rocksdb/db/db_test2.cc

5988 lines
208 KiB
C++
Raw Normal View History

// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
// This source code is licensed under both the GPLv2 (found in the
// COPYING file in the root directory) and Apache 2.0 License
// (found in the LICENSE.Apache file in the root directory).
//
// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file. See the AUTHORS file for names of contributors.
#include <atomic>
#include <cstdlib>
#include <functional>
#include <memory>
#include "db/db_test_util.h"
#include "db/read_callback.h"
#include "options/options_helper.h"
#include "port/port.h"
#include "port/stack_trace.h"
#include "rocksdb/persistent_cache.h"
2016-03-22 20:07:15 +01:00
#include "rocksdb/wal_filter.h"
#include "util/random.h"
#include "utilities/fault_injection_env.h"
namespace ROCKSDB_NAMESPACE {
class DBTest2 : public DBTestBase {
public:
DBTest2() : DBTestBase("db_test2", /*env_do_fsync=*/true) {}
};
#ifndef ROCKSDB_LITE
TEST_F(DBTest2, OpenForReadOnly) {
DB* db_ptr = nullptr;
std::string dbname = test::PerThreadDBPath("db_readonly");
Options options = CurrentOptions();
options.create_if_missing = true;
// OpenForReadOnly should fail but will create <dbname> in the file system
ASSERT_NOK(DB::OpenForReadOnly(options, dbname, &db_ptr));
// Since <dbname> is created, we should be able to delete the dir
// We first get the list files under <dbname>
// There should not be any subdirectories -- this is not checked here
std::vector<std::string> files;
ASSERT_OK(env_->GetChildren(dbname, &files));
for (auto& f : files) {
ASSERT_OK(env_->DeleteFile(dbname + "/" + f));
}
// <dbname> should be empty now and we should be able to delete it
ASSERT_OK(env_->DeleteDir(dbname));
options.create_if_missing = false;
// OpenForReadOnly should fail since <dbname> was successfully deleted
ASSERT_NOK(DB::OpenForReadOnly(options, dbname, &db_ptr));
// With create_if_missing false, there should not be a dir in the file system
ASSERT_NOK(env_->FileExists(dbname));
}
TEST_F(DBTest2, OpenForReadOnlyWithColumnFamilies) {
DB* db_ptr = nullptr;
std::string dbname = test::PerThreadDBPath("db_readonly");
Options options = CurrentOptions();
options.create_if_missing = true;
ColumnFamilyOptions cf_options(options);
std::vector<ColumnFamilyDescriptor> column_families;
column_families.push_back(
ColumnFamilyDescriptor(kDefaultColumnFamilyName, cf_options));
column_families.push_back(ColumnFamilyDescriptor("goku", cf_options));
std::vector<ColumnFamilyHandle*> handles;
// OpenForReadOnly should fail but will create <dbname> in the file system
ASSERT_NOK(
DB::OpenForReadOnly(options, dbname, column_families, &handles, &db_ptr));
// Since <dbname> is created, we should be able to delete the dir
// We first get the list files under <dbname>
// There should not be any subdirectories -- this is not checked here
std::vector<std::string> files;
ASSERT_OK(env_->GetChildren(dbname, &files));
for (auto& f : files) {
ASSERT_OK(env_->DeleteFile(dbname + "/" + f));
}
// <dbname> should be empty now and we should be able to delete it
ASSERT_OK(env_->DeleteDir(dbname));
options.create_if_missing = false;
// OpenForReadOnly should fail since <dbname> was successfully deleted
ASSERT_NOK(
DB::OpenForReadOnly(options, dbname, column_families, &handles, &db_ptr));
// With create_if_missing false, there should not be a dir in the file system
ASSERT_NOK(env_->FileExists(dbname));
}
class TestReadOnlyWithCompressedCache
: public DBTestBase,
public testing::WithParamInterface<std::tuple<int, bool>> {
public:
TestReadOnlyWithCompressedCache()
: DBTestBase("test_readonly_with_compressed_cache",
/*env_do_fsync=*/true) {
max_open_files_ = std::get<0>(GetParam());
use_mmap_ = std::get<1>(GetParam());
}
int max_open_files_;
bool use_mmap_;
};
TEST_P(TestReadOnlyWithCompressedCache, ReadOnlyWithCompressedCache) {
if (use_mmap_ && !IsMemoryMappedAccessSupported()) {
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566) Summary: This PR does a few things: 1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation. 2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed: - The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated - The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory). 3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10). I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged. Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently. Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566 Reviewed By: zhichao-cao Differential Revision: D24408980 Pulled By: jay-zhuang fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
2020-10-27 18:31:34 +01:00
ROCKSDB_GTEST_SKIP("Test requires MMAP support");
return;
}
ASSERT_OK(Put("foo", "bar"));
ASSERT_OK(Put("foo2", "barbarbarbarbarbarbarbar"));
ASSERT_OK(Flush());
DB* db_ptr = nullptr;
Options options = CurrentOptions();
options.allow_mmap_reads = use_mmap_;
options.max_open_files = max_open_files_;
options.compression = kSnappyCompression;
BlockBasedTableOptions table_options;
table_options.block_cache_compressed = NewLRUCache(8 * 1024 * 1024);
table_options.no_block_cache = true;
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
options.statistics = CreateDBStatistics();
ASSERT_OK(DB::OpenForReadOnly(options, dbname_, &db_ptr));
std::string v;
ASSERT_OK(db_ptr->Get(ReadOptions(), "foo", &v));
ASSERT_EQ("bar", v);
ASSERT_EQ(0, options.statistics->getTickerCount(BLOCK_CACHE_COMPRESSED_HIT));
ASSERT_OK(db_ptr->Get(ReadOptions(), "foo", &v));
ASSERT_EQ("bar", v);
if (Snappy_Supported()) {
if (use_mmap_) {
ASSERT_EQ(0,
options.statistics->getTickerCount(BLOCK_CACHE_COMPRESSED_HIT));
} else {
ASSERT_EQ(1,
options.statistics->getTickerCount(BLOCK_CACHE_COMPRESSED_HIT));
}
}
delete db_ptr;
}
INSTANTIATE_TEST_CASE_P(TestReadOnlyWithCompressedCache,
TestReadOnlyWithCompressedCache,
::testing::Combine(::testing::Values(-1, 100),
::testing::Bool()));
class PartitionedIndexTestListener : public EventListener {
public:
void OnFlushCompleted(DB* /*db*/, const FlushJobInfo& info) override {
ASSERT_GT(info.table_properties.index_partitions, 1);
ASSERT_EQ(info.table_properties.index_key_is_user_key, 0);
}
};
TEST_F(DBTest2, PartitionedIndexUserToInternalKey) {
const int kValueSize = 10500;
const int kNumEntriesPerFile = 1000;
const int kNumFiles = 3;
const int kNumDistinctKeys = 30;
BlockBasedTableOptions table_options;
Options options = CurrentOptions();
options.disable_auto_compactions = true;
table_options.index_type = BlockBasedTableOptions::kTwoLevelIndexSearch;
PartitionedIndexTestListener* listener = new PartitionedIndexTestListener();
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
options.listeners.emplace_back(listener);
std::vector<const Snapshot*> snapshots;
Reopen(options);
Random rnd(301);
for (int i = 0; i < kNumFiles; i++) {
for (int j = 0; j < kNumEntriesPerFile; j++) {
int key_id = (i * kNumEntriesPerFile + j) % kNumDistinctKeys;
std::string value = rnd.RandomString(kValueSize);
ASSERT_OK(Put("keykey_" + std::to_string(key_id), value));
snapshots.push_back(db_->GetSnapshot());
}
ASSERT_OK(Flush());
}
for (auto s : snapshots) {
db_->ReleaseSnapshot(s);
}
}
#endif // ROCKSDB_LITE
class PrefixFullBloomWithReverseComparator
: public DBTestBase,
public ::testing::WithParamInterface<bool> {
public:
PrefixFullBloomWithReverseComparator()
: DBTestBase("prefix_bloom_reverse", /*env_do_fsync=*/true) {}
void SetUp() override { if_cache_filter_ = GetParam(); }
bool if_cache_filter_;
};
TEST_P(PrefixFullBloomWithReverseComparator,
PrefixFullBloomWithReverseComparator) {
Options options = last_options_;
options.comparator = ReverseBytewiseComparator();
options.prefix_extractor.reset(NewCappedPrefixTransform(3));
options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
BlockBasedTableOptions bbto;
if (if_cache_filter_) {
bbto.no_block_cache = false;
bbto.cache_index_and_filter_blocks = true;
bbto.block_cache = NewLRUCache(1);
}
bbto.filter_policy.reset(NewBloomFilterPolicy(10, false));
bbto.whole_key_filtering = false;
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
DestroyAndReopen(options);
ASSERT_OK(dbfull()->Put(WriteOptions(), "bar123", "foo"));
ASSERT_OK(dbfull()->Put(WriteOptions(), "bar234", "foo2"));
ASSERT_OK(dbfull()->Put(WriteOptions(), "foo123", "foo3"));
dbfull()->Flush(FlushOptions());
if (bbto.block_cache) {
bbto.block_cache->EraseUnRefEntries();
}
std::unique_ptr<Iterator> iter(db_->NewIterator(ReadOptions()));
iter->Seek("bar345");
ASSERT_OK(iter->status());
ASSERT_TRUE(iter->Valid());
ASSERT_EQ("bar234", iter->key().ToString());
ASSERT_EQ("foo2", iter->value().ToString());
iter->Next();
ASSERT_TRUE(iter->Valid());
ASSERT_EQ("bar123", iter->key().ToString());
ASSERT_EQ("foo", iter->value().ToString());
iter->Seek("foo234");
ASSERT_OK(iter->status());
ASSERT_TRUE(iter->Valid());
ASSERT_EQ("foo123", iter->key().ToString());
ASSERT_EQ("foo3", iter->value().ToString());
iter->Seek("bar");
ASSERT_OK(iter->status());
ASSERT_TRUE(!iter->Valid());
}
INSTANTIATE_TEST_CASE_P(PrefixFullBloomWithReverseComparator,
PrefixFullBloomWithReverseComparator, testing::Bool());
TEST_F(DBTest2, IteratorPropertyVersionNumber) {
Put("", "");
Iterator* iter1 = db_->NewIterator(ReadOptions());
std::string prop_value;
ASSERT_OK(
iter1->GetProperty("rocksdb.iterator.super-version-number", &prop_value));
uint64_t version_number1 =
static_cast<uint64_t>(std::atoi(prop_value.c_str()));
Put("", "");
Flush();
Iterator* iter2 = db_->NewIterator(ReadOptions());
ASSERT_OK(
iter2->GetProperty("rocksdb.iterator.super-version-number", &prop_value));
uint64_t version_number2 =
static_cast<uint64_t>(std::atoi(prop_value.c_str()));
ASSERT_GT(version_number2, version_number1);
Put("", "");
Iterator* iter3 = db_->NewIterator(ReadOptions());
ASSERT_OK(
iter3->GetProperty("rocksdb.iterator.super-version-number", &prop_value));
uint64_t version_number3 =
static_cast<uint64_t>(std::atoi(prop_value.c_str()));
ASSERT_EQ(version_number2, version_number3);
iter1->SeekToFirst();
ASSERT_OK(
iter1->GetProperty("rocksdb.iterator.super-version-number", &prop_value));
uint64_t version_number1_new =
static_cast<uint64_t>(std::atoi(prop_value.c_str()));
ASSERT_EQ(version_number1, version_number1_new);
delete iter1;
delete iter2;
delete iter3;
}
TEST_F(DBTest2, CacheIndexAndFilterWithDBRestart) {
Options options = CurrentOptions();
options.create_if_missing = true;
options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
BlockBasedTableOptions table_options;
table_options.cache_index_and_filter_blocks = true;
table_options.filter_policy.reset(NewBloomFilterPolicy(20));
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566) Summary: This PR does a few things: 1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation. 2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed: - The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated - The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory). 3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10). I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged. Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently. Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566 Reviewed By: zhichao-cao Differential Revision: D24408980 Pulled By: jay-zhuang fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
2020-10-27 18:31:34 +01:00
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
CreateAndReopenWithCF({"pikachu"}, options);
Put(1, "a", "begin");
Put(1, "z", "end");
ASSERT_OK(Flush(1));
TryReopenWithColumnFamilies({"default", "pikachu"}, options);
std::string value;
value = Get(1, "a");
}
2016-03-22 20:07:15 +01:00
TEST_F(DBTest2, MaxSuccessiveMergesChangeWithDBRecovery) {
Options options = CurrentOptions();
options.create_if_missing = true;
options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
options.max_successive_merges = 3;
options.merge_operator = MergeOperators::CreatePutOperator();
options.disable_auto_compactions = true;
DestroyAndReopen(options);
Put("poi", "Finch");
db_->Merge(WriteOptions(), "poi", "Reese");
db_->Merge(WriteOptions(), "poi", "Shaw");
db_->Merge(WriteOptions(), "poi", "Root");
options.max_successive_merges = 2;
Reopen(options);
}
2016-03-22 20:07:15 +01:00
#ifndef ROCKSDB_LITE
class DBTestSharedWriteBufferAcrossCFs
: public DBTestBase,
public testing::WithParamInterface<std::tuple<bool, bool>> {
public:
DBTestSharedWriteBufferAcrossCFs()
: DBTestBase("db_test_shared_write_buffer", /*env_do_fsync=*/true) {}
void SetUp() override {
use_old_interface_ = std::get<0>(GetParam());
cost_cache_ = std::get<1>(GetParam());
}
bool use_old_interface_;
bool cost_cache_;
};
TEST_P(DBTestSharedWriteBufferAcrossCFs, SharedWriteBufferAcrossCFs) {
Options options = CurrentOptions();
options.arena_block_size = 4096;
auto flush_listener = std::make_shared<FlushCounterListener>();
options.listeners.push_back(flush_listener);
// Don't trip the listener at shutdown.
options.avoid_flush_during_shutdown = true;
// Avoid undeterministic value by malloc_usable_size();
// Force arena block size to 1
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"Arena::Arena:0", [&](void* arg) {
size_t* block_size = static_cast<size_t*>(arg);
*block_size = 1;
});
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"Arena::AllocateNewBlock:0", [&](void* arg) {
std::pair<size_t*, size_t*>* pair =
static_cast<std::pair<size_t*, size_t*>*>(arg);
*std::get<0>(*pair) = *std::get<1>(*pair);
});
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
// The total soft write buffer size is about 105000
std::shared_ptr<Cache> cache = NewLRUCache(4 * 1024 * 1024, 2);
ASSERT_LT(cache->GetUsage(), 256 * 1024);
if (use_old_interface_) {
options.db_write_buffer_size = 120000; // this is the real limit
} else if (!cost_cache_) {
options.write_buffer_manager.reset(new WriteBufferManager(114285));
} else {
options.write_buffer_manager.reset(new WriteBufferManager(114285, cache));
}
options.write_buffer_size = 500000; // this is never hit
CreateAndReopenWithCF({"pikachu", "dobrynia", "nikitich"}, options);
WriteOptions wo;
wo.disableWAL = true;
std::function<void()> wait_flush = [&]() {
dbfull()->TEST_WaitForFlushMemTable(handles_[0]);
dbfull()->TEST_WaitForFlushMemTable(handles_[1]);
dbfull()->TEST_WaitForFlushMemTable(handles_[2]);
dbfull()->TEST_WaitForFlushMemTable(handles_[3]);
};
// Create some data and flush "default" and "nikitich" so that they
// are newer CFs created.
flush_listener->expected_flush_reason = FlushReason::kManualFlush;
ASSERT_OK(Put(3, Key(1), DummyString(1), wo));
Flush(3);
ASSERT_OK(Put(3, Key(1), DummyString(1), wo));
ASSERT_OK(Put(0, Key(1), DummyString(1), wo));
Flush(0);
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "default"),
static_cast<uint64_t>(1));
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "nikitich"),
static_cast<uint64_t>(1));
flush_listener->expected_flush_reason = FlushReason::kWriteBufferManager;
ASSERT_OK(Put(3, Key(1), DummyString(30000), wo));
if (cost_cache_) {
ASSERT_GE(cache->GetUsage(), 256 * 1024);
ASSERT_LE(cache->GetUsage(), 2 * 256 * 1024);
}
wait_flush();
ASSERT_OK(Put(0, Key(1), DummyString(60000), wo));
if (cost_cache_) {
ASSERT_GE(cache->GetUsage(), 256 * 1024);
ASSERT_LE(cache->GetUsage(), 2 * 256 * 1024);
}
wait_flush();
ASSERT_OK(Put(2, Key(1), DummyString(1), wo));
// No flush should trigger
wait_flush();
{
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "default"),
static_cast<uint64_t>(1));
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "pikachu"),
static_cast<uint64_t>(0));
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "dobrynia"),
static_cast<uint64_t>(0));
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "nikitich"),
static_cast<uint64_t>(1));
}
// Trigger a flush. Flushing "nikitich".
ASSERT_OK(Put(3, Key(2), DummyString(30000), wo));
wait_flush();
ASSERT_OK(Put(0, Key(1), DummyString(1), wo));
wait_flush();
{
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "default"),
static_cast<uint64_t>(1));
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "pikachu"),
static_cast<uint64_t>(0));
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "dobrynia"),
static_cast<uint64_t>(0));
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "nikitich"),
static_cast<uint64_t>(2));
}
// Without hitting the threshold, no flush should trigger.
ASSERT_OK(Put(2, Key(1), DummyString(30000), wo));
wait_flush();
ASSERT_OK(Put(2, Key(1), DummyString(1), wo));
wait_flush();
ASSERT_OK(Put(2, Key(1), DummyString(1), wo));
wait_flush();
{
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "default"),
static_cast<uint64_t>(1));
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "pikachu"),
static_cast<uint64_t>(0));
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "dobrynia"),
static_cast<uint64_t>(0));
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "nikitich"),
static_cast<uint64_t>(2));
}
// Hit the write buffer limit again. "default"
// will have been flushed.
ASSERT_OK(Put(2, Key(2), DummyString(10000), wo));
wait_flush();
ASSERT_OK(Put(3, Key(1), DummyString(1), wo));
wait_flush();
ASSERT_OK(Put(0, Key(1), DummyString(1), wo));
wait_flush();
ASSERT_OK(Put(0, Key(1), DummyString(1), wo));
wait_flush();
ASSERT_OK(Put(0, Key(1), DummyString(1), wo));
wait_flush();
{
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "default"),
static_cast<uint64_t>(2));
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "pikachu"),
static_cast<uint64_t>(0));
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "dobrynia"),
static_cast<uint64_t>(0));
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "nikitich"),
static_cast<uint64_t>(2));
}
// Trigger another flush. This time "dobrynia". "pikachu" should not
// be flushed, althrough it was never flushed.
ASSERT_OK(Put(1, Key(1), DummyString(1), wo));
wait_flush();
ASSERT_OK(Put(2, Key(1), DummyString(80000), wo));
wait_flush();
ASSERT_OK(Put(1, Key(1), DummyString(1), wo));
wait_flush();
ASSERT_OK(Put(2, Key(1), DummyString(1), wo));
wait_flush();
{
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "default"),
static_cast<uint64_t>(2));
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "pikachu"),
static_cast<uint64_t>(0));
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "dobrynia"),
static_cast<uint64_t>(1));
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "nikitich"),
static_cast<uint64_t>(2));
}
if (cost_cache_) {
ASSERT_GE(cache->GetUsage(), 256 * 1024);
Close();
options.write_buffer_manager.reset();
last_options_.write_buffer_manager.reset();
ASSERT_LT(cache->GetUsage(), 256 * 1024);
}
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
}
INSTANTIATE_TEST_CASE_P(DBTestSharedWriteBufferAcrossCFs,
DBTestSharedWriteBufferAcrossCFs,
::testing::Values(std::make_tuple(true, false),
std::make_tuple(false, false),
std::make_tuple(false, true)));
TEST_F(DBTest2, SharedWriteBufferLimitAcrossDB) {
std::string dbname2 = test::PerThreadDBPath("db_shared_wb_db2");
Options options = CurrentOptions();
options.arena_block_size = 4096;
auto flush_listener = std::make_shared<FlushCounterListener>();
options.listeners.push_back(flush_listener);
// Don't trip the listener at shutdown.
options.avoid_flush_during_shutdown = true;
// Avoid undeterministic value by malloc_usable_size();
// Force arena block size to 1
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"Arena::Arena:0", [&](void* arg) {
size_t* block_size = static_cast<size_t*>(arg);
*block_size = 1;
});
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"Arena::AllocateNewBlock:0", [&](void* arg) {
std::pair<size_t*, size_t*>* pair =
static_cast<std::pair<size_t*, size_t*>*>(arg);
*std::get<0>(*pair) = *std::get<1>(*pair);
});
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
options.write_buffer_size = 500000; // this is never hit
// Use a write buffer total size so that the soft limit is about
// 105000.
options.write_buffer_manager.reset(new WriteBufferManager(120000));
CreateAndReopenWithCF({"cf1", "cf2"}, options);
ASSERT_OK(DestroyDB(dbname2, options));
DB* db2 = nullptr;
ASSERT_OK(DB::Open(options, dbname2, &db2));
WriteOptions wo;
wo.disableWAL = true;
std::function<void()> wait_flush = [&]() {
dbfull()->TEST_WaitForFlushMemTable(handles_[0]);
dbfull()->TEST_WaitForFlushMemTable(handles_[1]);
dbfull()->TEST_WaitForFlushMemTable(handles_[2]);
static_cast<DBImpl*>(db2)->TEST_WaitForFlushMemTable();
};
// Trigger a flush on cf2
flush_listener->expected_flush_reason = FlushReason::kWriteBufferManager;
ASSERT_OK(Put(2, Key(1), DummyString(70000), wo));
wait_flush();
ASSERT_OK(Put(0, Key(1), DummyString(20000), wo));
wait_flush();
// Insert to DB2
ASSERT_OK(db2->Put(wo, Key(2), DummyString(20000)));
wait_flush();
ASSERT_OK(Put(2, Key(1), DummyString(1), wo));
wait_flush();
static_cast<DBImpl*>(db2)->TEST_WaitForFlushMemTable();
{
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "default") +
GetNumberOfSstFilesForColumnFamily(db_, "cf1") +
GetNumberOfSstFilesForColumnFamily(db_, "cf2"),
static_cast<uint64_t>(1));
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db2, "default"),
static_cast<uint64_t>(0));
}
// Triggering to flush another CF in DB1
ASSERT_OK(db2->Put(wo, Key(2), DummyString(70000)));
wait_flush();
ASSERT_OK(Put(2, Key(1), DummyString(1), wo));
wait_flush();
{
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "default"),
static_cast<uint64_t>(1));
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "cf1"),
static_cast<uint64_t>(0));
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "cf2"),
static_cast<uint64_t>(1));
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db2, "default"),
static_cast<uint64_t>(0));
}
// Triggering flush in DB2.
ASSERT_OK(db2->Put(wo, Key(3), DummyString(40000)));
wait_flush();
ASSERT_OK(db2->Put(wo, Key(1), DummyString(1)));
wait_flush();
static_cast<DBImpl*>(db2)->TEST_WaitForFlushMemTable();
{
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "default"),
static_cast<uint64_t>(1));
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "cf1"),
static_cast<uint64_t>(0));
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db_, "cf2"),
static_cast<uint64_t>(1));
ASSERT_EQ(GetNumberOfSstFilesForColumnFamily(db2, "default"),
static_cast<uint64_t>(1));
}
delete db2;
ASSERT_OK(DestroyDB(dbname2, options));
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
}
TEST_F(DBTest2, TestWriteBufferNoLimitWithCache) {
Options options = CurrentOptions();
options.arena_block_size = 4096;
std::shared_ptr<Cache> cache =
NewLRUCache(LRUCacheOptions(10000000, 1, false, 0.0));
options.write_buffer_size = 50000; // this is never hit
// Use a write buffer total size so that the soft limit is about
// 105000.
options.write_buffer_manager.reset(new WriteBufferManager(0, cache));
Reopen(options);
ASSERT_OK(Put("foo", "bar"));
// One dummy entry is 256KB.
ASSERT_GT(cache->GetUsage(), 128000);
}
2016-03-22 20:07:15 +01:00
namespace {
void ValidateKeyExistence(DB* db, const std::vector<Slice>& keys_must_exist,
const std::vector<Slice>& keys_must_not_exist) {
// Ensure that expected keys exist
std::vector<std::string> values;
if (keys_must_exist.size() > 0) {
std::vector<Status> status_list =
db->MultiGet(ReadOptions(), keys_must_exist, &values);
for (size_t i = 0; i < keys_must_exist.size(); i++) {
ASSERT_OK(status_list[i]);
}
}
// Ensure that given keys don't exist
if (keys_must_not_exist.size() > 0) {
std::vector<Status> status_list =
db->MultiGet(ReadOptions(), keys_must_not_exist, &values);
for (size_t i = 0; i < keys_must_not_exist.size(); i++) {
ASSERT_TRUE(status_list[i].IsNotFound());
}
}
}
} // namespace
TEST_F(DBTest2, WalFilterTest) {
class TestWalFilter : public WalFilter {
private:
// Processing option that is requested to be applied at the given index
WalFilter::WalProcessingOption wal_processing_option_;
// Index at which to apply wal_processing_option_
// At other indexes default wal_processing_option::kContinueProcessing is
// returned.
size_t apply_option_at_record_index_;
// Current record index, incremented with each record encountered.
size_t current_record_index_;
public:
TestWalFilter(WalFilter::WalProcessingOption wal_processing_option,
size_t apply_option_for_record_index)
: wal_processing_option_(wal_processing_option),
apply_option_at_record_index_(apply_option_for_record_index),
current_record_index_(0) {}
WalProcessingOption LogRecord(const WriteBatch& /*batch*/,
WriteBatch* /*new_batch*/,
bool* /*batch_changed*/) const override {
2016-03-22 20:07:15 +01:00
WalFilter::WalProcessingOption option_to_return;
if (current_record_index_ == apply_option_at_record_index_) {
option_to_return = wal_processing_option_;
}
else {
option_to_return = WalProcessingOption::kContinueProcessing;
}
// Filter is passed as a const object for RocksDB to not modify the
// object, however we modify it for our own purpose here and hence
// cast the constness away.
(const_cast<TestWalFilter*>(this)->current_record_index_)++;
return option_to_return;
}
const char* Name() const override { return "TestWalFilter"; }
2016-03-22 20:07:15 +01:00
};
// Create 3 batches with two keys each
std::vector<std::vector<std::string>> batch_keys(3);
batch_keys[0].push_back("key1");
batch_keys[0].push_back("key2");
batch_keys[1].push_back("key3");
batch_keys[1].push_back("key4");
batch_keys[2].push_back("key5");
batch_keys[2].push_back("key6");
// Test with all WAL processing options
for (int option = 0;
option < static_cast<int>(
WalFilter::WalProcessingOption::kWalProcessingOptionMax);
option++) {
Options options = OptionsForLogIterTest();
DestroyAndReopen(options);
CreateAndReopenWithCF({ "pikachu" }, options);
// Write given keys in given batches
for (size_t i = 0; i < batch_keys.size(); i++) {
WriteBatch batch;
for (size_t j = 0; j < batch_keys[i].size(); j++) {
batch.Put(handles_[0], batch_keys[i][j], DummyString(1024));
}
dbfull()->Write(WriteOptions(), &batch);
}
WalFilter::WalProcessingOption wal_processing_option =
static_cast<WalFilter::WalProcessingOption>(option);
// Create a test filter that would apply wal_processing_option at the first
// record
size_t apply_option_for_record_index = 1;
TestWalFilter test_wal_filter(wal_processing_option,
apply_option_for_record_index);
// Reopen database with option to use WAL filter
options = OptionsForLogIterTest();
options.wal_filter = &test_wal_filter;
Status status =
TryReopenWithColumnFamilies({ "default", "pikachu" }, options);
if (wal_processing_option ==
WalFilter::WalProcessingOption::kCorruptedRecord) {
assert(!status.ok());
// In case of corruption we can turn off paranoid_checks to reopen
// databse
options.paranoid_checks = false;
ReopenWithColumnFamilies({ "default", "pikachu" }, options);
}
else {
assert(status.ok());
}
// Compute which keys we expect to be found
// and which we expect not to be found after recovery.
std::vector<Slice> keys_must_exist;
std::vector<Slice> keys_must_not_exist;
switch (wal_processing_option) {
case WalFilter::WalProcessingOption::kCorruptedRecord:
case WalFilter::WalProcessingOption::kContinueProcessing: {
fprintf(stderr, "Testing with complete WAL processing\n");
// we expect all records to be processed
for (size_t i = 0; i < batch_keys.size(); i++) {
for (size_t j = 0; j < batch_keys[i].size(); j++) {
keys_must_exist.push_back(Slice(batch_keys[i][j]));
}
}
break;
}
case WalFilter::WalProcessingOption::kIgnoreCurrentRecord: {
fprintf(stderr,
"Testing with ignoring record %" ROCKSDB_PRIszt " only\n",
apply_option_for_record_index);
// We expect the record with apply_option_for_record_index to be not
// found.
for (size_t i = 0; i < batch_keys.size(); i++) {
for (size_t j = 0; j < batch_keys[i].size(); j++) {
if (i == apply_option_for_record_index) {
keys_must_not_exist.push_back(Slice(batch_keys[i][j]));
}
else {
keys_must_exist.push_back(Slice(batch_keys[i][j]));
}
}
}
break;
}
case WalFilter::WalProcessingOption::kStopReplay: {
fprintf(stderr,
"Testing with stopping replay from record %" ROCKSDB_PRIszt
"\n",
apply_option_for_record_index);
// We expect records beyond apply_option_for_record_index to be not
// found.
for (size_t i = 0; i < batch_keys.size(); i++) {
for (size_t j = 0; j < batch_keys[i].size(); j++) {
if (i >= apply_option_for_record_index) {
keys_must_not_exist.push_back(Slice(batch_keys[i][j]));
}
else {
keys_must_exist.push_back(Slice(batch_keys[i][j]));
}
}
}
break;
}
default:
assert(false); // unhandled case
}
bool checked_after_reopen = false;
while (true) {
// Ensure that expected keys exists
// and not expected keys don't exist after recovery
ValidateKeyExistence(db_, keys_must_exist, keys_must_not_exist);
if (checked_after_reopen) {
break;
}
// reopen database again to make sure previous log(s) are not used
//(even if they were skipped)
// reopn database with option to use WAL filter
options = OptionsForLogIterTest();
ReopenWithColumnFamilies({ "default", "pikachu" }, options);
checked_after_reopen = true;
}
}
}
TEST_F(DBTest2, WalFilterTestWithChangeBatch) {
class ChangeBatchHandler : public WriteBatch::Handler {
private:
// Batch to insert keys in
WriteBatch* new_write_batch_;
// Number of keys to add in the new batch
size_t num_keys_to_add_in_new_batch_;
// Number of keys added to new batch
size_t num_keys_added_;
public:
ChangeBatchHandler(WriteBatch* new_write_batch,
size_t num_keys_to_add_in_new_batch)
: new_write_batch_(new_write_batch),
num_keys_to_add_in_new_batch_(num_keys_to_add_in_new_batch),
num_keys_added_(0) {}
void Put(const Slice& key, const Slice& value) override {
2016-03-22 20:07:15 +01:00
if (num_keys_added_ < num_keys_to_add_in_new_batch_) {
new_write_batch_->Put(key, value);
++num_keys_added_;
}
}
};
class TestWalFilterWithChangeBatch : public WalFilter {
private:
// Index at which to start changing records
size_t change_records_from_index_;
// Number of keys to add in the new batch
size_t num_keys_to_add_in_new_batch_;
// Current record index, incremented with each record encountered.
size_t current_record_index_;
public:
TestWalFilterWithChangeBatch(size_t change_records_from_index,
size_t num_keys_to_add_in_new_batch)
: change_records_from_index_(change_records_from_index),
num_keys_to_add_in_new_batch_(num_keys_to_add_in_new_batch),
current_record_index_(0) {}
WalProcessingOption LogRecord(const WriteBatch& batch,
WriteBatch* new_batch,
bool* batch_changed) const override {
2016-03-22 20:07:15 +01:00
if (current_record_index_ >= change_records_from_index_) {
ChangeBatchHandler handler(new_batch, num_keys_to_add_in_new_batch_);
batch.Iterate(&handler);
*batch_changed = true;
}
// Filter is passed as a const object for RocksDB to not modify the
// object, however we modify it for our own purpose here and hence
// cast the constness away.
