const std::vector<Slice>& keys, std::vector<std::string>* values) {
|
2017-11-28 20:42:28 +01:00
|
|
|
StopWatch multiget_sw(env_, statistics_, BLOB_DB_MULTIGET_MICROS);
|
|
|
|
RecordTick(statistics_, BLOB_DB_NUM_MULTIGET);
|
2017-08-21 03:12:38 +02:00
|
|
|
// Get a snapshot to avoid blob file get deleted between we
|
|
|
|
// fetch and index entry and reading from the file.
|
|
|
|
ReadOptions ro(read_options);
|
|
|
|
bool snapshot_created = SetSnapshotIfNeeded(&ro);
|
2017-10-18 02:24:25 +02:00
|
|
|
|
|
|
|
std::vector<Status> statuses;
|
|
|
|
statuses.reserve(keys.size());
|
|
|
|
values->clear();
|
|
|
|
values->reserve(keys.size());
|
|
|
|
PinnableSlice value;
|
|
|
|
for (size_t i = 0; i < keys.size(); i++) {
|
|
|
|
statuses.push_back(Get(ro, DefaultColumnFamily(), keys[i], &value));
|
|
|
|
values->push_back(value.ToString());
|
|
|
|
value.Reset();
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
2017-08-21 03:12:38 +02:00
|
|
|
if (snapshot_created) {
|
|
|
|
db_->ReleaseSnapshot(ro.snapshot);
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
return statuses;
|
|
|
|
}
|
|
|
|
|
2017-08-21 03:12:38 +02:00
|
|
|
bool BlobDBImpl::SetSnapshotIfNeeded(ReadOptions* read_options) {
|
|
|
|
assert(read_options != nullptr);
|
|
|
|
if (read_options->snapshot != nullptr) {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
read_options->snapshot = db_->GetSnapshot();
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2017-10-18 02:24:25 +02:00
|
|
|
Status BlobDBImpl::GetBlobValue(const Slice& key, const Slice& index_entry,
|
2018-08-07 02:28:40 +02:00
|
|
|
PinnableSlice* value, uint64_t* expiration) {
|
2017-10-17 21:11:52 +02:00
|
|
|
assert(value != nullptr);
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
BlobIndex blob_index;
|
|
|
|
Status s = blob_index.DecodeFrom(index_entry);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
if (blob_index.HasTTL() && blob_index.expiration() <= EpochNow()) {
|
|
|
|
return Status::NotFound("Key expired");
|
|
|
|
}
|
2018-08-07 02:28:40 +02:00
|
|
|
if (expiration != nullptr) {
|
|
|
|
if (blob_index.HasTTL()) {
|
|
|
|
*expiration = blob_index.expiration();
|
|
|
|
} else {
|
|
|
|
*expiration = kNoExpiration;
|
|
|
|
}
|
|
|
|
}
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
if (blob_index.IsInlined()) {
|
|
|
|
// TODO(yiwu): If index_entry is a PinnableSlice, we can also pin the same
|
|
|
|
// memory buffer to avoid extra copy.
|
|
|
|
value->PinSelf(blob_index.value());
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
if (blob_index.size() == 0) {
|
|
|
|
value->PinSelf("");
|
|
|
|
return Status::OK();
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
|
|
|
// offset has to have certain min, as we will read CRC
|
|
|
|
// later from the Blob Header, which needs to be also a
|
|
|
|
// valid offset.
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
if (blob_index.offset() <
|
2017-10-27 22:14:34 +02:00
|
|
|
(BlobLogHeader::kSize + BlobLogRecord::kHeaderSize + key.size())) {
|
2017-05-10 23:54:35 +02:00
|
|
|
if (debug_level_ >= 2) {
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
ROCKS_LOG_ERROR(db_options_.info_log,
|
|
|
|
"Invalid blob index file_number: %" PRIu64
|
|
|
|
" blob_offset: %" PRIu64 " blob_size: %" PRIu64
|
|
|
|
" key: %s",
|
|
|
|
blob_index.file_number(), blob_index.offset(),
|
|
|
|
blob_index.size(), key.data());
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
return Status::NotFound("Invalid blob offset");
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
std::shared_ptr<BlobFile> bfile;
|
|
|
|
{
|
|
|
|
ReadLock rl(&mutex_);
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
auto hitr = blob_files_.find(blob_index.file_number());
|
2017-05-10 23:54:35 +02:00
|
|
|
|
|
|
|
// file was deleted
|
|
|
|
if (hitr == blob_files_.end()) {
|
|
|
|
return Status::NotFound("Blob Not Found as blob file missing");
|
|
|
|
}
|
|
|
|
|
|
|
|
bfile = hitr->second;
|
|
|
|
}
|
|
|
|
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
if (blob_index.size() == 0 && value != nullptr) {
|
2017-10-18 02:24:25 +02:00
|
|
|
value->PinSelf("");
|
2017-05-10 23:54:35 +02:00
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
// takes locks when called
|
2018-09-21 01:50:07 +02:00
|
|
|
std::shared_ptr<RandomAccessFileReader> reader;
|
|
|
|
s = GetBlobFileReader(bfile, &reader);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2018-01-06 01:35:49 +01:00
|
|
|
assert(blob_index.offset() > key.size() + sizeof(uint32_t));
|
|
|
|
uint64_t record_offset = blob_index.offset() - key.size() - sizeof(uint32_t);
|
|
|
|
uint64_t record_size = sizeof(uint32_t) + key.size() + blob_index.size();
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2017-10-27 22:14:34 +02:00
|
|
|
// Allocate the buffer. This is safe in C++11
|
2018-09-06 03:07:53 +02:00
|
|
|
std::string buffer_str(static_cast<size_t>(record_size), static_cast<char>(0));
|
2018-01-06 01:35:49 +01:00
|
|
|
char* buffer = &buffer_str[0];
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2018-01-06 01:35:49 +01:00
|
|
|
// A partial blob record contain checksum, key and value.
|
|
|
|
Slice blob_record;
|
2017-11-28 20:42:28 +01:00
|
|
|
{
|
|
|
|
StopWatch read_sw(env_, statistics_, BLOB_DB_BLOB_FILE_READ_MICROS);
|
2018-09-06 03:07:53 +02:00
|
|
|
s = reader->Read(record_offset, static_cast<size_t>(record_size), &blob_record, buffer);
|
2018-01-06 01:35:49 +01:00
|
|
|
RecordTick(statistics_, BLOB_DB_BLOB_FILE_BYTES_READ, blob_record.size());
|
2017-11-28 20:42:28 +01:00
|
|
|
}
|
2018-01-06 01:35:49 +01:00
|
|
|
if (!s.ok()) {
|
|
|
|
ROCKS_LOG_DEBUG(db_options_.info_log,
|
|
|
|
"Failed to read blob from blob file %" PRIu64
|
|
|
|
", blob_offset: %" PRIu64 ", blob_size: %" PRIu64
|
2019-04-04 21:05:42 +02:00
|
|
|
", key_size: %" ROCKSDB_PRIszt ", status: '%s'",
|
2018-01-06 01:35:49 +01:00
|
|
|
bfile->BlobFileNumber(), blob_index.offset(),
|
|
|
|
blob_index.size(), key.size(), s.ToString().c_str());
|
|
|
|
return s;
|
2017-10-17 21:11:52 +02:00
|
|
|
}
|
2018-01-06 01:35:49 +01:00
|
|
|
if (blob_record.size() != record_size) {
|
2019-04-04 21:05:42 +02:00
|
|
|
ROCKS_LOG_DEBUG(
|
|
|
|
db_options_.info_log,
|
|
|
|
"Failed to read blob from blob file %" PRIu64 ", blob_offset: %" PRIu64
|
|
|
|
", blob_size: %" PRIu64 ", key_size: %" ROCKSDB_PRIszt
|
|
|
|
", read %" ROCKSDB_PRIszt " bytes, expected %" PRIu64 " bytes",
|
|
|
|
bfile->BlobFileNumber(), blob_index.offset(), blob_index.size(),
|
|
|
|
key.size(), blob_record.size(), record_size);
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2018-01-06 01:35:49 +01:00
|
|
|
return Status::Corruption("Failed to retrieve blob from blob index.");
|
2017-10-17 21:11:52 +02:00
|
|
|
}
|
2018-01-06 01:35:49 +01:00
|
|
|
Slice crc_slice(blob_record.data(), sizeof(uint32_t));
|
|
|
|
Slice blob_value(blob_record.data() + sizeof(uint32_t) + key.size(),
|
2018-09-06 03:07:53 +02:00
|
|
|
static_cast<size_t>(blob_index.size()));
|
2018-01-06 01:35:49 +01:00
|
|
|
uint32_t crc_exp;
|
|
|
|
if (!GetFixed32(&crc_slice, &crc_exp)) {
|
|
|
|
ROCKS_LOG_DEBUG(db_options_.info_log,
|
|
|
|
"Unable to decode CRC from blob file %" PRIu64
|
|
|
|
", blob_offset: %" PRIu64 ", blob_size: %" PRIu64
|
2019-04-04 21:05:42 +02:00
|
|
|
", key size: %" ROCKSDB_PRIszt ", status: '%s'",
|
2018-01-06 01:35:49 +01:00
|
|
|
bfile->BlobFileNumber(), blob_index.offset(),
|
|
|
|
blob_index.size(), key.size(), s.ToString().c_str());
|
|
|
|
return Status::Corruption("Unable to decode checksum.");
|
|
|
|
}
|
|
|
|
|
|
|
|
uint32_t crc = crc32c::Value(blob_record.data() + sizeof(uint32_t),
|
|
|
|
blob_record.size() - sizeof(uint32_t));
|
2017-10-17 21:11:52 +02:00
|
|
|
crc = crc32c::Mask(crc); // Adjust for storage
|
|
|
|
if (crc != crc_exp) {
|
|
|
|
if (debug_level_ >= 2) {
|
|
|
|
ROCKS_LOG_ERROR(db_options_.info_log,
|
|
|
|
"Blob crc mismatch file: %s blob_offset: %" PRIu64
|
|
|
|
" blob_size: %" PRIu64 " key: %s status: '%s'",
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
bfile->PathName().c_str(), blob_index.offset(),
|
|
|
|
blob_index.size(), key.data(), s.ToString().c_str());
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
2017-10-17 21:11:52 +02:00
|
|
|
return Status::Corruption("Corruption. Blob CRC mismatch");
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2018-01-06 01:35:49 +01:00
|
|
|
if (bfile->compression() == kNoCompression) {
|
|
|
|
value->PinSelf(blob_value);
|
|
|
|
} else {
|
2017-10-17 21:11:52 +02:00
|
|
|
BlockContents contents;
|
|
|
|
auto cfh = reinterpret_cast<ColumnFamilyHandleImpl*>(DefaultColumnFamily());
|
2017-11-28 20:42:28 +01:00
|
|
|
{
|
|
|
|
StopWatch decompression_sw(env_, statistics_,
|
|
|
|
BLOB_DB_DECOMPRESSION_MICROS);
|
2019-01-19 04:10:17 +01:00
|
|
|
UncompressionContext context(bfile->compression());
|
|
|
|
UncompressionInfo info(context, UncompressionDict::GetEmptyDict(),
|
|
|
|
bfile->compression());
|
2017-11-28 20:42:28 +01:00
|
|
|
s = UncompressBlockContentsForCompressionType(
|
2019-01-19 04:10:17 +01:00
|
|
|
info, blob_value.data(), blob_value.size(), &contents,
|
2018-06-05 21:51:05 +02:00
|
|
|
kBlockBasedTableVersionFormat, *(cfh->cfd()->ioptions()));
|
2017-11-28 20:42:28 +01:00
|
|
|
}
|
2018-01-06 01:35:49 +01:00
|
|
|
value->PinSelf(contents.data);
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2017-08-21 03:12:38 +02:00
|
|
|
Status BlobDBImpl::Get(const ReadOptions& read_options,
|
2017-05-10 23:54:35 +02:00
|
|
|
ColumnFamilyHandle* column_family, const Slice& key,
|
2017-07-13 02:56:40 +02:00
|
|
|
PinnableSlice* value) {
|
2018-08-07 02:28:40 +02:00
|
|
|
return Get(read_options, column_family, key, value, nullptr /*expiration*/);
|
|
|
|
}
|
|
|
|
|
|
|
|
Status BlobDBImpl::Get(const ReadOptions& read_options,
|
|
|
|
ColumnFamilyHandle* column_family, const Slice& key,
|
|
|
|
PinnableSlice* value, uint64_t* expiration) {
|
2017-11-28 20:42:28 +01:00
|
|
|
StopWatch get_sw(env_, statistics_, BLOB_DB_GET_MICROS);
|
|
|
|
RecordTick(statistics_, BLOB_DB_NUM_GET);
|
2018-08-07 02:28:40 +02:00
|
|
|
return GetImpl(read_options, column_family, key, value, expiration);
|
2017-11-28 20:42:28 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
Status BlobDBImpl::GetImpl(const ReadOptions& read_options,
|
|
|
|
ColumnFamilyHandle* column_family, const Slice& key,
|
2018-08-07 02:28:40 +02:00
|
|
|
PinnableSlice* value, uint64_t* expiration) {
|
2017-09-08 19:57:12 +02:00
|
|
|
if (column_family != DefaultColumnFamily()) {
|
|
|
|
return Status::NotSupported(
|
|
|
|
"Blob DB doesn't support non-default column family.");
|
|
|
|
}
|
2017-08-21 03:12:38 +02:00
|
|
|
// Get a snapshot to avoid blob file get deleted between we
|
|
|
|
// fetch and index entry and reading from the file.
|
|
|
|
// TODO(yiwu): For Get() retry if file not found would be a simpler strategy.
|
|
|
|
ReadOptions ro(read_options);
|
|
|
|
bool snapshot_created = SetSnapshotIfNeeded(&ro);
|
|
|
|
|
2018-08-28 01:23:48 +02:00
|
|
|
PinnableSlice index_entry;
|
2017-05-10 23:54:35 +02:00
|
|
|
Status s;
|
2017-10-18 02:24:25 +02:00
|
|
|
bool is_blob_index = false;
|
New API to get all merge operands for a Key (#5604)
Summary:
This is a new API added to db.h to allow for fetching all merge operands associated with a Key. The main motivation for this API is to support use cases where doing a full online merge is not necessary as it is performance sensitive. Example use-cases:
1. Update subset of columns and read subset of columns -
Imagine a SQL Table, a row is encoded as a K/V pair (as it is done in MyRocks). If there are many columns and users only updated one of them, we can use merge operator to reduce write amplification. While users only read one or two columns in the read query, this feature can avoid a full merging of the whole row, and save some CPU.
2. Updating very few attributes in a value which is a JSON-like document -
Updating one attribute can be done efficiently using merge operator, while reading back one attribute can be done more efficiently if we don't need to do a full merge.
----------------------------------------------------------------------------------------------------
API :
Status GetMergeOperands(
const ReadOptions& options, ColumnFamilyHandle* column_family,
const Slice& key, PinnableSlice* merge_operands,
GetMergeOperandsOptions* get_merge_operands_options,
int* number_of_operands)
Example usage :
int size = 100;
int number_of_operands = 0;
std::vector<PinnableSlice> values(size);
GetMergeOperandsOptions merge_operands_info;
db_->GetMergeOperands(ReadOptions(), db_->DefaultColumnFamily(), "k1", values.data(), merge_operands_info, &number_of_operands);
Description :
Returns all the merge operands corresponding to the key. If the number of merge operands in DB is greater than merge_operands_options.expected_max_number_of_operands no merge operands are returned and status is Incomplete. Merge operands returned are in the order of insertion.
merge_operands-> Points to an array of at-least merge_operands_options.expected_max_number_of_operands and the caller is responsible for allocating it. If the status returned is Incomplete then number_of_operands will contain the total number of merge operands found in DB for key.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5604
Test Plan:
Added unit test and perf test in db_bench that can be run using the command:
./db_bench -benchmarks=getmergeoperands --merge_operator=sortlist
Differential Revision: D16657366
Pulled By: vjnadimpalli
fbshipit-source-id: 0faadd752351745224ee12d4ae9ef3cb529951bf
2019-08-06 23:22:34 +02:00
|
|
|
DBImpl::GetImplOptions get_impl_options;
|
|
|
|
get_impl_options.column_family = column_family;
|
|
|
|
get_impl_options.value = &index_entry;
|
|
|
|
get_impl_options.is_blob_index = &is_blob_index;
|
|
|
|
s = db_impl_->GetImpl(ro, key, get_impl_options);
|
2017-08-21 03:12:38 +02:00
|
|
|
TEST_SYNC_POINT("BlobDBImpl::Get:AfterIndexEntryGet:1");
|
|
|
|
TEST_SYNC_POINT("BlobDBImpl::Get:AfterIndexEntryGet:2");
|
2018-08-07 02:28:40 +02:00
|
|
|
if (expiration != nullptr) {
|
|
|
|
*expiration = kNoExpiration;
|
|
|
|
}
|
2018-08-28 01:23:48 +02:00
|
|
|
RecordTick(statistics_, BLOB_DB_NUM_KEYS_READ);
|
|
|
|
if (s.ok()) {
|
|
|
|
if (is_blob_index) {
|
|
|
|
s = GetBlobValue(key, index_entry, value, expiration);
|
|
|
|
} else {
|
|
|
|
// The index entry is the value itself in this case.