(const_cast<TestWalFilterWithChangeBatch*>(this)
->current_record_index_)++;
return WalProcessingOption::kContinueProcessing;
}
const char* Name() const override { return "TestWalFilterWithChangeBatch"; }
2016-03-22 20:07:15 +01:00
};
std::vector<std::vector<std::string>> batch_keys(3);
batch_keys[0].push_back("key1");
batch_keys[0].push_back("key2");
batch_keys[1].push_back("key3");
batch_keys[1].push_back("key4");
batch_keys[2].push_back("key5");
batch_keys[2].push_back("key6");
Options options = OptionsForLogIterTest();
DestroyAndReopen(options);
CreateAndReopenWithCF({ "pikachu" }, options);
// Write given keys in given batches
for (size_t i = 0; i < batch_keys.size(); i++) {
WriteBatch batch;
for (size_t j = 0; j < batch_keys[i].size(); j++) {
batch.Put(handles_[0], batch_keys[i][j], DummyString(1024));
}
dbfull()->Write(WriteOptions(), &batch);
}
// Create a test filter that would apply wal_processing_option at the first
// record
size_t change_records_from_index = 1;
size_t num_keys_to_add_in_new_batch = 1;
TestWalFilterWithChangeBatch test_wal_filter_with_change_batch(
change_records_from_index, num_keys_to_add_in_new_batch);
// Reopen database with option to use WAL filter
options = OptionsForLogIterTest();
options.wal_filter = &test_wal_filter_with_change_batch;
ReopenWithColumnFamilies({ "default", "pikachu" }, options);
// Ensure that all keys exist before change_records_from_index_
// And after that index only single key exists
// as our filter adds only single key for each batch
std::vector<Slice> keys_must_exist;
std::vector<Slice> keys_must_not_exist;
for (size_t i = 0; i < batch_keys.size(); i++) {
for (size_t j = 0; j < batch_keys[i].size(); j++) {
if (i >= change_records_from_index && j >= num_keys_to_add_in_new_batch) {
2016-03-22 20:07:15 +01:00
keys_must_not_exist.push_back(Slice(batch_keys[i][j]));
}
else {
keys_must_exist.push_back(Slice(batch_keys[i][j]));
}
}
}
bool checked_after_reopen = false;
while (true) {
// Ensure that expected keys exists
// and not expected keys don't exist after recovery
ValidateKeyExistence(db_, keys_must_exist, keys_must_not_exist);
if (checked_after_reopen) {
break;
}
// reopen database again to make sure previous log(s) are not used
//(even if they were skipped)
// reopn database with option to use WAL filter
options = OptionsForLogIterTest();
ReopenWithColumnFamilies({ "default", "pikachu" }, options);
checked_after_reopen = true;
}
}
TEST_F(DBTest2, WalFilterTestWithChangeBatchExtraKeys) {
class TestWalFilterWithChangeBatchAddExtraKeys : public WalFilter {
public:
WalProcessingOption LogRecord(const WriteBatch& batch, WriteBatch* new_batch,
bool* batch_changed) const override {
*new_batch = batch;
new_batch->Put("key_extra", "value_extra");
*batch_changed = true;
return WalProcessingOption::kContinueProcessing;
}
const char* Name() const override {
return "WalFilterTestWithChangeBatchExtraKeys";
}
2016-03-22 20:07:15 +01:00
};
std::vector<std::vector<std::string>> batch_keys(3);
batch_keys[0].push_back("key1");
batch_keys[0].push_back("key2");
batch_keys[1].push_back("key3");
batch_keys[1].push_back("key4");
batch_keys[2].push_back("key5");
batch_keys[2].push_back("key6");
Options options = OptionsForLogIterTest();
DestroyAndReopen(options);
CreateAndReopenWithCF({ "pikachu" }, options);
// Write given keys in given batches
for (size_t i = 0; i < batch_keys.size(); i++) {
WriteBatch batch;
for (size_t j = 0; j < batch_keys[i].size(); j++) {
batch.Put(handles_[0], batch_keys[i][j], DummyString(1024));
}
dbfull()->Write(WriteOptions(), &batch);
}
// Create a test filter that would add extra keys
TestWalFilterWithChangeBatchAddExtraKeys test_wal_filter_extra_keys;
// Reopen database with option to use WAL filter
options = OptionsForLogIterTest();
options.wal_filter = &test_wal_filter_extra_keys;
Status status = TryReopenWithColumnFamilies({"default", "pikachu"}, options);
2016-03-22 20:07:15 +01:00
ASSERT_TRUE(status.IsNotSupported());
// Reopen without filter, now reopen should succeed - previous
// attempt to open must not have altered the db.
options = OptionsForLogIterTest();
ReopenWithColumnFamilies({ "default", "pikachu" }, options);
std::vector<Slice> keys_must_exist;
std::vector<Slice> keys_must_not_exist; // empty vector
for (size_t i = 0; i < batch_keys.size(); i++) {
for (size_t j = 0; j < batch_keys[i].size(); j++) {
keys_must_exist.push_back(Slice(batch_keys[i][j]));
}
}
ValidateKeyExistence(db_, keys_must_exist, keys_must_not_exist);
}
TEST_F(DBTest2, WalFilterTestWithColumnFamilies) {
class TestWalFilterWithColumnFamilies : public WalFilter {
private:
// column_family_id -> log_number map (provided to WALFilter)
std::map<uint32_t, uint64_t> cf_log_number_map_;
// column_family_name -> column_family_id map (provided to WALFilter)
std::map<std::string, uint32_t> cf_name_id_map_;
// column_family_name -> keys_found_in_wal map
// We store keys that are applicable to the column_family
// during recovery (i.e. aren't already flushed to SST file(s))
// for verification against the keys we expect.
std::map<uint32_t, std::vector<std::string>> cf_wal_keys_;
public:
void ColumnFamilyLogNumberMap(
const std::map<uint32_t, uint64_t>& cf_lognumber_map,
const std::map<std::string, uint32_t>& cf_name_id_map) override {
cf_log_number_map_ = cf_lognumber_map;
cf_name_id_map_ = cf_name_id_map;
}
WalProcessingOption LogRecordFound(unsigned long long log_number,
const std::string& /*log_file_name*/,
const WriteBatch& batch,
WriteBatch* /*new_batch*/,
bool* /*batch_changed*/) override {
class LogRecordBatchHandler : public WriteBatch::Handler {
2016-03-22 20:07:15 +01:00
private:
const std::map<uint32_t, uint64_t> & cf_log_number_map_;
std::map<uint32_t, std::vector<std::string>> & cf_wal_keys_;
unsigned long long log_number_;
public:
LogRecordBatchHandler(unsigned long long current_log_number,
const std::map<uint32_t, uint64_t> & cf_log_number_map,
std::map<uint32_t, std::vector<std::string>> & cf_wal_keys) :
cf_log_number_map_(cf_log_number_map),
cf_wal_keys_(cf_wal_keys),
log_number_(current_log_number){}
Status PutCF(uint32_t column_family_id, const Slice& key,
const Slice& /*value*/) override {
2016-03-22 20:07:15 +01:00
auto it = cf_log_number_map_.find(column_family_id);
assert(it != cf_log_number_map_.end());
unsigned long long log_number_for_cf = it->second;
// If the current record is applicable for column_family_id
// (i.e. isn't flushed to SST file(s) for column_family_id)
// add it to the cf_wal_keys_ map for verification.
if (log_number_ >= log_number_for_cf) {
cf_wal_keys_[column_family_id].push_back(std::string(key.data(),
key.size()));
}
return Status::OK();
}
} handler(log_number, cf_log_number_map_, cf_wal_keys_);
batch.Iterate(&handler);
return WalProcessingOption::kContinueProcessing;
}
2016-03-22 20:07:15 +01:00
const char* Name() const override {
return "WalFilterTestWithColumnFamilies";
}
2016-03-22 20:07:15 +01:00
const std::map<uint32_t, std::vector<std::string>>& GetColumnFamilyKeys() {
2016-03-22 20:07:15 +01:00
return cf_wal_keys_;
}
const std::map<std::string, uint32_t> & GetColumnFamilyNameIdMap() {
return cf_name_id_map_;
}
};
std::vector<std::vector<std::string>> batch_keys_pre_flush(3);
batch_keys_pre_flush[0].push_back("key1");
batch_keys_pre_flush[0].push_back("key2");
batch_keys_pre_flush[1].push_back("key3");
batch_keys_pre_flush[1].push_back("key4");
batch_keys_pre_flush[2].push_back("key5");
batch_keys_pre_flush[2].push_back("key6");
Options options = OptionsForLogIterTest();
DestroyAndReopen(options);
CreateAndReopenWithCF({ "pikachu" }, options);
// Write given keys in given batches
for (size_t i = 0; i < batch_keys_pre_flush.size(); i++) {
WriteBatch batch;
for (size_t j = 0; j < batch_keys_pre_flush[i].size(); j++) {
batch.Put(handles_[0], batch_keys_pre_flush[i][j], DummyString(1024));
batch.Put(handles_[1], batch_keys_pre_flush[i][j], DummyString(1024));
}
dbfull()->Write(WriteOptions(), &batch);
}
//Flush default column-family
db_->Flush(FlushOptions(), handles_[0]);
// Do some more writes
std::vector<std::vector<std::string>> batch_keys_post_flush(3);
batch_keys_post_flush[0].push_back("key7");
batch_keys_post_flush[0].push_back("key8");
batch_keys_post_flush[1].push_back("key9");
batch_keys_post_flush[1].push_back("key10");
batch_keys_post_flush[2].push_back("key11");
batch_keys_post_flush[2].push_back("key12");
// Write given keys in given batches
for (size_t i = 0; i < batch_keys_post_flush.size(); i++) {
WriteBatch batch;
for (size_t j = 0; j < batch_keys_post_flush[i].size(); j++) {
batch.Put(handles_[0], batch_keys_post_flush[i][j], DummyString(1024));
batch.Put(handles_[1], batch_keys_post_flush[i][j], DummyString(1024));
}
dbfull()->Write(WriteOptions(), &batch);
}
// On Recovery we should only find the second batch applicable to default CF
// But both batches applicable to pikachu CF
// Create a test filter that would add extra keys
TestWalFilterWithColumnFamilies test_wal_filter_column_families;
// Reopen database with option to use WAL filter
options = OptionsForLogIterTest();
options.wal_filter = &test_wal_filter_column_families;
Status status =
TryReopenWithColumnFamilies({ "default", "pikachu" }, options);
ASSERT_TRUE(status.ok());
// verify that handles_[0] only has post_flush keys
// while handles_[1] has pre and post flush keys
auto cf_wal_keys = test_wal_filter_column_families.GetColumnFamilyKeys();
auto name_id_map = test_wal_filter_column_families.GetColumnFamilyNameIdMap();
2016-03-22 20:07:15 +01:00
size_t index = 0;
auto keys_cf = cf_wal_keys[name_id_map[kDefaultColumnFamilyName]];
//default column-family, only post_flush keys are expected
for (size_t i = 0; i < batch_keys_post_flush.size(); i++) {
for (size_t j = 0; j < batch_keys_post_flush[i].size(); j++) {
Slice key_from_the_log(keys_cf[index++]);
Slice batch_key(batch_keys_post_flush[i][j]);
ASSERT_TRUE(key_from_the_log.compare(batch_key) == 0);
}
}
ASSERT_TRUE(index == keys_cf.size());
index = 0;
keys_cf = cf_wal_keys[name_id_map["pikachu"]];
//pikachu column-family, all keys are expected
for (size_t i = 0; i < batch_keys_pre_flush.size(); i++) {
for (size_t j = 0; j < batch_keys_pre_flush[i].size(); j++) {
Slice key_from_the_log(keys_cf[index++]);
Slice batch_key(batch_keys_pre_flush[i][j]);
ASSERT_TRUE(key_from_the_log.compare(batch_key) == 0);
}
}
for (size_t i = 0; i < batch_keys_post_flush.size(); i++) {
for (size_t j = 0; j < batch_keys_post_flush[i].size(); j++) {
Slice key_from_the_log(keys_cf[index++]);
Slice batch_key(batch_keys_post_flush[i][j]);
ASSERT_TRUE(key_from_the_log.compare(batch_key) == 0);
}
}
ASSERT_TRUE(index == keys_cf.size());
}
TEST_F(DBTest2, PresetCompressionDict) {
Reduce scope of compression dictionary to single SST (#4952) Summary: Our previous approach was to train one compression dictionary per compaction, using the first output SST to train a dictionary, and then applying it on subsequent SSTs in the same compaction. While this was great for minimizing CPU/memory/I/O overhead, it did not achieve good compression ratios in practice. In our most promising potential use case, moderate reductions in a dictionary's scope make a major difference on compression ratio. So, this PR changes compression dictionary to be scoped per-SST. It accepts the tradeoff during table building to use more memory and CPU. Important changes include: - The `BlockBasedTableBuilder` has a new state when dictionary compression is in-use: `kBuffered`. In that state it accumulates uncompressed data in-memory whenever `Add` is called. - After accumulating target file size bytes or calling `BlockBasedTableBuilder::Finish`, a `BlockBasedTableBuilder` moves to the `kUnbuffered` state. The transition (`EnterUnbuffered()`) involves sampling the buffered data, training a dictionary, and compressing/writing out all buffered data. In the `kUnbuffered` state, a `BlockBasedTableBuilder` behaves the same as before -- blocks are compressed/written out as soon as they fill up. - Samples are now whole uncompressed data blocks, except the final sample may be a partial data block so we don't breach the user's configured `max_dict_bytes` or `zstd_max_train_bytes`. The dictionary trainer is supposed to work better when we pass it real units of compression. Previously we were passing 64-byte KV samples which was not realistic. Pull Request resolved: https://github.com/facebook/rocksdb/pull/4952 Differential Revision: D13967980 Pulled By: ajkr fbshipit-source-id: 82bea6f7537e1529c7a1a4cdee84585f5949300f
2019-02-12 04:42:25 +01:00
// Verifies that compression ratio improves when dictionary is enabled, and
// improves even further when the dictionary is trained by ZSTD.
const size_t kBlockSizeBytes = 4 << 10;
const size_t kL0FileBytes = 128 << 10;
const size_t kApproxPerBlockOverheadBytes = 50;
const int kNumL0Files = 5;
Options options;
// Make sure to use any custom env that the test is configured with.
options.env = CurrentOptions().env;
options.allow_concurrent_memtable_write = false;
options.arena_block_size = kBlockSizeBytes;
options.create_if_missing = true;
options.disable_auto_compactions = true;
options.level0_file_num_compaction_trigger = kNumL0Files;
options.memtable_factory.reset(
new SpecialSkipListFactory(kL0FileBytes / kBlockSizeBytes));
options.num_levels = 2;
options.target_file_size_base = kL0FileBytes;
options.target_file_size_multiplier = 2;
options.write_buffer_size = kL0FileBytes;
BlockBasedTableOptions table_options;
table_options.block_size = kBlockSizeBytes;
std::vector<CompressionType> compression_types;
if (Zlib_Supported()) {
compression_types.push_back(kZlibCompression);
}
#if LZ4_VERSION_NUMBER >= 10400 // r124+
compression_types.push_back(kLZ4Compression);
compression_types.push_back(kLZ4HCCompression);
#endif // LZ4_VERSION_NUMBER >= 10400
if (ZSTD_Supported()) {
compression_types.push_back(kZSTD);
}
enum DictionaryTypes : int {
kWithoutDict,
kWithDict,
kWithZSTDTrainedDict,
kDictEnd,
};
for (auto compression_type : compression_types) {
options.compression = compression_type;
size_t bytes_without_dict = 0;
size_t bytes_with_dict = 0;
size_t bytes_with_zstd_trained_dict = 0;
for (int i = kWithoutDict; i < kDictEnd; i++) {
// First iteration: compress without preset dictionary
// Second iteration: compress with preset dictionary
// Third iteration (zstd only): compress with zstd-trained dictionary
//
// To make sure the compression dictionary has the intended effect, we
// verify the compressed size is smaller in successive iterations. Also in
// the non-first iterations, verify the data we get out is the same data
// we put in.
switch (i) {
case kWithoutDict:
options.compression_opts.max_dict_bytes = 0;
options.compression_opts.zstd_max_train_bytes = 0;
break;
case kWithDict:
options.compression_opts.max_dict_bytes = kBlockSizeBytes;
options.compression_opts.zstd_max_train_bytes = 0;
break;
case kWithZSTDTrainedDict:
if (compression_type != kZSTD) {
continue;
}
options.compression_opts.max_dict_bytes = kBlockSizeBytes;
Reduce scope of compression dictionary to single SST (#4952) Summary: Our previous approach was to train one compression dictionary per compaction, using the first output SST to train a dictionary, and then applying it on subsequent SSTs in the same compaction. While this was great for minimizing CPU/memory/I/O overhead, it did not achieve good compression ratios in practice. In our most promising potential use case, moderate reductions in a dictionary's scope make a major difference on compression ratio. So, this PR changes compression dictionary to be scoped per-SST. It accepts the tradeoff during table building to use more memory and CPU. Important changes include: - The `BlockBasedTableBuilder` has a new state when dictionary compression is in-use: `kBuffered`. In that state it accumulates uncompressed data in-memory whenever `Add` is called. - After accumulating target file size bytes or calling `BlockBasedTableBuilder::Finish`, a `BlockBasedTableBuilder` moves to the `kUnbuffered` state. The transition (`EnterUnbuffered()`) involves sampling the buffered data, training a dictionary, and compressing/writing out all buffered data. In the `kUnbuffered` state, a `BlockBasedTableBuilder` behaves the same as before -- blocks are compressed/written out as soon as they fill up. - Samples are now whole uncompressed data blocks, except the final sample may be a partial data block so we don't breach the user's configured `max_dict_bytes` or `zstd_max_train_bytes`. The dictionary trainer is supposed to work better when we pass it real units of compression. Previously we were passing 64-byte KV samples which was not realistic. Pull Request resolved: https://github.com/facebook/rocksdb/pull/4952 Differential Revision: D13967980 Pulled By: ajkr fbshipit-source-id: 82bea6f7537e1529c7a1a4cdee84585f5949300f
2019-02-12 04:42:25 +01:00
options.compression_opts.zstd_max_train_bytes = kL0FileBytes;
break;
default:
assert(false);
}
options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
CreateAndReopenWithCF({"pikachu"}, options);
Random rnd(301);
Reduce scope of compression dictionary to single SST (#4952) Summary: Our previous approach was to train one compression dictionary per compaction, using the first output SST to train a dictionary, and then applying it on subsequent SSTs in the same compaction. While this was great for minimizing CPU/memory/I/O overhead, it did not achieve good compression ratios in practice. In our most promising potential use case, moderate reductions in a dictionary's scope make a major difference on compression ratio. So, this PR changes compression dictionary to be scoped per-SST. It accepts the tradeoff during table building to use more memory and CPU. Important changes include: - The `BlockBasedTableBuilder` has a new state when dictionary compression is in-use: `kBuffered`. In that state it accumulates uncompressed data in-memory whenever `Add` is called. - After accumulating target file size bytes or calling `BlockBasedTableBuilder::Finish`, a `BlockBasedTableBuilder` moves to the `kUnbuffered` state. The transition (`EnterUnbuffered()`) involves sampling the buffered data, training a dictionary, and compressing/writing out all buffered data. In the `kUnbuffered` state, a `BlockBasedTableBuilder` behaves the same as before -- blocks are compressed/written out as soon as they fill up. - Samples are now whole uncompressed data blocks, except the final sample may be a partial data block so we don't breach the user's configured `max_dict_bytes` or `zstd_max_train_bytes`. The dictionary trainer is supposed to work better when we pass it real units of compression. Previously we were passing 64-byte KV samples which was not realistic. Pull Request resolved: https://github.com/facebook/rocksdb/pull/4952 Differential Revision: D13967980 Pulled By: ajkr fbshipit-source-id: 82bea6f7537e1529c7a1a4cdee84585f5949300f
2019-02-12 04:42:25 +01:00
std::string seq_datas[10];
for (int j = 0; j < 10; ++j) {
seq_datas[j] =
rnd.RandomString(kBlockSizeBytes - kApproxPerBlockOverheadBytes);
Reduce scope of compression dictionary to single SST (#4952) Summary: Our previous approach was to train one compression dictionary per compaction, using the first output SST to train a dictionary, and then applying it on subsequent SSTs in the same compaction. While this was great for minimizing CPU/memory/I/O overhead, it did not achieve good compression ratios in practice. In our most promising potential use case, moderate reductions in a dictionary's scope make a major difference on compression ratio. So, this PR changes compression dictionary to be scoped per-SST. It accepts the tradeoff during table building to use more memory and CPU. Important changes include: - The `BlockBasedTableBuilder` has a new state when dictionary compression is in-use: `kBuffered`. In that state it accumulates uncompressed data in-memory whenever `Add` is called. - After accumulating target file size bytes or calling `BlockBasedTableBuilder::Finish`, a `BlockBasedTableBuilder` moves to the `kUnbuffered` state. The transition (`EnterUnbuffered()`) involves sampling the buffered data, training a dictionary, and compressing/writing out all buffered data. In the `kUnbuffered` state, a `BlockBasedTableBuilder` behaves the same as before -- blocks are compressed/written out as soon as they fill up. - Samples are now whole uncompressed data blocks, except the final sample may be a partial data block so we don't breach the user's configured `max_dict_bytes` or `zstd_max_train_bytes`. The dictionary trainer is supposed to work better when we pass it real units of compression. Previously we were passing 64-byte KV samples which was not realistic. Pull Request resolved: https://github.com/facebook/rocksdb/pull/4952 Differential Revision: D13967980 Pulled By: ajkr fbshipit-source-id: 82bea6f7537e1529c7a1a4cdee84585f5949300f
2019-02-12 04:42:25 +01:00
}
ASSERT_EQ(0, NumTableFilesAtLevel(0, 1));
for (int j = 0; j < kNumL0Files; ++j) {
for (size_t k = 0; k < kL0FileBytes / kBlockSizeBytes + 1; ++k) {
Reduce scope of compression dictionary to single SST (#4952) Summary: Our previous approach was to train one compression dictionary per compaction, using the first output SST to train a dictionary, and then applying it on subsequent SSTs in the same compaction. While this was great for minimizing CPU/memory/I/O overhead, it did not achieve good compression ratios in practice. In our most promising potential use case, moderate reductions in a dictionary's scope make a major difference on compression ratio. So, this PR changes compression dictionary to be scoped per-SST. It accepts the tradeoff during table building to use more memory and CPU. Important changes include: - The `BlockBasedTableBuilder` has a new state when dictionary compression is in-use: `kBuffered`. In that state it accumulates uncompressed data in-memory whenever `Add` is called. - After accumulating target file size bytes or calling `BlockBasedTableBuilder::Finish`, a `BlockBasedTableBuilder` moves to the `kUnbuffered` state. The transition (`EnterUnbuffered()`) involves sampling the buffered data, training a dictionary, and compressing/writing out all buffered data. In the `kUnbuffered` state, a `BlockBasedTableBuilder` behaves the same as before -- blocks are compressed/written out as soon as they fill up. - Samples are now whole uncompressed data blocks, except the final sample may be a partial data block so we don't breach the user's configured `max_dict_bytes` or `zstd_max_train_bytes`. The dictionary trainer is supposed to work better when we pass it real units of compression. Previously we were passing 64-byte KV samples which was not realistic. Pull Request resolved: https://github.com/facebook/rocksdb/pull/4952 Differential Revision: D13967980 Pulled By: ajkr fbshipit-source-id: 82bea6f7537e1529c7a1a4cdee84585f5949300f
2019-02-12 04:42:25 +01:00
auto key_num = j * (kL0FileBytes / kBlockSizeBytes) + k;
ASSERT_OK(Put(1, Key(static_cast<int>(key_num)),
seq_datas[(key_num / 10) % 10]));
}
dbfull()->TEST_WaitForFlushMemTable(handles_[1]);
ASSERT_EQ(j + 1, NumTableFilesAtLevel(0, 1));
}
Reduce scope of compression dictionary to single SST (#4952) Summary: Our previous approach was to train one compression dictionary per compaction, using the first output SST to train a dictionary, and then applying it on subsequent SSTs in the same compaction. While this was great for minimizing CPU/memory/I/O overhead, it did not achieve good compression ratios in practice. In our most promising potential use case, moderate reductions in a dictionary's scope make a major difference on compression ratio. So, this PR changes compression dictionary to be scoped per-SST. It accepts the tradeoff during table building to use more memory and CPU. Important changes include: - The `BlockBasedTableBuilder` has a new state when dictionary compression is in-use: `kBuffered`. In that state it accumulates uncompressed data in-memory whenever `Add` is called. - After accumulating target file size bytes or calling `BlockBasedTableBuilder::Finish`, a `BlockBasedTableBuilder` moves to the `kUnbuffered` state. The transition (`EnterUnbuffered()`) involves sampling the buffered data, training a dictionary, and compressing/writing out all buffered data. In the `kUnbuffered` state, a `BlockBasedTableBuilder` behaves the same as before -- blocks are compressed/written out as soon as they fill up. - Samples are now whole uncompressed data blocks, except the final sample may be a partial data block so we don't breach the user's configured `max_dict_bytes` or `zstd_max_train_bytes`. The dictionary trainer is supposed to work better when we pass it real units of compression. Previously we were passing 64-byte KV samples which was not realistic. Pull Request resolved: https://github.com/facebook/rocksdb/pull/4952 Differential Revision: D13967980 Pulled By: ajkr fbshipit-source-id: 82bea6f7537e1529c7a1a4cdee84585f5949300f
2019-02-12 04:42:25 +01:00
dbfull()->TEST_CompactRange(0, nullptr, nullptr, handles_[1],
true /* disallow_trivial_move */);
ASSERT_EQ(0, NumTableFilesAtLevel(0, 1));
ASSERT_GT(NumTableFilesAtLevel(1, 1), 0);
// Get the live sst files size
size_t total_sst_bytes = TotalSize(1);
if (i == kWithoutDict) {
bytes_without_dict = total_sst_bytes;
} else if (i == kWithDict) {
bytes_with_dict = total_sst_bytes;
} else if (i == kWithZSTDTrainedDict) {
bytes_with_zstd_trained_dict = total_sst_bytes;
}
for (size_t j = 0; j < kNumL0Files * (kL0FileBytes / kBlockSizeBytes);
j++) {
Reduce scope of compression dictionary to single SST (#4952) Summary: Our previous approach was to train one compression dictionary per compaction, using the first output SST to train a dictionary, and then applying it on subsequent SSTs in the same compaction. While this was great for minimizing CPU/memory/I/O overhead, it did not achieve good compression ratios in practice. In our most promising potential use case, moderate reductions in a dictionary's scope make a major difference on compression ratio. So, this PR changes compression dictionary to be scoped per-SST. It accepts the tradeoff during table building to use more memory and CPU. Important changes include: - The `BlockBasedTableBuilder` has a new state when dictionary compression is in-use: `kBuffered`. In that state it accumulates uncompressed data in-memory whenever `Add` is called. - After accumulating target file size bytes or calling `BlockBasedTableBuilder::Finish`, a `BlockBasedTableBuilder` moves to the `kUnbuffered` state. The transition (`EnterUnbuffered()`) involves sampling the buffered data, training a dictionary, and compressing/writing out all buffered data. In the `kUnbuffered` state, a `BlockBasedTableBuilder` behaves the same as before -- blocks are compressed/written out as soon as they fill up. - Samples are now whole uncompressed data blocks, except the final sample may be a partial data block so we don't breach the user's configured `max_dict_bytes` or `zstd_max_train_bytes`. The dictionary trainer is supposed to work better when we pass it real units of compression. Previously we were passing 64-byte KV samples which was not realistic. Pull Request resolved: https://github.com/facebook/rocksdb/pull/4952 Differential Revision: D13967980 Pulled By: ajkr fbshipit-source-id: 82bea6f7537e1529c7a1a4cdee84585f5949300f
2019-02-12 04:42:25 +01:00
ASSERT_EQ(seq_datas[(j / 10) % 10], Get(1, Key(static_cast<int>(j))));
}
if (i == kWithDict) {
ASSERT_GT(bytes_without_dict, bytes_with_dict);
} else if (i == kWithZSTDTrainedDict) {
// In zstd compression, it is sometimes possible that using a trained
// dictionary does not get as good a compression ratio as without
// training.
// But using a dictionary (with or without training) should always get
// better compression ratio than not using one.
ASSERT_TRUE(bytes_with_dict > bytes_with_zstd_trained_dict ||
bytes_without_dict > bytes_with_zstd_trained_dict);
}
DestroyAndReopen(options);
}
}
}
Reduce scope of compression dictionary to single SST (#4952) Summary: Our previous approach was to train one compression dictionary per compaction, using the first output SST to train a dictionary, and then applying it on subsequent SSTs in the same compaction. While this was great for minimizing CPU/memory/I/O overhead, it did not achieve good compression ratios in practice. In our most promising potential use case, moderate reductions in a dictionary's scope make a major difference on compression ratio. So, this PR changes compression dictionary to be scoped per-SST. It accepts the tradeoff during table building to use more memory and CPU. Important changes include: - The `BlockBasedTableBuilder` has a new state when dictionary compression is in-use: `kBuffered`. In that state it accumulates uncompressed data in-memory whenever `Add` is called. - After accumulating target file size bytes or calling `BlockBasedTableBuilder::Finish`, a `BlockBasedTableBuilder` moves to the `kUnbuffered` state. The transition (`EnterUnbuffered()`) involves sampling the buffered data, training a dictionary, and compressing/writing out all buffered data. In the `kUnbuffered` state, a `BlockBasedTableBuilder` behaves the same as before -- blocks are compressed/written out as soon as they fill up. - Samples are now whole uncompressed data blocks, except the final sample may be a partial data block so we don't breach the user's configured `max_dict_bytes` or `zstd_max_train_bytes`. The dictionary trainer is supposed to work better when we pass it real units of compression. Previously we were passing 64-byte KV samples which was not realistic. Pull Request resolved: https://github.com/facebook/rocksdb/pull/4952 Differential Revision: D13967980 Pulled By: ajkr fbshipit-source-id: 82bea6f7537e1529c7a1a4cdee84585f5949300f
2019-02-12 04:42:25 +01:00
TEST_F(DBTest2, PresetCompressionDictLocality) {
if (!ZSTD_Supported()) {
return;
}
// Verifies that compression dictionary is generated from local data. The
// verification simply checks all output SSTs have different compression
// dictionaries. We do not verify effectiveness as that'd likely be flaky in
// the future.
const int kNumEntriesPerFile = 1 << 10; // 1KB
const int kNumBytesPerEntry = 1 << 10; // 1KB
const int kNumFiles = 4;
Options options = CurrentOptions();
options.compression = kZSTD;
options.compression_opts.max_dict_bytes = 1 << 14; // 16KB
options.compression_opts.zstd_max_train_bytes = 1 << 18; // 256KB
options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
Reduce scope of compression dictionary to single SST (#4952) Summary: Our previous approach was to train one compression dictionary per compaction, using the first output SST to train a dictionary, and then applying it on subsequent SSTs in the same compaction. While this was great for minimizing CPU/memory/I/O overhead, it did not achieve good compression ratios in practice. In our most promising potential use case, moderate reductions in a dictionary's scope make a major difference on compression ratio. So, this PR changes compression dictionary to be scoped per-SST. It accepts the tradeoff during table building to use more memory and CPU. Important changes include: - The `BlockBasedTableBuilder` has a new state when dictionary compression is in-use: `kBuffered`. In that state it accumulates uncompressed data in-memory whenever `Add` is called. - After accumulating target file size bytes or calling `BlockBasedTableBuilder::Finish`, a `BlockBasedTableBuilder` moves to the `kUnbuffered` state. The transition (`EnterUnbuffered()`) involves sampling the buffered data, training a dictionary, and compressing/writing out all buffered data. In the `kUnbuffered` state, a `BlockBasedTableBuilder` behaves the same as before -- blocks are compressed/written out as soon as they fill up. - Samples are now whole uncompressed data blocks, except the final sample may be a partial data block so we don't breach the user's configured `max_dict_bytes` or `zstd_max_train_bytes`. The dictionary trainer is supposed to work better when we pass it real units of compression. Previously we were passing 64-byte KV samples which was not realistic. Pull Request resolved: https://github.com/facebook/rocksdb/pull/4952 Differential Revision: D13967980 Pulled By: ajkr fbshipit-source-id: 82bea6f7537e1529c7a1a4cdee84585f5949300f
2019-02-12 04:42:25 +01:00
options.target_file_size_base = kNumEntriesPerFile * kNumBytesPerEntry;
BlockBasedTableOptions table_options;
table_options.cache_index_and_filter_blocks = true;
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566) Summary: This PR does a few things: 1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation. 2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed: - The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated - The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory). 3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10). I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged. Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently. Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566 Reviewed By: zhichao-cao Differential Revision: D24408980 Pulled By: jay-zhuang fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
2020-10-27 18:31:34 +01:00
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
Reduce scope of compression dictionary to single SST (#4952) Summary: Our previous approach was to train one compression dictionary per compaction, using the first output SST to train a dictionary, and then applying it on subsequent SSTs in the same compaction. While this was great for minimizing CPU/memory/I/O overhead, it did not achieve good compression ratios in practice. In our most promising potential use case, moderate reductions in a dictionary's scope make a major difference on compression ratio. So, this PR changes compression dictionary to be scoped per-SST. It accepts the tradeoff during table building to use more memory and CPU. Important changes include: - The `BlockBasedTableBuilder` has a new state when dictionary compression is in-use: `kBuffered`. In that state it accumulates uncompressed data in-memory whenever `Add` is called. - After accumulating target file size bytes or calling `BlockBasedTableBuilder::Finish`, a `BlockBasedTableBuilder` moves to the `kUnbuffered` state. The transition (`EnterUnbuffered()`) involves sampling the buffered data, training a dictionary, and compressing/writing out all buffered data. In the `kUnbuffered` state, a `BlockBasedTableBuilder` behaves the same as before -- blocks are compressed/written out as soon as they fill up. - Samples are now whole uncompressed data blocks, except the final sample may be a partial data block so we don't breach the user's configured `max_dict_bytes` or `zstd_max_train_bytes`. The dictionary trainer is supposed to work better when we pass it real units of compression. Previously we were passing 64-byte KV samples which was not realistic. Pull Request resolved: https://github.com/facebook/rocksdb/pull/4952 Differential Revision: D13967980 Pulled By: ajkr fbshipit-source-id: 82bea6f7537e1529c7a1a4cdee84585f5949300f
2019-02-12 04:42:25 +01:00
Reopen(options);
Random rnd(301);
for (int i = 0; i < kNumFiles; ++i) {
for (int j = 0; j < kNumEntriesPerFile; ++j) {
ASSERT_OK(Put(Key(i * kNumEntriesPerFile + j),
rnd.RandomString(kNumBytesPerEntry)));
Reduce scope of compression dictionary to single SST (#4952) Summary: Our previous approach was to train one compression dictionary per compaction, using the first output SST to train a dictionary, and then applying it on subsequent SSTs in the same compaction. While this was great for minimizing CPU/memory/I/O overhead, it did not achieve good compression ratios in practice. In our most promising potential use case, moderate reductions in a dictionary's scope make a major difference on compression ratio. So, this PR changes compression dictionary to be scoped per-SST. It accepts the tradeoff during table building to use more memory and CPU. Important changes include: - The `BlockBasedTableBuilder` has a new state when dictionary compression is in-use: `kBuffered`. In that state it accumulates uncompressed data in-memory whenever `Add` is called. - After accumulating target file size bytes or calling `BlockBasedTableBuilder::Finish`, a `BlockBasedTableBuilder` moves to the `kUnbuffered` state. The transition (`EnterUnbuffered()`) involves sampling the buffered data, training a dictionary, and compressing/writing out all buffered data. In the `kUnbuffered` state, a `BlockBasedTableBuilder` behaves the same as before -- blocks are compressed/written out as soon as they fill up. - Samples are now whole uncompressed data blocks, except the final sample may be a partial data block so we don't breach the user's configured `max_dict_bytes` or `zstd_max_train_bytes`. The dictionary trainer is supposed to work better when we pass it real units of compression. Previously we were passing 64-byte KV samples which was not realistic. Pull Request resolved: https://github.com/facebook/rocksdb/pull/4952 Differential Revision: D13967980 Pulled By: ajkr fbshipit-source-id: 82bea6f7537e1529c7a1a4cdee84585f5949300f
2019-02-12 04:42:25 +01:00
}
ASSERT_OK(Flush());
MoveFilesToLevel(1);
ASSERT_EQ(NumTableFilesAtLevel(1), i + 1);
}
// Store all the dictionaries generated during a full compaction.
std::vector<std::string> compression_dicts;
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
Reduce scope of compression dictionary to single SST (#4952) Summary: Our previous approach was to train one compression dictionary per compaction, using the first output SST to train a dictionary, and then applying it on subsequent SSTs in the same compaction. While this was great for minimizing CPU/memory/I/O overhead, it did not achieve good compression ratios in practice. In our most promising potential use case, moderate reductions in a dictionary's scope make a major difference on compression ratio. So, this PR changes compression dictionary to be scoped per-SST. It accepts the tradeoff during table building to use more memory and CPU. Important changes include: - The `BlockBasedTableBuilder` has a new state when dictionary compression is in-use: `kBuffered`. In that state it accumulates uncompressed data in-memory whenever `Add` is called. - After accumulating target file size bytes or calling `BlockBasedTableBuilder::Finish`, a `BlockBasedTableBuilder` moves to the `kUnbuffered` state. The transition (`EnterUnbuffered()`) involves sampling the buffered data, training a dictionary, and compressing/writing out all buffered data. In the `kUnbuffered` state, a `BlockBasedTableBuilder` behaves the same as before -- blocks are compressed/written out as soon as they fill up. - Samples are now whole uncompressed data blocks, except the final sample may be a partial data block so we don't breach the user's configured `max_dict_bytes` or `zstd_max_train_bytes`. The dictionary trainer is supposed to work better when we pass it real units of compression. Previously we were passing 64-byte KV samples which was not realistic. Pull Request resolved: https://github.com/facebook/rocksdb/pull/4952 Differential Revision: D13967980 Pulled By: ajkr fbshipit-source-id: 82bea6f7537e1529c7a1a4cdee84585f5949300f
2019-02-12 04:42:25 +01:00
"BlockBasedTableBuilder::WriteCompressionDictBlock:RawDict",
[&](void* arg) {
compression_dicts.emplace_back(static_cast<Slice*>(arg)->ToString());
});
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
Reduce scope of compression dictionary to single SST (#4952) Summary: Our previous approach was to train one compression dictionary per compaction, using the first output SST to train a dictionary, and then applying it on subsequent SSTs in the same compaction. While this was great for minimizing CPU/memory/I/O overhead, it did not achieve good compression ratios in practice. In our most promising potential use case, moderate reductions in a dictionary's scope make a major difference on compression ratio. So, this PR changes compression dictionary to be scoped per-SST. It accepts the tradeoff during table building to use more memory and CPU. Important changes include: - The `BlockBasedTableBuilder` has a new state when dictionary compression is in-use: `kBuffered`. In that state it accumulates uncompressed data in-memory whenever `Add` is called. - After accumulating target file size bytes or calling `BlockBasedTableBuilder::Finish`, a `BlockBasedTableBuilder` moves to the `kUnbuffered` state. The transition (`EnterUnbuffered()`) involves sampling the buffered data, training a dictionary, and compressing/writing out all buffered data. In the `kUnbuffered` state, a `BlockBasedTableBuilder` behaves the same as before -- blocks are compressed/written out as soon as they fill up. - Samples are now whole uncompressed data blocks, except the final sample may be a partial data block so we don't breach the user's configured `max_dict_bytes` or `zstd_max_train_bytes`. The dictionary trainer is supposed to work better when we pass it real units of compression. Previously we were passing 64-byte KV samples which was not realistic. Pull Request resolved: https://github.com/facebook/rocksdb/pull/4952 Differential Revision: D13967980 Pulled By: ajkr fbshipit-source-id: 82bea6f7537e1529c7a1a4cdee84585f5949300f
2019-02-12 04:42:25 +01:00
CompactRangeOptions compact_range_opts;
compact_range_opts.bottommost_level_compaction =
BottommostLevelCompaction::kForceOptimized;
Reduce scope of compression dictionary to single SST (#4952) Summary: Our previous approach was to train one compression dictionary per compaction, using the first output SST to train a dictionary, and then applying it on subsequent SSTs in the same compaction. While this was great for minimizing CPU/memory/I/O overhead, it did not achieve good compression ratios in practice. In our most promising potential use case, moderate reductions in a dictionary's scope make a major difference on compression ratio. So, this PR changes compression dictionary to be scoped per-SST. It accepts the tradeoff during table building to use more memory and CPU. Important changes include: - The `BlockBasedTableBuilder` has a new state when dictionary compression is in-use: `kBuffered`. In that state it accumulates uncompressed data in-memory whenever `Add` is called. - After accumulating target file size bytes or calling `BlockBasedTableBuilder::Finish`, a `BlockBasedTableBuilder` moves to the `kUnbuffered` state. The transition (`EnterUnbuffered()`) involves sampling the buffered data, training a dictionary, and compressing/writing out all buffered data. In the `kUnbuffered` state, a `BlockBasedTableBuilder` behaves the same as before -- blocks are compressed/written out as soon as they fill up. - Samples are now whole uncompressed data blocks, except the final sample may be a partial data block so we don't breach the user's configured `max_dict_bytes` or `zstd_max_train_bytes`. The dictionary trainer is supposed to work better when we pass it real units of compression. Previously we were passing 64-byte KV samples which was not realistic. Pull Request resolved: https://github.com/facebook/rocksdb/pull/4952 Differential Revision: D13967980 Pulled By: ajkr fbshipit-source-id: 82bea6f7537e1529c7a1a4cdee84585f5949300f
2019-02-12 04:42:25 +01:00
ASSERT_OK(db_->CompactRange(compact_range_opts, nullptr, nullptr));
// Dictionary compression should not be so good as to compress four totally
// random files into one. If it does then there's probably something wrong
// with the test.
ASSERT_GT(NumTableFilesAtLevel(1), 1);
// Furthermore, there should be one compression dictionary generated per file.
// And they should all be different from each other.
ASSERT_EQ(NumTableFilesAtLevel(1),
static_cast<int>(compression_dicts.size()));
for (size_t i = 1; i < compression_dicts.size(); ++i) {
std::string& a = compression_dicts[i - 1];
std::string& b = compression_dicts[i];
size_t alen = a.size();
size_t blen = b.size();
ASSERT_TRUE(alen != blen || memcmp(a.data(), b.data(), alen) != 0);
}
}
class PresetCompressionDictTest
: public DBTestBase,
public testing::WithParamInterface<std::tuple<CompressionType, bool>> {
public:
PresetCompressionDictTest()
: DBTestBase("db_test2", false /* env_do_fsync */),
compression_type_(std::get<0>(GetParam())),
bottommost_(std::get<1>(GetParam())) {}
protected:
const CompressionType compression_type_;
const bool bottommost_;
};
INSTANTIATE_TEST_CASE_P(
DBTest2, PresetCompressionDictTest,
::testing::Combine(::testing::ValuesIn(GetSupportedDictCompressions()),
::testing::Bool()));
TEST_P(PresetCompressionDictTest, Flush) {
// Verifies that dictionary is generated and written during flush only when
Limit buffering for collecting samples for compression dictionary (#7970) Summary: For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file. However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage. Related changes include: - Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks - Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary - Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970 Test Plan: - updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level - looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set. Reviewed By: pdillinger Differential Revision: D26467994 Pulled By: ajkr fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
2021-02-19 23:06:59 +01:00
// `ColumnFamilyOptions::compression` enables dictionary. Also verifies the
// size of the dictionary is within expectations according to the limit on
// buffering set by `CompressionOptions::max_dict_buffer_bytes`.
const size_t kValueLen = 256;
const size_t kKeysPerFile = 1 << 10;
Limit buffering for collecting samples for compression dictionary (#7970) Summary: For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file. However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage. Related changes include: - Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks - Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary - Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970 Test Plan: - updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level - looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set. Reviewed By: pdillinger Differential Revision: D26467994 Pulled By: ajkr fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
2021-02-19 23:06:59 +01:00
const size_t kDictLen = 16 << 10;
const size_t kBlockLen = 4 << 10;
Options options = CurrentOptions();
if (bottommost_) {
options.bottommost_compression = compression_type_;
options.bottommost_compression_opts.enabled = true;
options.bottommost_compression_opts.max_dict_bytes = kDictLen;
Limit buffering for collecting samples for compression dictionary (#7970) Summary: For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file. However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage. Related changes include: - Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks - Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary - Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970 Test Plan: - updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level - looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set. Reviewed By: pdillinger Differential Revision: D26467994 Pulled By: ajkr fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
2021-02-19 23:06:59 +01:00
options.bottommost_compression_opts.max_dict_buffer_bytes = kBlockLen;
} else {
options.compression = compression_type_;
options.compression_opts.max_dict_bytes = kDictLen;
Limit buffering for collecting samples for compression dictionary (#7970) Summary: For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file. However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage. Related changes include: - Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks - Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary - Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970 Test Plan: - updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level - looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set. Reviewed By: pdillinger Differential Revision: D26467994 Pulled By: ajkr fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
2021-02-19 23:06:59 +01:00
options.compression_opts.max_dict_buffer_bytes = kBlockLen;
}
options.memtable_factory.reset(new SpecialSkipListFactory(kKeysPerFile));
options.statistics = CreateDBStatistics();
BlockBasedTableOptions bbto;
Limit buffering for collecting samples for compression dictionary (#7970) Summary: For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file. However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage. Related changes include: - Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks - Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary - Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970 Test Plan: - updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level - looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set. Reviewed By: pdillinger Differential Revision: D26467994 Pulled By: ajkr fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
2021-02-19 23:06:59 +01:00
bbto.block_size = kBlockLen;
bbto.cache_index_and_filter_blocks = true;
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
Reopen(options);
Random rnd(301);
for (size_t i = 0; i <= kKeysPerFile; ++i) {
ASSERT_OK(Put(Key(static_cast<int>(i)), rnd.RandomString(kValueLen)));
}
ASSERT_OK(dbfull()->TEST_WaitForFlushMemTable());
Limit buffering for collecting samples for compression dictionary (#7970) Summary: For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file. However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage. Related changes include: - Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks - Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary - Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970 Test Plan: - updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level - looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set. Reviewed By: pdillinger Differential Revision: D26467994 Pulled By: ajkr fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
2021-02-19 23:06:59 +01:00
// We can use `BLOCK_CACHE_COMPRESSION_DICT_BYTES_INSERT` to detect whether a
// compression dictionary exists since dictionaries would be preloaded when
// the flush finishes.
if (bottommost_) {
Limit buffering for collecting samples for compression dictionary (#7970) Summary: For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file. However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage. Related changes include: - Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks - Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary - Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970 Test Plan: - updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level - looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set. Reviewed By: pdillinger Differential Revision: D26467994 Pulled By: ajkr fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
2021-02-19 23:06:59 +01:00
// Flush is never considered bottommost. This should change in the future
// since flushed files may have nothing underneath them, like the one in
// this test case.
ASSERT_EQ(
TestGetTickerCount(options, BLOCK_CACHE_COMPRESSION_DICT_BYTES_INSERT),
0);
} else {
Limit buffering for collecting samples for compression dictionary (#7970) Summary: For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file. However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage. Related changes include: - Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks - Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary - Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970 Test Plan: - updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level - looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set. Reviewed By: pdillinger Differential Revision: D26467994 Pulled By: ajkr fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
2021-02-19 23:06:59 +01:00
ASSERT_GT(
TestGetTickerCount(options, BLOCK_CACHE_COMPRESSION_DICT_BYTES_INSERT),
0);
// TODO(ajkr): fix the below assertion to work with ZSTD. The expectation on
// number of bytes needs to be adjusted in case the cached block is in
// ZSTD's digested dictionary format.
if (compression_type_ != kZSTD &&
compression_type_ != kZSTDNotFinalCompression) {
// Although we limited buffering to `kBlockLen`, there may be up to two
// blocks of data included in the dictionary since we only check limit
// after each block is built.
ASSERT_LE(TestGetTickerCount(options,
BLOCK_CACHE_COMPRESSION_DICT_BYTES_INSERT),
2 * kBlockLen);
}
}
}
TEST_P(PresetCompressionDictTest, CompactNonBottommost) {
// Verifies that dictionary is generated and written during compaction to
// non-bottommost level only when `ColumnFamilyOptions::compression` enables
Limit buffering for collecting samples for compression dictionary (#7970) Summary: For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file. However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage. Related changes include: - Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks - Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary - Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970 Test Plan: - updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level - looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set. Reviewed By: pdillinger Differential Revision: D26467994 Pulled By: ajkr fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
2021-02-19 23:06:59 +01:00
// dictionary. Also verifies the size of the dictionary is within expectations
// according to the limit on buffering set by
// `CompressionOptions::max_dict_buffer_bytes`.
const size_t kValueLen = 256;
const size_t kKeysPerFile = 1 << 10;
Limit buffering for collecting samples for compression dictionary (#7970) Summary: For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file. However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage. Related changes include: - Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks - Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary - Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970 Test Plan: - updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level - looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set. Reviewed By: pdillinger Differential Revision: D26467994 Pulled By: ajkr fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
2021-02-19 23:06:59 +01:00
const size_t kDictLen = 16 << 10;
const size_t kBlockLen = 4 << 10;
Options options = CurrentOptions();
if (bottommost_) {
options.bottommost_compression = compression_type_;
options.bottommost_compression_opts.enabled = true;
options.bottommost_compression_opts.max_dict_bytes = kDictLen;
Limit buffering for collecting samples for compression dictionary (#7970) Summary: For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file. However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage. Related changes include: - Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks - Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary - Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970 Test Plan: - updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level - looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set. Reviewed By: pdillinger Differential Revision: D26467994 Pulled By: ajkr fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
2021-02-19 23:06:59 +01:00
options.bottommost_compression_opts.max_dict_buffer_bytes = kBlockLen;
} else {
options.compression = compression_type_;
options.compression_opts.max_dict_bytes = kDictLen;
Limit buffering for collecting samples for compression dictionary (#7970) Summary: For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file. However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage. Related changes include: - Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks - Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary - Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970 Test Plan: - updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level - looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set. Reviewed By: pdillinger Differential Revision: D26467994 Pulled By: ajkr fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
2021-02-19 23:06:59 +01:00
options.compression_opts.max_dict_buffer_bytes = kBlockLen;
}
options.disable_auto_compactions = true;
options.statistics = CreateDBStatistics();
BlockBasedTableOptions bbto;
Limit buffering for collecting samples for compression dictionary (#7970) Summary: For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file. However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage. Related changes include: - Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks - Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary - Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970 Test Plan: - updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level - looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set. Reviewed By: pdillinger Differential Revision: D26467994 Pulled By: ajkr fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
2021-02-19 23:06:59 +01:00
bbto.block_size = kBlockLen;
bbto.cache_index_and_filter_blocks = true;
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
Reopen(options);
Random rnd(301);
for (size_t j = 0; j <= kKeysPerFile; ++j) {
ASSERT_OK(Put(Key(static_cast<int>(j)), rnd.RandomString(kValueLen)));
}
ASSERT_OK(Flush());
MoveFilesToLevel(2);
for (int i = 0; i < 2; ++i) {
for (size_t j = 0; j <= kKeysPerFile; ++j) {
ASSERT_OK(Put(Key(static_cast<int>(j)), rnd.RandomString(kValueLen)));
}
ASSERT_OK(Flush());
}
#ifndef ROCKSDB_LITE
ASSERT_EQ("2,0,1", FilesPerLevel(0));
#endif // ROCKSDB_LITE
Limit buffering for collecting samples for compression dictionary (#7970) Summary: For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file. However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage. Related changes include: - Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks - Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary - Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970 Test Plan: - updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level - looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set. Reviewed By: pdillinger Differential Revision: D26467994 Pulled By: ajkr fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
2021-02-19 23:06:59 +01:00
uint64_t prev_compression_dict_bytes_inserted =
TestGetTickerCount(options, BLOCK_CACHE_COMPRESSION_DICT_BYTES_INSERT);
// This L0->L1 compaction merges the two L0 files into L1. The produced L1
// file is not bottommost due to the existing L2 file covering the same key-
// range.
ASSERT_OK(dbfull()->TEST_CompactRange(0, nullptr, nullptr));
#ifndef ROCKSDB_LITE
ASSERT_EQ("0,1,1", FilesPerLevel(0));
#endif // ROCKSDB_LITE
Limit buffering for collecting samples for compression dictionary (#7970) Summary: For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file. However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage. Related changes include: - Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks - Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary - Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970 Test Plan: - updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level - looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set. Reviewed By: pdillinger Differential Revision: D26467994 Pulled By: ajkr fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
2021-02-19 23:06:59 +01:00
// We can use `BLOCK_CACHE_COMPRESSION_DICT_BYTES_INSERT` to detect whether a
// compression dictionary exists since dictionaries would be preloaded when
// the compaction finishes.
if (bottommost_) {
Limit buffering for collecting samples for compression dictionary (#7970) Summary: For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file. However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage. Related changes include: - Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks - Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary - Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970 Test Plan: - updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level - looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set. Reviewed By: pdillinger Differential Revision: D26467994 Pulled By: ajkr fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
2021-02-19 23:06:59 +01:00
ASSERT_EQ(
TestGetTickerCount(options, BLOCK_CACHE_COMPRESSION_DICT_BYTES_INSERT),
prev_compression_dict_bytes_inserted);
} else {
Limit buffering for collecting samples for compression dictionary (#7970) Summary: For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file. However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage. Related changes include: - Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks - Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary - Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970 Test Plan: - updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level - looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set. Reviewed By: pdillinger Differential Revision: D26467994 Pulled By: ajkr fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
2021-02-19 23:06:59 +01:00
ASSERT_GT(
TestGetTickerCount(options, BLOCK_CACHE_COMPRESSION_DICT_BYTES_INSERT),
prev_compression_dict_bytes_inserted);
// TODO(ajkr): fix the below assertion to work with ZSTD. The expectation on
// number of bytes needs to be adjusted in case the cached block is in
// ZSTD's digested dictionary format.
if (compression_type_ != kZSTD &&
compression_type_ != kZSTDNotFinalCompression) {
// Although we limited buffering to `kBlockLen`, there may be up to two
// blocks of data included in the dictionary since we only check limit
// after each block is built.
ASSERT_LE(TestGetTickerCount(options,
BLOCK_CACHE_COMPRESSION_DICT_BYTES_INSERT),
prev_compression_dict_bytes_inserted + 2 * kBlockLen);
}
}
}
TEST_P(PresetCompressionDictTest, CompactBottommost) {
// Verifies that dictionary is generated and written during compaction to
// non-bottommost level only when either `ColumnFamilyOptions::compression` or
Limit buffering for collecting samples for compression dictionary (#7970) Summary: For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file. However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage. Related changes include: - Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks - Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary - Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970 Test Plan: - updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level - looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set. Reviewed By: pdillinger Differential Revision: D26467994 Pulled By: ajkr fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
2021-02-19 23:06:59 +01:00
// `ColumnFamilyOptions::bottommost_compression` enables dictionary. Also
// verifies the size of the dictionary is within expectations according to the
// limit on buffering set by `CompressionOptions::max_dict_buffer_bytes`.
const size_t kValueLen = 256;
const size_t kKeysPerFile = 1 << 10;
Limit buffering for collecting samples for compression dictionary (#7970) Summary: For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file. However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage. Related changes include: - Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks - Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary - Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970 Test Plan: - updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level - looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set. Reviewed By: pdillinger Differential Revision: D26467994 Pulled By: ajkr fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
2021-02-19 23:06:59 +01:00
const size_t kDictLen = 16 << 10;
const size_t kBlockLen = 4 << 10;
Options options = CurrentOptions();
if (bottommost_) {
options.bottommost_compression = compression_type_;
options.bottommost_compression_opts.enabled = true;
options.bottommost_compression_opts.max_dict_bytes = kDictLen;
Limit buffering for collecting samples for compression dictionary (#7970) Summary: For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file. However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage. Related changes include: - Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks - Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary - Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970 Test Plan: - updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level - looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set. Reviewed By: pdillinger Differential Revision: D26467994 Pulled By: ajkr fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
2021-02-19 23:06:59 +01:00
options.bottommost_compression_opts.max_dict_buffer_bytes = kBlockLen;
} else {
options.compression = compression_type_;
options.compression_opts.max_dict_bytes = kDictLen;
Limit buffering for collecting samples for compression dictionary (#7970) Summary: For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file. However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage. Related changes include: - Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks - Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary - Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970 Test Plan: - updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level - looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set. Reviewed By: pdillinger Differential Revision: D26467994 Pulled By: ajkr fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
2021-02-19 23:06:59 +01:00
options.compression_opts.max_dict_buffer_bytes = kBlockLen;
}
options.disable_auto_compactions = true;
options.statistics = CreateDBStatistics();
BlockBasedTableOptions bbto;
Limit buffering for collecting samples for compression dictionary (#7970) Summary: For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file. However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage. Related changes include: - Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks - Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary - Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970 Test Plan: - updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level - looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set. Reviewed By: pdillinger Differential Revision: D26467994 Pulled By: ajkr fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
2021-02-19 23:06:59 +01:00
bbto.block_size = kBlockLen;
bbto.cache_index_and_filter_blocks = true;
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
Reopen(options);
Random rnd(301);
for (int i = 0; i < 2; ++i) {
for (size_t j = 0; j <= kKeysPerFile; ++j) {
ASSERT_OK(Put(Key(static_cast<int>(j)), rnd.RandomString(kValueLen)));
}
ASSERT_OK(Flush());
}
#ifndef ROCKSDB_LITE
ASSERT_EQ("2", FilesPerLevel(0));
#endif // ROCKSDB_LITE
Limit buffering for collecting samples for compression dictionary (#7970) Summary: For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file. However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage. Related changes include: - Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks - Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary - Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970 Test Plan: - updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level - looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set. Reviewed By: pdillinger Differential Revision: D26467994 Pulled By: ajkr fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
2021-02-19 23:06:59 +01:00
uint64_t prev_compression_dict_bytes_inserted =
TestGetTickerCount(options, BLOCK_CACHE_COMPRESSION_DICT_BYTES_INSERT);
CompactRangeOptions cro;
ASSERT_OK(db_->CompactRange(cro, nullptr, nullptr));
#ifndef ROCKSDB_LITE
ASSERT_EQ("0,1", FilesPerLevel(0));
#endif // ROCKSDB_LITE
Limit buffering for collecting samples for compression dictionary (#7970) Summary: For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file. However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage. Related changes include: - Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks - Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary - Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970 Test Plan: - updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level - looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set. Reviewed By: pdillinger Differential Revision: D26467994 Pulled By: ajkr fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
2021-02-19 23:06:59 +01:00
ASSERT_GT(
TestGetTickerCount(options, BLOCK_CACHE_COMPRESSION_DICT_BYTES_INSERT),
prev_compression_dict_bytes_inserted);
// TODO(ajkr): fix the below assertion to work with ZSTD. The expectation on
// number of bytes needs to be adjusted in case the cached block is in ZSTD's
// digested dictionary format.
if (compression_type_ != kZSTD &&
compression_type_ != kZSTDNotFinalCompression) {
// Although we limited buffering to `kBlockLen`, there may be up to two
// blocks of data included in the dictionary since we only check limit after
// each block is built.
ASSERT_LE(
TestGetTickerCount(options, BLOCK_CACHE_COMPRESSION_DICT_BYTES_INSERT),
prev_compression_dict_bytes_inserted + 2 * kBlockLen);
}
}
class CompactionCompressionListener : public EventListener {
public:
explicit CompactionCompressionListener(Options* db_options)
: db_options_(db_options) {}
void OnCompactionCompleted(DB* db, const CompactionJobInfo& ci) override {
// Figure out last level with files
int bottommost_level = 0;
for (int level = 0; level < db->NumberLevels(); level++) {
std::string files_at_level;
ASSERT_TRUE(
db->GetProperty("rocksdb.num-files-at-level" + NumberToString(level),
&files_at_level));
if (files_at_level != "0") {
bottommost_level = level;
}
}
if (db_options_->bottommost_compression != kDisableCompressionOption &&
ci.output_level == bottommost_level) {
ASSERT_EQ(ci.compression, db_options_->bottommost_compression);
} else if (db_options_->compression_per_level.size() != 0) {
ASSERT_EQ(ci.compression,
db_options_->compression_per_level[ci.output_level]);
} else {
ASSERT_EQ(ci.compression, db_options_->compression);
}
max_level_checked = std::max(max_level_checked, ci.output_level);
}
int max_level_checked = 0;
const Options* db_options_;
};
enum CompressionFailureType {
kTestCompressionFail,
kTestDecompressionFail,
kTestDecompressionCorruption
};
class CompressionFailuresTest
: public DBTest2,
public testing::WithParamInterface<std::tuple<
CompressionFailureType, CompressionType, uint32_t, uint32_t>> {
public:
CompressionFailuresTest() {
std::tie(compression_failure_type_, compression_type_,
compression_max_dict_bytes_, compression_parallel_threads_) =
GetParam();
}
CompressionFailureType compression_failure_type_ = kTestCompressionFail;
CompressionType compression_type_ = kNoCompression;
uint32_t compression_max_dict_bytes_ = 0;
uint32_t compression_parallel_threads_ = 0;
};
INSTANTIATE_TEST_CASE_P(
DBTest2, CompressionFailuresTest,
::testing::Combine(::testing::Values(kTestCompressionFail,
kTestDecompressionFail,
kTestDecompressionCorruption),
::testing::ValuesIn(GetSupportedCompressions()),
::testing::Values(0, 10), ::testing::Values(1, 4)));
TEST_P(CompressionFailuresTest, CompressionFailures) {
if (compression_type_ == kNoCompression) {
return;
}
Options options = CurrentOptions();
options.level0_file_num_compaction_trigger = 2;
options.max_bytes_for_level_base = 1024;
options.max_bytes_for_level_multiplier = 2;
options.num_levels = 7;
options.max_background_compactions = 1;
options.target_file_size_base = 512;
BlockBasedTableOptions table_options;
table_options.block_size = 512;
table_options.verify_compression = true;
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566) Summary: This PR does a few things: 1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation. 2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed: - The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated - The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory). 3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10). I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged. Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently. Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566 Reviewed By: zhichao-cao Differential Revision: D24408980 Pulled By: jay-zhuang fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
2020-10-27 18:31:34 +01:00
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
options.compression = compression_type_;
options.compression_opts.parallel_threads = compression_parallel_threads_;
options.compression_opts.max_dict_bytes = compression_max_dict_bytes_;
options.bottommost_compression_opts.parallel_threads =
compression_parallel_threads_;
options.bottommost_compression_opts.max_dict_bytes =
compression_max_dict_bytes_;
if (compression_failure_type_ == kTestCompressionFail) {
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"CompressData:TamperWithReturnValue", [](void* arg) {
bool* ret = static_cast<bool*>(arg);
*ret = false;
});
} else if (compression_failure_type_ == kTestDecompressionFail) {
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"UncompressBlockContentsForCompressionType:TamperWithReturnValue",
[](void* arg) {
Status* ret = static_cast<Status*>(arg);
ASSERT_OK(*ret);
*ret = Status::Corruption("kTestDecompressionFail");
});
} else if (compression_failure_type_ == kTestDecompressionCorruption) {
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"UncompressBlockContentsForCompressionType:"
"TamperWithDecompressionOutput",
[](void* arg) {
BlockContents* contents = static_cast<BlockContents*>(arg);
// Ensure uncompressed data != original data
const size_t len = contents->data.size() + 1;
std::unique_ptr<char[]> fake_data(new char[len]());
*contents = BlockContents(std::move(fake_data), len);
});
}
std::map<std::string, std::string> key_value_written;
const int kKeySize = 5;
const int kValUnitSize = 16;
const int kValSize = 256;
Random rnd(405);
Status s = Status::OK();
DestroyAndReopen(options);
// Write 10 random files
for (int i = 0; i < 10; i++) {
for (int j = 0; j < 5; j++) {
std::string key = rnd.RandomString(kKeySize);
// Ensure good compression ratio
std::string valueUnit = rnd.RandomString(kValUnitSize);
std::string value;
for (int k = 0; k < kValSize; k += kValUnitSize) {
value += valueUnit;
}
s = Put(key, value);
if (compression_failure_type_ == kTestCompressionFail) {
key_value_written[key] = value;
ASSERT_OK(s);
}
}
s = Flush();
if (compression_failure_type_ == kTestCompressionFail) {
ASSERT_OK(s);
}
s = dbfull()->TEST_WaitForCompact();
if (compression_failure_type_ == kTestCompressionFail) {
ASSERT_OK(s);
}
if (i == 4) {
// Make compression fail at the mid of table building
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
}
}
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
if (compression_failure_type_ == kTestCompressionFail) {
// Should be kNoCompression, check content consistency
std::unique_ptr<Iterator> db_iter(db_->NewIterator(ReadOptions()));
for (db_iter->SeekToFirst(); db_iter->Valid(); db_iter->Next()) {
std::string key = db_iter->key().ToString();
std::string value = db_iter->value().ToString();
ASSERT_NE(key_value_written.find(key), key_value_written.end());
ASSERT_EQ(key_value_written[key], value);
key_value_written.erase(key);
}
ASSERT_EQ(0, key_value_written.size());
} else if (compression_failure_type_ == kTestDecompressionFail) {
ASSERT_EQ(std::string(s.getState()),
"Could not decompress: kTestDecompressionFail");
} else if (compression_failure_type_ == kTestDecompressionCorruption) {
ASSERT_EQ(std::string(s.getState()),
"Decompressed block did not match raw block");
}
}
TEST_F(DBTest2, CompressionOptions) {
if (!Zlib_Supported() || !Snappy_Supported()) {
return;
}
Options options = CurrentOptions();
options.level0_file_num_compaction_trigger = 2;
options.max_bytes_for_level_base = 100;
options.max_bytes_for_level_multiplier = 2;
options.num_levels = 7;
options.max_background_compactions = 1;
CompactionCompressionListener* listener =
new CompactionCompressionListener(&options);
options.listeners.emplace_back(listener);
const int kKeySize = 5;
const int kValSize = 20;
Random rnd(301);
std::vector<uint32_t> compression_parallel_threads = {1, 4};
std::map<std::string, std::string> key_value_written;
for (int iter = 0; iter <= 2; iter++) {
listener->max_level_checked = 0;
if (iter == 0) {
// Use different compression algorithms for different levels but
// always use Zlib for bottommost level
options.compression_per_level = {kNoCompression, kNoCompression,
kNoCompression, kSnappyCompression,
kSnappyCompression, kSnappyCompression,
kZlibCompression};
options.compression = kNoCompression;
options.bottommost_compression = kZlibCompression;
} else if (iter == 1) {
// Use Snappy except for bottommost level use ZLib
options.compression_per_level = {};
options.compression = kSnappyCompression;
options.bottommost_compression = kZlibCompression;
} else if (iter == 2) {
// Use Snappy everywhere
options.compression_per_level = {};
options.compression = kSnappyCompression;
options.bottommost_compression = kDisableCompressionOption;
}
for (auto num_threads : compression_parallel_threads) {
options.compression_opts.parallel_threads = num_threads;
options.bottommost_compression_opts.parallel_threads = num_threads;
DestroyAndReopen(options);
// Write 10 random files
for (int i = 0; i < 10; i++) {
for (int j = 0; j < 5; j++) {
std::string key = rnd.RandomString(kKeySize);
std::string value = rnd.RandomString(kValSize);
key_value_written[key] = value;
ASSERT_OK(Put(key, value));
}
ASSERT_OK(Flush());
dbfull()->TEST_WaitForCompact();
}
// Make sure that we wrote enough to check all 7 levels
ASSERT_EQ(listener->max_level_checked, 6);
// Make sure database content is the same as key_value_written
std::unique_ptr<Iterator> db_iter(db_->NewIterator(ReadOptions()));
for (db_iter->SeekToFirst(); db_iter->Valid(); db_iter->Next()) {
std::string key = db_iter->key().ToString();
std::string value = db_iter->value().ToString();
ASSERT_NE(key_value_written.find(key), key_value_written.end());
ASSERT_EQ(key_value_written[key], value);
key_value_written.erase(key);
}
ASSERT_EQ(0, key_value_written.size());
}
}
}
class CompactionStallTestListener : public EventListener {
public:
CompactionStallTestListener() : compacting_files_cnt_(0), compacted_files_cnt_(0) {}
void OnCompactionBegin(DB* /*db*/, const CompactionJobInfo& ci) override {
ASSERT_EQ(ci.cf_name, "default");
ASSERT_EQ(ci.base_input_level, 0);
ASSERT_EQ(ci.compaction_reason, CompactionReason::kLevelL0FilesNum);
compacting_files_cnt_ += ci.input_files.size();
}
void OnCompactionCompleted(DB* /*db*/, const CompactionJobInfo& ci) override {
ASSERT_EQ(ci.cf_name, "default");
ASSERT_EQ(ci.base_input_level, 0);
ASSERT_EQ(ci.compaction_reason, CompactionReason::kLevelL0FilesNum);
compacted_files_cnt_ += ci.input_files.size();
}
std::atomic<size_t> compacting_files_cnt_;
std::atomic<size_t> compacted_files_cnt_;
};
TEST_F(DBTest2, CompactionStall) {
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
{{"DBImpl::BGWorkCompaction", "DBTest2::CompactionStall:0"},
{"DBImpl::BGWorkCompaction", "DBTest2::CompactionStall:1"},
{"DBTest2::CompactionStall:2",
"DBImpl::NotifyOnCompactionBegin::UnlockMutex"},
{"DBTest2::CompactionStall:3",
"DBImpl::NotifyOnCompactionCompleted::UnlockMutex"}});
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
Options options = CurrentOptions();
options.level0_file_num_compaction_trigger = 4;
options.max_background_compactions = 40;
CompactionStallTestListener* listener = new CompactionStallTestListener();
options.listeners.emplace_back(listener);
DestroyAndReopen(options);
// make sure all background compaction jobs can be scheduled
auto stop_token =
dbfull()->TEST_write_controler().GetCompactionPressureToken();
Random rnd(301);
// 4 Files in L0
for (int i = 0; i < 4; i++) {
for (int j = 0; j < 10; j++) {
ASSERT_OK(Put(rnd.RandomString(10), rnd.RandomString(10)));
}
ASSERT_OK(Flush());
}
// Wait for compaction to be triggered
TEST_SYNC_POINT("DBTest2::CompactionStall:0");
// Clear "DBImpl::BGWorkCompaction" SYNC_POINT since we want to hold it again
// at DBTest2::CompactionStall::1
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearTrace();
// Another 6 L0 files to trigger compaction again
for (int i = 0; i < 6; i++) {
for (int j = 0; j < 10; j++) {
ASSERT_OK(Put(rnd.RandomString(10), rnd.RandomString(10)));
}
ASSERT_OK(Flush());
}
// Wait for another compaction to be triggered
TEST_SYNC_POINT("DBTest2::CompactionStall:1");
// Hold NotifyOnCompactionBegin in the unlock mutex section
TEST_SYNC_POINT("DBTest2::CompactionStall:2");
// Hold NotifyOnCompactionCompleted in the unlock mutex section
TEST_SYNC_POINT("DBTest2::CompactionStall:3");
dbfull()->TEST_WaitForCompact();
ASSERT_LT(NumTableFilesAtLevel(0),
options.level0_file_num_compaction_trigger);
ASSERT_GT(listener->compacted_files_cnt_.load(),
10 - options.level0_file_num_compaction_trigger);
ASSERT_EQ(listener->compacting_files_cnt_.load(), listener->compacted_files_cnt_.load());
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
}
#endif // ROCKSDB_LITE
TEST_F(DBTest2, FirstSnapshotTest) {
Options options;
options.write_buffer_size = 100000; // Small write buffer
options = CurrentOptions(options);
CreateAndReopenWithCF({"pikachu"}, options);
// This snapshot will have sequence number 0 what is expected behaviour.
const Snapshot* s1 = db_->GetSnapshot();
Put(1, "k1", std::string(100000, 'x')); // Fill memtable
Put(1, "k2", std::string(100000, 'y')); // Trigger flush
db_->ReleaseSnapshot(s1);
}
#ifndef ROCKSDB_LITE
TEST_F(DBTest2, DuplicateSnapshot) {
Options options;
options = CurrentOptions(options);
std::vector<const Snapshot*> snapshots;
DBImpl* dbi = static_cast_with_check<DBImpl>(db_);
SequenceNumber oldest_ww_snap, first_ww_snap;
Put("k", "v"); // inc seq
snapshots.push_back(db_->GetSnapshot());
snapshots.push_back(db_->GetSnapshot());
Put("k", "v"); // inc seq
snapshots.push_back(db_->GetSnapshot());
snapshots.push_back(dbi->GetSnapshotForWriteConflictBoundary());
first_ww_snap = snapshots.back()->GetSequenceNumber();
Put("k", "v"); // inc seq
snapshots.push_back(dbi->GetSnapshotForWriteConflictBoundary());
snapshots.push_back(db_->GetSnapshot());
Put("k", "v"); // inc seq
snapshots.push_back(db_->GetSnapshot());
{
InstrumentedMutexLock l(dbi->mutex());
auto seqs = dbi->snapshots().GetAll(&oldest_ww_snap);
ASSERT_EQ(seqs.size(), 4); // duplicates are not counted
ASSERT_EQ(oldest_ww_snap, first_ww_snap);
}
for (auto s : snapshots) {
db_->ReleaseSnapshot(s);
}
}
#endif // ROCKSDB_LITE
class PinL0IndexAndFilterBlocksTest
: public DBTestBase,
public testing::WithParamInterface<std::tuple<bool, bool>> {
public:
PinL0IndexAndFilterBlocksTest()
: DBTestBase("db_pin_l0_index_bloom_test", /*env_do_fsync=*/true) {}
void SetUp() override {
infinite_max_files_ = std::get<0>(GetParam());
disallow_preload_ = std::get<1>(GetParam());
}
void CreateTwoLevels(Options* options, bool close_afterwards) {
if (infinite_max_files_) {
options->max_open_files = -1;
}
options->create_if_missing = true;
options->statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
BlockBasedTableOptions table_options;
table_options.cache_index_and_filter_blocks = true;
table_options.pin_l0_filter_and_index_blocks_in_cache = true;
table_options.filter_policy.reset(NewBloomFilterPolicy(20));
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566) Summary: This PR does a few things: 1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation. 2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed: - The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated - The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory). 3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10). I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged. Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently. Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566 Reviewed By: zhichao-cao Differential Revision: D24408980 Pulled By: jay-zhuang fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
2020-10-27 18:31:34 +01:00
options->table_factory.reset(NewBlockBasedTableFactory(table_options));
CreateAndReopenWithCF({"pikachu"}, *options);
Put(1, "a", "begin");
Put(1, "z", "end");
ASSERT_OK(Flush(1));
// move this table to L1
dbfull()->TEST_CompactRange(0, nullptr, nullptr, handles_[1]);
// reset block cache
table_options.block_cache = NewLRUCache(64 * 1024);
options->table_factory.reset(NewBlockBasedTableFactory(table_options));
TryReopenWithColumnFamilies({"default", "pikachu"}, *options);
// create new table at L0
Put(1, "a2", "begin2");
Put(1, "z2", "end2");
ASSERT_OK(Flush(1));
if (close_afterwards) {
Close(); // This ensures that there is no ref to block cache entries
}
table_options.block_cache->EraseUnRefEntries();
}
bool infinite_max_files_;
bool disallow_preload_;
};
TEST_P(PinL0IndexAndFilterBlocksTest,
IndexAndFilterBlocksOfNewTableAddedToCacheWithPinning) {
Options options = CurrentOptions();
if (infinite_max_files_) {
options.max_open_files = -1;
}
options.create_if_missing = true;
options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
BlockBasedTableOptions table_options;
table_options.cache_index_and_filter_blocks = true;
table_options.pin_l0_filter_and_index_blocks_in_cache = true;
table_options.filter_policy.reset(NewBloomFilterPolicy(20));
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566) Summary: This PR does a few things: 1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation. 2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed: - The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated - The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory). 3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10). I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged. Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently. Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566 Reviewed By: zhichao-cao Differential Revision: D24408980 Pulled By: jay-zhuang fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
2020-10-27 18:31:34 +01:00
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
CreateAndReopenWithCF({"pikachu"}, options);
ASSERT_OK(Put(1, "key", "val"));
// Create a new table.
ASSERT_OK(Flush(1));
// index/filter blocks added to block cache right after table creation.
ASSERT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_FILTER_MISS));
ASSERT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_FILTER_HIT));
ASSERT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_INDEX_MISS));
ASSERT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_INDEX_HIT));
// only index/filter were added
ASSERT_EQ(2, TestGetTickerCount(options, BLOCK_CACHE_ADD));
ASSERT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_DATA_MISS));
std::string value;
// Miss and hit count should remain the same, they're all pinned.
db_->KeyMayExist(ReadOptions(), handles_[1], "key", &value);
ASSERT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_FILTER_MISS));
ASSERT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_FILTER_HIT));
ASSERT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_INDEX_MISS));
ASSERT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_INDEX_HIT));
// Miss and hit count should remain the same, they're all pinned.
value = Get(1, "key");
ASSERT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_FILTER_MISS));
ASSERT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_FILTER_HIT));
ASSERT_EQ(1, TestGetTickerCount(options, BLOCK_CACHE_INDEX_MISS));
ASSERT_EQ(0, TestGetTickerCount(options, BLOCK_CACHE_INDEX_HIT));
}
TEST_P(PinL0IndexAndFilterBlocksTest,
MultiLevelIndexAndFilterBlocksCachedWithPinning) {
Options options = CurrentOptions();
PinL0IndexAndFilterBlocksTest::CreateTwoLevels(&options, false);
// get base cache values
uint64_t fm = TestGetTickerCount(options, BLOCK_CACHE_FILTER_MISS);
uint64_t fh = TestGetTickerCount(options, BLOCK_CACHE_FILTER_HIT);
uint64_t im = TestGetTickerCount(options, BLOCK_CACHE_INDEX_MISS);
uint64_t ih = TestGetTickerCount(options, BLOCK_CACHE_INDEX_HIT);
std::string value;
// this should be read from L0
// so cache values don't change
value = Get(1, "a2");
ASSERT_EQ(fm, TestGetTickerCount(options, BLOCK_CACHE_FILTER_MISS));
ASSERT_EQ(fh, TestGetTickerCount(options, BLOCK_CACHE_FILTER_HIT));
ASSERT_EQ(im, TestGetTickerCount(options, BLOCK_CACHE_INDEX_MISS));
ASSERT_EQ(ih, TestGetTickerCount(options, BLOCK_CACHE_INDEX_HIT));
// this should be read from L1
// the file is opened, prefetching results in a cache filter miss
// the block is loaded and added to the cache,
// then the get results in a cache hit for L1
// When we have inifinite max_files, there is still cache miss because we have
// reset the block cache
value = Get(1, "a");
ASSERT_EQ(fm + 1, TestGetTickerCount(options, BLOCK_CACHE_FILTER_MISS));
ASSERT_EQ(im + 1, TestGetTickerCount(options, BLOCK_CACHE_INDEX_MISS));
}
TEST_P(PinL0IndexAndFilterBlocksTest, DisablePrefetchingNonL0IndexAndFilter) {
Options options = CurrentOptions();
// This ensures that db does not ref anything in the block cache, so
// EraseUnRefEntries could clear them up.
bool close_afterwards = true;
PinL0IndexAndFilterBlocksTest::CreateTwoLevels(&options, close_afterwards);
// Get base cache values
uint64_t fm = TestGetTickerCount(options, BLOCK_CACHE_FILTER_MISS);
uint64_t fh = TestGetTickerCount(options, BLOCK_CACHE_FILTER_HIT);
uint64_t im = TestGetTickerCount(options, BLOCK_CACHE_INDEX_MISS);
uint64_t ih = TestGetTickerCount(options, BLOCK_CACHE_INDEX_HIT);
if (disallow_preload_) {
// Now we have two files. We narrow the max open files to allow 3 entries
// so that preloading SST files won't happen.
options.max_open_files = 13;
// RocksDB sanitize max open files to at least 20. Modify it back.