|
|
|
|
value->PinSelf(index_entry);
|
|
|
|
}
|
|
|
|
RecordTick(statistics_, BLOB_DB_BYTES_READ, value->size());
|
2017-08-21 03:12:38 +02:00
|
|
|
}
|
|
|
|
if (snapshot_created) {
|
|
|
|
db_->ReleaseSnapshot(ro.snapshot);
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
std::pair<bool, int64_t> BlobDBImpl::SanityCheck(bool aborted) {
|
2018-08-30 20:43:39 +02:00
|
|
|
if (aborted) {
|
|
|
|
return std::make_pair(false, -1);
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2019-08-15 01:07:03 +02:00
|
|
|
ReadLock rl(&mutex_);
|
|
|
|
|
2017-07-13 19:03:39 +02:00
|
|
|
ROCKS_LOG_INFO(db_options_.info_log, "Starting Sanity Check");
|
2019-04-04 21:05:42 +02:00
|
|
|
ROCKS_LOG_INFO(db_options_.info_log, "Number of files %" ROCKSDB_PRIszt,
|
2017-07-13 19:03:39 +02:00
|
|
|
blob_files_.size());
|
2019-04-04 21:05:42 +02:00
|
|
|
ROCKS_LOG_INFO(db_options_.info_log, "Number of open files %" ROCKSDB_PRIszt,
|
2017-11-01 00:33:55 +01:00
|
|
|
open_ttl_files_.size());
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2017-11-01 00:33:55 +01:00
|
|
|
for (auto bfile : open_ttl_files_) {
|
2017-05-10 23:54:35 +02:00
|
|
|
assert(!bfile->Immutable());
|
|
|
|
}
|
|
|
|
|
2018-08-30 20:43:39 +02:00
|
|
|
uint64_t now = EpochNow();
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2018-08-30 20:43:39 +02:00
|
|
|
for (auto blob_file_pair : blob_files_) {
|
|
|
|
auto blob_file = blob_file_pair.second;
|
|
|
|
char buf[1000];
|
|
|
|
int pos = snprintf(buf, sizeof(buf),
|
|
|
|
"Blob file %" PRIu64 ", size %" PRIu64
|
|
|
|
", blob count %" PRIu64 ", immutable %d",
|
|
|
|
blob_file->BlobFileNumber(), blob_file->GetFileSize(),
|
|
|
|
blob_file->BlobCount(), blob_file->Immutable());
|
|
|
|
if (blob_file->HasTTL()) {
|
2019-08-15 01:07:03 +02:00
|
|
|
ExpirationRange expiration_range;
|
|
|
|
|
|
|
|
{
|
|
|
|
ReadLock file_lock(&blob_file->mutex_);
|
|
|
|
expiration_range = blob_file->GetExpirationRange();
|
|
|
|
}
|
|
|
|
|
2018-08-30 20:43:39 +02:00
|
|
|
pos += snprintf(buf + pos, sizeof(buf) - pos,
|
|
|
|
", expiration range (%" PRIu64 ", %" PRIu64 ")",
|
|
|
|
expiration_range.first, expiration_range.second);
|
|
|
|
if (!blob_file->Obsolete()) {
|
|
|
|
pos += snprintf(buf + pos, sizeof(buf) - pos,
|
|
|
|
", expire in %" PRIu64 " seconds",
|
|
|
|
expiration_range.second - now);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (blob_file->Obsolete()) {
|
|
|
|
pos += snprintf(buf + pos, sizeof(buf) - pos, ", obsolete at %" PRIu64,
|
|
|
|
blob_file->GetObsoleteSequence());
|
|
|
|
}
|
|
|
|
snprintf(buf + pos, sizeof(buf) - pos, ".");
|
|
|
|
ROCKS_LOG_INFO(db_options_.info_log, "%s", buf);
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
// reschedule
|
|
|
|
return std::make_pair(true, -1);
|
|
|
|
}
|
|
|
|
|
2018-03-06 20:46:20 +01:00
|
|
|
Status BlobDBImpl::CloseBlobFile(std::shared_ptr<BlobFile> bfile,
|
|
|
|
bool need_lock) {
|
2017-11-01 00:33:55 +01:00
|
|
|
assert(bfile != nullptr);
|
2018-03-06 20:46:20 +01:00
|
|
|
write_mutex_.AssertHeld();
|
2017-08-25 19:40:25 +02:00
|
|
|
Status s;
|
2017-11-21 06:06:30 +01:00
|
|
|
ROCKS_LOG_INFO(db_options_.info_log,
|
|
|
|
"Closing blob file %" PRIu64 ". Path: %s",
|
|
|
|
bfile->BlobFileNumber(), bfile->PathName().c_str());
|
2017-05-10 23:54:35 +02:00
|
|
|
{
|
2018-03-06 20:46:20 +01:00
|
|
|
std::unique_ptr<WriteLock> lock;
|
|
|
|
if (need_lock) {
|
|
|
|
lock.reset(new WriteLock(&mutex_));
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2017-08-25 19:40:25 +02:00
|
|
|
if (bfile->HasTTL()) {
|
2017-10-23 23:20:53 +02:00
|
|
|
size_t erased __attribute__((__unused__));
|
2017-11-01 00:33:55 +01:00
|
|
|
erased = open_ttl_files_.erase(bfile);
|
2018-03-22 00:16:05 +01:00
|
|
|
} else if (bfile == open_non_ttl_file_) {
|
2017-11-01 00:33:55 +01:00
|
|
|
open_non_ttl_file_ = nullptr;
|
2017-08-25 19:40:25 +02:00
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
if (!bfile->closed_.load()) {
|
2018-03-06 20:46:20 +01:00
|
|
|
std::unique_ptr<WriteLock> file_lock;
|
|
|
|
if (need_lock) {
|
|
|
|
file_lock.reset(new WriteLock(&bfile->mutex_));
|
|
|
|
}
|
2017-08-25 19:40:25 +02:00
|
|
|
s = bfile->WriteFooterAndCloseLocked();
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2018-03-06 20:46:20 +01:00
|
|
|
if (s.ok()) {
|
|
|
|
total_blob_size_ += BlobLogFooter::kSize;
|
|
|
|
} else {
|
2017-08-25 19:40:25 +02:00
|
|
|
ROCKS_LOG_ERROR(db_options_.info_log,
|
|
|
|
"Failed to close blob file %" PRIu64 "with error: %s",
|
|
|
|
bfile->BlobFileNumber(), s.ToString().c_str());
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2017-08-25 19:40:25 +02:00
|
|
|
return s;
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2017-08-25 19:40:25 +02:00
|
|
|
Status BlobDBImpl::CloseBlobFileIfNeeded(std::shared_ptr<BlobFile>& bfile) {
|
|
|
|
// atomic read
|
|
|
|
if (bfile->GetFileSize() < bdb_options_.blob_file_size) {
|
|
|
|
return Status::OK();
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
2017-08-25 19:40:25 +02:00
|
|
|
return CloseBlobFile(bfile);
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2018-03-06 20:46:20 +01:00
|
|
|
void BlobDBImpl::ObsoleteBlobFile(std::shared_ptr<BlobFile> blob_file,
|
|
|
|
SequenceNumber obsolete_seq,
|
|
|
|
bool update_size) {
|
|
|
|
// Should hold write lock of mutex_ or during DB open.
|
|
|
|
blob_file->MarkObsolete(obsolete_seq);
|
|
|
|
obsolete_files_.push_back(blob_file);
|
|
|
|
assert(total_blob_size_.load() >= blob_file->GetFileSize());
|
|
|
|
if (update_size) {
|
|
|
|
total_blob_size_ -= blob_file->GetFileSize();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-11-02 23:50:30 +01:00
|
|
|
bool BlobDBImpl::VisibleToActiveSnapshot(
|
2017-05-10 23:54:35 +02:00
|
|
|
const std::shared_ptr<BlobFile>& bfile) {
|
|
|
|
assert(bfile->Obsolete());
|
2017-12-15 22:18:32 +01:00
|
|
|
|
|
|
|
// We check whether the oldest snapshot is no less than the last sequence
|
|
|
|
// by the time the blob file become obsolete. If so, the blob file is not
|
|
|
|
// visible to all existing snapshots.
|
|
|
|
//
|
|
|
|
// If we keep track of the earliest sequence of the keys in the blob file,
|
|
|
|
// we could instead check if there's a snapshot falls in range
|
|
|
|
// [earliest_sequence, obsolete_sequence). But doing so will make the
|
|
|
|
// implementation more complicated.
|
2017-11-02 23:50:30 +01:00
|
|
|
SequenceNumber obsolete_sequence = bfile->GetObsoleteSequence();
|
2018-08-07 01:45:02 +02:00
|
|
|
SequenceNumber oldest_snapshot = kMaxSequenceNumber;
|
2017-12-15 22:18:32 +01:00
|
|
|
{
|
|
|
|
// Need to lock DBImpl mutex before access snapshot list.
|
|
|
|
InstrumentedMutexLock l(db_impl_->mutex());
|
2018-08-07 01:45:02 +02:00
|
|
|
auto& snapshots = db_impl_->snapshots();
|
2017-12-15 22:18:32 +01:00
|
|
|
if (!snapshots.empty()) {
|
|
|
|
oldest_snapshot = snapshots.oldest()->GetSequenceNumber();
|
|
|
|
}
|
|
|
|
}
|
2018-08-30 20:43:39 +02:00
|
|
|
bool visible = oldest_snapshot < obsolete_sequence;
|
|
|
|
if (visible) {
|
|
|
|
ROCKS_LOG_INFO(db_options_.info_log,
|
|
|
|
"Obsolete blob file %" PRIu64 " (obsolete at %" PRIu64
|
|
|
|
") visible to oldest snapshot %" PRIu64 ".",
|
|
|
|
bfile->BlobFileNumber(), obsolete_sequence, oldest_snapshot);
|
|
|
|
}
|
|
|
|
return visible;
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2018-08-21 07:38:34 +02:00
|
|
|
std::pair<bool, int64_t> BlobDBImpl::EvictExpiredFiles(bool aborted) {
|
|
|
|
if (aborted) {
|
|
|
|
return std::make_pair(false, -1);
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2019-02-22 02:23:05 +01:00
|
|
|
TEST_SYNC_POINT("BlobDBImpl::EvictExpiredFiles:0");
|
|
|
|
TEST_SYNC_POINT("BlobDBImpl::EvictExpiredFiles:1");
|
|
|
|
|
2017-05-10 23:54:35 +02:00
|
|
|
std::vector<std::shared_ptr<BlobFile>> process_files;
|
2018-08-21 07:38:34 +02:00
|
|
|
uint64_t now = EpochNow();
|
2017-05-10 23:54:35 +02:00
|
|
|
{
|
|
|
|
ReadLock rl(&mutex_);
|
2018-08-21 07:38:34 +02:00
|
|
|
for (auto p : blob_files_) {
|
|
|
|
auto& blob_file = p.second;
|
|
|
|
ReadLock file_lock(&blob_file->mutex_);
|
|
|
|
if (blob_file->HasTTL() && !blob_file->Obsolete() &&
|
|
|
|
blob_file->GetExpirationRange().second <= now) {
|
|
|
|
process_files.push_back(blob_file);
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2019-02-22 02:23:05 +01:00
|
|
|
TEST_SYNC_POINT("BlobDBImpl::EvictExpiredFiles:2");
|
|
|
|
TEST_SYNC_POINT("BlobDBImpl::EvictExpiredFiles:3");
|
|
|
|
TEST_SYNC_POINT_CALLBACK("BlobDBImpl::EvictExpiredFiles:cb", nullptr);
|
|
|
|
|
2018-08-21 07:38:34 +02:00
|
|
|
SequenceNumber seq = GetLatestSequenceNumber();
|
|
|
|
{
|
|
|
|
MutexLock l(&write_mutex_);
|
|
|
|
for (auto& blob_file : process_files) {
|
|
|
|
WriteLock file_lock(&blob_file->mutex_);
|
|
|
|
if (!blob_file->Immutable()) {
|
|
|
|
CloseBlobFile(blob_file, false /*need_lock*/);
|
|
|
|
}
|
|
|
|
// Need to double check if the file is obsolete.
|
|
|
|
if (!blob_file->Obsolete()) {
|
|
|
|
ObsoleteBlobFile(blob_file, seq, true /*update_size*/);
|
|
|
|
}
|
|
|
|
}
|
2017-08-21 01:56:01 +02:00
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
|
|
|
return std::make_pair(true, -1);
|
|
|
|
}
|
|
|
|
|
2017-12-11 21:01:22 +01:00
|
|
|
Status BlobDBImpl::SyncBlobFiles() {
|
2017-08-04 22:09:56 +02:00
|
|
|
MutexLock l(&write_mutex_);
|
|
|
|
|
2017-05-10 23:54:35 +02:00
|
|
|
std::vector<std::shared_ptr<BlobFile>> process_files;
|
|
|
|
{
|
|
|
|
ReadLock rl(&mutex_);
|
2017-11-01 00:33:55 +01:00
|
|
|
for (auto fitr : open_ttl_files_) {
|
2017-12-20 01:34:25 +01:00
|
|
|
process_files.push_back(fitr);
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
2017-12-20 01:34:25 +01:00
|
|
|
if (open_non_ttl_file_ != nullptr) {
|
2017-11-01 00:33:55 +01:00
|
|
|
process_files.push_back(open_non_ttl_file_);
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-12-11 21:01:22 +01:00
|
|
|
Status s;
|
|
|
|
for (auto& blob_file : process_files) {
|
2017-12-20 01:34:25 +01:00
|
|
|
s = blob_file->Fsync();
|
|
|
|
if (!s.ok()) {
|
|
|
|
ROCKS_LOG_ERROR(db_options_.info_log,
|
|
|
|
"Failed to sync blob file %" PRIu64 ", status: %s",
|
|
|
|
blob_file->BlobFileNumber(), s.ToString().c_str());
|
|
|
|
return s;
|
2017-12-11 21:01:22 +01:00
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2017-12-20 01:34:25 +01:00
|
|
|
s = dir_ent_->Fsync();
|
|
|
|
if (!s.ok()) {
|
|
|
|
ROCKS_LOG_ERROR(db_options_.info_log,
|
|
|
|
"Failed to sync blob directory, status: %s",
|
|
|
|
s.ToString().c_str());
|
2017-12-11 21:01:22 +01:00
|
|
|
}
|
|
|
|
return s;
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
std::pair<bool, int64_t> BlobDBImpl::ReclaimOpenFiles(bool aborted) {
|
|
|
|
if (aborted) return std::make_pair(false, -1);
|
|
|
|
|
2017-08-01 21:48:22 +02:00
|
|
|
if (open_file_count_.load() < kOpenFilesTrigger) {
|
2017-05-10 23:54:35 +02:00
|
|
|
return std::make_pair(true, -1);
|
2017-08-01 21:48:22 +02:00
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
|
|
|
// in the future, we should sort by last_access_
|
|
|
|
// instead of closing every file
|
|
|
|
ReadLock rl(&mutex_);
|
|
|
|
for (auto const& ent : blob_files_) {
|
|
|
|
auto bfile = ent.second;
|
|
|
|
if (bfile->last_access_.load() == -1) continue;
|
|
|
|
|
|
|
|
WriteLock lockbfile_w(&bfile->mutex_);
|
|
|
|
CloseRandomAccessLocked(bfile);
|
|
|
|
}
|
|
|
|
|
|
|
|
return std::make_pair(true, -1);
|
|
|
|
}
|
|
|
|
|
2017-10-18 02:24:25 +02:00
|
|
|
// Write callback for garbage collection to check if key has been updated
|
|
|
|
// since last read. Similar to how OptimisticTransaction works. See inline
|
|
|
|
// comment in GCFileAndUpdateLSM().
|
|
|
|
class BlobDBImpl::GarbageCollectionWriteCallback : public WriteCallback {
|
|
|
|
public:
|
|
|
|
GarbageCollectionWriteCallback(ColumnFamilyData* cfd, const Slice& key,
|
|
|
|
SequenceNumber upper_bound)
|
|
|
|
: cfd_(cfd), key_(key), upper_bound_(upper_bound) {}
|
|
|
|
|
2019-02-14 22:52:47 +01:00
|
|
|
Status Callback(DB* db) override {
|
2019-08-27 19:57:28 +02:00
|
|
|
auto* db_impl = static_cast_with_check<DBImpl, DB>(db);
|
2017-10-18 02:24:25 +02:00
|
|
|
auto* sv = db_impl->GetAndRefSuperVersion(cfd_);
|
|
|
|
SequenceNumber latest_seq = 0;
|
|
|
|
bool found_record_for_key = false;
|
|
|
|
bool is_blob_index = false;
|
|
|
|
Status s = db_impl->GetLatestSequenceForKey(
|
2019-06-11 20:42:19 +02:00
|
|
|
sv, key_, false /*cache_only*/, 0 /*lower_bound_seq*/, &latest_seq,
|
|
|
|
&found_record_for_key, &is_blob_index);
|
2017-10-18 02:24:25 +02:00
|
|
|
db_impl->ReturnAndCleanupSuperVersion(cfd_, sv);
|
|
|
|
if (!s.ok() && !s.IsNotFound()) {
|
|
|
|
// Error.
|
|
|
|
assert(!s.IsBusy());
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
if (s.IsNotFound()) {
|
|
|
|
assert(!found_record_for_key);
|
|
|
|
return Status::Busy("Key deleted");
|
|
|
|
}
|
|
|
|
assert(found_record_for_key);
|
|
|
|
assert(is_blob_index);
|
|
|
|
if (latest_seq > upper_bound_) {
|
|
|
|
return Status::Busy("Key overwritten");
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2019-02-14 22:52:47 +01:00
|
|
|
bool AllowWriteBatching() override { return false; }
|
2017-10-18 02:24:25 +02:00
|
|
|
|
|
|
|
private:
|
|
|
|
ColumnFamilyData* cfd_;
|
|
|
|
// Key to check
|
|
|
|
Slice key_;
|
|
|
|
// Upper bound of sequence number to proceed.