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"SanitizeOptions::AfterChangeMaxOpenFiles", [&](void* arg) {
int* max_open_files = static_cast<int*>(arg);
*max_open_files = 13;
});
}
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
// Reopen database. If max_open_files is set as -1, table readers will be
// preloaded. This will trigger a BlockBasedTable::Open() and prefetch
// L0 index and filter. Level 1's prefetching is disabled in DB::Open()
TryReopenWithColumnFamilies({"default", "pikachu"}, options);
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
if (!disallow_preload_) {
// After reopen, cache miss are increased by one because we read (and only
// read) filter and index on L0
ASSERT_EQ(fm + 1, TestGetTickerCount(options, BLOCK_CACHE_FILTER_MISS));
ASSERT_EQ(fh, TestGetTickerCount(options, BLOCK_CACHE_FILTER_HIT));
ASSERT_EQ(im + 1, TestGetTickerCount(options, BLOCK_CACHE_INDEX_MISS));
ASSERT_EQ(ih, TestGetTickerCount(options, BLOCK_CACHE_INDEX_HIT));
} else {
// If max_open_files is not -1, we do not preload table readers, so there is
// no change.
ASSERT_EQ(fm, TestGetTickerCount(options, BLOCK_CACHE_FILTER_MISS));
ASSERT_EQ(fh, TestGetTickerCount(options, BLOCK_CACHE_FILTER_HIT));
ASSERT_EQ(im, TestGetTickerCount(options, BLOCK_CACHE_INDEX_MISS));
ASSERT_EQ(ih, TestGetTickerCount(options, BLOCK_CACHE_INDEX_HIT));
}
std::string value;
// this should be read from L0
value = Get(1, "a2");
// If max_open_files is -1, we have pinned index and filter in Rep, so there
// will not be changes in index and filter misses or hits. If max_open_files
// is not -1, Get() will open a TableReader and prefetch index and filter.
ASSERT_EQ(fm + 1, TestGetTickerCount(options, BLOCK_CACHE_FILTER_MISS));
ASSERT_EQ(fh, TestGetTickerCount(options, BLOCK_CACHE_FILTER_HIT));
ASSERT_EQ(im + 1, TestGetTickerCount(options, BLOCK_CACHE_INDEX_MISS));
ASSERT_EQ(ih, TestGetTickerCount(options, BLOCK_CACHE_INDEX_HIT));
// this should be read from L1
value = Get(1, "a");
if (!disallow_preload_) {
// In inifinite max files case, there's a cache miss in executing Get()
// because index and filter are not prefetched before.
ASSERT_EQ(fm + 2, TestGetTickerCount(options, BLOCK_CACHE_FILTER_MISS));
ASSERT_EQ(fh, TestGetTickerCount(options, BLOCK_CACHE_FILTER_HIT));
ASSERT_EQ(im + 2, TestGetTickerCount(options, BLOCK_CACHE_INDEX_MISS));
ASSERT_EQ(ih, TestGetTickerCount(options, BLOCK_CACHE_INDEX_HIT));
} else {
// In this case, cache miss will be increased by one in
// BlockBasedTable::Open() because this is not in DB::Open() code path so we
// will prefetch L1's index and filter. Cache hit will also be increased by
// one because Get() will read index and filter from the block cache
// prefetched in previous Open() call.
ASSERT_EQ(fm + 2, TestGetTickerCount(options, BLOCK_CACHE_FILTER_MISS));
ASSERT_EQ(fh + 1, TestGetTickerCount(options, BLOCK_CACHE_FILTER_HIT));
ASSERT_EQ(im + 2, TestGetTickerCount(options, BLOCK_CACHE_INDEX_MISS));
ASSERT_EQ(ih + 1, TestGetTickerCount(options, BLOCK_CACHE_INDEX_HIT));
}
// Force a full compaction to one single file. There will be a block
// cache read for both of index and filter. If prefetch doesn't explicitly
// happen, it will happen when verifying the file.
Compact(1, "a", "zzzzz");
dbfull()->TEST_WaitForCompact();
if (!disallow_preload_) {
ASSERT_EQ(fm + 3, TestGetTickerCount(options, BLOCK_CACHE_FILTER_MISS));
ASSERT_EQ(fh, TestGetTickerCount(options, BLOCK_CACHE_FILTER_HIT));
ASSERT_EQ(im + 3, TestGetTickerCount(options, BLOCK_CACHE_INDEX_MISS));
ASSERT_EQ(ih + 3, TestGetTickerCount(options, BLOCK_CACHE_INDEX_HIT));
} else {
ASSERT_EQ(fm + 3, TestGetTickerCount(options, BLOCK_CACHE_FILTER_MISS));
ASSERT_EQ(fh + 1, TestGetTickerCount(options, BLOCK_CACHE_FILTER_HIT));
ASSERT_EQ(im + 3, TestGetTickerCount(options, BLOCK_CACHE_INDEX_MISS));
ASSERT_EQ(ih + 4, TestGetTickerCount(options, BLOCK_CACHE_INDEX_HIT));
}
// Bloom and index hit will happen when a Get() happens.
value = Get(1, "a");
if (!disallow_preload_) {
ASSERT_EQ(fm + 3, TestGetTickerCount(options, BLOCK_CACHE_FILTER_MISS));
ASSERT_EQ(fh + 1, TestGetTickerCount(options, BLOCK_CACHE_FILTER_HIT));
ASSERT_EQ(im + 3, TestGetTickerCount(options, BLOCK_CACHE_INDEX_MISS));
ASSERT_EQ(ih + 4, TestGetTickerCount(options, BLOCK_CACHE_INDEX_HIT));
} else {
ASSERT_EQ(fm + 3, TestGetTickerCount(options, BLOCK_CACHE_FILTER_MISS));
ASSERT_EQ(fh + 2, TestGetTickerCount(options, BLOCK_CACHE_FILTER_HIT));
ASSERT_EQ(im + 3, TestGetTickerCount(options, BLOCK_CACHE_INDEX_MISS));
ASSERT_EQ(ih + 5, TestGetTickerCount(options, BLOCK_CACHE_INDEX_HIT));
}
}
INSTANTIATE_TEST_CASE_P(PinL0IndexAndFilterBlocksTest,
PinL0IndexAndFilterBlocksTest,
::testing::Values(std::make_tuple(true, false),
std::make_tuple(false, false),
std::make_tuple(false, true)));
#ifndef ROCKSDB_LITE
TEST_F(DBTest2, MaxCompactionBytesTest) {
Options options = CurrentOptions();
options.memtable_factory.reset(
new SpecialSkipListFactory(DBTestBase::kNumKeysByGenerateNewRandomFile));
options.compaction_style = kCompactionStyleLevel;
options.write_buffer_size = 200 << 10;
options.arena_block_size = 4 << 10;
options.level0_file_num_compaction_trigger = 4;
options.num_levels = 4;
options.compression = kNoCompression;
options.max_bytes_for_level_base = 450 << 10;
options.target_file_size_base = 100 << 10;
// Infinite for full compaction.
options.max_compaction_bytes = options.target_file_size_base * 100;
Reopen(options);
Random rnd(301);
for (int num = 0; num < 8; num++) {
GenerateNewRandomFile(&rnd);
}
CompactRangeOptions cro;
cro.bottommost_level_compaction = BottommostLevelCompaction::kForceOptimized;
ASSERT_OK(db_->CompactRange(cro, nullptr, nullptr));
ASSERT_EQ("0,0,8", FilesPerLevel(0));
// When compact from Ln -> Ln+1, cut a file if the file overlaps with
// more than three files in Ln+1.
options.max_compaction_bytes = options.target_file_size_base * 3;
Reopen(options);
GenerateNewRandomFile(&rnd);
// Add three more small files that overlap with the previous file
for (int i = 0; i < 3; i++) {
Put("a", "z");
ASSERT_OK(Flush());
}
dbfull()->TEST_WaitForCompact();
// Output files to L1 are cut to three pieces, according to
// options.max_compaction_bytes
ASSERT_EQ("0,3,8", FilesPerLevel(0));
}
static void UniqueIdCallback(void* arg) {
int* result = reinterpret_cast<int*>(arg);
if (*result == -1) {
*result = 0;
}
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearTrace();
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"GetUniqueIdFromFile:FS_IOC_GETVERSION", UniqueIdCallback);
}
class MockPersistentCache : public PersistentCache {
public:
explicit MockPersistentCache(const bool is_compressed, const size_t max_size)
: is_compressed_(is_compressed), max_size_(max_size) {
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"GetUniqueIdFromFile:FS_IOC_GETVERSION", UniqueIdCallback);
}
~MockPersistentCache() override {}
PersistentCache::StatsType Stats() override {
return PersistentCache::StatsType();
}
uint64_t NewId() override {
return last_id_.fetch_add(1, std::memory_order_relaxed);
}
Status Insert(const Slice& page_key, const char* data,
const size_t size) override {
MutexLock _(&lock_);
if (size_ > max_size_) {
size_ -= data_.begin()->second.size();
data_.erase(data_.begin());
}
data_.insert(std::make_pair(page_key.ToString(), std::string(data, size)));
size_ += size;
return Status::OK();
}
Status Lookup(const Slice& page_key, std::unique_ptr<char[]>* data,
size_t* size) override {
MutexLock _(&lock_);
auto it = data_.find(page_key.ToString());
if (it == data_.end()) {
return Status::NotFound();
}
assert(page_key.ToString() == it->first);
data->reset(new char[it->second.size()]);
memcpy(data->get(), it->second.c_str(), it->second.size());
*size = it->second.size();
return Status::OK();
}
bool IsCompressed() override { return is_compressed_; }
std::string GetPrintableOptions() const override {
return "MockPersistentCache";
}
port::Mutex lock_;
std::map<std::string, std::string> data_;
const bool is_compressed_ = true;
size_t size_ = 0;
const size_t max_size_ = 10 * 1024; // 10KiB
std::atomic<uint64_t> last_id_{1};
};
#ifdef OS_LINUX
// Make sure that in CPU time perf context counters, Env::NowCPUNanos()
// is used, rather than Env::CPUNanos();
TEST_F(DBTest2, TestPerfContextGetCpuTime) {
// force resizing table cache so table handle is not preloaded so that
// we can measure find_table_nanos during Get().
dbfull()->TEST_table_cache()->SetCapacity(0);
ASSERT_OK(Put("foo", "bar"));
ASSERT_OK(Flush());
env_->now_cpu_count_.store(0);
Fix+clean up handling of mock sleeps (#7101) Summary: We have a number of tests hanging on MacOS and windows due to mishandling of code for mock sleeps. In addition, the code was in terrible shape because the same variable (addon_time_) would sometimes refer to microseconds and sometimes to seconds. One test even assumed it was nanoseconds but was written to pass anyway. This has been cleaned up so that DB tests generally use a SpecialEnv function to mock sleep, for either some number of microseconds or seconds depending on the function called. But to call one of these, the test must first call SetMockSleep (precondition enforced with assertion), which also turns sleeps in RocksDB into mock sleeps. To also removes accounting for actual clock time, call SetTimeElapseOnlySleepOnReopen, which implies SetMockSleep (on DB re-open). This latter setting only works by applying on DB re-open, otherwise havoc can ensue if Env goes back in time with DB open. More specifics: Removed some unused test classes, and updated comments on the general problem. Fixed DBSSTTest.GetTotalSstFilesSize using a sync point callback instead of mock time. For this we have the only modification to production code, inserting a sync point callback in flush_job.cc, which is not a change to production behavior. Removed unnecessary resetting of mock times to 0 in many tests. RocksDB deals in relative time. Any behaviors relying on absolute date/time are likely a bug. (The above test DBSSTTest.GetTotalSstFilesSize was the only one clearly injecting a specific absolute time for actual testing convenience.) Just in case I misunderstood some test, I put this note in each replacement: // NOTE: Presumed unnecessary and removed: resetting mock time in env Strengthened some tests like MergeTestTime, MergeCompactionTimeTest, and FilterCompactionTimeTest in db_test.cc stats_history_test and blob_db_test are each their own beast, rather deeply dependent on MockTimeEnv. Each gets its own variant of a work-around for TimedWait in a mock time environment. (Reduces redundancy and inconsistency in stats_history_test.) Intended follow-up: Remove TimedWait from the public API of InstrumentedCondVar, and only make that accessible through Env by passing in an InstrumentedCondVar and a deadline. Then the Env implementations mocking time can fix this problem without using sync points. (Test infrastructure using sync points interferes with individual tests' control over sync points.) With that change, we can simplify/consolidate the scattered work-arounds. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7101 Test Plan: make check on Linux and MacOS Reviewed By: zhichao-cao Differential Revision: D23032815 Pulled By: pdillinger fbshipit-source-id: 7f33967ada8b83011fb54e8279365c008bd6610b
2020-08-11 21:39:49 +02:00
env_->SetMockSleep();
// NOTE: Presumed unnecessary and removed: resetting mock time in env
// CPU timing is not enabled with kEnableTimeExceptForMutex
SetPerfLevel(PerfLevel::kEnableTimeExceptForMutex);
ASSERT_EQ("bar", Get("foo"));
ASSERT_EQ(0, get_perf_context()->get_cpu_nanos);
ASSERT_EQ(0, env_->now_cpu_count_.load());
Fix+clean up handling of mock sleeps (#7101) Summary: We have a number of tests hanging on MacOS and windows due to mishandling of code for mock sleeps. In addition, the code was in terrible shape because the same variable (addon_time_) would sometimes refer to microseconds and sometimes to seconds. One test even assumed it was nanoseconds but was written to pass anyway. This has been cleaned up so that DB tests generally use a SpecialEnv function to mock sleep, for either some number of microseconds or seconds depending on the function called. But to call one of these, the test must first call SetMockSleep (precondition enforced with assertion), which also turns sleeps in RocksDB into mock sleeps. To also removes accounting for actual clock time, call SetTimeElapseOnlySleepOnReopen, which implies SetMockSleep (on DB re-open). This latter setting only works by applying on DB re-open, otherwise havoc can ensue if Env goes back in time with DB open. More specifics: Removed some unused test classes, and updated comments on the general problem. Fixed DBSSTTest.GetTotalSstFilesSize using a sync point callback instead of mock time. For this we have the only modification to production code, inserting a sync point callback in flush_job.cc, which is not a change to production behavior. Removed unnecessary resetting of mock times to 0 in many tests. RocksDB deals in relative time. Any behaviors relying on absolute date/time are likely a bug. (The above test DBSSTTest.GetTotalSstFilesSize was the only one clearly injecting a specific absolute time for actual testing convenience.) Just in case I misunderstood some test, I put this note in each replacement: // NOTE: Presumed unnecessary and removed: resetting mock time in env Strengthened some tests like MergeTestTime, MergeCompactionTimeTest, and FilterCompactionTimeTest in db_test.cc stats_history_test and blob_db_test are each their own beast, rather deeply dependent on MockTimeEnv. Each gets its own variant of a work-around for TimedWait in a mock time environment. (Reduces redundancy and inconsistency in stats_history_test.) Intended follow-up: Remove TimedWait from the public API of InstrumentedCondVar, and only make that accessible through Env by passing in an InstrumentedCondVar and a deadline. Then the Env implementations mocking time can fix this problem without using sync points. (Test infrastructure using sync points interferes with individual tests' control over sync points.) With that change, we can simplify/consolidate the scattered work-arounds. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7101 Test Plan: make check on Linux and MacOS Reviewed By: zhichao-cao Differential Revision: D23032815 Pulled By: pdillinger fbshipit-source-id: 7f33967ada8b83011fb54e8279365c008bd6610b
2020-08-11 21:39:49 +02:00
constexpr uint64_t kDummyAddonSeconds = uint64_t{1000000};
constexpr uint64_t kDummyAddonNanos = 1000000000U * kDummyAddonSeconds;
// Add time to NowNanos() reading.
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"TableCache::FindTable:0",
Fix+clean up handling of mock sleeps (#7101) Summary: We have a number of tests hanging on MacOS and windows due to mishandling of code for mock sleeps. In addition, the code was in terrible shape because the same variable (addon_time_) would sometimes refer to microseconds and sometimes to seconds. One test even assumed it was nanoseconds but was written to pass anyway. This has been cleaned up so that DB tests generally use a SpecialEnv function to mock sleep, for either some number of microseconds or seconds depending on the function called. But to call one of these, the test must first call SetMockSleep (precondition enforced with assertion), which also turns sleeps in RocksDB into mock sleeps. To also removes accounting for actual clock time, call SetTimeElapseOnlySleepOnReopen, which implies SetMockSleep (on DB re-open). This latter setting only works by applying on DB re-open, otherwise havoc can ensue if Env goes back in time with DB open. More specifics: Removed some unused test classes, and updated comments on the general problem. Fixed DBSSTTest.GetTotalSstFilesSize using a sync point callback instead of mock time. For this we have the only modification to production code, inserting a sync point callback in flush_job.cc, which is not a change to production behavior. Removed unnecessary resetting of mock times to 0 in many tests. RocksDB deals in relative time. Any behaviors relying on absolute date/time are likely a bug. (The above test DBSSTTest.GetTotalSstFilesSize was the only one clearly injecting a specific absolute time for actual testing convenience.) Just in case I misunderstood some test, I put this note in each replacement: // NOTE: Presumed unnecessary and removed: resetting mock time in env Strengthened some tests like MergeTestTime, MergeCompactionTimeTest, and FilterCompactionTimeTest in db_test.cc stats_history_test and blob_db_test are each their own beast, rather deeply dependent on MockTimeEnv. Each gets its own variant of a work-around for TimedWait in a mock time environment. (Reduces redundancy and inconsistency in stats_history_test.) Intended follow-up: Remove TimedWait from the public API of InstrumentedCondVar, and only make that accessible through Env by passing in an InstrumentedCondVar and a deadline. Then the Env implementations mocking time can fix this problem without using sync points. (Test infrastructure using sync points interferes with individual tests' control over sync points.) With that change, we can simplify/consolidate the scattered work-arounds. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7101 Test Plan: make check on Linux and MacOS Reviewed By: zhichao-cao Differential Revision: D23032815 Pulled By: pdillinger fbshipit-source-id: 7f33967ada8b83011fb54e8279365c008bd6610b
2020-08-11 21:39:49 +02:00
[&](void* /*arg*/) { env_->MockSleepForSeconds(kDummyAddonSeconds); });
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
SetPerfLevel(PerfLevel::kEnableTimeAndCPUTimeExceptForMutex);
ASSERT_EQ("bar", Get("foo"));
ASSERT_GT(env_->now_cpu_count_.load(), 2);
Fix+clean up handling of mock sleeps (#7101) Summary: We have a number of tests hanging on MacOS and windows due to mishandling of code for mock sleeps. In addition, the code was in terrible shape because the same variable (addon_time_) would sometimes refer to microseconds and sometimes to seconds. One test even assumed it was nanoseconds but was written to pass anyway. This has been cleaned up so that DB tests generally use a SpecialEnv function to mock sleep, for either some number of microseconds or seconds depending on the function called. But to call one of these, the test must first call SetMockSleep (precondition enforced with assertion), which also turns sleeps in RocksDB into mock sleeps. To also removes accounting for actual clock time, call SetTimeElapseOnlySleepOnReopen, which implies SetMockSleep (on DB re-open). This latter setting only works by applying on DB re-open, otherwise havoc can ensue if Env goes back in time with DB open. More specifics: Removed some unused test classes, and updated comments on the general problem. Fixed DBSSTTest.GetTotalSstFilesSize using a sync point callback instead of mock time. For this we have the only modification to production code, inserting a sync point callback in flush_job.cc, which is not a change to production behavior. Removed unnecessary resetting of mock times to 0 in many tests. RocksDB deals in relative time. Any behaviors relying on absolute date/time are likely a bug. (The above test DBSSTTest.GetTotalSstFilesSize was the only one clearly injecting a specific absolute time for actual testing convenience.) Just in case I misunderstood some test, I put this note in each replacement: // NOTE: Presumed unnecessary and removed: resetting mock time in env Strengthened some tests like MergeTestTime, MergeCompactionTimeTest, and FilterCompactionTimeTest in db_test.cc stats_history_test and blob_db_test are each their own beast, rather deeply dependent on MockTimeEnv. Each gets its own variant of a work-around for TimedWait in a mock time environment. (Reduces redundancy and inconsistency in stats_history_test.) Intended follow-up: Remove TimedWait from the public API of InstrumentedCondVar, and only make that accessible through Env by passing in an InstrumentedCondVar and a deadline. Then the Env implementations mocking time can fix this problem without using sync points. (Test infrastructure using sync points interferes with individual tests' control over sync points.) With that change, we can simplify/consolidate the scattered work-arounds. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7101 Test Plan: make check on Linux and MacOS Reviewed By: zhichao-cao Differential Revision: D23032815 Pulled By: pdillinger fbshipit-source-id: 7f33967ada8b83011fb54e8279365c008bd6610b
2020-08-11 21:39:49 +02:00
ASSERT_LT(get_perf_context()->get_cpu_nanos, kDummyAddonNanos);
ASSERT_GT(get_perf_context()->find_table_nanos, kDummyAddonNanos);
SetPerfLevel(PerfLevel::kDisable);
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
}
TEST_F(DBTest2, TestPerfContextIterCpuTime) {
DestroyAndReopen(CurrentOptions());
// force resizing table cache so table handle is not preloaded so that
// we can measure find_table_nanos during iteration
dbfull()->TEST_table_cache()->SetCapacity(0);
const size_t kNumEntries = 10;
for (size_t i = 0; i < kNumEntries; ++i) {
ASSERT_OK(Put("k" + ToString(i), "v" + ToString(i)));
}
ASSERT_OK(Flush());
for (size_t i = 0; i < kNumEntries; ++i) {
ASSERT_EQ("v" + ToString(i), Get("k" + ToString(i)));
}
std::string last_key = "k" + ToString(kNumEntries - 1);
std::string last_value = "v" + ToString(kNumEntries - 1);
env_->now_cpu_count_.store(0);
Fix+clean up handling of mock sleeps (#7101) Summary: We have a number of tests hanging on MacOS and windows due to mishandling of code for mock sleeps. In addition, the code was in terrible shape because the same variable (addon_time_) would sometimes refer to microseconds and sometimes to seconds. One test even assumed it was nanoseconds but was written to pass anyway. This has been cleaned up so that DB tests generally use a SpecialEnv function to mock sleep, for either some number of microseconds or seconds depending on the function called. But to call one of these, the test must first call SetMockSleep (precondition enforced with assertion), which also turns sleeps in RocksDB into mock sleeps. To also removes accounting for actual clock time, call SetTimeElapseOnlySleepOnReopen, which implies SetMockSleep (on DB re-open). This latter setting only works by applying on DB re-open, otherwise havoc can ensue if Env goes back in time with DB open. More specifics: Removed some unused test classes, and updated comments on the general problem. Fixed DBSSTTest.GetTotalSstFilesSize using a sync point callback instead of mock time. For this we have the only modification to production code, inserting a sync point callback in flush_job.cc, which is not a change to production behavior. Removed unnecessary resetting of mock times to 0 in many tests. RocksDB deals in relative time. Any behaviors relying on absolute date/time are likely a bug. (The above test DBSSTTest.GetTotalSstFilesSize was the only one clearly injecting a specific absolute time for actual testing convenience.) Just in case I misunderstood some test, I put this note in each replacement: // NOTE: Presumed unnecessary and removed: resetting mock time in env Strengthened some tests like MergeTestTime, MergeCompactionTimeTest, and FilterCompactionTimeTest in db_test.cc stats_history_test and blob_db_test are each their own beast, rather deeply dependent on MockTimeEnv. Each gets its own variant of a work-around for TimedWait in a mock time environment. (Reduces redundancy and inconsistency in stats_history_test.) Intended follow-up: Remove TimedWait from the public API of InstrumentedCondVar, and only make that accessible through Env by passing in an InstrumentedCondVar and a deadline. Then the Env implementations mocking time can fix this problem without using sync points. (Test infrastructure using sync points interferes with individual tests' control over sync points.) With that change, we can simplify/consolidate the scattered work-arounds. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7101 Test Plan: make check on Linux and MacOS Reviewed By: zhichao-cao Differential Revision: D23032815 Pulled By: pdillinger fbshipit-source-id: 7f33967ada8b83011fb54e8279365c008bd6610b
2020-08-11 21:39:49 +02:00
env_->SetMockSleep();
// NOTE: Presumed unnecessary and removed: resetting mock time in env
// CPU timing is not enabled with kEnableTimeExceptForMutex
SetPerfLevel(PerfLevel::kEnableTimeExceptForMutex);
Iterator* iter = db_->NewIterator(ReadOptions());
iter->Seek("k0");
ASSERT_TRUE(iter->Valid());
ASSERT_EQ("v0", iter->value().ToString());
iter->SeekForPrev(last_key);
ASSERT_TRUE(iter->Valid());
iter->SeekToLast();
ASSERT_TRUE(iter->Valid());
ASSERT_EQ(last_value, iter->value().ToString());
iter->SeekToFirst();
ASSERT_TRUE(iter->Valid());
ASSERT_EQ("v0", iter->value().ToString());
ASSERT_EQ(0, get_perf_context()->iter_seek_cpu_nanos);
iter->Next();
ASSERT_TRUE(iter->Valid());
ASSERT_EQ("v1", iter->value().ToString());
ASSERT_EQ(0, get_perf_context()->iter_next_cpu_nanos);
iter->Prev();
ASSERT_TRUE(iter->Valid());
ASSERT_EQ("v0", iter->value().ToString());
ASSERT_EQ(0, get_perf_context()->iter_prev_cpu_nanos);
ASSERT_EQ(0, env_->now_cpu_count_.load());
delete iter;
Fix+clean up handling of mock sleeps (#7101) Summary: We have a number of tests hanging on MacOS and windows due to mishandling of code for mock sleeps. In addition, the code was in terrible shape because the same variable (addon_time_) would sometimes refer to microseconds and sometimes to seconds. One test even assumed it was nanoseconds but was written to pass anyway. This has been cleaned up so that DB tests generally use a SpecialEnv function to mock sleep, for either some number of microseconds or seconds depending on the function called. But to call one of these, the test must first call SetMockSleep (precondition enforced with assertion), which also turns sleeps in RocksDB into mock sleeps. To also removes accounting for actual clock time, call SetTimeElapseOnlySleepOnReopen, which implies SetMockSleep (on DB re-open). This latter setting only works by applying on DB re-open, otherwise havoc can ensue if Env goes back in time with DB open. More specifics: Removed some unused test classes, and updated comments on the general problem. Fixed DBSSTTest.GetTotalSstFilesSize using a sync point callback instead of mock time. For this we have the only modification to production code, inserting a sync point callback in flush_job.cc, which is not a change to production behavior. Removed unnecessary resetting of mock times to 0 in many tests. RocksDB deals in relative time. Any behaviors relying on absolute date/time are likely a bug. (The above test DBSSTTest.GetTotalSstFilesSize was the only one clearly injecting a specific absolute time for actual testing convenience.) Just in case I misunderstood some test, I put this note in each replacement: // NOTE: Presumed unnecessary and removed: resetting mock time in env Strengthened some tests like MergeTestTime, MergeCompactionTimeTest, and FilterCompactionTimeTest in db_test.cc stats_history_test and blob_db_test are each their own beast, rather deeply dependent on MockTimeEnv. Each gets its own variant of a work-around for TimedWait in a mock time environment. (Reduces redundancy and inconsistency in stats_history_test.) Intended follow-up: Remove TimedWait from the public API of InstrumentedCondVar, and only make that accessible through Env by passing in an InstrumentedCondVar and a deadline. Then the Env implementations mocking time can fix this problem without using sync points. (Test infrastructure using sync points interferes with individual tests' control over sync points.) With that change, we can simplify/consolidate the scattered work-arounds. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7101 Test Plan: make check on Linux and MacOS Reviewed By: zhichao-cao Differential Revision: D23032815 Pulled By: pdillinger fbshipit-source-id: 7f33967ada8b83011fb54e8279365c008bd6610b
2020-08-11 21:39:49 +02:00
constexpr uint64_t kDummyAddonSeconds = uint64_t{1000000};
constexpr uint64_t kDummyAddonNanos = 1000000000U * kDummyAddonSeconds;
// Add time to NowNanos() reading.
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"TableCache::FindTable:0",
Fix+clean up handling of mock sleeps (#7101) Summary: We have a number of tests hanging on MacOS and windows due to mishandling of code for mock sleeps. In addition, the code was in terrible shape because the same variable (addon_time_) would sometimes refer to microseconds and sometimes to seconds. One test even assumed it was nanoseconds but was written to pass anyway. This has been cleaned up so that DB tests generally use a SpecialEnv function to mock sleep, for either some number of microseconds or seconds depending on the function called. But to call one of these, the test must first call SetMockSleep (precondition enforced with assertion), which also turns sleeps in RocksDB into mock sleeps. To also removes accounting for actual clock time, call SetTimeElapseOnlySleepOnReopen, which implies SetMockSleep (on DB re-open). This latter setting only works by applying on DB re-open, otherwise havoc can ensue if Env goes back in time with DB open. More specifics: Removed some unused test classes, and updated comments on the general problem. Fixed DBSSTTest.GetTotalSstFilesSize using a sync point callback instead of mock time. For this we have the only modification to production code, inserting a sync point callback in flush_job.cc, which is not a change to production behavior. Removed unnecessary resetting of mock times to 0 in many tests. RocksDB deals in relative time. Any behaviors relying on absolute date/time are likely a bug. (The above test DBSSTTest.GetTotalSstFilesSize was the only one clearly injecting a specific absolute time for actual testing convenience.) Just in case I misunderstood some test, I put this note in each replacement: // NOTE: Presumed unnecessary and removed: resetting mock time in env Strengthened some tests like MergeTestTime, MergeCompactionTimeTest, and FilterCompactionTimeTest in db_test.cc stats_history_test and blob_db_test are each their own beast, rather deeply dependent on MockTimeEnv. Each gets its own variant of a work-around for TimedWait in a mock time environment. (Reduces redundancy and inconsistency in stats_history_test.) Intended follow-up: Remove TimedWait from the public API of InstrumentedCondVar, and only make that accessible through Env by passing in an InstrumentedCondVar and a deadline. Then the Env implementations mocking time can fix this problem without using sync points. (Test infrastructure using sync points interferes with individual tests' control over sync points.) With that change, we can simplify/consolidate the scattered work-arounds. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7101 Test Plan: make check on Linux and MacOS Reviewed By: zhichao-cao Differential Revision: D23032815 Pulled By: pdillinger fbshipit-source-id: 7f33967ada8b83011fb54e8279365c008bd6610b
2020-08-11 21:39:49 +02:00
[&](void* /*arg*/) { env_->MockSleepForSeconds(kDummyAddonSeconds); });
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
SetPerfLevel(PerfLevel::kEnableTimeAndCPUTimeExceptForMutex);
iter = db_->NewIterator(ReadOptions());
iter->Seek("k0");
ASSERT_TRUE(iter->Valid());
ASSERT_EQ("v0", iter->value().ToString());
iter->SeekForPrev(last_key);
ASSERT_TRUE(iter->Valid());
iter->SeekToLast();
ASSERT_TRUE(iter->Valid());
ASSERT_EQ(last_value, iter->value().ToString());
iter->SeekToFirst();
ASSERT_TRUE(iter->Valid());
ASSERT_EQ("v0", iter->value().ToString());
ASSERT_GT(get_perf_context()->iter_seek_cpu_nanos, 0);
Fix+clean up handling of mock sleeps (#7101) Summary: We have a number of tests hanging on MacOS and windows due to mishandling of code for mock sleeps. In addition, the code was in terrible shape because the same variable (addon_time_) would sometimes refer to microseconds and sometimes to seconds. One test even assumed it was nanoseconds but was written to pass anyway. This has been cleaned up so that DB tests generally use a SpecialEnv function to mock sleep, for either some number of microseconds or seconds depending on the function called. But to call one of these, the test must first call SetMockSleep (precondition enforced with assertion), which also turns sleeps in RocksDB into mock sleeps. To also removes accounting for actual clock time, call SetTimeElapseOnlySleepOnReopen, which implies SetMockSleep (on DB re-open). This latter setting only works by applying on DB re-open, otherwise havoc can ensue if Env goes back in time with DB open. More specifics: Removed some unused test classes, and updated comments on the general problem. Fixed DBSSTTest.GetTotalSstFilesSize using a sync point callback instead of mock time. For this we have the only modification to production code, inserting a sync point callback in flush_job.cc, which is not a change to production behavior. Removed unnecessary resetting of mock times to 0 in many tests. RocksDB deals in relative time. Any behaviors relying on absolute date/time are likely a bug. (The above test DBSSTTest.GetTotalSstFilesSize was the only one clearly injecting a specific absolute time for actual testing convenience.) Just in case I misunderstood some test, I put this note in each replacement: // NOTE: Presumed unnecessary and removed: resetting mock time in env Strengthened some tests like MergeTestTime, MergeCompactionTimeTest, and FilterCompactionTimeTest in db_test.cc stats_history_test and blob_db_test are each their own beast, rather deeply dependent on MockTimeEnv. Each gets its own variant of a work-around for TimedWait in a mock time environment. (Reduces redundancy and inconsistency in stats_history_test.) Intended follow-up: Remove TimedWait from the public API of InstrumentedCondVar, and only make that accessible through Env by passing in an InstrumentedCondVar and a deadline. Then the Env implementations mocking time can fix this problem without using sync points. (Test infrastructure using sync points interferes with individual tests' control over sync points.) With that change, we can simplify/consolidate the scattered work-arounds. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7101 Test Plan: make check on Linux and MacOS Reviewed By: zhichao-cao Differential Revision: D23032815 Pulled By: pdillinger fbshipit-source-id: 7f33967ada8b83011fb54e8279365c008bd6610b
2020-08-11 21:39:49 +02:00
ASSERT_LT(get_perf_context()->iter_seek_cpu_nanos, kDummyAddonNanos);
iter->Next();
ASSERT_TRUE(iter->Valid());
ASSERT_EQ("v1", iter->value().ToString());
ASSERT_GT(get_perf_context()->iter_next_cpu_nanos, 0);
Fix+clean up handling of mock sleeps (#7101) Summary: We have a number of tests hanging on MacOS and windows due to mishandling of code for mock sleeps. In addition, the code was in terrible shape because the same variable (addon_time_) would sometimes refer to microseconds and sometimes to seconds. One test even assumed it was nanoseconds but was written to pass anyway. This has been cleaned up so that DB tests generally use a SpecialEnv function to mock sleep, for either some number of microseconds or seconds depending on the function called. But to call one of these, the test must first call SetMockSleep (precondition enforced with assertion), which also turns sleeps in RocksDB into mock sleeps. To also removes accounting for actual clock time, call SetTimeElapseOnlySleepOnReopen, which implies SetMockSleep (on DB re-open). This latter setting only works by applying on DB re-open, otherwise havoc can ensue if Env goes back in time with DB open. More specifics: Removed some unused test classes, and updated comments on the general problem. Fixed DBSSTTest.GetTotalSstFilesSize using a sync point callback instead of mock time. For this we have the only modification to production code, inserting a sync point callback in flush_job.cc, which is not a change to production behavior. Removed unnecessary resetting of mock times to 0 in many tests. RocksDB deals in relative time. Any behaviors relying on absolute date/time are likely a bug. (The above test DBSSTTest.GetTotalSstFilesSize was the only one clearly injecting a specific absolute time for actual testing convenience.) Just in case I misunderstood some test, I put this note in each replacement: // NOTE: Presumed unnecessary and removed: resetting mock time in env Strengthened some tests like MergeTestTime, MergeCompactionTimeTest, and FilterCompactionTimeTest in db_test.cc stats_history_test and blob_db_test are each their own beast, rather deeply dependent on MockTimeEnv. Each gets its own variant of a work-around for TimedWait in a mock time environment. (Reduces redundancy and inconsistency in stats_history_test.) Intended follow-up: Remove TimedWait from the public API of InstrumentedCondVar, and only make that accessible through Env by passing in an InstrumentedCondVar and a deadline. Then the Env implementations mocking time can fix this problem without using sync points. (Test infrastructure using sync points interferes with individual tests' control over sync points.) With that change, we can simplify/consolidate the scattered work-arounds. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7101 Test Plan: make check on Linux and MacOS Reviewed By: zhichao-cao Differential Revision: D23032815 Pulled By: pdillinger fbshipit-source-id: 7f33967ada8b83011fb54e8279365c008bd6610b
2020-08-11 21:39:49 +02:00
ASSERT_LT(get_perf_context()->iter_next_cpu_nanos, kDummyAddonNanos);
iter->Prev();
ASSERT_TRUE(iter->Valid());
ASSERT_EQ("v0", iter->value().ToString());
ASSERT_GT(get_perf_context()->iter_prev_cpu_nanos, 0);
Fix+clean up handling of mock sleeps (#7101) Summary: We have a number of tests hanging on MacOS and windows due to mishandling of code for mock sleeps. In addition, the code was in terrible shape because the same variable (addon_time_) would sometimes refer to microseconds and sometimes to seconds. One test even assumed it was nanoseconds but was written to pass anyway. This has been cleaned up so that DB tests generally use a SpecialEnv function to mock sleep, for either some number of microseconds or seconds depending on the function called. But to call one of these, the test must first call SetMockSleep (precondition enforced with assertion), which also turns sleeps in RocksDB into mock sleeps. To also removes accounting for actual clock time, call SetTimeElapseOnlySleepOnReopen, which implies SetMockSleep (on DB re-open). This latter setting only works by applying on DB re-open, otherwise havoc can ensue if Env goes back in time with DB open. More specifics: Removed some unused test classes, and updated comments on the general problem. Fixed DBSSTTest.GetTotalSstFilesSize using a sync point callback instead of mock time. For this we have the only modification to production code, inserting a sync point callback in flush_job.cc, which is not a change to production behavior. Removed unnecessary resetting of mock times to 0 in many tests. RocksDB deals in relative time. Any behaviors relying on absolute date/time are likely a bug. (The above test DBSSTTest.GetTotalSstFilesSize was the only one clearly injecting a specific absolute time for actual testing convenience.) Just in case I misunderstood some test, I put this note in each replacement: // NOTE: Presumed unnecessary and removed: resetting mock time in env Strengthened some tests like MergeTestTime, MergeCompactionTimeTest, and FilterCompactionTimeTest in db_test.cc stats_history_test and blob_db_test are each their own beast, rather deeply dependent on MockTimeEnv. Each gets its own variant of a work-around for TimedWait in a mock time environment. (Reduces redundancy and inconsistency in stats_history_test.) Intended follow-up: Remove TimedWait from the public API of InstrumentedCondVar, and only make that accessible through Env by passing in an InstrumentedCondVar and a deadline. Then the Env implementations mocking time can fix this problem without using sync points. (Test infrastructure using sync points interferes with individual tests' control over sync points.) With that change, we can simplify/consolidate the scattered work-arounds. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7101 Test Plan: make check on Linux and MacOS Reviewed By: zhichao-cao Differential Revision: D23032815 Pulled By: pdillinger fbshipit-source-id: 7f33967ada8b83011fb54e8279365c008bd6610b
2020-08-11 21:39:49 +02:00
ASSERT_LT(get_perf_context()->iter_prev_cpu_nanos, kDummyAddonNanos);
ASSERT_GE(env_->now_cpu_count_.load(), 12);
Fix+clean up handling of mock sleeps (#7101) Summary: We have a number of tests hanging on MacOS and windows due to mishandling of code for mock sleeps. In addition, the code was in terrible shape because the same variable (addon_time_) would sometimes refer to microseconds and sometimes to seconds. One test even assumed it was nanoseconds but was written to pass anyway. This has been cleaned up so that DB tests generally use a SpecialEnv function to mock sleep, for either some number of microseconds or seconds depending on the function called. But to call one of these, the test must first call SetMockSleep (precondition enforced with assertion), which also turns sleeps in RocksDB into mock sleeps. To also removes accounting for actual clock time, call SetTimeElapseOnlySleepOnReopen, which implies SetMockSleep (on DB re-open). This latter setting only works by applying on DB re-open, otherwise havoc can ensue if Env goes back in time with DB open. More specifics: Removed some unused test classes, and updated comments on the general problem. Fixed DBSSTTest.GetTotalSstFilesSize using a sync point callback instead of mock time. For this we have the only modification to production code, inserting a sync point callback in flush_job.cc, which is not a change to production behavior. Removed unnecessary resetting of mock times to 0 in many tests. RocksDB deals in relative time. Any behaviors relying on absolute date/time are likely a bug. (The above test DBSSTTest.GetTotalSstFilesSize was the only one clearly injecting a specific absolute time for actual testing convenience.) Just in case I misunderstood some test, I put this note in each replacement: // NOTE: Presumed unnecessary and removed: resetting mock time in env Strengthened some tests like MergeTestTime, MergeCompactionTimeTest, and FilterCompactionTimeTest in db_test.cc stats_history_test and blob_db_test are each their own beast, rather deeply dependent on MockTimeEnv. Each gets its own variant of a work-around for TimedWait in a mock time environment. (Reduces redundancy and inconsistency in stats_history_test.) Intended follow-up: Remove TimedWait from the public API of InstrumentedCondVar, and only make that accessible through Env by passing in an InstrumentedCondVar and a deadline. Then the Env implementations mocking time can fix this problem without using sync points. (Test infrastructure using sync points interferes with individual tests' control over sync points.) With that change, we can simplify/consolidate the scattered work-arounds. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7101 Test Plan: make check on Linux and MacOS Reviewed By: zhichao-cao Differential Revision: D23032815 Pulled By: pdillinger fbshipit-source-id: 7f33967ada8b83011fb54e8279365c008bd6610b
2020-08-11 21:39:49 +02:00
ASSERT_GT(get_perf_context()->find_table_nanos, kDummyAddonNanos);
SetPerfLevel(PerfLevel::kDisable);
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
delete iter;
}
#endif // OS_LINUX
#if !defined OS_SOLARIS
TEST_F(DBTest2, PersistentCache) {
int num_iter = 80;
Options options;
options.write_buffer_size = 64 * 1024; // small write buffer
options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
options = CurrentOptions(options);
auto bsizes = {/*no block cache*/ 0, /*1M*/ 1 * 1024 * 1024};
auto types = {/*compressed*/ 1, /*uncompressed*/ 0};
for (auto bsize : bsizes) {
for (auto type : types) {
BlockBasedTableOptions table_options;
table_options.persistent_cache.reset(
new MockPersistentCache(type, 10 * 1024));
table_options.no_block_cache = true;
table_options.block_cache = bsize ? NewLRUCache(bsize) : nullptr;
table_options.block_cache_compressed = nullptr;
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
DestroyAndReopen(options);
CreateAndReopenWithCF({"pikachu"}, options);
// default column family doesn't have block cache
Options no_block_cache_opts;
no_block_cache_opts.statistics = options.statistics;
no_block_cache_opts = CurrentOptions(no_block_cache_opts);
BlockBasedTableOptions table_options_no_bc;
table_options_no_bc.no_block_cache = true;
no_block_cache_opts.table_factory.reset(
NewBlockBasedTableFactory(table_options_no_bc));
ReopenWithColumnFamilies(
{"default", "pikachu"},
std::vector<Options>({no_block_cache_opts, options}));
Random rnd(301);
// Write 8MB (80 values, each 100K)
ASSERT_EQ(NumTableFilesAtLevel(0, 1), 0);
std::vector<std::string> values;
std::string str;
for (int i = 0; i < num_iter; i++) {
if (i % 4 == 0) { // high compression ratio
str = rnd.RandomString(1000);
}
values.push_back(str);
ASSERT_OK(Put(1, Key(i), values[i]));
}
// flush all data from memtable so that reads are from block cache
ASSERT_OK(Flush(1));
for (int i = 0; i < num_iter; i++) {
ASSERT_EQ(Get(1, Key(i)), values[i]);
}
auto hit = options.statistics->getTickerCount(PERSISTENT_CACHE_HIT);
auto miss = options.statistics->getTickerCount(PERSISTENT_CACHE_MISS);
ASSERT_GT(hit, 0);
ASSERT_GT(miss, 0);
}
}
}
#endif // !defined OS_SOLARIS
namespace {
void CountSyncPoint() {
TEST_SYNC_POINT_CALLBACK("DBTest2::MarkedPoint", nullptr /* arg */);
}
} // namespace
TEST_F(DBTest2, SyncPointMarker) {
std::atomic<int> sync_point_called(0);
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"DBTest2::MarkedPoint",
[&](void* /*arg*/) { sync_point_called.fetch_add(1); });
// The first dependency enforces Marker can be loaded before MarkedPoint.
// The second checks that thread 1's MarkedPoint should be disabled here.
// Execution order:
// | Thread 1 | Thread 2 |
// | | Marker |
// | MarkedPoint | |
// | Thread1First | |
// | | MarkedPoint |
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependencyAndMarkers(
{{"DBTest2::SyncPointMarker:Thread1First", "DBTest2::MarkedPoint"}},
{{"DBTest2::SyncPointMarker:Marker", "DBTest2::MarkedPoint"}});
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
std::function<void()> func1 = [&]() {
CountSyncPoint();
TEST_SYNC_POINT("DBTest2::SyncPointMarker:Thread1First");
};
std::function<void()> func2 = [&]() {
TEST_SYNC_POINT("DBTest2::SyncPointMarker:Marker");
CountSyncPoint();
};
auto thread1 = port::Thread(func1);
auto thread2 = port::Thread(func2);
thread1.join();
thread2.join();
// Callback is only executed once
ASSERT_EQ(sync_point_called.load(), 1);
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
}
#endif
Introduce FullMergeV2 (eliminate memcpy from merge operators) Summary: This diff update the code to pin the merge operator operands while the merge operation is done, so that we can eliminate the memcpy cost, to do that we need a new public API for FullMerge that replace the std::deque<std::string> with std::vector<Slice> This diff is stacked on top of D56493 and D56511 In this diff we - Update FullMergeV2 arguments to be encapsulated in MergeOperationInput and MergeOperationOutput which will make it easier to add new arguments in the future - Replace std::deque<std::string> with std::vector<Slice> to pass operands - Replace MergeContext std::deque with std::vector (based on a simple benchmark I ran https://gist.github.com/IslamAbdelRahman/78fc86c9ab9f52b1df791e58943fb187) - Allow FullMergeV2 output to be an existing operand ``` [Everything in Memtable | 10K operands | 10 KB each | 1 operand per key] DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=10000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000 [FullMergeV2] readseq : 0.607 micros/op 1648235 ops/sec; 16121.2 MB/s readseq : 0.478 micros/op 2091546 ops/sec; 20457.2 MB/s readseq : 0.252 micros/op 3972081 ops/sec; 38850.5 MB/s readseq : 0.237 micros/op 4218328 ops/sec; 41259.0 MB/s readseq : 0.247 micros/op 4043927 ops/sec; 39553.2 MB/s [master] readseq : 3.935 micros/op 254140 ops/sec; 2485.7 MB/s readseq : 3.722 micros/op 268657 ops/sec; 2627.7 MB/s readseq : 3.149 micros/op 317605 ops/sec; 3106.5 MB/s readseq : 3.125 micros/op 320024 ops/sec; 3130.1 MB/s readseq : 4.075 micros/op 245374 ops/sec; 2400.0 MB/s ``` ``` [Everything in Memtable | 10K operands | 10 KB each | 10 operand per key] DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=1000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000 [FullMergeV2] readseq : 3.472 micros/op 288018 ops/sec; 2817.1 MB/s readseq : 2.304 micros/op 434027 ops/sec; 4245.2 MB/s readseq : 1.163 micros/op 859845 ops/sec; 8410.0 MB/s readseq : 1.192 micros/op 838926 ops/sec; 8205.4 MB/s readseq : 1.250 micros/op 800000 ops/sec; 7824.7 MB/s [master] readseq : 24.025 micros/op 41623 ops/sec; 407.1 MB/s readseq : 18.489 micros/op 54086 ops/sec; 529.0 MB/s readseq : 18.693 micros/op 53495 ops/sec; 523.2 MB/s readseq : 23.621 micros/op 42335 ops/sec; 414.1 MB/s readseq : 18.775 micros/op 53262 ops/sec; 521.0 MB/s ``` ``` [Everything in Block cache | 10K operands | 10 KB each | 1 operand per key] [FullMergeV2] $ DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions readseq : 14.741 micros/op 67837 ops/sec; 663.5 MB/s readseq : 1.029 micros/op 971446 ops/sec; 9501.6 MB/s readseq : 0.974 micros/op 1026229 ops/sec; 10037.4 MB/s readseq : 0.965 micros/op 1036080 ops/sec; 10133.8 MB/s readseq : 0.943 micros/op 1060657 ops/sec; 10374.2 MB/s [master] readseq : 16.735 micros/op 59755 ops/sec; 584.5 MB/s readseq : 3.029 micros/op 330151 ops/sec; 3229.2 MB/s readseq : 3.136 micros/op 318883 ops/sec; 3119.0 MB/s readseq : 3.065 micros/op 326245 ops/sec; 3191.0 MB/s readseq : 3.014 micros/op 331813 ops/sec; 3245.4 MB/s ``` ``` [Everything in Block cache | 10K operands | 10 KB each | 10 operand per key] DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10-operands-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions [FullMergeV2] readseq : 24.325 micros/op 41109 ops/sec; 402.1 MB/s readseq : 1.470 micros/op 680272 ops/sec; 6653.7 MB/s readseq : 1.231 micros/op 812347 ops/sec; 7945.5 MB/s readseq : 1.091 micros/op 916590 ops/sec; 8965.1 MB/s readseq : 1.109 micros/op 901713 ops/sec; 8819.6 MB/s [master] readseq : 27.257 micros/op 36687 ops/sec; 358.8 MB/s readseq : 4.443 micros/op 225073 ops/sec; 2201.4 MB/s readseq : 5.830 micros/op 171526 ops/sec; 1677.7 MB/s readseq : 4.173 micros/op 239635 ops/sec; 2343.8 MB/s readseq : 4.150 micros/op 240963 ops/sec; 2356.8 MB/s ``` Test Plan: COMPILE_WITH_ASAN=1 make check -j64 Reviewers: yhchiang, andrewkr, sdong Reviewed By: sdong Subscribers: lovro, andrewkr, dhruba Differential Revision: https://reviews.facebook.net/D57075
2016-07-20 18:49:03 +02:00
size_t GetEncodedEntrySize(size_t key_size, size_t value_size) {
std::string buffer;
PutVarint32(&buffer, static_cast<uint32_t>(0));
PutVarint32(&buffer, static_cast<uint32_t>(key_size));
PutVarint32(&buffer, static_cast<uint32_t>(value_size));
return buffer.size() + key_size + value_size;
}
TEST_F(DBTest2, ReadAmpBitmap) {
Options options = CurrentOptions();
BlockBasedTableOptions bbto;
uint32_t bytes_per_bit[2] = {1, 16};
for (size_t k = 0; k < 2; k++) {
// Disable delta encoding to make it easier to calculate read amplification
bbto.use_delta_encoding = false;
// Huge block cache to make it easier to calculate read amplification
bbto.block_cache = NewLRUCache(1024 * 1024 * 1024);
bbto.read_amp_bytes_per_bit = bytes_per_bit[k];
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
DestroyAndReopen(options);
const size_t kNumEntries = 10000;
Random rnd(301);
for (size_t i = 0; i < kNumEntries; i++) {
ASSERT_OK(Put(Key(static_cast<int>(i)), rnd.RandomString(100)));
}
ASSERT_OK(Flush());
Close();
Reopen(options);
// Read keys/values randomly and verify that reported read amp error
// is less than 2%
uint64_t total_useful_bytes = 0;
std::set<int> read_keys;
std::string value;
for (size_t i = 0; i < kNumEntries * 5; i++) {
int key_idx = rnd.Next() % kNumEntries;
std::string key = Key(key_idx);
ASSERT_OK(db_->Get(ReadOptions(), key, &value));
if (read_keys.find(key_idx) == read_keys.end()) {
auto internal_key = InternalKey(key, 0, ValueType::kTypeValue);
total_useful_bytes +=
GetEncodedEntrySize(internal_key.size(), value.size());
read_keys.insert(key_idx);
}
double expected_read_amp =
static_cast<double>(total_useful_bytes) /
options.statistics->getTickerCount(READ_AMP_TOTAL_READ_BYTES);
double read_amp =
static_cast<double>(options.statistics->getTickerCount(
READ_AMP_ESTIMATE_USEFUL_BYTES)) /
options.statistics->getTickerCount(READ_AMP_TOTAL_READ_BYTES);
double error_pct = fabs(expected_read_amp - read_amp) * 100;
// Error between reported read amp and real read amp should be less than
// 2%
EXPECT_LE(error_pct, 2);
}
// Make sure we read every thing in the DB (which is smaller than our cache)
Iterator* iter = db_->NewIterator(ReadOptions());
for (iter->SeekToFirst(); iter->Valid(); iter->Next()) {
ASSERT_EQ(iter->value().ToString(), Get(iter->key().ToString()));
}
delete iter;
// Read amp is on average 100% since we read all what we loaded in memory
if (k == 0) {
ASSERT_EQ(
options.statistics->getTickerCount(READ_AMP_ESTIMATE_USEFUL_BYTES),
options.statistics->getTickerCount(READ_AMP_TOTAL_READ_BYTES));
} else {
ASSERT_NEAR(
options.statistics->getTickerCount(READ_AMP_ESTIMATE_USEFUL_BYTES) *
1.0f /
options.statistics->getTickerCount(READ_AMP_TOTAL_READ_BYTES),
1, .01);
}
}
}
#ifndef OS_SOLARIS // GetUniqueIdFromFile is not implemented
TEST_F(DBTest2, ReadAmpBitmapLiveInCacheAfterDBClose) {
{
const int kIdBufLen = 100;
char id_buf[kIdBufLen];
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566) Summary: This PR does a few things: 1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation. 2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed: - The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated - The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory). 3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10). I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged. Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently. Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566 Reviewed By: zhichao-cao Differential Revision: D24408980 Pulled By: jay-zhuang fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
2020-10-27 18:31:34 +01:00
Status s = Status::NotSupported();
#ifndef OS_WIN
// You can't open a directory on windows using random access file
std::unique_ptr<RandomAccessFile> file;
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566) Summary: This PR does a few things: 1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation. 2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed: - The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated - The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory). 3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10). I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged. Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently. Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566 Reviewed By: zhichao-cao Differential Revision: D24408980 Pulled By: jay-zhuang fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
2020-10-27 18:31:34 +01:00
s = env_->NewRandomAccessFile(dbname_, &file, EnvOptions());
if (s.ok()) {
if (file->GetUniqueId(id_buf, kIdBufLen) == 0) {
// fs holding db directory doesn't support getting a unique file id,
// this means that running this test will fail because lru_cache will
// load the blocks again regardless of them being already in the cache
return;
}
}
#endif
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566) Summary: This PR does a few things: 1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation. 2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed: - The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated - The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory). 3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10). I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged. Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently. Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566 Reviewed By: zhichao-cao Differential Revision: D24408980 Pulled By: jay-zhuang fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
2020-10-27 18:31:34 +01:00
if (!s.ok()) {
std::unique_ptr<Directory> dir;
ASSERT_OK(env_->NewDirectory(dbname_, &dir));
if (dir->GetUniqueId(id_buf, kIdBufLen) == 0) {
// fs holding db directory doesn't support getting a unique file id,
// this means that running this test will fail because lru_cache will
// load the blocks again regardless of them being already in the cache
return;
}
}
}
uint32_t bytes_per_bit[2] = {1, 16};
for (size_t k = 0; k < 2; k++) {
std::shared_ptr<Cache> lru_cache = NewLRUCache(1024 * 1024 * 1024);
std::shared_ptr<Statistics> stats = ROCKSDB_NAMESPACE::CreateDBStatistics();
Options options = CurrentOptions();
BlockBasedTableOptions bbto;
// Disable delta encoding to make it easier to calculate read amplification
bbto.use_delta_encoding = false;
// Huge block cache to make it easier to calculate read amplification
bbto.block_cache = lru_cache;
bbto.read_amp_bytes_per_bit = bytes_per_bit[k];
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
options.statistics = stats;
DestroyAndReopen(options);
const int kNumEntries = 10000;
Random rnd(301);
for (int i = 0; i < kNumEntries; i++) {
ASSERT_OK(Put(Key(i), rnd.RandomString(100)));
}
ASSERT_OK(Flush());
Close();
Reopen(options);
uint64_t total_useful_bytes = 0;
std::set<int> read_keys;
std::string value;
// Iter1: Read half the DB, Read even keys
// Key(0), Key(2), Key(4), Key(6), Key(8), ...
for (int i = 0; i < kNumEntries; i += 2) {
std::string key = Key(i);
ASSERT_OK(db_->Get(ReadOptions(), key, &value));
if (read_keys.find(i) == read_keys.end()) {
auto internal_key = InternalKey(key, 0, ValueType::kTypeValue);
total_useful_bytes +=
GetEncodedEntrySize(internal_key.size(), value.size());
read_keys.insert(i);
}
}
size_t total_useful_bytes_iter1 =
options.statistics->getTickerCount(READ_AMP_ESTIMATE_USEFUL_BYTES);
size_t total_loaded_bytes_iter1 =
options.statistics->getTickerCount(READ_AMP_TOTAL_READ_BYTES);
Close();
std::shared_ptr<Statistics> new_statistics =
ROCKSDB_NAMESPACE::CreateDBStatistics();
// Destroy old statistics obj that the blocks in lru_cache are pointing to
options.statistics.reset();
// Use the statistics object that we just created
options.statistics = new_statistics;
Reopen(options);
// Iter2: Read half the DB, Read odd keys
// Key(1), Key(3), Key(5), Key(7), Key(9), ...
for (int i = 1; i < kNumEntries; i += 2) {
std::string key = Key(i);
ASSERT_OK(db_->Get(ReadOptions(), key, &value));
if (read_keys.find(i) == read_keys.end()) {
auto internal_key = InternalKey(key, 0, ValueType::kTypeValue);
total_useful_bytes +=
GetEncodedEntrySize(internal_key.size(), value.size());
read_keys.insert(i);
}
}
size_t total_useful_bytes_iter2 =
options.statistics->getTickerCount(READ_AMP_ESTIMATE_USEFUL_BYTES);
size_t total_loaded_bytes_iter2 =
options.statistics->getTickerCount(READ_AMP_TOTAL_READ_BYTES);
// Read amp is on average 100% since we read all what we loaded in memory
if (k == 0) {
ASSERT_EQ(total_useful_bytes_iter1 + total_useful_bytes_iter2,
total_loaded_bytes_iter1 + total_loaded_bytes_iter2);
} else {
ASSERT_NEAR((total_useful_bytes_iter1 + total_useful_bytes_iter2) * 1.0f /
(total_loaded_bytes_iter1 + total_loaded_bytes_iter2),
1, .01);
}
}
}
#endif // !OS_SOLARIS
#ifndef ROCKSDB_LITE
TEST_F(DBTest2, AutomaticCompactionOverlapManualCompaction) {
Options options = CurrentOptions();
options.num_levels = 3;
options.IncreaseParallelism(20);
DestroyAndReopen(options);
ASSERT_OK(Put(Key(0), "a"));
ASSERT_OK(Put(Key(5), "a"));
ASSERT_OK(Flush());
ASSERT_OK(Put(Key(10), "a"));
ASSERT_OK(Put(Key(15), "a"));
ASSERT_OK(Flush());
CompactRangeOptions cro;
cro.change_level = true;
cro.target_level = 2;
ASSERT_OK(db_->CompactRange(cro, nullptr, nullptr));
auto get_stat = [](std::string level_str, LevelStatType type,
std::map<std::string, std::string> props) {
auto prop_str =
"compaction." + level_str + "." +
InternalStats::compaction_level_stats.at(type).property_name.c_str();
auto prop_item = props.find(prop_str);
return prop_item == props.end() ? 0 : std::stod(prop_item->second);
};
// Trivial move 2 files to L2
ASSERT_EQ("0,0,2", FilesPerLevel());
// Also test that the stats GetMapProperty API reporting the same result
{
std::map<std::string, std::string> prop;
ASSERT_TRUE(dbfull()->GetMapProperty("rocksdb.cfstats", &prop));
ASSERT_EQ(0, get_stat("L0", LevelStatType::NUM_FILES, prop));
ASSERT_EQ(0, get_stat("L1", LevelStatType::NUM_FILES, prop));
ASSERT_EQ(2, get_stat("L2", LevelStatType::NUM_FILES, prop));
ASSERT_EQ(2, get_stat("Sum", LevelStatType::NUM_FILES, prop));
}
// While the compaction is running, we will create 2 new files that
// can fit in L2, these 2 files will be moved to L2 and overlap with
// the running compaction and break the LSM consistency.
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"CompactionJob::Run():Start", [&](void* /*arg*/) {
ASSERT_OK(
dbfull()->SetOptions({{"level0_file_num_compaction_trigger", "2"},
{"max_bytes_for_level_base", "1"}}));
ASSERT_OK(Put(Key(6), "a"));
ASSERT_OK(Put(Key(7), "a"));
ASSERT_OK(Flush());
ASSERT_OK(Put(Key(8), "a"));
ASSERT_OK(Put(Key(9), "a"));
ASSERT_OK(Flush());
});
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
// Run a manual compaction that will compact the 2 files in L2
// into 1 file in L2
cro.exclusive_manual_compaction = false;
cro.bottommost_level_compaction = BottommostLevelCompaction::kForceOptimized;
ASSERT_OK(db_->CompactRange(cro, nullptr, nullptr));
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
// Test that the stats GetMapProperty API reporting 1 file in L2
{
std::map<std::string, std::string> prop;
ASSERT_TRUE(dbfull()->GetMapProperty("rocksdb.cfstats", &prop));
ASSERT_EQ(1, get_stat("L2", LevelStatType::NUM_FILES, prop));
}
}
TEST_F(DBTest2, ManualCompactionOverlapManualCompaction) {
Options options = CurrentOptions();
options.num_levels = 2;
options.IncreaseParallelism(20);
options.disable_auto_compactions = true;
DestroyAndReopen(options);
ASSERT_OK(Put(Key(0), "a"));
ASSERT_OK(Put(Key(5), "a"));
ASSERT_OK(Flush());
ASSERT_OK(Put(Key(10), "a"));
ASSERT_OK(Put(Key(15), "a"));
ASSERT_OK(Flush());
ASSERT_OK(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
// Trivial move 2 files to L1
ASSERT_EQ("0,2", FilesPerLevel());
std::function<void()> bg_manual_compact = [&]() {
std::string k1 = Key(6);
std::string k2 = Key(9);
Slice k1s(k1);
Slice k2s(k2);
CompactRangeOptions cro;
cro.exclusive_manual_compaction = false;
ASSERT_OK(db_->CompactRange(cro, &k1s, &k2s));
};
ROCKSDB_NAMESPACE::port::Thread bg_thread;
// While the compaction is running, we will create 2 new files that
// can fit in L1, these 2 files will be moved to L1 and overlap with
// the running compaction and break the LSM consistency.
std::atomic<bool> flag(false);
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"CompactionJob::Run():Start", [&](void* /*arg*/) {
if (flag.exchange(true)) {
// We want to make sure to call this callback only once
return;
}
ASSERT_OK(Put(Key(6), "a"));
ASSERT_OK(Put(Key(7), "a"));
ASSERT_OK(Flush());
ASSERT_OK(Put(Key(8), "a"));
ASSERT_OK(Put(Key(9), "a"));
ASSERT_OK(Flush());
// Start a non-exclusive manual compaction in a bg thread
bg_thread = port::Thread(bg_manual_compact);
// This manual compaction conflict with the other manual compaction
// so it should wait until the first compaction finish
env_->SleepForMicroseconds(1000000);
});
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
// Run a manual compaction that will compact the 2 files in L1
// into 1 file in L1
CompactRangeOptions cro;
cro.exclusive_manual_compaction = false;
cro.bottommost_level_compaction = BottommostLevelCompaction::kForceOptimized;
ASSERT_OK(db_->CompactRange(cro, nullptr, nullptr));
bg_thread.join();
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
}
TEST_F(DBTest2, PausingManualCompaction1) {
Options options = CurrentOptions();
options.disable_auto_compactions = true;
options.num_levels = 7;
DestroyAndReopen(options);
Random rnd(301);
// Generate a file containing 10 keys.
for (int i = 0; i < 10; i++) {
ASSERT_OK(Put(Key(i), rnd.RandomString(50)));
}
ASSERT_OK(Flush());
// Generate another file containing same keys
for (int i = 0; i < 10; i++) {
ASSERT_OK(Put(Key(i), rnd.RandomString(50)));
}
ASSERT_OK(Flush());
int manual_compactions_paused = 0;
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"CompactionJob::Run():PausingManualCompaction:1", [&](void* arg) {
Disable manual compaction during `ReFitLevel()` (#7250) Summary: Manual compaction with `CompactRangeOptions::change_levels` set could refit to a level targeted by another manual compaction. If force_consistency_checks were disabled, it could be possible for overlapping files to be written at that target level. This PR prevents the possibility by calling `DisableManualCompaction()` prior to `ReFitLevel()`. It also improves the manual compaction disabling mechanism to wait for pending manual compactions to complete before returning, and support disabling from multiple threads. Fixes https://github.com/facebook/rocksdb/issues/6432. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7250 Test Plan: crash test command that repro'd the bug reliably: ``` $ TEST_TMPDIR=/dev/shm python tools/db_crashtest.py blackbox --simple -target_file_size_base=524288 -write_buffer_size=1048576 -clear_column_family_one_in=0 -reopen=0 -max_key=10000000 -column_families=1 -max_background_compactions=8 -compact_range_one_in=100000 -compression_type=none -compaction_style=1 -num_levels=5 -universal_min_merge_width=4 -universal_max_merge_width=8 -level0_file_num_compaction_trigger=12 -rate_limiter_bytes_per_sec=1048576000 -universal_max_size_amplification_percent=100 --duration=3600 --interval=60 --use_direct_io_for_flush_and_compaction=0 --use_direct_reads=0 --enable_compaction_filter=0 ``` Reviewed By: ltamasi Differential Revision: D23090800 Pulled By: ajkr fbshipit-source-id: afcbcd51b42ce76789fdb907d8b9ada790709c13
2020-08-14 20:28:12 +02:00
auto paused = static_cast<std::atomic<int>*>(arg);
ASSERT_EQ(0, paused->load(std::memory_order_acquire));
paused->fetch_add(1, std::memory_order_release);
manual_compactions_paused += 1;
});
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
std::vector<std::string> files_before_compact, files_after_compact;
// Remember file name before compaction is triggered
std::vector<LiveFileMetaData> files_meta;
dbfull()->GetLiveFilesMetaData(&files_meta);
for (auto file : files_meta) {
files_before_compact.push_back(file.name);
}
// OK, now trigger a manual compaction
dbfull()->CompactRange(CompactRangeOptions(), nullptr, nullptr);
// Wait for compactions to get scheduled and stopped
dbfull()->TEST_WaitForCompact(true);
// Get file names after compaction is stopped
files_meta.clear();
dbfull()->GetLiveFilesMetaData(&files_meta);
for (auto file : files_meta) {
files_after_compact.push_back(file.name);
}
// Like nothing happened
ASSERT_EQ(files_before_compact, files_after_compact);
ASSERT_EQ(manual_compactions_paused, 1);
manual_compactions_paused = 0;
// Now make sure CompactFiles also not run
dbfull()->CompactFiles(ROCKSDB_NAMESPACE::CompactionOptions(),
files_before_compact, 0);
// Wait for manual compaction to get scheduled and finish
dbfull()->TEST_WaitForCompact(true);
files_meta.clear();
files_after_compact.clear();
dbfull()->GetLiveFilesMetaData(&files_meta);
for (auto file : files_meta) {
files_after_compact.push_back(file.name);
}
ASSERT_EQ(files_before_compact, files_after_compact);
// CompactFiles returns at entry point
ASSERT_EQ(manual_compactions_paused, 0);
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
}
// PausingManualCompaction does not affect auto compaction
TEST_F(DBTest2, PausingManualCompaction2) {
Options options = CurrentOptions();
options.level0_file_num_compaction_trigger = 2;
options.disable_auto_compactions = false;
DestroyAndReopen(options);
dbfull()->DisableManualCompaction();
Random rnd(301);
for (int i = 0; i < 2; i++) {
// Generate a file containing 10 keys.
for (int j = 0; j < 100; j++) {
ASSERT_OK(Put(Key(j), rnd.RandomString(50)));
}
ASSERT_OK(Flush());
}
ASSERT_OK(dbfull()->TEST_WaitForCompact(true));
std::vector<LiveFileMetaData> files_meta;
dbfull()->GetLiveFilesMetaData(&files_meta);
ASSERT_EQ(files_meta.size(), 1);
}
TEST_F(DBTest2, PausingManualCompaction3) {
CompactRangeOptions compact_options;
Options options = CurrentOptions();
options.disable_auto_compactions = true;
options.num_levels = 7;
Random rnd(301);
auto generate_files = [&]() {
for (int i = 0; i < options.num_levels; i++) {
for (int j = 0; j < options.num_levels - i + 1; j++) {
for (int k = 0; k < 1000; k++) {
ASSERT_OK(Put(Key(k + j * 1000), rnd.RandomString(50)));
}
Flush();
}
for (int l = 1; l < options.num_levels - i; l++) {
MoveFilesToLevel(l);
}
}
};
DestroyAndReopen(options);
generate_files();
#ifndef ROCKSDB_LITE
ASSERT_EQ("2,3,4,5,6,7,8", FilesPerLevel());
#endif // !ROCKSDB_LITE
int run_manual_compactions = 0;
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"CompactionJob::Run():PausingManualCompaction:1",
[&](void* /*arg*/) { run_manual_compactions++; });
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
dbfull()->DisableManualCompaction();
dbfull()->CompactRange(compact_options, nullptr, nullptr);
dbfull()->TEST_WaitForCompact(true);
// As manual compaction disabled, not even reach sync point
ASSERT_EQ(run_manual_compactions, 0);
#ifndef ROCKSDB_LITE
ASSERT_EQ("2,3,4,5,6,7,8", FilesPerLevel());
#endif // !ROCKSDB_LITE
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearCallBack(
"CompactionJob::Run():PausingManualCompaction:1");
dbfull()->EnableManualCompaction();
dbfull()->CompactRange(compact_options, nullptr, nullptr);
dbfull()->TEST_WaitForCompact(true);
#ifndef ROCKSDB_LITE
ASSERT_EQ("0,0,0,0,0,0,2", FilesPerLevel());
#endif // !ROCKSDB_LITE
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
}
TEST_F(DBTest2, PausingManualCompaction4) {
CompactRangeOptions compact_options;
Options options = CurrentOptions();
options.disable_auto_compactions = true;
options.num_levels = 7;
Random rnd(301);
auto generate_files = [&]() {
for (int i = 0; i < options.num_levels; i++) {
for (int j = 0; j < options.num_levels - i + 1; j++) {
for (int k = 0; k < 1000; k++) {
ASSERT_OK(Put(Key(k + j * 1000), rnd.RandomString(50)));
}
Flush();
}
for (int l = 1; l < options.num_levels - i; l++) {
MoveFilesToLevel(l);
}
}
};
DestroyAndReopen(options);
generate_files();
#ifndef ROCKSDB_LITE
ASSERT_EQ("2,3,4,5,6,7,8", FilesPerLevel());
#endif // !ROCKSDB_LITE
int run_manual_compactions = 0;
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"CompactionJob::Run():PausingManualCompaction:2", [&](void* arg) {
Disable manual compaction during `ReFitLevel()` (#7250) Summary: Manual compaction with `CompactRangeOptions::change_levels` set could refit to a level targeted by another manual compaction. If force_consistency_checks were disabled, it could be possible for overlapping files to be written at that target level. This PR prevents the possibility by calling `DisableManualCompaction()` prior to `ReFitLevel()`. It also improves the manual compaction disabling mechanism to wait for pending manual compactions to complete before returning, and support disabling from multiple threads. Fixes https://github.com/facebook/rocksdb/issues/6432. Pull Request resolved: https://github.com/facebook/rocksdb/pull/7250 Test Plan: crash test command that repro'd the bug reliably: ``` $ TEST_TMPDIR=/dev/shm python tools/db_crashtest.py blackbox --simple -target_file_size_base=524288 -write_buffer_size=1048576 -clear_column_family_one_in=0 -reopen=0 -max_key=10000000 -column_families=1 -max_background_compactions=8 -compact_range_one_in=100000 -compression_type=none -compaction_style=1 -num_levels=5 -universal_min_merge_width=4 -universal_max_merge_width=8 -level0_file_num_compaction_trigger=12 -rate_limiter_bytes_per_sec=1048576000 -universal_max_size_amplification_percent=100 --duration=3600 --interval=60 --use_direct_io_for_flush_and_compaction=0 --use_direct_reads=0 --enable_compaction_filter=0 ``` Reviewed By: ltamasi Differential Revision: D23090800 Pulled By: ajkr fbshipit-source-id: afcbcd51b42ce76789fdb907d8b9ada790709c13
2020-08-14 20:28:12 +02:00
auto paused = static_cast<std::atomic<int>*>(arg);
ASSERT_EQ(0, paused->load(std::memory_order_acquire));
paused->fetch_add(1, std::memory_order_release);
run_manual_compactions++;
});
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
dbfull()->CompactRange(compact_options, nullptr, nullptr);
dbfull()->TEST_WaitForCompact(true);
ASSERT_EQ(run_manual_compactions, 1);
#ifndef ROCKSDB_LITE
ASSERT_EQ("2,3,4,5,6,7,8", FilesPerLevel());
#endif // !ROCKSDB_LITE
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearCallBack(
"CompactionJob::Run():PausingManualCompaction:2");
dbfull()->EnableManualCompaction();
dbfull()->CompactRange(compact_options, nullptr, nullptr);
dbfull()->TEST_WaitForCompact(true);
#ifndef ROCKSDB_LITE
ASSERT_EQ("0,0,0,0,0,0,2", FilesPerLevel());
#endif // !ROCKSDB_LITE
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
}
TEST_F(DBTest2, CancelManualCompaction1) {
CompactRangeOptions compact_options;
auto canceledPtr =
std::unique_ptr<std::atomic<bool>>(new std::atomic<bool>{true});
compact_options.canceled = canceledPtr.get();
Options options = CurrentOptions();
options.disable_auto_compactions = true;
options.num_levels = 7;
Random rnd(301);
auto generate_files = [&]() {
for (int i = 0; i < options.num_levels; i++) {
for (int j = 0; j < options.num_levels - i + 1; j++) {
for (int k = 0; k < 1000; k++) {
ASSERT_OK(Put(Key(k + j * 1000), rnd.RandomString(50)));
}
Flush();
}
for (int l = 1; l < options.num_levels - i; l++) {
MoveFilesToLevel(l);
}
}
};
DestroyAndReopen(options);
generate_files();
#ifndef ROCKSDB_LITE
ASSERT_EQ("2,3,4,5,6,7,8", FilesPerLevel());
#endif // !ROCKSDB_LITE
int run_manual_compactions = 0;
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"CompactionJob::Run():PausingManualCompaction:1",
[&](void* /*arg*/) { run_manual_compactions++; });
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
// Setup a callback to disable compactions after a couple of levels are
// compacted
int compactions_run = 0;
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"DBImpl::RunManualCompaction()::1",
[&](void* /*arg*/) { ++compactions_run; });
dbfull()->CompactRange(compact_options, nullptr, nullptr);
dbfull()->TEST_WaitForCompact(true);
// Since compactions are disabled, we shouldn't start compacting.