|
|
|
|
SequenceNumber upper_bound_;
|
|
|
|
};
|
|
|
|
|
2017-05-10 23:54:35 +02:00
|
|
|
// iterate over the blobs sequentially and check if the blob sequence number
|
|
|
|
// is the latest. If it is the latest, preserve it, otherwise delete it
|
|
|
|
// if it is TTL based, and the TTL has expired, then
|
|
|
|
// we can blow the entity if the key is still the latest or the Key is not
|
|
|
|
// found
|
|
|
|
// WHAT HAPPENS IF THE KEY HAS BEEN OVERRIDEN. Then we can drop the blob
|
|
|
|
// without doing anything if the earliest snapshot is not
|
|
|
|
// referring to that sequence number, i.e. it is later than the sequence number
|
|
|
|
// of the new key
|
|
|
|
//
|
|
|
|
// if it is not TTL based, then we can blow the key if the key has been
|
|
|
|
// DELETED in the LSM
|
|
|
|
Status BlobDBImpl::GCFileAndUpdateLSM(const std::shared_ptr<BlobFile>& bfptr,
|
2017-08-11 21:30:02 +02:00
|
|
|
GCStats* gc_stats) {
|
2017-11-28 20:42:28 +01:00
|
|
|
StopWatch gc_sw(env_, statistics_, BLOB_DB_GC_MICROS);
|
2017-08-11 21:30:02 +02:00
|
|
|
uint64_t now = EpochNow();
|
2017-05-10 23:54:35 +02:00
|
|
|
|
|
|
|
std::shared_ptr<Reader> reader =
|
2018-06-24 08:08:33 +02:00
|
|
|
bfptr->OpenRandomAccessReader(env_, db_options_, env_options_);
|
2017-05-10 23:54:35 +02:00
|
|
|
if (!reader) {
|
2017-07-13 19:03:39 +02:00
|
|
|
ROCKS_LOG_ERROR(db_options_.info_log,
|
2019-04-04 21:05:42 +02:00
|
|
|
"File sequential reader could not be opened for %s",
|
2017-07-13 19:03:39 +02:00
|
|
|
bfptr->PathName().c_str());
|
2017-05-10 23:54:35 +02:00
|
|
|
return Status::IOError("failed to create sequential reader");
|
|
|
|
}
|
|
|
|
|
|
|
|
BlobLogHeader header;
|
|
|
|
Status s = reader->ReadHeader(&header);
|
|
|
|
if (!s.ok()) {
|
2017-07-13 19:03:39 +02:00
|
|
|
ROCKS_LOG_ERROR(db_options_.info_log,
|
|
|
|
"Failure to read header for blob-file %s",
|
|
|
|
bfptr->PathName().c_str());
|
2017-05-10 23:54:35 +02:00
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2018-10-08 23:22:06 +02:00
|
|
|
auto cfh = db_impl_->DefaultColumnFamily();
|
2017-10-18 02:24:25 +02:00
|
|
|
auto* cfd = reinterpret_cast<ColumnFamilyHandleImpl*>(cfh)->cfd();
|
|
|
|
auto column_family_id = cfd->GetID();
|
2017-10-27 22:14:34 +02:00
|
|
|
bool has_ttl = header.has_ttl;
|
2017-05-10 23:54:35 +02:00
|
|
|
|
|
|
|
// this reads the key but skips the blob
|
2017-10-17 21:11:52 +02:00
|
|
|
Reader::ReadLevel shallow = Reader::kReadHeaderKey;
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2019-08-15 01:07:03 +02:00
|
|
|
ExpirationRange expiration_range;
|
|
|
|
|
|
|
|
{
|
|
|
|
ReadLock file_lock(&bfptr->mutex_);
|
|
|
|
expiration_range = bfptr->GetExpirationRange();
|
|
|
|
}
|
|
|
|
|
|
|
|
bool file_expired = has_ttl && now >= expiration_range.second;
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2018-03-02 21:54:24 +01:00
|
|
|
if (!file_expired) {
|
2017-05-10 23:54:35 +02:00
|
|
|
// read the blob because you have to write it back to new file
|
2017-10-17 21:11:52 +02:00
|
|
|
shallow = Reader::kReadHeaderKeyBlob;
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
BlobLogRecord record;
|
|
|
|
std::shared_ptr<BlobFile> newfile;
|
|
|
|
std::shared_ptr<Writer> new_writer;
|
2017-08-11 21:30:02 +02:00
|
|
|
uint64_t blob_offset = 0;
|
|
|
|
|
|
|
|
while (true) {
|
|
|
|
assert(s.ok());
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2017-10-18 02:24:25 +02:00
|
|
|
// Read the next blob record.
|
|
|
|
Status read_record_status =
|
|
|
|
reader->ReadRecord(&record, shallow, &blob_offset);
|
|
|
|
// Exit if we reach the end of blob file.
|
|
|
|
// TODO(yiwu): properly handle ReadRecord error.
|
|
|
|
if (!read_record_status.ok()) {
|
|
|
|
break;
|
2017-08-11 21:30:02 +02:00
|
|
|
}
|
2017-10-18 02:24:25 +02:00
|
|
|
gc_stats->blob_count++;
|
|
|
|
|
|
|
|
// Similar to OptimisticTransaction, we obtain latest_seq from
|
|
|
|
// base DB, which is guaranteed to be no smaller than the sequence of
|
|
|
|
// current key. We use a WriteCallback on write to check the key sequence
|
|
|
|
// on write. If the key sequence is larger than latest_seq, we know
|
|
|
|
// a new versions is inserted and the old blob can be disgard.
|
|
|
|
//
|
|
|
|
// We cannot use OptimisticTransaction because we need to pass
|
|
|
|
// is_blob_index flag to GetImpl.
|
|
|
|
SequenceNumber latest_seq = GetLatestSequenceNumber();
|
|
|
|
bool is_blob_index = false;
|
|
|
|
PinnableSlice index_entry;
|
New API to get all merge operands for a Key (#5604)
Summary:
This is a new API added to db.h to allow for fetching all merge operands associated with a Key. The main motivation for this API is to support use cases where doing a full online merge is not necessary as it is performance sensitive. Example use-cases:
1. Update subset of columns and read subset of columns -
Imagine a SQL Table, a row is encoded as a K/V pair (as it is done in MyRocks). If there are many columns and users only updated one of them, we can use merge operator to reduce write amplification. While users only read one or two columns in the read query, this feature can avoid a full merging of the whole row, and save some CPU.
2. Updating very few attributes in a value which is a JSON-like document -
Updating one attribute can be done efficiently using merge operator, while reading back one attribute can be done more efficiently if we don't need to do a full merge.
----------------------------------------------------------------------------------------------------
API :
Status GetMergeOperands(
const ReadOptions& options, ColumnFamilyHandle* column_family,
const Slice& key, PinnableSlice* merge_operands,
GetMergeOperandsOptions* get_merge_operands_options,
int* number_of_operands)
Example usage :
int size = 100;
int number_of_operands = 0;
std::vector<PinnableSlice> values(size);
GetMergeOperandsOptions merge_operands_info;
db_->GetMergeOperands(ReadOptions(), db_->DefaultColumnFamily(), "k1", values.data(), merge_operands_info, &number_of_operands);
Description :
Returns all the merge operands corresponding to the key. If the number of merge operands in DB is greater than merge_operands_options.expected_max_number_of_operands no merge operands are returned and status is Incomplete. Merge operands returned are in the order of insertion.
merge_operands-> Points to an array of at-least merge_operands_options.expected_max_number_of_operands and the caller is responsible for allocating it. If the status returned is Incomplete then number_of_operands will contain the total number of merge operands found in DB for key.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5604
Test Plan:
Added unit test and perf test in db_bench that can be run using the command:
./db_bench -benchmarks=getmergeoperands --merge_operator=sortlist
Differential Revision: D16657366
Pulled By: vjnadimpalli
fbshipit-source-id: 0faadd752351745224ee12d4ae9ef3cb529951bf
2019-08-06 23:22:34 +02:00
|
|
|
DBImpl::GetImplOptions get_impl_options;
|
|
|
|
get_impl_options.column_family = cfh;
|
|
|
|
get_impl_options.value = &index_entry;
|
|
|
|
get_impl_options.is_blob_index = &is_blob_index;
|
|
|
|
Status get_status =
|
|
|
|
db_impl_->GetImpl(ReadOptions(), record.key, get_impl_options);
|
2017-10-18 02:24:25 +02:00
|
|
|
TEST_SYNC_POINT("BlobDBImpl::GCFileAndUpdateLSM:AfterGetFromBaseDB");
|
2017-11-02 23:50:30 +01:00
|
|
|
if (!get_status.ok() && !get_status.IsNotFound()) {
|
2017-10-18 02:24:25 +02:00
|
|
|
// error
|
2017-08-11 21:30:02 +02:00
|
|
|
s = get_status;
|
|
|
|
ROCKS_LOG_ERROR(db_options_.info_log,
|
|
|
|
"Error while getting index entry: %s",
|
|
|
|
s.ToString().c_str());
|
|
|
|
break;
|
|
|
|
}
|
2017-10-18 02:24:25 +02:00
|
|
|
if (get_status.IsNotFound() || !is_blob_index) {
|
|
|
|
// Either the key is deleted or updated with a newer version whish is
|
|
|
|
// inlined in LSM.