// E.g. we should call the compaction function exactly one time.
ASSERT_EQ(compactions_run, 0);
ASSERT_EQ(run_manual_compactions, 0);
#ifndef ROCKSDB_LITE
ASSERT_EQ("2,3,4,5,6,7,8", FilesPerLevel());
#endif // !ROCKSDB_LITE
compactions_run = 0;
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearCallBack(
"DBImpl::RunManualCompaction()::1");
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"DBImpl::RunManualCompaction()::1", [&](void* /*arg*/) {
++compactions_run;
// After 3 compactions disable
if (compactions_run == 3) {
compact_options.canceled->store(true, std::memory_order_release);
}
});
compact_options.canceled->store(false, std::memory_order_release);
dbfull()->CompactRange(compact_options, nullptr, nullptr);
dbfull()->TEST_WaitForCompact(true);
ASSERT_EQ(compactions_run, 3);
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearCallBack(
"DBImpl::RunManualCompaction()::1");
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearCallBack(
"CompactionJob::Run():PausingManualCompaction:1");
// Compactions should work again if we re-enable them..
compact_options.canceled->store(false, std::memory_order_relaxed);
dbfull()->CompactRange(compact_options, nullptr, nullptr);
dbfull()->TEST_WaitForCompact(true);
#ifndef ROCKSDB_LITE
ASSERT_EQ("0,0,0,0,0,0,2", FilesPerLevel());
#endif // !ROCKSDB_LITE
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
}
TEST_F(DBTest2, CancelManualCompaction2) {
CompactRangeOptions compact_options;
auto canceledPtr =
std::unique_ptr<std::atomic<bool>>(new std::atomic<bool>{true});
compact_options.canceled = canceledPtr.get();
compact_options.max_subcompactions = 1;
Options options = CurrentOptions();
options.disable_auto_compactions = true;
options.num_levels = 7;
Random rnd(301);
auto generate_files = [&]() {
for (int i = 0; i < options.num_levels; i++) {
for (int j = 0; j < options.num_levels - i + 1; j++) {
for (int k = 0; k < 1000; k++) {
ASSERT_OK(Put(Key(k + j * 1000), rnd.RandomString(50)));
}
Flush();
}
for (int l = 1; l < options.num_levels - i; l++) {
MoveFilesToLevel(l);
}
}
};
DestroyAndReopen(options);
generate_files();
#ifndef ROCKSDB_LITE
ASSERT_EQ("2,3,4,5,6,7,8", FilesPerLevel());
#endif // !ROCKSDB_LITE
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
int compactions_run = 0;
std::atomic<int> kv_compactions{0};
int compactions_stopped_at = 0;
int kv_compactions_stopped_at = 0;
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"DBImpl::RunManualCompaction()::1", [&](void* /*arg*/) {
++compactions_run;
// After 3 compactions disable
});
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"CompactionIterator:ProcessKV", [&](void* /*arg*/) {
int kv_compactions_run =
kv_compactions.fetch_add(1, std::memory_order_release);
if (kv_compactions_run == 5) {
compact_options.canceled->store(true, std::memory_order_release);
kv_compactions_stopped_at = kv_compactions_run;
compactions_stopped_at = compactions_run;
}
});
compact_options.canceled->store(false, std::memory_order_release);
dbfull()->CompactRange(compact_options, nullptr, nullptr);
dbfull()->TEST_WaitForCompact(true);
// NOTE: as we set compact_options.max_subcompacitons = 1, and store true to
// the canceled variable from the single compacting thread (via callback),
// this value is deterministically kv_compactions_stopped_at + 1.
ASSERT_EQ(kv_compactions, kv_compactions_stopped_at + 1);
ASSERT_EQ(compactions_run, compactions_stopped_at);
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearCallBack(
"CompactionIterator::ProcessKV");
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearCallBack(
"DBImpl::RunManualCompaction()::1");
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearCallBack(
"CompactionJob::Run():PausingManualCompaction:1");
// Compactions should work again if we re-enable them..
compact_options.canceled->store(false, std::memory_order_relaxed);
dbfull()->CompactRange(compact_options, nullptr, nullptr);
dbfull()->TEST_WaitForCompact(true);
#ifndef ROCKSDB_LITE
ASSERT_EQ("0,0,0,0,0,0,2", FilesPerLevel());
#endif // !ROCKSDB_LITE
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
}
class CancelCompactionListener : public EventListener {
public:
CancelCompactionListener()
: num_compaction_started_(0), num_compaction_ended_(0) {}
void OnCompactionBegin(DB* /*db*/, const CompactionJobInfo& ci) override {
ASSERT_EQ(ci.cf_name, "default");
ASSERT_EQ(ci.base_input_level, 0);
num_compaction_started_++;
}
void OnCompactionCompleted(DB* /*db*/, const CompactionJobInfo& ci) override {
ASSERT_EQ(ci.cf_name, "default");
ASSERT_EQ(ci.base_input_level, 0);
ASSERT_EQ(ci.status.code(), code_);
ASSERT_EQ(ci.status.subcode(), subcode_);
num_compaction_ended_++;
}
std::atomic<size_t> num_compaction_started_;
std::atomic<size_t> num_compaction_ended_;
Status::Code code_;
Status::SubCode subcode_;
};
TEST_F(DBTest2, CancelManualCompactionWithListener) {
CompactRangeOptions compact_options;
auto canceledPtr =
std::unique_ptr<std::atomic<bool>>(new std::atomic<bool>{true});
compact_options.canceled = canceledPtr.get();
compact_options.max_subcompactions = 1;
Options options = CurrentOptions();
options.disable_auto_compactions = true;
CancelCompactionListener* listener = new CancelCompactionListener();
options.listeners.emplace_back(listener);
DestroyAndReopen(options);
Random rnd(301);
for (int i = 0; i < 10; i++) {
for (int j = 0; j < 10; j++) {
ASSERT_OK(Put(Key(i + j * 10), rnd.RandomString(50)));
}
Flush();
}
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"CompactionIterator:ProcessKV", [&](void* /*arg*/) {
compact_options.canceled->store(true, std::memory_order_release);
});
int running_compaction = 0;
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"CompactionJob::FinishCompactionOutputFile1",
[&](void* /*arg*/) { running_compaction++; });
// Case I: 1 Notify begin compaction, 2 DisableManualCompaction, 3 Compaction
// not run, 4 Notify compaction end.
listener->code_ = Status::kIncomplete;
listener->subcode_ = Status::SubCode::kManualCompactionPaused;
compact_options.canceled->store(false, std::memory_order_release);
dbfull()->CompactRange(compact_options, nullptr, nullptr);
dbfull()->TEST_WaitForCompact(true);
ASSERT_GT(listener->num_compaction_started_, 0);
ASSERT_EQ(listener->num_compaction_started_, listener->num_compaction_ended_);
ASSERT_EQ(running_compaction, 0);
listener->num_compaction_started_ = 0;
listener->num_compaction_ended_ = 0;
// Case II: 1 DisableManualCompaction, 2 Notify begin compaction (return
// without notifying), 3 Notify compaction end (return without notifying).
dbfull()->CompactRange(compact_options, nullptr, nullptr);
dbfull()->TEST_WaitForCompact(true);
ASSERT_EQ(listener->num_compaction_started_, 0);
ASSERT_EQ(listener->num_compaction_started_, listener->num_compaction_ended_);
ASSERT_EQ(running_compaction, 0);
// Case III: 1 Notify begin compaction, 2 Compaction in between
// 3. DisableManualCompaction, , 4 Notify compaction end.
// compact_options.canceled->store(false, std::memory_order_release);
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearCallBack(
"CompactionIterator:ProcessKV");
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"CompactionJob::Run:BeforeVerify", [&](void* /*arg*/) {
compact_options.canceled->store(true, std::memory_order_release);
});
listener->code_ = Status::kOk;
listener->subcode_ = Status::SubCode::kNone;
compact_options.canceled->store(false, std::memory_order_release);
dbfull()->CompactRange(compact_options, nullptr, nullptr);
dbfull()->TEST_WaitForCompact(true);
ASSERT_GT(listener->num_compaction_started_, 0);
ASSERT_EQ(listener->num_compaction_started_, listener->num_compaction_ended_);
// Compaction job will succeed.
ASSERT_GT(running_compaction, 0);
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
}
TEST_F(DBTest2, CompactionOnBottomPriorityWithListener) {
int num_levels = 3;
const int kNumFilesTrigger = 4;
Options options = CurrentOptions();
env_->SetBackgroundThreads(0, Env::Priority::HIGH);
env_->SetBackgroundThreads(0, Env::Priority::LOW);
env_->SetBackgroundThreads(1, Env::Priority::BOTTOM);
options.env = env_;
options.compaction_style = kCompactionStyleUniversal;
options.num_levels = num_levels;
options.write_buffer_size = 100 << 10; // 100KB
options.target_file_size_base = 32 << 10; // 32KB
options.level0_file_num_compaction_trigger = kNumFilesTrigger;
// Trigger compaction if size amplification exceeds 110%
options.compaction_options_universal.max_size_amplification_percent = 110;
CancelCompactionListener* listener = new CancelCompactionListener();
options.listeners.emplace_back(listener);
DestroyAndReopen(options);
int num_bottom_thread_compaction_scheduled = 0;
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"DBImpl::BackgroundCompaction:ForwardToBottomPriPool",
[&](void* /*arg*/) { num_bottom_thread_compaction_scheduled++; });
int num_compaction_jobs = 0;
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"CompactionJob::Run():End",
[&](void* /*arg*/) { num_compaction_jobs++; });
listener->code_ = Status::kOk;
listener->subcode_ = Status::SubCode::kNone;
Random rnd(301);
for (int i = 0; i < 1; ++i) {
for (int num = 0; num < kNumFilesTrigger; num++) {
int key_idx = 0;
GenerateNewFile(&rnd, &key_idx, true /* no_wait */);
// use no_wait above because that one waits for flush and compaction. We
// don't want to wait for compaction because the full compaction is
// intentionally blocked while more files are flushed.
ASSERT_OK(dbfull()->TEST_WaitForFlushMemTable());
}
}
ASSERT_OK(dbfull()->TEST_WaitForCompact());
ASSERT_GT(num_bottom_thread_compaction_scheduled, 0);
ASSERT_EQ(num_compaction_jobs, 1);
ASSERT_GT(listener->num_compaction_started_, 0);
ASSERT_EQ(listener->num_compaction_started_, listener->num_compaction_ended_);
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
}
TEST_F(DBTest2, OptimizeForPointLookup) {
Options options = CurrentOptions();
Close();
options.OptimizeForPointLookup(2);
ASSERT_OK(DB::Open(options, dbname_, &db_));
ASSERT_OK(Put("foo", "v1"));
ASSERT_EQ("v1", Get("foo"));
Flush();
ASSERT_EQ("v1", Get("foo"));
}
TEST_F(DBTest2, OptimizeForSmallDB) {
Options options = CurrentOptions();
Close();
options.OptimizeForSmallDb();
// Find the cache object
ASSERT_TRUE(options.table_factory->IsInstanceOf(
TableFactory::kBlockBasedTableName()));
auto table_options =
options.table_factory->GetOptions<BlockBasedTableOptions>();
ASSERT_TRUE(table_options != nullptr);
std::shared_ptr<Cache> cache = table_options->block_cache;
ASSERT_EQ(0, cache->GetUsage());
ASSERT_OK(DB::Open(options, dbname_, &db_));
ASSERT_OK(Put("foo", "v1"));
// memtable size is costed to the block cache
ASSERT_NE(0, cache->GetUsage());
ASSERT_EQ("v1", Get("foo"));
Flush();
size_t prev_size = cache->GetUsage();
// Remember block cache size, so that we can find that
// it is filled after Get().
// Use pinnable slice so that it can ping the block so that
// when we check the size it is not evicted.
PinnableSlice value;
ASSERT_OK(db_->Get(ReadOptions(), db_->DefaultColumnFamily(), "foo", &value));
ASSERT_GT(cache->GetUsage(), prev_size);
value.Reset();
}
#endif // ROCKSDB_LITE
TEST_F(DBTest2, IterRaceFlush1) {
ASSERT_OK(Put("foo", "v1"));
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
{{"DBImpl::NewIterator:1", "DBTest2::IterRaceFlush:1"},
{"DBTest2::IterRaceFlush:2", "DBImpl::NewIterator:2"}});
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
ROCKSDB_NAMESPACE::port::Thread t1([&] {
TEST_SYNC_POINT("DBTest2::IterRaceFlush:1");
ASSERT_OK(Put("foo", "v2"));
Flush();
TEST_SYNC_POINT("DBTest2::IterRaceFlush:2");
});
// iterator is created after the first Put(), so it should see either
// "v1" or "v2".
{
std::unique_ptr<Iterator> it(db_->NewIterator(ReadOptions()));
it->Seek("foo");
ASSERT_TRUE(it->Valid());
ASSERT_EQ("foo", it->key().ToString());
}
t1.join();
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
}
TEST_F(DBTest2, IterRaceFlush2) {
ASSERT_OK(Put("foo", "v1"));
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
{{"DBImpl::NewIterator:3", "DBTest2::IterRaceFlush2:1"},
{"DBTest2::IterRaceFlush2:2", "DBImpl::NewIterator:4"}});
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
ROCKSDB_NAMESPACE::port::Thread t1([&] {
TEST_SYNC_POINT("DBTest2::IterRaceFlush2:1");
ASSERT_OK(Put("foo", "v2"));
Flush();
TEST_SYNC_POINT("DBTest2::IterRaceFlush2:2");
});
// iterator is created after the first Put(), so it should see either
// "v1" or "v2".
{
std::unique_ptr<Iterator> it(db_->NewIterator(ReadOptions()));
it->Seek("foo");
ASSERT_TRUE(it->Valid());
ASSERT_EQ("foo", it->key().ToString());
}
t1.join();
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
}
TEST_F(DBTest2, IterRefreshRaceFlush) {
ASSERT_OK(Put("foo", "v1"));
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
{{"ArenaWrappedDBIter::Refresh:1", "DBTest2::IterRefreshRaceFlush:1"},
{"DBTest2::IterRefreshRaceFlush:2", "ArenaWrappedDBIter::Refresh:2"}});
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
ROCKSDB_NAMESPACE::port::Thread t1([&] {
TEST_SYNC_POINT("DBTest2::IterRefreshRaceFlush:1");
ASSERT_OK(Put("foo", "v2"));
Flush();
TEST_SYNC_POINT("DBTest2::IterRefreshRaceFlush:2");
});
// iterator is created after the first Put(), so it should see either
// "v1" or "v2".
{
std::unique_ptr<Iterator> it(db_->NewIterator(ReadOptions()));
it->Refresh();
it->Seek("foo");
ASSERT_TRUE(it->Valid());
ASSERT_EQ("foo", it->key().ToString());
}
t1.join();
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
}
TEST_F(DBTest2, GetRaceFlush1) {
ASSERT_OK(Put("foo", "v1"));
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
{{"DBImpl::GetImpl:1", "DBTest2::GetRaceFlush:1"},
{"DBTest2::GetRaceFlush:2", "DBImpl::GetImpl:2"}});
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
ROCKSDB_NAMESPACE::port::Thread t1([&] {
TEST_SYNC_POINT("DBTest2::GetRaceFlush:1");
ASSERT_OK(Put("foo", "v2"));
Flush();
TEST_SYNC_POINT("DBTest2::GetRaceFlush:2");
});
// Get() is issued after the first Put(), so it should see either
// "v1" or "v2".
ASSERT_NE("NOT_FOUND", Get("foo"));
t1.join();
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
}
TEST_F(DBTest2, GetRaceFlush2) {
ASSERT_OK(Put("foo", "v1"));
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
{{"DBImpl::GetImpl:3", "DBTest2::GetRaceFlush:1"},
{"DBTest2::GetRaceFlush:2", "DBImpl::GetImpl:4"}});
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
port::Thread t1([&] {
TEST_SYNC_POINT("DBTest2::GetRaceFlush:1");
ASSERT_OK(Put("foo", "v2"));
Flush();
TEST_SYNC_POINT("DBTest2::GetRaceFlush:2");
});
// Get() is issued after the first Put(), so it should see either
// "v1" or "v2".
ASSERT_NE("NOT_FOUND", Get("foo"));
t1.join();
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
}
TEST_F(DBTest2, DirectIO) {
if (!IsDirectIOSupported()) {
return;
}
Options options = CurrentOptions();
options.use_direct_reads = options.use_direct_io_for_flush_and_compaction =
true;
options.allow_mmap_reads = options.allow_mmap_writes = false;
DestroyAndReopen(options);
ASSERT_OK(Put(Key(0), "a"));
ASSERT_OK(Put(Key(5), "a"));
ASSERT_OK(Flush());
ASSERT_OK(Put(Key(10), "a"));
ASSERT_OK(Put(Key(15), "a"));
ASSERT_OK(Flush());
ASSERT_OK(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
Reopen(options);
}
TEST_F(DBTest2, MemtableOnlyIterator) {
Options options = CurrentOptions();
CreateAndReopenWithCF({"pikachu"}, options);
ASSERT_OK(Put(1, "foo", "first"));
ASSERT_OK(Put(1, "bar", "second"));
ReadOptions ropt;
ropt.read_tier = kMemtableTier;
std::string value;
Iterator* it = nullptr;
// Before flushing
// point lookups
ASSERT_OK(db_->Get(ropt, handles_[1], "foo", &value));
ASSERT_EQ("first", value);
ASSERT_OK(db_->Get(ropt, handles_[1], "bar", &value));
ASSERT_EQ("second", value);
// Memtable-only iterator (read_tier=kMemtableTier); data not flushed yet.
it = db_->NewIterator(ropt, handles_[1]);
int count = 0;
for (it->SeekToFirst(); it->Valid(); it->Next()) {
ASSERT_TRUE(it->Valid());
count++;
}
ASSERT_TRUE(!it->Valid());
ASSERT_EQ(2, count);
delete it;
Flush(1);
// After flushing
// point lookups
ASSERT_OK(db_->Get(ropt, handles_[1], "foo", &value));
ASSERT_EQ("first", value);
ASSERT_OK(db_->Get(ropt, handles_[1], "bar", &value));
ASSERT_EQ("second", value);
// nothing should be returned using memtable-only iterator after flushing.
it = db_->NewIterator(ropt, handles_[1]);
count = 0;
for (it->SeekToFirst(); it->Valid(); it->Next()) {
ASSERT_TRUE(it->Valid());
count++;
}
ASSERT_TRUE(!it->Valid());
ASSERT_EQ(0, count);
delete it;
// Add a key to memtable
ASSERT_OK(Put(1, "foobar", "third"));
it = db_->NewIterator(ropt, handles_[1]);
count = 0;
for (it->SeekToFirst(); it->Valid(); it->Next()) {
ASSERT_TRUE(it->Valid());
ASSERT_EQ("foobar", it->key().ToString());
ASSERT_EQ("third", it->value().ToString());
count++;
}
ASSERT_TRUE(!it->Valid());
ASSERT_EQ(1, count);
delete it;
}
TEST_F(DBTest2, LowPriWrite) {
Options options = CurrentOptions();
// Compaction pressure should trigger since 6 files
options.level0_file_num_compaction_trigger = 4;
options.level0_slowdown_writes_trigger = 12;
options.level0_stop_writes_trigger = 30;
options.delayed_write_rate = 8 * 1024 * 1024;
Reopen(options);
std::atomic<int> rate_limit_count(0);
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"GenericRateLimiter::Request:1", [&](void* arg) {
rate_limit_count.fetch_add(1);
int64_t* rate_bytes_per_sec = static_cast<int64_t*>(arg);
ASSERT_EQ(1024 * 1024, *rate_bytes_per_sec);
});
// Block compaction
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency({
{"DBTest.LowPriWrite:0", "DBImpl::BGWorkCompaction"},
});
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
WriteOptions wo;
for (int i = 0; i < 6; i++) {
wo.low_pri = false;
Put("", "", wo);
wo.low_pri = true;
Put("", "", wo);
Flush();
}
ASSERT_EQ(0, rate_limit_count.load());
wo.low_pri = true;
Put("", "", wo);
ASSERT_EQ(1, rate_limit_count.load());
wo.low_pri = false;
Put("", "", wo);
ASSERT_EQ(1, rate_limit_count.load());
TEST_SYNC_POINT("DBTest.LowPriWrite:0");
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
dbfull()->TEST_WaitForCompact();
wo.low_pri = true;
Put("", "", wo);
ASSERT_EQ(1, rate_limit_count.load());
wo.low_pri = false;
Put("", "", wo);
ASSERT_EQ(1, rate_limit_count.load());
}
#ifndef ROCKSDB_LITE
TEST_F(DBTest2, RateLimitedCompactionReads) {
// compaction input has 512KB data
const int kNumKeysPerFile = 128;
const int kBytesPerKey = 1024;
const int kNumL0Files = 4;
for (auto use_direct_io : {false, true}) {
if (use_direct_io && !IsDirectIOSupported()) {
continue;
}
Options options = CurrentOptions();
options.compression = kNoCompression;
options.level0_file_num_compaction_trigger = kNumL0Files;
options.memtable_factory.reset(new SpecialSkipListFactory(kNumKeysPerFile));
options.new_table_reader_for_compaction_inputs = true;
// takes roughly one second, split into 100 x 10ms intervals. Each interval
// permits 5.12KB, which is smaller than the block size, so this test
// exercises the code for chunking reads.
options.rate_limiter.reset(NewGenericRateLimiter(
static_cast<int64_t>(kNumL0Files * kNumKeysPerFile *
kBytesPerKey) /* rate_bytes_per_sec */,
10 * 1000 /* refill_period_us */, 10 /* fairness */,
RateLimiter::Mode::kReadsOnly));
options.use_direct_reads = options.use_direct_io_for_flush_and_compaction =
use_direct_io;
BlockBasedTableOptions bbto;
bbto.block_size = 16384;
bbto.no_block_cache = true;
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566) Summary: This PR does a few things: 1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation. 2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed: - The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated - The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory). 3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10). I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged. Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently. Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566 Reviewed By: zhichao-cao Differential Revision: D24408980 Pulled By: jay-zhuang fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
2020-10-27 18:31:34 +01:00
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
DestroyAndReopen(options);
for (int i = 0; i < kNumL0Files; ++i) {
for (int j = 0; j <= kNumKeysPerFile; ++j) {
ASSERT_OK(Put(Key(j), DummyString(kBytesPerKey)));
}
dbfull()->TEST_WaitForFlushMemTable();
ASSERT_EQ(i + 1, NumTableFilesAtLevel(0));
}
dbfull()->TEST_WaitForCompact();
ASSERT_EQ(0, NumTableFilesAtLevel(0));
ASSERT_EQ(0, options.rate_limiter->GetTotalBytesThrough(Env::IO_HIGH));
// should be slightly above 512KB due to non-data blocks read. Arbitrarily
// chose 1MB as the upper bound on the total bytes read.
size_t rate_limited_bytes =
options.rate_limiter->GetTotalBytesThrough(Env::IO_LOW);
// Include the explicit prefetch of the footer in direct I/O case.
size_t direct_io_extra = use_direct_io ? 512 * 1024 : 0;
ASSERT_GE(
rate_limited_bytes,
static_cast<size_t>(kNumKeysPerFile * kBytesPerKey * kNumL0Files));
ASSERT_LT(
rate_limited_bytes,
static_cast<size_t>(2 * kNumKeysPerFile * kBytesPerKey * kNumL0Files +
direct_io_extra));
Iterator* iter = db_->NewIterator(ReadOptions());
for (iter->SeekToFirst(); iter->Valid(); iter->Next()) {
ASSERT_EQ(iter->value().ToString(), DummyString(kBytesPerKey));
}
delete iter;
// bytes read for user iterator shouldn't count against the rate limit.
ASSERT_EQ(rate_limited_bytes,
static_cast<size_t>(
options.rate_limiter->GetTotalBytesThrough(Env::IO_LOW)));
}
}
#endif // ROCKSDB_LITE
// Make sure DB can be reopen with reduced number of levels, given no file
// is on levels higher than the new num_levels.
TEST_F(DBTest2, ReduceLevel) {
Options options;
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566) Summary: This PR does a few things: 1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation. 2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed: - The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated - The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory). 3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10). I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged. Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently. Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566 Reviewed By: zhichao-cao Differential Revision: D24408980 Pulled By: jay-zhuang fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
2020-10-27 18:31:34 +01:00
options.env = env_;
options.disable_auto_compactions = true;
options.num_levels = 7;
Reopen(options);
Put("foo", "bar");
Flush();
MoveFilesToLevel(6);
#ifndef ROCKSDB_LITE
ASSERT_EQ("0,0,0,0,0,0,1", FilesPerLevel());
#endif // !ROCKSDB_LITE
CompactRangeOptions compact_options;
compact_options.change_level = true;
compact_options.target_level = 1;
dbfull()->CompactRange(compact_options, nullptr, nullptr);
#ifndef ROCKSDB_LITE
ASSERT_EQ("0,1", FilesPerLevel());
#endif // !ROCKSDB_LITE
options.num_levels = 3;
Reopen(options);
#ifndef ROCKSDB_LITE
ASSERT_EQ("0,1", FilesPerLevel());
#endif // !ROCKSDB_LITE
}
// Test that ReadCallback is actually used in both memtbale and sst tables
TEST_F(DBTest2, ReadCallbackTest) {
Options options;
options.disable_auto_compactions = true;
options.num_levels = 7;
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566) Summary: This PR does a few things: 1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation. 2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed: - The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated - The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory). 3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10). I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged. Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently. Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566 Reviewed By: zhichao-cao Differential Revision: D24408980 Pulled By: jay-zhuang fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
2020-10-27 18:31:34 +01:00
options.env = env_;
Reopen(options);
std::vector<const Snapshot*> snapshots;
// Try to create a db with multiple layers and a memtable
const std::string key = "foo";
const std::string value = "bar";
// This test assumes that the seq start with 1 and increased by 1 after each
// write batch of size 1. If that behavior changes, the test needs to be
// updated as well.
// TODO(myabandeh): update this test to use the seq number that is returned by
// the DB instead of assuming what seq the DB used.
int i = 1;
for (; i < 10; i++) {
Put(key, value + std::to_string(i));
// Take a snapshot to avoid the value being removed during compaction
auto snapshot = dbfull()->GetSnapshot();
snapshots.push_back(snapshot);
}
Flush();
for (; i < 20; i++) {
Put(key, value + std::to_string(i));
// Take a snapshot to avoid the value being removed during compaction
auto snapshot = dbfull()->GetSnapshot();
snapshots.push_back(snapshot);
}
Flush();
MoveFilesToLevel(6);
#ifndef ROCKSDB_LITE
ASSERT_EQ("0,0,0,0,0,0,2", FilesPerLevel());
#endif // !ROCKSDB_LITE
for (; i < 30; i++) {
Put(key, value + std::to_string(i));
auto snapshot = dbfull()->GetSnapshot();
snapshots.push_back(snapshot);
}
Flush();
#ifndef ROCKSDB_LITE
ASSERT_EQ("1,0,0,0,0,0,2", FilesPerLevel());
#endif // !ROCKSDB_LITE
// And also add some values to the memtable
for (; i < 40; i++) {
Put(key, value + std::to_string(i));
auto snapshot = dbfull()->GetSnapshot();
snapshots.push_back(snapshot);
}
class TestReadCallback : public ReadCallback {
public:
WriteUnPrepared: less virtual in iterator callback (#5049) Summary: WriteUnPrepared adds a virtual function, MaxUnpreparedSequenceNumber, to ReadCallback, which returns 0 unless WriteUnPrepared is enabled and the transaction has uncommitted data written to the DB. Together with snapshot sequence number, this determines the last sequence that is visible to reads. The patch clarifies the guarantees of the GetIterator API in WriteUnPrepared transactions and make use of that to statically initialize the read callback and thus avoid the virtual call. Furthermore it increases the minimum value for min_uncommitted from 0 to 1 as seq 0 is used only for last level keys that are committed in all snapshots. The following benchmark shows +0.26% higher throughput in seekrandom benchmark. Benchmark: ./db_bench --benchmarks=fillrandom --use_existing_db=0 --num=1000000 --db=/dev/shm/dbbench ./db_bench --benchmarks=seekrandom[X10] --use_existing_db=1 --db=/dev/shm/dbbench --num=1000000 --duration=60 --seek_nexts=100 seekrandom [AVG 10 runs] : 20355 ops/sec; 225.2 MB/sec seekrandom [MEDIAN 10 runs] : 20425 ops/sec; 225.9 MB/sec ./db_bench_lessvirtual3 --benchmarks=seekrandom[X10] --use_existing_db=1 --db=/dev/shm/dbbench --num=1000000 --duration=60 --seek_nexts=100 seekrandom [AVG 10 runs] : 20409 ops/sec; 225.8 MB/sec seekrandom [MEDIAN 10 runs] : 20487 ops/sec; 226.6 MB/sec Pull Request resolved: https://github.com/facebook/rocksdb/pull/5049 Differential Revision: D14366459 Pulled By: maysamyabandeh fbshipit-source-id: ebaff8908332a5ae9af7defeadabcb624be660ef
2019-04-02 23:43:03 +02:00
explicit TestReadCallback(SequenceNumber snapshot)
: ReadCallback(snapshot), snapshot_(snapshot) {}
bool IsVisibleFullCheck(SequenceNumber seq) override {
return seq <= snapshot_;
}
private:
SequenceNumber snapshot_;
};
for (int seq = 1; seq < i; seq++) {
PinnableSlice pinnable_val;
ReadOptions roptions;
TestReadCallback callback(seq);
bool dont_care = true;
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 = dbfull()->DefaultColumnFamily();
get_impl_options.value = &pinnable_val;
get_impl_options.value_found = &dont_care;
get_impl_options.callback = &callback;
Status s = dbfull()->GetImpl(roptions, key, get_impl_options);
ASSERT_TRUE(s.ok());
// Assuming that after each Put the DB increased seq by one, the value and
// seq number must be equal since we also inc value by 1 after each Put.
ASSERT_EQ(value + std::to_string(seq), pinnable_val.ToString());
}
for (auto snapshot : snapshots) {
dbfull()->ReleaseSnapshot(snapshot);
}
}
#ifndef ROCKSDB_LITE
TEST_F(DBTest2, LiveFilesOmitObsoleteFiles) {
// Regression test for race condition where an obsolete file is returned to
// user as a "live file" but then deleted, all while file deletions are
// disabled.
//
// It happened like this:
//
// 1. [flush thread] Log file "x.log" found by FindObsoleteFiles
// 2. [user thread] DisableFileDeletions, GetSortedWalFiles are called and the
// latter returned "x.log"
// 3. [flush thread] PurgeObsoleteFiles deleted "x.log"
// 4. [user thread] Reading "x.log" failed
//
// Unfortunately the only regression test I can come up with involves sleep.
// We cannot set SyncPoints to repro since, once the fix is applied, the
// SyncPoints would cause a deadlock as the repro's sequence of events is now
// prohibited.
//
// Instead, if we sleep for a second between Find and Purge, and ensure the
// read attempt happens after purge, then the sequence of events will almost
// certainly happen on the old code.
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency({
{"DBImpl::BackgroundCallFlush:FilesFound",
"DBTest2::LiveFilesOmitObsoleteFiles:FlushTriggered"},
{"DBImpl::PurgeObsoleteFiles:End",
"DBTest2::LiveFilesOmitObsoleteFiles:LiveFilesCaptured"},
});
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"DBImpl::PurgeObsoleteFiles:Begin",
[&](void* /*arg*/) { env_->SleepForMicroseconds(1000000); });
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
Put("key", "val");
FlushOptions flush_opts;
flush_opts.wait = false;
db_->Flush(flush_opts);
TEST_SYNC_POINT("DBTest2::LiveFilesOmitObsoleteFiles:FlushTriggered");
db_->DisableFileDeletions();
VectorLogPtr log_files;
db_->GetSortedWalFiles(log_files);
TEST_SYNC_POINT("DBTest2::LiveFilesOmitObsoleteFiles:LiveFilesCaptured");
for (const auto& log_file : log_files) {
ASSERT_OK(env_->FileExists(LogFileName(dbname_, log_file->LogNumber())));
}
db_->EnableFileDeletions();
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
}
TEST_F(DBTest2, TestNumPread) {
Options options = CurrentOptions();
Create a CustomEnv class; Add WinFileSystem; Make LegacyFileSystemWrapper private (#7703) Summary: This PR does the following: -> Creates a WinFileSystem class. This class is the Windows equivalent of the PosixFileSystem and will be used on Windows systems. -> Introduces a CustomEnv class. A CustomEnv is an Env that takes a FileSystem as constructor argument. I believe there will only ever be two implementations of this class (PosixEnv and WinEnv). There is still a CustomEnvWrapper class that takes an Env and a FileSystem and wraps the Env calls with the input Env but uses the FileSystem for the FileSystem calls -> Eliminates the public uses of the LegacyFileSystemWrapper. With this change in place, there are effectively the following patterns of Env: - "Base Env classes" (PosixEnv, WinEnv). These classes implement the core Env functions (e.g. Threads) and have a hard-coded input FileSystem. These classes inherit from CompositeEnv, implement the core Env functions (threads) and delegate the FileSystem-like calls to the input file system. - Wrapped Composite Env classes (MemEnv). These classes take in an Env and a FileSystem. The core env functions are re-directed to the wrapped env. The file system calls are redirected to the input file system - Legacy Wrapped Env classes. These classes take in an Env input (but no FileSystem). The core env functions are re-directed to the wrapped env. A "Legacy File System" is created using this env and the file system calls directed to the env itself. With these changes in place, the PosixEnv becomes a singleton -- there is only ever one created. Any other use of the PosixEnv is via another wrapped env. This cleans up some of the issues with the env construction and destruction. Additionally, there were places in the code that required had an Env when they required a FileSystem. Many of these places would wrap the Env with a LegacyFileSystemWrapper instead of using the env->GetFileSystem(). These places were changed, thereby removing layers of additional redirection (LegacyFileSystem --> Env --> Env::FileSystem). Pull Request resolved: https://github.com/facebook/rocksdb/pull/7703 Reviewed By: zhichao-cao Differential Revision: D25762190 Pulled By: anand1976 fbshipit-source-id: 1a088e97fc916f28ac69c149cd1dcad0ab31704b
2021-01-06 19:48:24 +01:00
bool prefetch_supported =
test::IsPrefetchSupported(env_->GetFileSystem(), dbname_);
// disable block cache
BlockBasedTableOptions table_options;
table_options.no_block_cache = true;
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
Reopen(options);
env_->count_random_reads_ = true;
env_->random_file_open_counter_.store(0);
ASSERT_OK(Put("bar", "foo"));
ASSERT_OK(Put("foo", "bar"));
ASSERT_OK(Flush());
Create a CustomEnv class; Add WinFileSystem; Make LegacyFileSystemWrapper private (#7703) Summary: This PR does the following: -> Creates a WinFileSystem class. This class is the Windows equivalent of the PosixFileSystem and will be used on Windows systems. -> Introduces a CustomEnv class. A CustomEnv is an Env that takes a FileSystem as constructor argument. I believe there will only ever be two implementations of this class (PosixEnv and WinEnv). There is still a CustomEnvWrapper class that takes an Env and a FileSystem and wraps the Env calls with the input Env but uses the FileSystem for the FileSystem calls -> Eliminates the public uses of the LegacyFileSystemWrapper. With this change in place, there are effectively the following patterns of Env: - "Base Env classes" (PosixEnv, WinEnv). These classes implement the core Env functions (e.g. Threads) and have a hard-coded input FileSystem. These classes inherit from CompositeEnv, implement the core Env functions (threads) and delegate the FileSystem-like calls to the input file system. - Wrapped Composite Env classes (MemEnv). These classes take in an Env and a FileSystem. The core env functions are re-directed to the wrapped env. The file system calls are redirected to the input file system - Legacy Wrapped Env classes. These classes take in an Env input (but no FileSystem). The core env functions are re-directed to the wrapped env. A "Legacy File System" is created using this env and the file system calls directed to the env itself. With these changes in place, the PosixEnv becomes a singleton -- there is only ever one created. Any other use of the PosixEnv is via another wrapped env. This cleans up some of the issues with the env construction and destruction. Additionally, there were places in the code that required had an Env when they required a FileSystem. Many of these places would wrap the Env with a LegacyFileSystemWrapper instead of using the env->GetFileSystem(). These places were changed, thereby removing layers of additional redirection (LegacyFileSystem --> Env --> Env::FileSystem). Pull Request resolved: https://github.com/facebook/rocksdb/pull/7703 Reviewed By: zhichao-cao Differential Revision: D25762190 Pulled By: anand1976 fbshipit-source-id: 1a088e97fc916f28ac69c149cd1dcad0ab31704b
2021-01-06 19:48:24 +01:00
if (prefetch_supported) {
// After flush, we'll open the file and read footer, meta block,
// property block and index block.
ASSERT_EQ(4, env_->random_read_counter_.Read());
} else {
// With prefetch not supported, we will do a single read into a buffer
ASSERT_EQ(1, env_->random_read_counter_.Read());
}
ASSERT_EQ(1, env_->random_file_open_counter_.load());
// One pread per a normal data block read
env_->random_file_open_counter_.store(0);
env_->random_read_counter_.Reset();
ASSERT_EQ("bar", Get("foo"));
ASSERT_EQ(1, env_->random_read_counter_.Read());
// All files are already opened.
ASSERT_EQ(0, env_->random_file_open_counter_.load());
env_->random_file_open_counter_.store(0);
env_->random_read_counter_.Reset();
ASSERT_OK(Put("bar2", "foo2"));
ASSERT_OK(Put("foo2", "bar2"));
ASSERT_OK(Flush());
Create a CustomEnv class; Add WinFileSystem; Make LegacyFileSystemWrapper private (#7703) Summary: This PR does the following: -> Creates a WinFileSystem class. This class is the Windows equivalent of the PosixFileSystem and will be used on Windows systems. -> Introduces a CustomEnv class. A CustomEnv is an Env that takes a FileSystem as constructor argument. I believe there will only ever be two implementations of this class (PosixEnv and WinEnv). There is still a CustomEnvWrapper class that takes an Env and a FileSystem and wraps the Env calls with the input Env but uses the FileSystem for the FileSystem calls -> Eliminates the public uses of the LegacyFileSystemWrapper. With this change in place, there are effectively the following patterns of Env: - "Base Env classes" (PosixEnv, WinEnv). These classes implement the core Env functions (e.g. Threads) and have a hard-coded input FileSystem. These classes inherit from CompositeEnv, implement the core Env functions (threads) and delegate the FileSystem-like calls to the input file system. - Wrapped Composite Env classes (MemEnv). These classes take in an Env and a FileSystem. The core env functions are re-directed to the wrapped env. The file system calls are redirected to the input file system - Legacy Wrapped Env classes. These classes take in an Env input (but no FileSystem). The core env functions are re-directed to the wrapped env. A "Legacy File System" is created using this env and the file system calls directed to the env itself. With these changes in place, the PosixEnv becomes a singleton -- there is only ever one created. Any other use of the PosixEnv is via another wrapped env. This cleans up some of the issues with the env construction and destruction. Additionally, there were places in the code that required had an Env when they required a FileSystem. Many of these places would wrap the Env with a LegacyFileSystemWrapper instead of using the env->GetFileSystem(). These places were changed, thereby removing layers of additional redirection (LegacyFileSystem --> Env --> Env::FileSystem). Pull Request resolved: https://github.com/facebook/rocksdb/pull/7703 Reviewed By: zhichao-cao Differential Revision: D25762190 Pulled By: anand1976 fbshipit-source-id: 1a088e97fc916f28ac69c149cd1dcad0ab31704b
2021-01-06 19:48:24 +01:00
if (prefetch_supported) {
// After flush, we'll open the file and read footer, meta block,
// property block and index block.
ASSERT_EQ(4, env_->random_read_counter_.Read());
} else {
// With prefetch not supported, we will do a single read into a buffer
ASSERT_EQ(1, env_->random_read_counter_.Read());
}
ASSERT_EQ(1, env_->random_file_open_counter_.load());
env_->random_file_open_counter_.store(0);
env_->random_read_counter_.Reset();
ASSERT_OK(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
Create a CustomEnv class; Add WinFileSystem; Make LegacyFileSystemWrapper private (#7703) Summary: This PR does the following: -> Creates a WinFileSystem class. This class is the Windows equivalent of the PosixFileSystem and will be used on Windows systems. -> Introduces a CustomEnv class. A CustomEnv is an Env that takes a FileSystem as constructor argument. I believe there will only ever be two implementations of this class (PosixEnv and WinEnv). There is still a CustomEnvWrapper class that takes an Env and a FileSystem and wraps the Env calls with the input Env but uses the FileSystem for the FileSystem calls -> Eliminates the public uses of the LegacyFileSystemWrapper. With this change in place, there are effectively the following patterns of Env: - "Base Env classes" (PosixEnv, WinEnv). These classes implement the core Env functions (e.g. Threads) and have a hard-coded input FileSystem. These classes inherit from CompositeEnv, implement the core Env functions (threads) and delegate the FileSystem-like calls to the input file system. - Wrapped Composite Env classes (MemEnv). These classes take in an Env and a FileSystem. The core env functions are re-directed to the wrapped env. The file system calls are redirected to the input file system - Legacy Wrapped Env classes. These classes take in an Env input (but no FileSystem). The core env functions are re-directed to the wrapped env. A "Legacy File System" is created using this env and the file system calls directed to the env itself. With these changes in place, the PosixEnv becomes a singleton -- there is only ever one created. Any other use of the PosixEnv is via another wrapped env. This cleans up some of the issues with the env construction and destruction. Additionally, there were places in the code that required had an Env when they required a FileSystem. Many of these places would wrap the Env with a LegacyFileSystemWrapper instead of using the env->GetFileSystem(). These places were changed, thereby removing layers of additional redirection (LegacyFileSystem --> Env --> Env::FileSystem). Pull Request resolved: https://github.com/facebook/rocksdb/pull/7703 Reviewed By: zhichao-cao Differential Revision: D25762190 Pulled By: anand1976 fbshipit-source-id: 1a088e97fc916f28ac69c149cd1dcad0ab31704b
2021-01-06 19:48:24 +01:00
if (prefetch_supported) {
// Compaction needs two input blocks, which requires 2 preads, and
// generate a new SST file which needs 4 preads (footer, meta block,
// property block and index block). In total 6.
ASSERT_EQ(6, env_->random_read_counter_.Read());
} else {
// With prefetch off, compaction needs two input blocks,
// followed by a single buffered read. In total 3.
ASSERT_EQ(3, env_->random_read_counter_.Read());
}
// All compaction input files should have already been opened.
ASSERT_EQ(1, env_->random_file_open_counter_.load());
// One pread per a normal data block read
env_->random_file_open_counter_.store(0);
env_->random_read_counter_.Reset();
ASSERT_EQ("foo2", Get("bar2"));
ASSERT_EQ(1, env_->random_read_counter_.Read());
// SST files are already opened.
ASSERT_EQ(0, env_->random_file_open_counter_.load());
}
TEST_F(DBTest2, TraceAndReplay) {
Options options = CurrentOptions();
options.merge_operator = MergeOperators::CreatePutOperator();
ReadOptions ro;
WriteOptions wo;
TraceOptions trace_opts;
EnvOptions env_opts;
CreateAndReopenWithCF({"pikachu"}, options);
Random rnd(301);
Iterator* single_iter = nullptr;
ASSERT_TRUE(db_->EndTrace().IsIOError());
std::string trace_filename = dbname_ + "/rocksdb.trace";
std::unique_ptr<TraceWriter> trace_writer;
ASSERT_OK(NewFileTraceWriter(env_, env_opts, trace_filename, &trace_writer));
ASSERT_OK(db_->StartTrace(trace_opts, std::move(trace_writer)));
ASSERT_OK(Put(0, "a", "1"));
ASSERT_OK(Merge(0, "b", "2"));
ASSERT_OK(Delete(0, "c"));
ASSERT_OK(SingleDelete(0, "d"));
ASSERT_OK(db_->DeleteRange(wo, dbfull()->DefaultColumnFamily(), "e", "f"));
WriteBatch batch;
ASSERT_OK(batch.Put("f", "11"));
ASSERT_OK(batch.Merge("g", "12"));
ASSERT_OK(batch.Delete("h"));
ASSERT_OK(batch.SingleDelete("i"));
ASSERT_OK(batch.DeleteRange("j", "k"));
ASSERT_OK(db_->Write(wo, &batch));
single_iter = db_->NewIterator(ro);
single_iter->Seek("f");
single_iter->SeekForPrev("g");
delete single_iter;
ASSERT_EQ("1", Get(0, "a"));
ASSERT_EQ("12", Get(0, "g"));
ASSERT_OK(Put(1, "foo", "bar"));
ASSERT_OK(Put(1, "rocksdb", "rocks"));
ASSERT_EQ("NOT_FOUND", Get(1, "leveldb"));
ASSERT_OK(db_->EndTrace());
// These should not get into the trace file as it is after EndTrace.
Put("hello", "world");
Merge("foo", "bar");
// Open another db, replay, and verify the data
std::string value;
std::string dbname2 = test::PerThreadDBPath(env_, "/db_replay");
ASSERT_OK(DestroyDB(dbname2, options));
// Using a different name than db2, to pacify infer's use-after-lifetime
// warnings (http://fbinfer.com).
DB* db2_init = nullptr;
options.create_if_missing = true;
ASSERT_OK(DB::Open(options, dbname2, &db2_init));
ColumnFamilyHandle* cf;
ASSERT_OK(
db2_init->CreateColumnFamily(ColumnFamilyOptions(), "pikachu", &cf));
delete cf;
delete db2_init;
DB* db2 = nullptr;
std::vector<ColumnFamilyDescriptor> column_families;
ColumnFamilyOptions cf_options;
cf_options.merge_operator = MergeOperators::CreatePutOperator();
column_families.push_back(ColumnFamilyDescriptor("default", cf_options));
column_families.push_back(
ColumnFamilyDescriptor("pikachu", ColumnFamilyOptions()));
std::vector<ColumnFamilyHandle*> handles;
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566) Summary: This PR does a few things: 1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation. 2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed: - The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated - The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory). 3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10). I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged. Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently. Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566 Reviewed By: zhichao-cao Differential Revision: D24408980 Pulled By: jay-zhuang fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
2020-10-27 18:31:34 +01:00
DBOptions db_opts;
db_opts.env = env_;
ASSERT_OK(DB::Open(db_opts, dbname2, column_families, &handles, &db2));
env_->SleepForMicroseconds(100);
// Verify that the keys don't already exist
ASSERT_TRUE(db2->Get(ro, handles[0], "a", &value).IsNotFound());
ASSERT_TRUE(db2->Get(ro, handles[0], "g", &value).IsNotFound());
std::unique_ptr<TraceReader> trace_reader;
ASSERT_OK(NewFileTraceReader(env_, env_opts, trace_filename, &trace_reader));
Replayer replayer(db2, handles_, std::move(trace_reader));
ASSERT_OK(replayer.Replay());
ASSERT_OK(db2->Get(ro, handles[0], "a", &value));
ASSERT_EQ("1", value);
ASSERT_OK(db2->Get(ro, handles[0], "g", &value));
ASSERT_EQ("12", value);
ASSERT_TRUE(db2->Get(ro, handles[0], "hello", &value).IsNotFound());
ASSERT_TRUE(db2->Get(ro, handles[0], "world", &value).IsNotFound());
ASSERT_OK(db2->Get(ro, handles[1], "foo", &value));
ASSERT_EQ("bar", value);
ASSERT_OK(db2->Get(ro, handles[1], "rocksdb", &value));
ASSERT_EQ("rocks", value);
for (auto handle : handles) {
delete handle;
}
delete db2;
ASSERT_OK(DestroyDB(dbname2, options));
}
TEST_F(DBTest2, TraceWithLimit) {
Options options = CurrentOptions();
options.merge_operator = MergeOperators::CreatePutOperator();
ReadOptions ro;
WriteOptions wo;
TraceOptions trace_opts;
EnvOptions env_opts;
CreateAndReopenWithCF({"pikachu"}, options);
Random rnd(301);
// test the max trace file size options
trace_opts.max_trace_file_size = 5;
std::string trace_filename = dbname_ + "/rocksdb.trace1";
std::unique_ptr<TraceWriter> trace_writer;
ASSERT_OK(NewFileTraceWriter(env_, env_opts, trace_filename, &trace_writer));
ASSERT_OK(db_->StartTrace(trace_opts, std::move(trace_writer)));
ASSERT_OK(Put(0, "a", "1"));
ASSERT_OK(Put(0, "b", "1"));
ASSERT_OK(Put(0, "c", "1"));
ASSERT_OK(db_->EndTrace());
std::string dbname2 = test::PerThreadDBPath(env_, "/db_replay2");
std::string value;
ASSERT_OK(DestroyDB(dbname2, options));
// Using a different name than db2, to pacify infer's use-after-lifetime
// warnings (http://fbinfer.com).
DB* db2_init = nullptr;
options.create_if_missing = true;
ASSERT_OK(DB::Open(options, dbname2, &db2_init));
ColumnFamilyHandle* cf;
ASSERT_OK(
db2_init->CreateColumnFamily(ColumnFamilyOptions(), "pikachu", &cf));
delete cf;
delete db2_init;
DB* db2 = nullptr;
std::vector<ColumnFamilyDescriptor> column_families;
ColumnFamilyOptions cf_options;
cf_options.merge_operator = MergeOperators::CreatePutOperator();
column_families.push_back(ColumnFamilyDescriptor("default", cf_options));
column_families.push_back(
ColumnFamilyDescriptor("pikachu", ColumnFamilyOptions()));
std::vector<ColumnFamilyHandle*> handles;
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566) Summary: This PR does a few things: 1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation. 2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed: - The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated - The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory). 3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10). I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged. Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently. Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566 Reviewed By: zhichao-cao Differential Revision: D24408980 Pulled By: jay-zhuang fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
2020-10-27 18:31:34 +01:00
DBOptions db_opts;
db_opts.env = env_;
ASSERT_OK(DB::Open(db_opts, dbname2, column_families, &handles, &db2));
env_->SleepForMicroseconds(100);
// Verify that the keys don't already exist
ASSERT_TRUE(db2->Get(ro, handles[0], "a", &value).IsNotFound());
ASSERT_TRUE(db2->Get(ro, handles[0], "b", &value).IsNotFound());
ASSERT_TRUE(db2->Get(ro, handles[0], "c", &value).IsNotFound());
std::unique_ptr<TraceReader> trace_reader;
ASSERT_OK(NewFileTraceReader(env_, env_opts, trace_filename, &trace_reader));
Replayer replayer(db2, handles_, std::move(trace_reader));
ASSERT_OK(replayer.Replay());
ASSERT_TRUE(db2->Get(ro, handles[0], "a", &value).IsNotFound());
ASSERT_TRUE(db2->Get(ro, handles[0], "b", &value).IsNotFound());
ASSERT_TRUE(db2->Get(ro, handles[0], "c", &value).IsNotFound());
for (auto handle : handles) {
delete handle;
}
delete db2;
ASSERT_OK(DestroyDB(dbname2, options));
}
TEST_F(DBTest2, TraceWithSampling) {
Options options = CurrentOptions();
ReadOptions ro;
WriteOptions wo;
TraceOptions trace_opts;
EnvOptions env_opts;
CreateAndReopenWithCF({"pikachu"}, options);
Random rnd(301);
// test the trace file sampling options
trace_opts.sampling_frequency = 2;
std::string trace_filename = dbname_ + "/rocksdb.trace_sampling";
std::unique_ptr<TraceWriter> trace_writer;
ASSERT_OK(NewFileTraceWriter(env_, env_opts, trace_filename, &trace_writer));
ASSERT_OK(db_->StartTrace(trace_opts, std::move(trace_writer)));
ASSERT_OK(Put(0, "a", "1"));
ASSERT_OK(Put(0, "b", "2"));
ASSERT_OK(Put(0, "c", "3"));
ASSERT_OK(Put(0, "d", "4"));
ASSERT_OK(Put(0, "e", "5"));
ASSERT_OK(db_->EndTrace());
std::string dbname2 = test::PerThreadDBPath(env_, "/db_replay_sampling");
std::string value;
ASSERT_OK(DestroyDB(dbname2, options));
// Using a different name than db2, to pacify infer's use-after-lifetime
// warnings (http://fbinfer.com).
DB* db2_init = nullptr;
options.create_if_missing = true;
ASSERT_OK(DB::Open(options, dbname2, &db2_init));
ColumnFamilyHandle* cf;
ASSERT_OK(
db2_init->CreateColumnFamily(ColumnFamilyOptions(), "pikachu", &cf));
delete cf;
delete db2_init;
DB* db2 = nullptr;
std::vector<ColumnFamilyDescriptor> column_families;
ColumnFamilyOptions cf_options;
column_families.push_back(ColumnFamilyDescriptor("default", cf_options));
column_families.push_back(
ColumnFamilyDescriptor("pikachu", ColumnFamilyOptions()));
std::vector<ColumnFamilyHandle*> handles;
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566) Summary: This PR does a few things: 1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation. 2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed: - The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated - The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory). 3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10). I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged. Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently. Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566 Reviewed By: zhichao-cao Differential Revision: D24408980 Pulled By: jay-zhuang fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
2020-10-27 18:31:34 +01:00
DBOptions db_opts;
db_opts.env = env_;
ASSERT_OK(DB::Open(db_opts, dbname2, column_families, &handles, &db2));
env_->SleepForMicroseconds(100);
ASSERT_TRUE(db2->Get(ro, handles[0], "a", &value).IsNotFound());
ASSERT_TRUE(db2->Get(ro, handles[0], "b", &value).IsNotFound());
ASSERT_TRUE(db2->Get(ro, handles[0], "c", &value).IsNotFound());
ASSERT_TRUE(db2->Get(ro, handles[0], "d", &value).IsNotFound());
ASSERT_TRUE(db2->Get(ro, handles[0], "e", &value).IsNotFound());
std::unique_ptr<TraceReader> trace_reader;
ASSERT_OK(NewFileTraceReader(env_, env_opts, trace_filename, &trace_reader));
Replayer replayer(db2, handles_, std::move(trace_reader));
ASSERT_OK(replayer.Replay());
ASSERT_TRUE(db2->Get(ro, handles[0], "a", &value).IsNotFound());
ASSERT_FALSE(db2->Get(ro, handles[0], "b", &value).IsNotFound());
ASSERT_TRUE(db2->Get(ro, handles[0], "c", &value).IsNotFound());
ASSERT_FALSE(db2->Get(ro, handles[0], "d", &value).IsNotFound());
ASSERT_TRUE(db2->Get(ro, handles[0], "e", &value).IsNotFound());
for (auto handle : handles) {
delete handle;
}
delete db2;
ASSERT_OK(DestroyDB(dbname2, options));
}
TEST_F(DBTest2, TraceWithFilter) {
Options options = CurrentOptions();
options.merge_operator = MergeOperators::CreatePutOperator();
ReadOptions ro;
WriteOptions wo;
TraceOptions trace_opts;
EnvOptions env_opts;
CreateAndReopenWithCF({"pikachu"}, options);
Random rnd(301);
Iterator* single_iter = nullptr;
trace_opts.filter = TraceFilterType::kTraceFilterWrite;
std::string trace_filename = dbname_ + "/rocksdb.trace";
std::unique_ptr<TraceWriter> trace_writer;
ASSERT_OK(NewFileTraceWriter(env_, env_opts, trace_filename, &trace_writer));
ASSERT_OK(db_->StartTrace(trace_opts, std::move(trace_writer)));
ASSERT_OK(Put(0, "a", "1"));
ASSERT_OK(Merge(0, "b", "2"));
ASSERT_OK(Delete(0, "c"));
ASSERT_OK(SingleDelete(0, "d"));
ASSERT_OK(db_->DeleteRange(wo, dbfull()->DefaultColumnFamily(), "e", "f"));
WriteBatch batch;
ASSERT_OK(batch.Put("f", "11"));
ASSERT_OK(batch.Merge("g", "12"));
ASSERT_OK(batch.Delete("h"));
ASSERT_OK(batch.SingleDelete("i"));
ASSERT_OK(batch.DeleteRange("j", "k"));
ASSERT_OK(db_->Write(wo, &batch));
single_iter = db_->NewIterator(ro);
single_iter->Seek("f");
single_iter->SeekForPrev("g");
delete single_iter;
ASSERT_EQ("1", Get(0, "a"));
ASSERT_EQ("12", Get(0, "g"));
ASSERT_OK(Put(1, "foo", "bar"));
ASSERT_OK(Put(1, "rocksdb", "rocks"));
ASSERT_EQ("NOT_FOUND", Get(1, "leveldb"));
ASSERT_OK(db_->EndTrace());
// These should not get into the trace file as it is after EndTrace.
Put("hello", "world");
Merge("foo", "bar");
// Open another db, replay, and verify the data
std::string value;
std::string dbname2 = test::PerThreadDBPath(env_, "db_replay");
ASSERT_OK(DestroyDB(dbname2, options));
// Using a different name than db2, to pacify infer's use-after-lifetime
// warnings (http://fbinfer.com).
DB* db2_init = nullptr;
options.create_if_missing = true;
ASSERT_OK(DB::Open(options, dbname2, &db2_init));
ColumnFamilyHandle* cf;
ASSERT_OK(
db2_init->CreateColumnFamily(ColumnFamilyOptions(), "pikachu", &cf));
delete cf;
delete db2_init;
DB* db2 = nullptr;
std::vector<ColumnFamilyDescriptor> column_families;
ColumnFamilyOptions cf_options;
cf_options.merge_operator = MergeOperators::CreatePutOperator();
column_families.push_back(ColumnFamilyDescriptor("default", cf_options));
column_families.push_back(
ColumnFamilyDescriptor("pikachu", ColumnFamilyOptions()));
std::vector<ColumnFamilyHandle*> handles;
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566) Summary: This PR does a few things: 1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation. 2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed: - The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated - The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory). 3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10). I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged. Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently. Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566 Reviewed By: zhichao-cao Differential Revision: D24408980 Pulled By: jay-zhuang fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
2020-10-27 18:31:34 +01:00
DBOptions db_opts;
db_opts.env = env_;
ASSERT_OK(DB::Open(db_opts, dbname2, column_families, &handles, &db2));
env_->SleepForMicroseconds(100);
// Verify that the keys don't already exist
ASSERT_TRUE(db2->Get(ro, handles[0], "a", &value).IsNotFound());
ASSERT_TRUE(db2->Get(ro, handles[0], "g", &value).IsNotFound());
std::unique_ptr<TraceReader> trace_reader;
ASSERT_OK(NewFileTraceReader(env_, env_opts, trace_filename, &trace_reader));
Replayer replayer(db2, handles_, std::move(trace_reader));
ASSERT_OK(replayer.Replay());
// All the key-values should not present since we filter out the WRITE ops.
ASSERT_TRUE(db2->Get(ro, handles[0], "a", &value).IsNotFound());
ASSERT_TRUE(db2->Get(ro, handles[0], "g", &value).IsNotFound());
ASSERT_TRUE(db2->Get(ro, handles[0], "hello", &value).IsNotFound());
ASSERT_TRUE(db2->Get(ro, handles[0], "world", &value).IsNotFound());
ASSERT_TRUE(db2->Get(ro, handles[0], "foo", &value).IsNotFound());
ASSERT_TRUE(db2->Get(ro, handles[0], "rocksdb", &value).IsNotFound());
for (auto handle : handles) {
delete handle;
}
delete db2;
ASSERT_OK(DestroyDB(dbname2, options));
// Set up a new db.
std::string dbname3 = test::PerThreadDBPath(env_, "db_not_trace_read");
ASSERT_OK(DestroyDB(dbname3, options));
DB* db3_init = nullptr;
options.create_if_missing = true;
ColumnFamilyHandle* cf3;
ASSERT_OK(DB::Open(options, dbname3, &db3_init));
ASSERT_OK(
db3_init->CreateColumnFamily(ColumnFamilyOptions(), "pikachu", &cf3));
delete cf3;
delete db3_init;
column_families.clear();
column_families.push_back(ColumnFamilyDescriptor("default", cf_options));
column_families.push_back(
ColumnFamilyDescriptor("pikachu", ColumnFamilyOptions()));
handles.clear();
DB* db3 = nullptr;
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566) Summary: This PR does a few things: 1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation. 2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed: - The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated - The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory). 3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10). I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged. Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently. Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566 Reviewed By: zhichao-cao Differential Revision: D24408980 Pulled By: jay-zhuang fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
2020-10-27 18:31:34 +01:00
ASSERT_OK(DB::Open(db_opts, dbname3, column_families, &handles, &db3));
env_->SleepForMicroseconds(100);
// Verify that the keys don't already exist
ASSERT_TRUE(db3->Get(ro, handles[0], "a", &value).IsNotFound());
ASSERT_TRUE(db3->Get(ro, handles[0], "g", &value).IsNotFound());
//The tracer will not record the READ ops.
trace_opts.filter = TraceFilterType::kTraceFilterGet;
std::string trace_filename3 = dbname_ + "/rocksdb.trace_3";
std::unique_ptr<TraceWriter> trace_writer3;
ASSERT_OK(
NewFileTraceWriter(env_, env_opts, trace_filename3, &trace_writer3));
ASSERT_OK(db3->StartTrace(trace_opts, std::move(trace_writer3)));
ASSERT_OK(db3->Put(wo, handles[0], "a", "1"));
ASSERT_OK(db3->Merge(wo, handles[0], "b", "2"));
ASSERT_OK(db3->Delete(wo, handles[0], "c"));
ASSERT_OK(db3->SingleDelete(wo, handles[0], "d"));
ASSERT_OK(db3->Get(ro, handles[0], "a", &value));
ASSERT_EQ(value, "1");
ASSERT_TRUE(db3->Get(ro, handles[0], "c", &value).IsNotFound());
ASSERT_OK(db3->EndTrace());
for (auto handle : handles) {
delete handle;
}
delete db3;
ASSERT_OK(DestroyDB(dbname3, options));
std::unique_ptr<TraceReader> trace_reader3;
ASSERT_OK(
NewFileTraceReader(env_, env_opts, trace_filename3, &trace_reader3));
// Count the number of records in the trace file;
int count = 0;
std::string data;
Status s;
while (true) {
s = trace_reader3->Read(&data);
if (!s.ok()) {
break;
}
count += 1;
}
// We also need to count the header and footer
// 4 WRITE + HEADER + FOOTER = 6
ASSERT_EQ(count, 6);
}
#endif // ROCKSDB_LITE
Copy Get() result when file reads use mmap Summary: For iterator reads, a `SuperVersion` is pinned to preserve a snapshot of SST files, and `Block`s are pinned to allow `key()` and `value()` to return pointers directly into a RocksDB memory region. This works for both non-mmap reads, where the block owns the memory region, and mmap reads, where the file owns the memory region. For point reads with `PinnableSlice`, only the `Block` object is pinned. This works for non-mmap reads because the block owns the memory region, so even if the file is deleted after compaction, the memory region survives. However, for mmap reads, file deletion causes the memory region to which the `PinnableSlice` refers to be unmapped. The result is usually a segfault upon accessing the `PinnableSlice`, although sometimes it returned wrong results (I repro'd this a bunch of times with `db_stress`). This PR copies the value into the `PinnableSlice` when it comes from mmap'd memory. We can tell whether the `Block` owns its memory using `Block::cachable()`, which is unset when reads do not use the provided buffer as is the case with mmap file reads. When that is false we ensure the result of `Get()` is copied. This feels like a short-term solution as ideally we'd have the `PinnableSlice` pin the mmap'd memory so we can do zero-copy reads. It seemed hard so I chose this approach to fix correctness in the meantime. Closes https://github.com/facebook/rocksdb/pull/3881 Differential Revision: D8076288 Pulled By: ajkr fbshipit-source-id: 31d78ec010198723522323dbc6ea325122a46b08
2018-06-02 01:46:32 +02:00
TEST_F(DBTest2, PinnableSliceAndMmapReads) {
Options options = CurrentOptions();
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566) Summary: This PR does a few things: 1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation. 2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed: - The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated - The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory). 3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10). I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged. Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently. Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566 Reviewed By: zhichao-cao Differential Revision: D24408980 Pulled By: jay-zhuang fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
2020-10-27 18:31:34 +01:00
options.env = env_;
if (!IsMemoryMappedAccessSupported()) {
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566) Summary: This PR does a few things: 1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation. 2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed: - The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated - The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory). 3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10). I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged. Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently. Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566 Reviewed By: zhichao-cao Differential Revision: D24408980 Pulled By: jay-zhuang fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
2020-10-27 18:31:34 +01:00
ROCKSDB_GTEST_SKIP("Test requires default environment");
return;
}
Copy Get() result when file reads use mmap Summary: For iterator reads, a `SuperVersion` is pinned to preserve a snapshot of SST files, and `Block`s are pinned to allow `key()` and `value()` to return pointers directly into a RocksDB memory region. This works for both non-mmap reads, where the block owns the memory region, and mmap reads, where the file owns the memory region. For point reads with `PinnableSlice`, only the `Block` object is pinned. This works for non-mmap reads because the block owns the memory region, so even if the file is deleted after compaction, the memory region survives. However, for mmap reads, file deletion causes the memory region to which the `PinnableSlice` refers to be unmapped. The result is usually a segfault upon accessing the `PinnableSlice`, although sometimes it returned wrong results (I repro'd this a bunch of times with `db_stress`). This PR copies the value into the `PinnableSlice` when it comes from mmap'd memory. We can tell whether the `Block` owns its memory using `Block::cachable()`, which is unset when reads do not use the provided buffer as is the case with mmap file reads. When that is false we ensure the result of `Get()` is copied. This feels like a short-term solution as ideally we'd have the `PinnableSlice` pin the mmap'd memory so we can do zero-copy reads. It seemed hard so I chose this approach to fix correctness in the meantime. Closes https://github.com/facebook/rocksdb/pull/3881 Differential Revision: D8076288 Pulled By: ajkr fbshipit-source-id: 31d78ec010198723522323dbc6ea325122a46b08
2018-06-02 01:46:32 +02:00
options.allow_mmap_reads = true;
options.max_open_files = 100;
options.compression = kNoCompression;
Copy Get() result when file reads use mmap Summary: For iterator reads, a `SuperVersion` is pinned to preserve a snapshot of SST files, and `Block`s are pinned to allow `key()` and `value()` to return pointers directly into a RocksDB memory region. This works for both non-mmap reads, where the block owns the memory region, and mmap reads, where the file owns the memory region. For point reads with `PinnableSlice`, only the `Block` object is pinned. This works for non-mmap reads because the block owns the memory region, so even if the file is deleted after compaction, the memory region survives. However, for mmap reads, file deletion causes the memory region to which the `PinnableSlice` refers to be unmapped. The result is usually a segfault upon accessing the `PinnableSlice`, although sometimes it returned wrong results (I repro'd this a bunch of times with `db_stress`). This PR copies the value into the `PinnableSlice` when it comes from mmap'd memory. We can tell whether the `Block` owns its memory using `Block::cachable()`, which is unset when reads do not use the provided buffer as is the case with mmap file reads. When that is false we ensure the result of `Get()` is copied. This feels like a short-term solution as ideally we'd have the `PinnableSlice` pin the mmap'd memory so we can do zero-copy reads. It seemed hard so I chose this approach to fix correctness in the meantime. Closes https://github.com/facebook/rocksdb/pull/3881 Differential Revision: D8076288 Pulled By: ajkr fbshipit-source-id: 31d78ec010198723522323dbc6ea325122a46b08
2018-06-02 01:46:32 +02:00
Reopen(options);
ASSERT_OK(Put("foo", "bar"));
ASSERT_OK(Flush());
PinnableSlice pinned_value;
ASSERT_EQ(Get("foo", &pinned_value), Status::OK());
// It is not safe to pin mmap files as they might disappear by compaction
ASSERT_FALSE(pinned_value.IsPinned());
Copy Get() result when file reads use mmap Summary: For iterator reads, a `SuperVersion` is pinned to preserve a snapshot of SST files, and `Block`s are pinned to allow `key()` and `value()` to return pointers directly into a RocksDB memory region. This works for both non-mmap reads, where the block owns the memory region, and mmap reads, where the file owns the memory region. For point reads with `PinnableSlice`, only the `Block` object is pinned. This works for non-mmap reads because the block owns the memory region, so even if the file is deleted after compaction, the memory region survives. However, for mmap reads, file deletion causes the memory region to which the `PinnableSlice` refers to be unmapped. The result is usually a segfault upon accessing the `PinnableSlice`, although sometimes it returned wrong results (I repro'd this a bunch of times with `db_stress`). This PR copies the value into the `PinnableSlice` when it comes from mmap'd memory. We can tell whether the `Block` owns its memory using `Block::cachable()`, which is unset when reads do not use the provided buffer as is the case with mmap file reads. When that is false we ensure the result of `Get()` is copied. This feels like a short-term solution as ideally we'd have the `PinnableSlice` pin the mmap'd memory so we can do zero-copy reads. It seemed hard so I chose this approach to fix correctness in the meantime. Closes https://github.com/facebook/rocksdb/pull/3881 Differential Revision: D8076288 Pulled By: ajkr fbshipit-source-id: 31d78ec010198723522323dbc6ea325122a46b08
2018-06-02 01:46:32 +02:00
ASSERT_EQ(pinned_value.ToString(), "bar");
dbfull()->TEST_CompactRange(0 /* level */, nullptr /* begin */,
nullptr /* end */, nullptr /* column_family */,
true /* disallow_trivial_move */);
// Ensure pinned_value doesn't rely on memory munmap'd by the above
// compaction. It crashes if it does.