|
2017-11-28 20:42:28 +01:00
|
|
|
gc_stats->num_keys_overwritten++;
|
|
|
|
gc_stats->bytes_overwritten += record.record_size();
|
2017-10-18 02:24:25 +02:00
|
|
|
continue;
|
|
|
|
}
|
2017-08-03 21:56:44 +02:00
|
|
|
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
BlobIndex blob_index;
|
|
|
|
s = blob_index.DecodeFrom(index_entry);
|
2017-08-11 21:30:02 +02:00
|
|
|
if (!s.ok()) {
|
|
|
|
ROCKS_LOG_ERROR(db_options_.info_log,
|
|
|
|
"Error while decoding index entry: %s",
|
|
|
|
s.ToString().c_str());
|
|
|
|
break;
|
|
|
|
}
|
2017-11-24 20:50:42 +01:00
|
|
|
if (blob_index.IsInlined() ||
|
|
|
|
blob_index.file_number() != bfptr->BlobFileNumber() ||
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
blob_index.offset() != blob_offset) {
|
2017-08-11 21:30:02 +02:00
|
|
|
// Key has been overwritten. Drop the blob record.
|
2017-11-28 20:42:28 +01:00
|
|
|
gc_stats->num_keys_overwritten++;
|
|
|
|
gc_stats->bytes_overwritten += record.record_size();
|
2017-08-11 21:30:02 +02:00
|
|
|
continue;
|
|
|
|
}
|
2017-08-03 21:56:44 +02:00
|
|
|
|
2017-10-27 22:14:34 +02:00
|
|
|
GarbageCollectionWriteCallback callback(cfd, record.key, latest_seq);
|
2017-10-18 02:24:25 +02:00
|
|
|
|
2017-08-11 21:30:02 +02:00
|
|
|
// If key has expired, remove it from base DB.
|
2017-11-03 01:26:46 +01:00
|
|
|
// TODO(yiwu): Blob indexes will be remove by BlobIndexCompactionFilter.
|
|
|
|
// We can just drop the blob record.
|
2018-03-02 21:54:24 +01:00
|
|
|
if (file_expired || (has_ttl && now >= record.expiration)) {
|
2017-11-28 20:42:28 +01:00
|
|
|
gc_stats->num_keys_expired++;
|
|
|
|
gc_stats->bytes_expired += record.record_size();
|
2017-08-11 21:30:02 +02:00
|
|
|
TEST_SYNC_POINT("BlobDBImpl::GCFileAndUpdateLSM:BeforeDelete");
|
2017-10-18 02:24:25 +02:00
|
|
|
WriteBatch delete_batch;
|
2017-10-27 22:14:34 +02:00
|
|
|
Status delete_status = delete_batch.Delete(record.key);
|
2017-10-18 02:24:25 +02:00
|
|
|
if (delete_status.ok()) {
|
|
|
|
delete_status = db_impl_->WriteWithCallback(WriteOptions(),
|
|
|
|
&delete_batch, &callback);
|
|
|
|
}
|
2017-11-28 20:42:28 +01:00
|
|
|
if (!delete_status.ok() && !delete_status.IsBusy()) {
|
2017-08-11 21:30:02 +02:00
|
|
|
// We hit an error.
|
|
|
|
s = delete_status;
|
|
|
|
ROCKS_LOG_ERROR(db_options_.info_log,
|
|
|
|
"Error while deleting expired key: %s",
|
|
|
|
s.ToString().c_str());
|
|
|
|
break;
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
2017-08-11 21:30:02 +02:00
|
|
|
// Continue to next blob record or retry.
|
|
|
|
continue;
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2017-08-11 21:30:02 +02:00
|
|
|
// Relocate the blob record to new file.
|
|
|
|
if (!newfile) {
|
|
|
|
// new file
|
|
|
|
std::string reason("GC of ");
|
|
|
|
reason += bfptr->PathName();
|
|
|
|
newfile = NewBlobFile(reason);
|
|
|
|
|
2018-10-24 00:01:36 +02:00
|
|
|
s = CheckOrCreateWriterLocked(newfile, &new_writer);
|
|
|
|
if (!s.ok()) {
|
|
|
|
ROCKS_LOG_ERROR(db_options_.info_log,
|
|
|
|
"Failed to open file %s for writer, error: %s",
|
|
|
|
newfile->PathName().c_str(), s.ToString().c_str());
|
|
|
|
break;
|
|
|
|
}
|
2017-08-11 21:30:02 +02:00
|
|
|
// Can't use header beyond this point
|
2018-03-22 00:16:05 +01:00
|
|
|
newfile->header_ = std::move(header);
|
2017-08-11 21:30:02 +02:00
|
|
|
newfile->header_valid_ = true;
|
2017-10-27 22:14:34 +02:00
|
|
|
newfile->file_size_ = BlobLogHeader::kSize;
|
2018-03-22 00:16:05 +01:00
|
|
|
newfile->SetColumnFamilyId(bfptr->column_family_id());
|
|
|
|
newfile->SetHasTTL(bfptr->HasTTL());
|
|
|
|
newfile->SetCompression(bfptr->compression());
|
|
|
|
newfile->expiration_range_ = bfptr->expiration_range_;
|
2017-08-11 21:30:02 +02:00
|
|
|
|
2018-03-22 00:16:05 +01:00
|
|
|
s = new_writer->WriteHeader(newfile->header_);
|
2017-08-11 21:30:02 +02:00
|
|
|
if (!s.ok()) {
|
|
|
|
ROCKS_LOG_ERROR(db_options_.info_log,
|
|
|
|
"File: %s - header writing failed",
|
|
|
|
newfile->PathName().c_str());
|
|
|
|
break;
|
2017-08-03 21:56:44 +02:00
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2018-03-22 00:16:05 +01:00
|
|
|
// We don't add the file to open_ttl_files_ or open_non_ttl_files_, to
|
2018-08-21 07:38:34 +02:00
|
|
|
// avoid user writes writing to the file, and avoid
|
|
|
|
// EvictExpiredFiles close the file by mistake.
|
2017-08-11 21:30:02 +02:00
|
|
|
WriteLock wl(&mutex_);
|
|
|
|
blob_files_.insert(std::make_pair(newfile->BlobFileNumber(), newfile));
|
|
|
|
}
|
|
|
|
|
|
|
|
std::string new_index_entry;
|
|
|
|
uint64_t new_blob_offset = 0;
|
|
|
|
uint64_t new_key_offset = 0;
|
|
|
|
// write the blob to the blob log.
|
2017-10-27 22:14:34 +02:00
|
|
|
s = new_writer->AddRecord(record.key, record.value, record.expiration,
|
|
|
|
&new_key_offset, &new_blob_offset);
|
2017-08-11 21:30:02 +02:00
|
|
|
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
BlobIndex::EncodeBlob(&new_index_entry, newfile->BlobFileNumber(),
|
2017-10-27 22:14:34 +02:00
|
|
|
new_blob_offset, record.value.size(),
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
bdb_options_.compression);
|
2017-08-11 21:30:02 +02:00
|
|
|
|
|
|
|
newfile->blob_count_++;
|
2017-10-17 21:11:52 +02:00
|
|
|
newfile->file_size_ +=
|
2017-10-27 22:14:34 +02:00
|
|
|
BlobLogRecord::kHeaderSize + record.key.size() + record.value.size();
|
2017-08-11 21:30:02 +02:00
|
|
|
|
|
|
|
TEST_SYNC_POINT("BlobDBImpl::GCFileAndUpdateLSM:BeforeRelocate");
|
2017-10-18 02:24:25 +02:00
|
|
|
WriteBatch rewrite_batch;
|
|
|
|
Status rewrite_status = WriteBatchInternal::PutBlobIndex(
|
2017-10-27 22:14:34 +02:00
|
|
|
&rewrite_batch, column_family_id, record.key, new_index_entry);
|
2017-10-18 02:24:25 +02:00
|
|
|
if (rewrite_status.ok()) {
|
|
|
|
rewrite_status = db_impl_->WriteWithCallback(WriteOptions(),
|
|
|
|
&rewrite_batch, &callback);
|
|
|
|
}
|
|
|
|
if (rewrite_status.ok()) {
|
2017-11-28 20:42:28 +01:00
|
|
|
gc_stats->num_keys_relocated++;
|
|
|
|
gc_stats->bytes_relocated += record.record_size();
|
2017-10-18 02:24:25 +02:00
|
|
|
} else if (rewrite_status.IsBusy()) {
|
2017-08-11 21:30:02 +02:00
|
|
|
// The key is overwritten in the meanwhile. Drop the blob record.
|
2017-11-28 20:42:28 +01:00
|
|
|
gc_stats->num_keys_overwritten++;
|
|
|
|
gc_stats->bytes_overwritten += record.record_size();
|
2017-08-11 21:30:02 +02:00
|
|
|
} else {
|
|
|
|
// We hit an error.
|
2017-10-18 02:24:25 +02:00
|
|
|
s = rewrite_status;
|
2017-08-11 21:30:02 +02:00
|
|
|
ROCKS_LOG_ERROR(db_options_.info_log, "Error while relocating key: %s",
|
|
|
|
s.ToString().c_str());
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
} // end of ReadRecord loop
|
|
|
|
|
2018-03-06 20:46:20 +01:00
|
|
|
{
|
|
|
|
WriteLock wl(&mutex_);
|
|
|
|
ObsoleteBlobFile(bfptr, GetLatestSequenceNumber(), true /*update_size*/);
|
2017-11-02 23:50:30 +01:00
|
|
|
}
|
|
|
|
|
2017-08-11 21:30:02 +02:00
|
|
|
ROCKS_LOG_INFO(
|
2017-08-25 19:40:25 +02:00
|
|
|
db_options_.info_log,
|
2017-11-28 20:42:28 +01:00
|
|
|
"%s blob file %" PRIu64 ". Total blob records: %" PRIu64
|
2018-02-01 02:59:32 +01:00
|
|
|
", expired: %" PRIu64 " keys/%" PRIu64
|
|
|
|
" bytes, updated or deleted by user: %" PRIu64 " keys/%" PRIu64
|
|
|
|
" bytes, rewrite to new file: %" PRIu64 " keys/%" PRIu64 " bytes.",
|
2017-08-11 21:30:02 +02:00
|
|
|
s.ok() ? "Successfully garbage collected" : "Failed to garbage collect",
|
2017-11-28 20:42:28 +01:00
|
|
|
bfptr->BlobFileNumber(), gc_stats->blob_count, gc_stats->num_keys_expired,
|
|
|
|
gc_stats->bytes_expired, gc_stats->num_keys_overwritten,
|
|
|
|
gc_stats->bytes_overwritten, gc_stats->num_keys_relocated,
|
|
|
|
gc_stats->bytes_relocated);
|
|
|
|
RecordTick(statistics_, BLOB_DB_GC_NUM_FILES);
|
|
|
|
RecordTick(statistics_, BLOB_DB_GC_NUM_KEYS_OVERWRITTEN,
|
|
|
|
gc_stats->num_keys_overwritten);
|
|
|
|
RecordTick(statistics_, BLOB_DB_GC_NUM_KEYS_EXPIRED,
|
|
|
|
gc_stats->num_keys_expired);
|
|
|
|
RecordTick(statistics_, BLOB_DB_GC_BYTES_OVERWRITTEN,
|
|
|
|
gc_stats->bytes_overwritten);
|
|
|
|
RecordTick(statistics_, BLOB_DB_GC_BYTES_EXPIRED, gc_stats->bytes_expired);
|
2017-08-11 21:30:02 +02:00
|
|
|
if (newfile != nullptr) {
|
2018-03-22 00:16:05 +01:00
|
|
|
{
|
|
|
|
MutexLock l(&write_mutex_);
|
|
|
|
CloseBlobFile(newfile);
|
|
|
|
}
|
2018-03-06 20:46:20 +01:00
|
|
|
total_blob_size_ += newfile->file_size_;
|
2017-08-11 21:30:02 +02:00
|
|
|
ROCKS_LOG_INFO(db_options_.info_log, "New blob file %" PRIu64 ".",
|
|
|
|
newfile->BlobFileNumber());
|
2017-11-28 20:42:28 +01:00
|
|
|
RecordTick(statistics_, BLOB_DB_GC_NUM_NEW_FILES);
|
|
|
|
RecordTick(statistics_, BLOB_DB_GC_NUM_KEYS_RELOCATED,
|
|
|
|
gc_stats->num_keys_relocated);
|
|
|
|
RecordTick(statistics_, BLOB_DB_GC_BYTES_RELOCATED,
|
|
|
|
gc_stats->bytes_relocated);
|
|
|
|
}
|
|
|
|
if (!s.ok()) {
|
|
|
|
RecordTick(statistics_, BLOB_DB_GC_FAILURES);
|
2017-08-11 21:30:02 +02:00
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2017-08-21 03:12:38 +02:00
|
|
|
std::pair<bool, int64_t> BlobDBImpl::DeleteObsoleteFiles(bool aborted) {
|
2018-08-27 19:56:28 +02:00
|
|
|
if (aborted) {
|
|
|
|
return std::make_pair(false, -1);
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2018-08-27 19:56:28 +02:00
|
|
|
MutexLock delete_file_lock(&delete_file_mutex_);
|
|
|
|
if (disable_file_deletions_ > 0) {
|
|
|
|
return std::make_pair(true, -1);
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
std::list<std::shared_ptr<BlobFile>> tobsolete;
|
|
|
|
{
|
|
|
|
WriteLock wl(&mutex_);
|
2018-08-27 19:56:28 +02:00
|
|
|
if (obsolete_files_.empty()) {
|
|
|
|
return std::make_pair(true, -1);
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
tobsolete.swap(obsolete_files_);
|
|
|
|
}
|
|
|
|
|
|
|
|
bool file_deleted = false;
|
|
|
|
for (auto iter = tobsolete.begin(); iter != tobsolete.end();) {
|
|
|
|
auto bfile = *iter;
|
|
|
|
{
|
|
|
|
ReadLock lockbfile_r(&bfile->mutex_);
|
2017-11-02 23:50:30 +01:00
|
|
|
if (VisibleToActiveSnapshot(bfile)) {
|
|
|
|
ROCKS_LOG_INFO(db_options_.info_log,
|
|
|
|
"Could not delete file due to snapshot failure %s",
|
|
|
|
bfile->PathName().c_str());
|
2017-05-10 23:54:35 +02:00
|
|
|
++iter;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
}
|
2017-11-02 23:50:30 +01:00
|
|
|
ROCKS_LOG_INFO(db_options_.info_log,
|
|
|
|
"Will delete file due to snapshot success %s",
|
|
|
|
bfile->PathName().c_str());
|
2017-05-10 23:54:35 +02:00
|
|
|
|
2019-08-15 01:07:03 +02:00
|
|
|
{
|
|
|
|
WriteLock wl(&mutex_);
|
|
|
|
blob_files_.erase(bfile->BlobFileNumber());
|
|
|
|
}
|
|
|
|
|
2019-01-29 23:27:30 +01:00
|
|
|
Status s = DeleteDBFile(&(db_impl_->immutable_db_options()),
|
2019-07-07 06:04:22 +02:00
|
|
|
bfile->PathName(), blob_dir_, true,
|
|
|
|
/*force_fg=*/false);
|
2017-05-10 23:54:35 +02:00
|
|
|
if (!s.ok()) {
|
2017-07-13 19:03:39 +02:00
|
|
|
ROCKS_LOG_ERROR(db_options_.info_log,
|
|
|
|
"File failed to be deleted as obsolete %s",
|
|
|
|
bfile->PathName().c_str());
|
2017-05-10 23:54:35 +02:00
|
|
|
++iter;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
file_deleted = true;
|
2017-07-13 19:03:39 +02:00
|
|
|
ROCKS_LOG_INFO(db_options_.info_log,
|
|
|
|
"File deleted as obsolete from blob dir %s",
|
|
|
|
bfile->PathName().c_str());
|
2017-05-10 23:54:35 +02:00
|
|
|
|
|
|
|
iter = tobsolete.erase(iter);
|
|
|
|
}
|
|
|
|
|
|
|
|
// directory change. Fsync
|
2017-11-02 20:02:42 +01:00
|
|
|
if (file_deleted) {
|
2018-09-14 22:17:12 +02:00
|
|
|
Status s = dir_ent_->Fsync();
|
|
|
|
if (!s.ok()) {
|
|
|
|
ROCKS_LOG_ERROR(db_options_.info_log, "Failed to sync dir %s: %s",
|
|
|
|
blob_dir_.c_str(), s.ToString().c_str());
|
|
|
|
}
|
2017-11-02 20:02:42 +01:00
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
|
|
|
// put files back into obsolete if for some reason, delete failed
|
|
|
|
if (!tobsolete.empty()) {
|
|
|
|
WriteLock wl(&mutex_);
|
2017-08-21 03:12:38 +02:00
|
|
|
for (auto bfile : tobsolete) {
|
2019-08-15 01:07:03 +02:00
|
|
|
blob_files_.insert(std::make_pair(bfile->BlobFileNumber(), bfile));
|
2017-08-21 03:12:38 +02:00
|
|
|
obsolete_files_.push_front(bfile);
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
return std::make_pair(!aborted, -1);
|
|
|
|
}
|
|
|
|
|
|
|
|
void BlobDBImpl::CopyBlobFiles(
|
2018-06-26 07:32:29 +02:00
|
|
|
std::vector<std::shared_ptr<BlobFile>>* bfiles_copy) {
|
2017-05-10 23:54:35 +02:00
|
|
|
ReadLock rl(&mutex_);
|
2017-08-21 01:56:01 +02:00
|
|
|
for (auto const& p : blob_files_) {
|
2018-06-26 07:32:29 +02:00
|
|
|
bfiles_copy->push_back(p.second);
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
std::pair<bool, int64_t> BlobDBImpl::RunGC(bool aborted) {
|
2018-03-02 21:54:24 +01:00
|
|
|
if (aborted) {
|
|
|
|
return std::make_pair(false, -1);
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2018-03-02 21:54:24 +01:00
|
|
|
// TODO(yiwu): Garbage collection implementation.
|
|
|
|
|
2017-05-10 23:54:35 +02:00
|
|
|
// reschedule
|
|
|
|
return std::make_pair(true, -1);
|
|
|
|
}
|
|
|
|
|
2017-09-08 19:57:12 +02:00
|
|
|
Iterator* BlobDBImpl::NewIterator(const ReadOptions& read_options) {
|
2017-10-18 02:24:25 +02:00
|
|
|
auto* cfd =
|
|
|
|
reinterpret_cast<ColumnFamilyHandleImpl*>(DefaultColumnFamily())->cfd();
|
2017-08-21 03:12:38 +02:00
|
|
|
// Get a snapshot to avoid blob file get deleted between we
|
|
|
|
// fetch and index entry and reading from the file.
|
2017-10-18 02:24:25 +02:00
|
|
|
ManagedSnapshot* own_snapshot = nullptr;
|
|
|
|
const Snapshot* snapshot = read_options.snapshot;
|
|
|
|
if (snapshot == nullptr) {
|
|
|
|
own_snapshot = new ManagedSnapshot(db_);
|
|
|
|
snapshot = own_snapshot->snapshot();
|
|
|
|
}
|
|
|
|
auto* iter = db_impl_->NewIteratorImpl(
|
|
|
|
read_options, cfd, snapshot->GetSequenceNumber(),
|
|
|
|
nullptr /*read_callback*/, true /*allow_blob*/);
|
2017-11-28 20:42:28 +01:00
|
|
|
return new BlobDBIterator(own_snapshot, iter, this, env_, statistics_);
|
2017-05-10 23:54:35 +02:00
|
|
|
}
|
|
|
|
|
2017-06-14 22:08:54 +02:00
|
|
|
Status DestroyBlobDB(const std::string& dbname, const Options& options,
|
|
|
|
const BlobDBOptions& bdb_options) {
|
|
|
|
const ImmutableDBOptions soptions(SanitizeOptions(dbname, options));
|
|
|
|
Env* env = soptions.env;
|
|
|
|
|
|
|
|
Status status;
|
|
|
|
std::string blobdir;
|
|
|
|
blobdir = (bdb_options.path_relative) ? dbname + "/" + bdb_options.blob_dir
|
|
|
|
: bdb_options.blob_dir;
|
|
|
|
|
|
|
|
std::vector<std::string> filenames;
|
|
|
|
env->GetChildren(blobdir, &filenames);
|
|
|
|
|
|
|
|
for (const auto& f : filenames) {
|
|
|
|
uint64_t number;
|
|
|
|
FileType type;
|
|
|
|
if (ParseFileName(f, &number, &type) && type == kBlobFile) {
|
2019-07-07 06:04:22 +02:00
|
|
|
Status del = DeleteDBFile(&soptions, blobdir + "/" + f, blobdir, true,
|
|
|
|
/*force_fg=*/false);
|
2017-06-14 22:08:54 +02:00
|
|
|
if (status.ok() && !del.ok()) {
|
|
|
|
status = del;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
env->DeleteDir(blobdir);
|
|
|
|
|
|
|
|
Status destroy = DestroyDB(dbname, options);
|
|
|
|
if (status.ok() && !destroy.ok()) {
|
|
|
|
status = destroy;
|
|
|
|
}
|
|
|
|
|
|
|
|
return status;
|
|
|
|
}
|
|
|
|
|
2017-05-31 19:45:47 +02:00
|
|
|
#ifndef NDEBUG
|
Blob DB: Inline small values in base DB
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
2017-10-26 21:19:43 +02:00
|
|
|
Status BlobDBImpl::TEST_GetBlobValue(const Slice& key, const Slice& index_entry,
|
|
|
|
PinnableSlice* value) {
|
|
|
|
return GetBlobValue(key, index_entry, value);
|
|
|
|
}
|
|
|
|
|
2017-07-28 08:16:18 +02:00
|
|
|
std::vector<std::shared_ptr<BlobFile>> BlobDBImpl::TEST_GetBlobFiles() const {
|
2017-08-21 03:12:38 +02:00
|
|
|
ReadLock l(&mutex_);
|
2017-07-28 08:16:18 +02:00
|
|
|
std::vector<std::shared_ptr<BlobFile>> blob_files;
|
|
|
|
for (auto& p : blob_files_) {
|
|
|
|
blob_files.emplace_back(p.second);
|
|
|
|
}
|
|
|
|
return blob_files;
|
|
|
|
}
|
|
|
|
|
2017-08-21 01:56:01 +02:00
|
|
|
std::vector<std::shared_ptr<BlobFile>> BlobDBImpl::TEST_GetObsoleteFiles()
|
|
|
|
const {
|
|
|
|
ReadLock l(&mutex_);
|
|
|
|
std::vector<std::shared_ptr<BlobFile>> obsolete_files;
|
|
|
|
for (auto& bfile : obsolete_files_) {
|
|
|
|
obsolete_files.emplace_back(bfile);
|
|
|
|
}
|
|
|
|
return obsolete_files;
|
|
|
|
}
|
|
|
|
|
2017-08-21 03:12:38 +02:00
|
|
|
void BlobDBImpl::TEST_DeleteObsoleteFiles() {
|
|
|
|
DeleteObsoleteFiles(false /*abort*/);
|
|
|
|
}
|
|
|
|
|
2017-08-25 19:40:25 +02:00
|
|
|
Status BlobDBImpl::TEST_CloseBlobFile(std::shared_ptr<BlobFile>& bfile) {
|
2018-03-06 20:46:20 +01:00
|
|
|
MutexLock l(&write_mutex_);
|
2017-08-25 19:40:25 +02:00
|
|
|
return CloseBlobFile(bfile);
|
2017-07-28 08:16:18 +02:00
|
|
|
}
|
|
|
|
|
2018-03-06 20:46:20 +01:00
|
|
|
void BlobDBImpl::TEST_ObsoleteBlobFile(std::shared_ptr<BlobFile>& blob_file,
|
|
|
|
SequenceNumber obsolete_seq,
|
|
|
|
bool update_size) {
|
|
|
|
return ObsoleteBlobFile(blob_file, obsolete_seq, update_size);
|
|
|
|
}
|
|
|
|
|
2017-07-28 08:16:18 +02:00
|
|
|
Status BlobDBImpl::TEST_GCFileAndUpdateLSM(std::shared_ptr<BlobFile>& bfile,
|
|
|
|
GCStats* gc_stats) {
|
|
|
|
return GCFileAndUpdateLSM(bfile, gc_stats);
|
|
|
|
}
|
2017-08-21 01:56:01 +02:00
|
|
|
|
|
|
|
void BlobDBImpl::TEST_RunGC() { RunGC(false /*abort*/); }
|
2018-03-06 20:46:20 +01:00
|
|
|
|
2018-08-21 07:38:34 +02:00
|
|
|
void BlobDBImpl::TEST_EvictExpiredFiles() {
|
|
|
|
EvictExpiredFiles(false /*abort*/);
|
|
|
|
}
|
|
|
|
|
2018-03-06 20:46:20 +01:00
|
|
|
uint64_t BlobDBImpl::TEST_live_sst_size() { return live_sst_size_.load(); }
|
2017-05-31 19:45:47 +02:00
|
|
|
#endif // !NDEBUG
|
|
|
|
|
2017-05-10 23:54:35 +02:00
|
|
|
} // namespace blob_db
|
|
|
|
} // namespace rocksdb
|
|
|
|
#endif // ROCKSDB_LITE
|