ASSERT_EQ(pinned_value.ToString(), "bar");
#ifndef ROCKSDB_LITE
pinned_value.Reset();
// Unsafe to pin mmap files when they could be kicked out of table cache
Close();
ASSERT_OK(ReadOnlyReopen(options));
ASSERT_EQ(Get("foo", &pinned_value), Status::OK());
ASSERT_FALSE(pinned_value.IsPinned());
ASSERT_EQ(pinned_value.ToString(), "bar");
pinned_value.Reset();
// In read-only mode with infinite capacity on table cache it should pin the
// value and avoid the memcpy
Close();
options.max_open_files = -1;
ASSERT_OK(ReadOnlyReopen(options));
ASSERT_EQ(Get("foo", &pinned_value), Status::OK());
ASSERT_TRUE(pinned_value.IsPinned());
Copy Get() result when file reads use mmap Summary: For iterator reads, a `SuperVersion` is pinned to preserve a snapshot of SST files, and `Block`s are pinned to allow `key()` and `value()` to return pointers directly into a RocksDB memory region. This works for both non-mmap reads, where the block owns the memory region, and mmap reads, where the file owns the memory region. For point reads with `PinnableSlice`, only the `Block` object is pinned. This works for non-mmap reads because the block owns the memory region, so even if the file is deleted after compaction, the memory region survives. However, for mmap reads, file deletion causes the memory region to which the `PinnableSlice` refers to be unmapped. The result is usually a segfault upon accessing the `PinnableSlice`, although sometimes it returned wrong results (I repro'd this a bunch of times with `db_stress`). This PR copies the value into the `PinnableSlice` when it comes from mmap'd memory. We can tell whether the `Block` owns its memory using `Block::cachable()`, which is unset when reads do not use the provided buffer as is the case with mmap file reads. When that is false we ensure the result of `Get()` is copied. This feels like a short-term solution as ideally we'd have the `PinnableSlice` pin the mmap'd memory so we can do zero-copy reads. It seemed hard so I chose this approach to fix correctness in the meantime. Closes https://github.com/facebook/rocksdb/pull/3881 Differential Revision: D8076288 Pulled By: ajkr fbshipit-source-id: 31d78ec010198723522323dbc6ea325122a46b08
2018-06-02 01:46:32 +02:00
ASSERT_EQ(pinned_value.ToString(), "bar");
#endif
Copy Get() result when file reads use mmap Summary: For iterator reads, a `SuperVersion` is pinned to preserve a snapshot of SST files, and `Block`s are pinned to allow `key()` and `value()` to return pointers directly into a RocksDB memory region. This works for both non-mmap reads, where the block owns the memory region, and mmap reads, where the file owns the memory region. For point reads with `PinnableSlice`, only the `Block` object is pinned. This works for non-mmap reads because the block owns the memory region, so even if the file is deleted after compaction, the memory region survives. However, for mmap reads, file deletion causes the memory region to which the `PinnableSlice` refers to be unmapped. The result is usually a segfault upon accessing the `PinnableSlice`, although sometimes it returned wrong results (I repro'd this a bunch of times with `db_stress`). This PR copies the value into the `PinnableSlice` when it comes from mmap'd memory. We can tell whether the `Block` owns its memory using `Block::cachable()`, which is unset when reads do not use the provided buffer as is the case with mmap file reads. When that is false we ensure the result of `Get()` is copied. This feels like a short-term solution as ideally we'd have the `PinnableSlice` pin the mmap'd memory so we can do zero-copy reads. It seemed hard so I chose this approach to fix correctness in the meantime. Closes https://github.com/facebook/rocksdb/pull/3881 Differential Revision: D8076288 Pulled By: ajkr fbshipit-source-id: 31d78ec010198723522323dbc6ea325122a46b08
2018-06-02 01:46:32 +02:00
}
TEST_F(DBTest2, DISABLED_IteratorPinnedMemory) {
Options options = CurrentOptions();
options.create_if_missing = true;
options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
BlockBasedTableOptions bbto;
bbto.no_block_cache = false;
bbto.cache_index_and_filter_blocks = false;
bbto.block_cache = NewLRUCache(100000);
bbto.block_size = 400; // small block size
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566) Summary: This PR does a few things: 1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation. 2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed: - The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated - The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory). 3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10). I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged. Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently. Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566 Reviewed By: zhichao-cao Differential Revision: D24408980 Pulled By: jay-zhuang fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
2020-10-27 18:31:34 +01:00
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
Reopen(options);
Random rnd(301);
std::string v = rnd.RandomString(400);
// Since v is the size of a block, each key should take a block
// of 400+ bytes.
Put("1", v);
Put("3", v);
Put("5", v);
Put("7", v);
ASSERT_OK(Flush());
ASSERT_EQ(0, bbto.block_cache->GetPinnedUsage());
// Verify that iterators don't pin more than one data block in block cache
// at each time.
{
std::unique_ptr<Iterator> iter(db_->NewIterator(ReadOptions()));
iter->SeekToFirst();
for (int i = 0; i < 4; i++) {
ASSERT_TRUE(iter->Valid());
// Block cache should contain exactly one block.
ASSERT_GT(bbto.block_cache->GetPinnedUsage(), 0);
ASSERT_LT(bbto.block_cache->GetPinnedUsage(), 800);
iter->Next();
}
ASSERT_FALSE(iter->Valid());
iter->Seek("4");
ASSERT_TRUE(iter->Valid());
ASSERT_GT(bbto.block_cache->GetPinnedUsage(), 0);
ASSERT_LT(bbto.block_cache->GetPinnedUsage(), 800);
iter->Seek("3");
ASSERT_TRUE(iter->Valid());
ASSERT_GT(bbto.block_cache->GetPinnedUsage(), 0);
ASSERT_LT(bbto.block_cache->GetPinnedUsage(), 800);
}
ASSERT_EQ(0, bbto.block_cache->GetPinnedUsage());
// Test compaction case
Put("2", v);
Put("5", v);
Put("6", v);
Put("8", v);
ASSERT_OK(Flush());
// Clear existing data in block cache
bbto.block_cache->SetCapacity(0);
bbto.block_cache->SetCapacity(100000);
// Verify compaction input iterators don't hold more than one data blocks at
// one time.
std::atomic<bool> finished(false);
std::atomic<int> block_newed(0);
std::atomic<int> block_destroyed(0);
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"Block::Block:0", [&](void* /*arg*/) {
if (finished) {
return;
}
// Two iterators. At most 2 outstanding blocks.
EXPECT_GE(block_newed.load(), block_destroyed.load());
EXPECT_LE(block_newed.load(), block_destroyed.load() + 1);
block_newed.fetch_add(1);
});
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"Block::~Block", [&](void* /*arg*/) {
if (finished) {
return;
}
// Two iterators. At most 2 outstanding blocks.
EXPECT_GE(block_newed.load(), block_destroyed.load() + 1);
EXPECT_LE(block_newed.load(), block_destroyed.load() + 2);
block_destroyed.fetch_add(1);
});
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"CompactionJob::Run:BeforeVerify",
[&](void* /*arg*/) { finished = true; });
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
ASSERT_OK(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
// Two input files. Each of them has 4 data blocks.
ASSERT_EQ(8, block_newed.load());
ASSERT_EQ(8, block_destroyed.load());
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
}
TEST_F(DBTest2, TestBBTTailPrefetch) {
std::atomic<bool> called(false);
size_t expected_lower_bound = 512 * 1024;
size_t expected_higher_bound = 512 * 1024;
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"BlockBasedTable::Open::TailPrefetchLen", [&](void* arg) {
size_t* prefetch_size = static_cast<size_t*>(arg);
EXPECT_LE(expected_lower_bound, *prefetch_size);
EXPECT_GE(expected_higher_bound, *prefetch_size);
called = true;
});
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
Put("1", "1");
Put("9", "1");
Flush();
expected_lower_bound = 0;
expected_higher_bound = 8 * 1024;
Put("1", "1");
Put("9", "1");
Flush();
Put("1", "1");
Put("9", "1");
Flush();
// Full compaction to make sure there is no L0 file after the open.
ASSERT_OK(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
ASSERT_TRUE(called.load());
called = false;
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
std::atomic<bool> first_call(true);
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
"BlockBasedTable::Open::TailPrefetchLen", [&](void* arg) {
size_t* prefetch_size = static_cast<size_t*>(arg);
if (first_call) {
EXPECT_EQ(4 * 1024, *prefetch_size);
first_call = false;
} else {
EXPECT_GE(4 * 1024, *prefetch_size);
}
called = true;
});
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
Options options = CurrentOptions();
options.max_file_opening_threads = 1; // one thread
BlockBasedTableOptions table_options;
table_options.cache_index_and_filter_blocks = true;
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
options.max_open_files = -1;
Reopen(options);
Put("1", "1");
Put("9", "1");
Flush();
Put("1", "1");
Put("9", "1");
Flush();
ASSERT_TRUE(called.load());
called = false;
// Parallel loading SST files
options.max_file_opening_threads = 16;
Reopen(options);
ASSERT_OK(db_->CompactRange(CompactRangeOptions(), nullptr, nullptr));
ASSERT_TRUE(called.load());
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
}
TEST_F(DBTest2, TestGetColumnFamilyHandleUnlocked) {
// Setup sync point dependency to reproduce the race condition of
// DBImpl::GetColumnFamilyHandleUnlocked
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency({
{"TestGetColumnFamilyHandleUnlocked::GetColumnFamilyHandleUnlocked1",
"TestGetColumnFamilyHandleUnlocked::PreGetColumnFamilyHandleUnlocked2"},
{"TestGetColumnFamilyHandleUnlocked::GetColumnFamilyHandleUnlocked2",
"TestGetColumnFamilyHandleUnlocked::ReadColumnFamilyHandle1"},
});
SyncPoint::GetInstance()->EnableProcessing();
CreateColumnFamilies({"test1", "test2"}, Options());
ASSERT_EQ(handles_.size(), 2);
DBImpl* dbi = static_cast_with_check<DBImpl>(db_);
port::Thread user_thread1([&]() {
auto cfh = dbi->GetColumnFamilyHandleUnlocked(handles_[0]->GetID());
ASSERT_EQ(cfh->GetID(), handles_[0]->GetID());
TEST_SYNC_POINT("TestGetColumnFamilyHandleUnlocked::GetColumnFamilyHandleUnlocked1");
TEST_SYNC_POINT("TestGetColumnFamilyHandleUnlocked::ReadColumnFamilyHandle1");
ASSERT_EQ(cfh->GetID(), handles_[0]->GetID());
});
port::Thread user_thread2([&]() {
TEST_SYNC_POINT("TestGetColumnFamilyHandleUnlocked::PreGetColumnFamilyHandleUnlocked2");
auto cfh = dbi->GetColumnFamilyHandleUnlocked(handles_[1]->GetID());
ASSERT_EQ(cfh->GetID(), handles_[1]->GetID());
TEST_SYNC_POINT("TestGetColumnFamilyHandleUnlocked::GetColumnFamilyHandleUnlocked2");
ASSERT_EQ(cfh->GetID(), handles_[1]->GetID());
});
user_thread1.join();
user_thread2.join();
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
}
#ifndef ROCKSDB_LITE
TEST_F(DBTest2, TestCompactFiles) {
// Setup sync point dependency to reproduce the race condition of
// DBImpl::GetColumnFamilyHandleUnlocked
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency({
{"TestCompactFiles::IngestExternalFile1",
"TestCompactFiles::IngestExternalFile2"},
});
SyncPoint::GetInstance()->EnableProcessing();
Options options;
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566) Summary: This PR does a few things: 1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation. 2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed: - The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated - The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory). 3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10). I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged. Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently. Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566 Reviewed By: zhichao-cao Differential Revision: D24408980 Pulled By: jay-zhuang fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
2020-10-27 18:31:34 +01:00
options.env = env_;
options.num_levels = 2;
options.disable_auto_compactions = true;
Reopen(options);
auto* handle = db_->DefaultColumnFamily();
ASSERT_EQ(db_->NumberLevels(handle), 2);
ROCKSDB_NAMESPACE::SstFileWriter sst_file_writer{
ROCKSDB_NAMESPACE::EnvOptions(), options};
std::string external_file1 = dbname_ + "/test_compact_files1.sst_t";
std::string external_file2 = dbname_ + "/test_compact_files2.sst_t";
std::string external_file3 = dbname_ + "/test_compact_files3.sst_t";
ASSERT_OK(sst_file_writer.Open(external_file1));
ASSERT_OK(sst_file_writer.Put("1", "1"));
ASSERT_OK(sst_file_writer.Put("2", "2"));
ASSERT_OK(sst_file_writer.Finish());
ASSERT_OK(sst_file_writer.Open(external_file2));
ASSERT_OK(sst_file_writer.Put("3", "3"));
ASSERT_OK(sst_file_writer.Put("4", "4"));
ASSERT_OK(sst_file_writer.Finish());
ASSERT_OK(sst_file_writer.Open(external_file3));
ASSERT_OK(sst_file_writer.Put("5", "5"));
ASSERT_OK(sst_file_writer.Put("6", "6"));
ASSERT_OK(sst_file_writer.Finish());
ASSERT_OK(db_->IngestExternalFile(handle, {external_file1, external_file3},
IngestExternalFileOptions()));
ASSERT_EQ(NumTableFilesAtLevel(1, 0), 2);
std::vector<std::string> files;
GetSstFiles(env_, dbname_, &files);
ASSERT_EQ(files.size(), 2);
port::Thread user_thread1(
[&]() { db_->CompactFiles(CompactionOptions(), handle, files, 1); });
port::Thread user_thread2([&]() {
ASSERT_OK(db_->IngestExternalFile(handle, {external_file2},
IngestExternalFileOptions()));
TEST_SYNC_POINT("TestCompactFiles::IngestExternalFile1");
});
user_thread1.join();
user_thread2.join();
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
}
#endif // ROCKSDB_LITE
// TODO: figure out why this test fails in appveyor
#ifndef OS_WIN
TEST_F(DBTest2, MultiDBParallelOpenTest) {
const int kNumDbs = 2;
Options options = CurrentOptions();
std::vector<std::string> dbnames;
for (int i = 0; i < kNumDbs; ++i) {
dbnames.emplace_back(test::PerThreadDBPath(env_, "db" + ToString(i)));
ASSERT_OK(DestroyDB(dbnames.back(), options));
}
// Verify empty DBs can be created in parallel
std::vector<std::thread> open_threads;
std::vector<DB*> dbs{static_cast<unsigned int>(kNumDbs), nullptr};
options.create_if_missing = true;
for (int i = 0; i < kNumDbs; ++i) {
open_threads.emplace_back(
[&](int dbnum) {
ASSERT_OK(DB::Open(options, dbnames[dbnum], &dbs[dbnum]));
},
i);
}
// Now add some data and close, so next we can verify non-empty DBs can be
// recovered in parallel
for (int i = 0; i < kNumDbs; ++i) {
open_threads[i].join();
ASSERT_OK(dbs[i]->Put(WriteOptions(), "xi", "gua"));
delete dbs[i];
}
// Verify non-empty DBs can be recovered in parallel
dbs.clear();
open_threads.clear();
for (int i = 0; i < kNumDbs; ++i) {
open_threads.emplace_back(
[&](int dbnum) {
ASSERT_OK(DB::Open(options, dbnames[dbnum], &dbs[dbnum]));
},
i);
}
// Wait and cleanup
for (int i = 0; i < kNumDbs; ++i) {
open_threads[i].join();
delete dbs[i];
ASSERT_OK(DestroyDB(dbnames[i], options));
}
}
#endif // OS_WIN
namespace {
class DummyOldStats : public Statistics {
public:
uint64_t getTickerCount(uint32_t /*ticker_type*/) const override { return 0; }
void recordTick(uint32_t /* ticker_type */, uint64_t /* count */) override {
num_rt++;
}
void setTickerCount(uint32_t /*ticker_type*/, uint64_t /*count*/) override {}
uint64_t getAndResetTickerCount(uint32_t /*ticker_type*/) override {
return 0;
}
void measureTime(uint32_t /*histogram_type*/, uint64_t /*count*/) override {
num_mt++;
}
void histogramData(
uint32_t /*histogram_type*/,
ROCKSDB_NAMESPACE::HistogramData* const /*data*/) const override {}
std::string getHistogramString(uint32_t /*type*/) const override {
return "";
}
bool HistEnabledForType(uint32_t /*type*/) const override { return false; }
std::string ToString() const override { return ""; }
std::atomic<int> num_rt{0};
std::atomic<int> num_mt{0};
};
} // namespace
TEST_F(DBTest2, OldStatsInterface) {
DummyOldStats* dos = new DummyOldStats();
std::shared_ptr<Statistics> stats(dos);
Options options = CurrentOptions();
options.create_if_missing = true;
options.statistics = stats;
Reopen(options);
Put("foo", "bar");
ASSERT_EQ("bar", Get("foo"));
ASSERT_OK(Flush());
ASSERT_EQ("bar", Get("foo"));
ASSERT_GT(dos->num_rt, 0);
ASSERT_GT(dos->num_mt, 0);
}
TEST_F(DBTest2, CloseWithUnreleasedSnapshot) {
const Snapshot* ss = db_->GetSnapshot();
for (auto h : handles_) {
db_->DestroyColumnFamilyHandle(h);
}
handles_.clear();
ASSERT_NOK(db_->Close());
db_->ReleaseSnapshot(ss);
ASSERT_OK(db_->Close());
delete db_;
db_ = nullptr;
}
TEST_F(DBTest2, PrefixBloomReseek) {
Options options = CurrentOptions();
options.create_if_missing = true;
options.prefix_extractor.reset(NewCappedPrefixTransform(3));
BlockBasedTableOptions bbto;
bbto.filter_policy.reset(NewBloomFilterPolicy(10, false));
bbto.whole_key_filtering = false;
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
DestroyAndReopen(options);
// Construct two L1 files with keys:
// f1:[aaa1 ccc1] f2:[ddd0]
ASSERT_OK(Put("aaa1", ""));
ASSERT_OK(Put("ccc1", ""));
ASSERT_OK(Flush());
ASSERT_OK(Put("ddd0", ""));
ASSERT_OK(Flush());
CompactRangeOptions cro;
cro.bottommost_level_compaction = BottommostLevelCompaction::kSkip;
ASSERT_OK(db_->CompactRange(cro, nullptr, nullptr));
ASSERT_OK(Put("bbb1", ""));
Iterator* iter = db_->NewIterator(ReadOptions());
// Seeking into f1, the iterator will check bloom filter which returns the
// file iterator ot be invalidate, and the cursor will put into f2, with
// the next key to be "ddd0".
iter->Seek("bbb1");
ASSERT_TRUE(iter->Valid());
ASSERT_EQ("bbb1", iter->key().ToString());
// Reseek ccc1, the L1 iterator needs to go back to f1 and reseek.
iter->Seek("ccc1");
ASSERT_TRUE(iter->Valid());
ASSERT_EQ("ccc1", iter->key().ToString());
delete iter;
}
TEST_F(DBTest2, PrefixBloomFilteredOut) {
Options options = CurrentOptions();
options.create_if_missing = true;
options.prefix_extractor.reset(NewCappedPrefixTransform(3));
BlockBasedTableOptions bbto;
bbto.filter_policy.reset(NewBloomFilterPolicy(10, false));
bbto.whole_key_filtering = false;
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
DestroyAndReopen(options);
// Construct two L1 files with keys:
// f1:[aaa1 ccc1] f2:[ddd0]
ASSERT_OK(Put("aaa1", ""));
ASSERT_OK(Put("ccc1", ""));
ASSERT_OK(Flush());
ASSERT_OK(Put("ddd0", ""));
ASSERT_OK(Flush());
CompactRangeOptions cro;
cro.bottommost_level_compaction = BottommostLevelCompaction::kSkip;
ASSERT_OK(db_->CompactRange(cro, nullptr, nullptr));
Iterator* iter = db_->NewIterator(ReadOptions());
// Bloom filter is filterd out by f1.
// This is just one of several valid position following the contract.
// Postioning to ccc1 or ddd0 is also valid. This is just to validate
// the behavior of the current implementation. If underlying implementation
// changes, the test might fail here.
iter->Seek("bbb1");
ASSERT_FALSE(iter->Valid());
delete iter;
}
#ifndef ROCKSDB_LITE
TEST_F(DBTest2, RowCacheSnapshot) {
Options options = CurrentOptions();
options.statistics = ROCKSDB_NAMESPACE::CreateDBStatistics();
options.row_cache = NewLRUCache(8 * 8192);
DestroyAndReopen(options);
ASSERT_OK(Put("foo", "bar1"));
const Snapshot* s1 = db_->GetSnapshot();
ASSERT_OK(Put("foo", "bar2"));
ASSERT_OK(Flush());
ASSERT_OK(Put("foo2", "bar"));
const Snapshot* s2 = db_->GetSnapshot();
ASSERT_OK(Put("foo3", "bar"));
const Snapshot* s3 = db_->GetSnapshot();
ASSERT_EQ(TestGetTickerCount(options, ROW_CACHE_HIT), 0);
ASSERT_EQ(TestGetTickerCount(options, ROW_CACHE_MISS), 0);
ASSERT_EQ(Get("foo"), "bar2");
ASSERT_EQ(TestGetTickerCount(options, ROW_CACHE_HIT), 0);
ASSERT_EQ(TestGetTickerCount(options, ROW_CACHE_MISS), 1);
ASSERT_EQ(Get("foo"), "bar2");
ASSERT_EQ(TestGetTickerCount(options, ROW_CACHE_HIT), 1);
ASSERT_EQ(TestGetTickerCount(options, ROW_CACHE_MISS), 1);
ASSERT_EQ(Get("foo", s1), "bar1");
ASSERT_EQ(TestGetTickerCount(options, ROW_CACHE_HIT), 1);
ASSERT_EQ(TestGetTickerCount(options, ROW_CACHE_MISS), 2);
ASSERT_EQ(Get("foo", s2), "bar2");
ASSERT_EQ(TestGetTickerCount(options, ROW_CACHE_HIT), 2);
ASSERT_EQ(TestGetTickerCount(options, ROW_CACHE_MISS), 2);
ASSERT_EQ(Get("foo", s1), "bar1");
ASSERT_EQ(TestGetTickerCount(options, ROW_CACHE_HIT), 3);
ASSERT_EQ(TestGetTickerCount(options, ROW_CACHE_MISS), 2);
ASSERT_EQ(Get("foo", s3), "bar2");
ASSERT_EQ(TestGetTickerCount(options, ROW_CACHE_HIT), 4);
ASSERT_EQ(TestGetTickerCount(options, ROW_CACHE_MISS), 2);
db_->ReleaseSnapshot(s1);
db_->ReleaseSnapshot(s2);
db_->ReleaseSnapshot(s3);
}
#endif // ROCKSDB_LITE
// When DB is reopened with multiple column families, the manifest file
// is written after the first CF is flushed, and it is written again
// after each flush. If DB crashes between the flushes, the flushed CF
// flushed will pass the latest log file, and now we require it not
// to be corrupted, and triggering a corruption report.
// We need to fix the bug and enable the test.
TEST_F(DBTest2, CrashInRecoveryMultipleCF) {
const std::vector<std::string> sync_points = {
"DBImpl::RecoverLogFiles:BeforeFlushFinalMemtable",
"VersionSet::ProcessManifestWrites:BeforeWriteLastVersionEdit:0"};
for (const auto& test_sync_point : sync_points) {
Options options = CurrentOptions();
// First destroy original db to ensure a clean start.
DestroyAndReopen(options);
options.create_if_missing = true;
options.wal_recovery_mode = WALRecoveryMode::kPointInTimeRecovery;
CreateAndReopenWithCF({"pikachu"}, options);
ASSERT_OK(Put("foo", "bar"));
ASSERT_OK(Flush());
ASSERT_OK(Put(1, "foo", "bar"));
ASSERT_OK(Flush(1));
ASSERT_OK(Put("foo", "bar"));
ASSERT_OK(Put(1, "foo", "bar"));
// The value is large enough to be divided to two blocks.
std::string large_value(400, ' ');
ASSERT_OK(Put("foo1", large_value));
ASSERT_OK(Put("foo2", large_value));
Close();
// Corrupt the log file in the middle, so that it is not corrupted
// in the tail.
std::vector<std::string> filenames;
ASSERT_OK(env_->GetChildren(dbname_, &filenames));
for (const auto& f : filenames) {
uint64_t number;
FileType type;
if (ParseFileName(f, &number, &type) && type == FileType::kWalFile) {
std::string fname = dbname_ + "/" + f;
std::string file_content;
ASSERT_OK(ReadFileToString(env_, fname, &file_content));
file_content[400] = 'h';
file_content[401] = 'a';
ASSERT_OK(WriteStringToFile(env_, file_content, fname));
break;
}
}
// Reopen and freeze the file system after the first manifest write.
FaultInjectionTestEnv fit_env(options.env);
options.env = &fit_env;
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->SetCallBack(
test_sync_point,
[&](void* /*arg*/) { fit_env.SetFilesystemActive(false); });
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
ASSERT_NOK(TryReopenWithColumnFamilies(
{kDefaultColumnFamilyName, "pikachu"}, options));
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
fit_env.SetFilesystemActive(true);
// If we continue using failure ingestion Env, it will conplain something
// when renaming current file, which is not expected. Need to investigate
// why.
options.env = env_;
ASSERT_OK(TryReopenWithColumnFamilies({kDefaultColumnFamilyName, "pikachu"},
options));
}
}
TEST_F(DBTest2, SeekFileRangeDeleteTail) {
Options options = CurrentOptions();
options.prefix_extractor.reset(NewCappedPrefixTransform(1));
options.num_levels = 3;
DestroyAndReopen(options);
ASSERT_OK(Put("a", "a"));
const Snapshot* s1 = db_->GetSnapshot();
ASSERT_OK(
db_->DeleteRange(WriteOptions(), db_->DefaultColumnFamily(), "a", "f"));
ASSERT_OK(Put("b", "a"));
ASSERT_OK(Flush());
ASSERT_OK(Put("x", "a"));
ASSERT_OK(Put("z", "a"));
ASSERT_OK(Flush());
CompactRangeOptions cro;
cro.change_level = true;
cro.target_level = 2;
ASSERT_OK(db_->CompactRange(cro, nullptr, nullptr));
{
ReadOptions ro;
ro.total_order_seek = true;
std::unique_ptr<Iterator> iter(db_->NewIterator(ro));
iter->Seek("e");
ASSERT_TRUE(iter->Valid());
ASSERT_EQ("x", iter->key().ToString());
}
db_->ReleaseSnapshot(s1);
}
TEST_F(DBTest2, BackgroundPurgeTest) {
Options options = CurrentOptions();
options.write_buffer_manager =
std::make_shared<ROCKSDB_NAMESPACE::WriteBufferManager>(1 << 20);
options.avoid_unnecessary_blocking_io = true;
DestroyAndReopen(options);
size_t base_value = options.write_buffer_manager->memory_usage();
ASSERT_OK(Put("a", "a"));
Iterator* iter = db_->NewIterator(ReadOptions());
ASSERT_OK(Flush());
size_t value = options.write_buffer_manager->memory_usage();
ASSERT_GT(value, base_value);
db_->GetEnv()->SetBackgroundThreads(1, Env::Priority::HIGH);
test::SleepingBackgroundTask sleeping_task_after;
db_->GetEnv()->Schedule(&test::SleepingBackgroundTask::DoSleepTask,
&sleeping_task_after, Env::Priority::HIGH);
delete iter;
Env::Default()->SleepForMicroseconds(100000);
value = options.write_buffer_manager->memory_usage();
ASSERT_GT(value, base_value);
sleeping_task_after.WakeUp();
sleeping_task_after.WaitUntilDone();
test::SleepingBackgroundTask sleeping_task_after2;
db_->GetEnv()->Schedule(&test::SleepingBackgroundTask::DoSleepTask,
&sleeping_task_after2, Env::Priority::HIGH);
sleeping_task_after2.WakeUp();
sleeping_task_after2.WaitUntilDone();
value = options.write_buffer_manager->memory_usage();
ASSERT_EQ(base_value, value);
}
Fix a data race for cfd->log_number_ (#6249) Summary: A thread calling LogAndApply may release db mutex when calling WriteCurrentStateToManifest() which reads cfd->log_number_. Another thread can call SwitchMemtable() and writes to cfd->log_number_. Solution is to cache the cfd->log_number_ before releasing mutex in LogAndApply. Test Plan (on devserver): ``` $COMPILE_WITH_TSAN=1 make db_stress $./db_stress --acquire_snapshot_one_in=10000 --avoid_unnecessary_blocking_io=1 --block_size=16384 --bloom_bits=16 --bottommost_compression_type=zstd --cache_index_and_filter_blocks=1 --cache_size=1048576 --checkpoint_one_in=1000000 --checksum_type=kxxHash --clear_column_family_one_in=0 --compact_files_one_in=1000000 --compact_range_one_in=1000000 --compaction_ttl=0 --compression_max_dict_bytes=16384 --compression_type=zstd --compression_zstd_max_train_bytes=0 --continuous_verification_interval=0 --db=/dev/shm/rocksdb/rocksdb_crashtest_blackbox --db_write_buffer_size=1048576 --delpercent=5 --delrangepercent=0 --destroy_db_initially=0 --enable_pipelined_write=0 --flush_one_in=1000000 --format_version=5 --get_live_files_and_wal_files_one_in=1000000 --index_block_restart_interval=5 --index_type=0 --log2_keys_per_lock=22 --long_running_snapshots=0 --max_background_compactions=20 --max_bytes_for_level_base=10485760 --max_key=1000000 --max_manifest_file_size=16384 --max_write_batch_group_size_bytes=16 --max_write_buffer_number=3 --memtablerep=skip_list --mmap_read=0 --nooverwritepercent=1 --open_files=500000 --ops_per_thread=100000000 --partition_filters=0 --pause_background_one_in=1000000 --periodic_compaction_seconds=0 --prefixpercent=5 --progress_reports=0 --readpercent=45 --recycle_log_file_num=0 --reopen=20 --set_options_one_in=10000 --snapshot_hold_ops=100000 --subcompactions=2 --sync=1 --target_file_size_base=2097152 --target_file_size_multiplier=2 --test_batches_snapshots=1 --use_direct_io_for_flush_and_compaction=0 --use_direct_reads=0 --use_full_merge_v1=0 --use_merge=0 --use_multiget=1 --verify_checksum=1 --verify_checksum_one_in=1000000 --verify_db_one_in=100000 --write_buffer_size=4194304 --write_dbid_to_manifest=1 --writepercent=35 ``` Then repeat the following multiple times, e.g. 100 after compiling with tsan. ``` $./db_test2 --gtest_filter=DBTest2.SwitchMemtableRaceWithNewManifest ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6249 Differential Revision: D19235077 Pulled By: riversand963 fbshipit-source-id: 79467b52f48739ce7c27e440caa2447a40653173
2020-01-07 05:08:24 +01:00
TEST_F(DBTest2, SwitchMemtableRaceWithNewManifest) {
Options options = CurrentOptions();
DestroyAndReopen(options);
options.max_manifest_file_size = 10;
options.create_if_missing = true;
CreateAndReopenWithCF({"pikachu"}, options);
ASSERT_EQ(2, handles_.size());
Fix a data race for cfd->log_number_ (#6249) Summary: A thread calling LogAndApply may release db mutex when calling WriteCurrentStateToManifest() which reads cfd->log_number_. Another thread can call SwitchMemtable() and writes to cfd->log_number_. Solution is to cache the cfd->log_number_ before releasing mutex in LogAndApply. Test Plan (on devserver): ``` $COMPILE_WITH_TSAN=1 make db_stress $./db_stress --acquire_snapshot_one_in=10000 --avoid_unnecessary_blocking_io=1 --block_size=16384 --bloom_bits=16 --bottommost_compression_type=zstd --cache_index_and_filter_blocks=1 --cache_size=1048576 --checkpoint_one_in=1000000 --checksum_type=kxxHash --clear_column_family_one_in=0 --compact_files_one_in=1000000 --compact_range_one_in=1000000 --compaction_ttl=0 --compression_max_dict_bytes=16384 --compression_type=zstd --compression_zstd_max_train_bytes=0 --continuous_verification_interval=0 --db=/dev/shm/rocksdb/rocksdb_crashtest_blackbox --db_write_buffer_size=1048576 --delpercent=5 --delrangepercent=0 --destroy_db_initially=0 --enable_pipelined_write=0 --flush_one_in=1000000 --format_version=5 --get_live_files_and_wal_files_one_in=1000000 --index_block_restart_interval=5 --index_type=0 --log2_keys_per_lock=22 --long_running_snapshots=0 --max_background_compactions=20 --max_bytes_for_level_base=10485760 --max_key=1000000 --max_manifest_file_size=16384 --max_write_batch_group_size_bytes=16 --max_write_buffer_number=3 --memtablerep=skip_list --mmap_read=0 --nooverwritepercent=1 --open_files=500000 --ops_per_thread=100000000 --partition_filters=0 --pause_background_one_in=1000000 --periodic_compaction_seconds=0 --prefixpercent=5 --progress_reports=0 --readpercent=45 --recycle_log_file_num=0 --reopen=20 --set_options_one_in=10000 --snapshot_hold_ops=100000 --subcompactions=2 --sync=1 --target_file_size_base=2097152 --target_file_size_multiplier=2 --test_batches_snapshots=1 --use_direct_io_for_flush_and_compaction=0 --use_direct_reads=0 --use_full_merge_v1=0 --use_merge=0 --use_multiget=1 --verify_checksum=1 --verify_checksum_one_in=1000000 --verify_db_one_in=100000 --write_buffer_size=4194304 --write_dbid_to_manifest=1 --writepercent=35 ``` Then repeat the following multiple times, e.g. 100 after compiling with tsan. ``` $./db_test2 --gtest_filter=DBTest2.SwitchMemtableRaceWithNewManifest ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6249 Differential Revision: D19235077 Pulled By: riversand963 fbshipit-source-id: 79467b52f48739ce7c27e440caa2447a40653173
2020-01-07 05:08:24 +01:00
ASSERT_OK(Put("foo", "value"));
const int kL0Files = options.level0_file_num_compaction_trigger;
for (int i = 0; i < kL0Files; ++i) {
ASSERT_OK(Put(/*cf=*/1, "a", std::to_string(i)));
ASSERT_OK(Flush(/*cf=*/1));
}
Fix a data race for cfd->log_number_ (#6249) Summary: A thread calling LogAndApply may release db mutex when calling WriteCurrentStateToManifest() which reads cfd->log_number_. Another thread can call SwitchMemtable() and writes to cfd->log_number_. Solution is to cache the cfd->log_number_ before releasing mutex in LogAndApply. Test Plan (on devserver): ``` $COMPILE_WITH_TSAN=1 make db_stress $./db_stress --acquire_snapshot_one_in=10000 --avoid_unnecessary_blocking_io=1 --block_size=16384 --bloom_bits=16 --bottommost_compression_type=zstd --cache_index_and_filter_blocks=1 --cache_size=1048576 --checkpoint_one_in=1000000 --checksum_type=kxxHash --clear_column_family_one_in=0 --compact_files_one_in=1000000 --compact_range_one_in=1000000 --compaction_ttl=0 --compression_max_dict_bytes=16384 --compression_type=zstd --compression_zstd_max_train_bytes=0 --continuous_verification_interval=0 --db=/dev/shm/rocksdb/rocksdb_crashtest_blackbox --db_write_buffer_size=1048576 --delpercent=5 --delrangepercent=0 --destroy_db_initially=0 --enable_pipelined_write=0 --flush_one_in=1000000 --format_version=5 --get_live_files_and_wal_files_one_in=1000000 --index_block_restart_interval=5 --index_type=0 --log2_keys_per_lock=22 --long_running_snapshots=0 --max_background_compactions=20 --max_bytes_for_level_base=10485760 --max_key=1000000 --max_manifest_file_size=16384 --max_write_batch_group_size_bytes=16 --max_write_buffer_number=3 --memtablerep=skip_list --mmap_read=0 --nooverwritepercent=1 --open_files=500000 --ops_per_thread=100000000 --partition_filters=0 --pause_background_one_in=1000000 --periodic_compaction_seconds=0 --prefixpercent=5 --progress_reports=0 --readpercent=45 --recycle_log_file_num=0 --reopen=20 --set_options_one_in=10000 --snapshot_hold_ops=100000 --subcompactions=2 --sync=1 --target_file_size_base=2097152 --target_file_size_multiplier=2 --test_batches_snapshots=1 --use_direct_io_for_flush_and_compaction=0 --use_direct_reads=0 --use_full_merge_v1=0 --use_merge=0 --use_multiget=1 --verify_checksum=1 --verify_checksum_one_in=1000000 --verify_db_one_in=100000 --write_buffer_size=4194304 --write_dbid_to_manifest=1 --writepercent=35 ``` Then repeat the following multiple times, e.g. 100 after compiling with tsan. ``` $./db_test2 --gtest_filter=DBTest2.SwitchMemtableRaceWithNewManifest ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6249 Differential Revision: D19235077 Pulled By: riversand963 fbshipit-source-id: 79467b52f48739ce7c27e440caa2447a40653173
2020-01-07 05:08:24 +01:00
port::Thread thread([&]() { ASSERT_OK(Flush()); });
ASSERT_OK(dbfull()->TEST_WaitForCompact());
Fix a data race for cfd->log_number_ (#6249) Summary: A thread calling LogAndApply may release db mutex when calling WriteCurrentStateToManifest() which reads cfd->log_number_. Another thread can call SwitchMemtable() and writes to cfd->log_number_. Solution is to cache the cfd->log_number_ before releasing mutex in LogAndApply. Test Plan (on devserver): ``` $COMPILE_WITH_TSAN=1 make db_stress $./db_stress --acquire_snapshot_one_in=10000 --avoid_unnecessary_blocking_io=1 --block_size=16384 --bloom_bits=16 --bottommost_compression_type=zstd --cache_index_and_filter_blocks=1 --cache_size=1048576 --checkpoint_one_in=1000000 --checksum_type=kxxHash --clear_column_family_one_in=0 --compact_files_one_in=1000000 --compact_range_one_in=1000000 --compaction_ttl=0 --compression_max_dict_bytes=16384 --compression_type=zstd --compression_zstd_max_train_bytes=0 --continuous_verification_interval=0 --db=/dev/shm/rocksdb/rocksdb_crashtest_blackbox --db_write_buffer_size=1048576 --delpercent=5 --delrangepercent=0 --destroy_db_initially=0 --enable_pipelined_write=0 --flush_one_in=1000000 --format_version=5 --get_live_files_and_wal_files_one_in=1000000 --index_block_restart_interval=5 --index_type=0 --log2_keys_per_lock=22 --long_running_snapshots=0 --max_background_compactions=20 --max_bytes_for_level_base=10485760 --max_key=1000000 --max_manifest_file_size=16384 --max_write_batch_group_size_bytes=16 --max_write_buffer_number=3 --memtablerep=skip_list --mmap_read=0 --nooverwritepercent=1 --open_files=500000 --ops_per_thread=100000000 --partition_filters=0 --pause_background_one_in=1000000 --periodic_compaction_seconds=0 --prefixpercent=5 --progress_reports=0 --readpercent=45 --recycle_log_file_num=0 --reopen=20 --set_options_one_in=10000 --snapshot_hold_ops=100000 --subcompactions=2 --sync=1 --target_file_size_base=2097152 --target_file_size_multiplier=2 --test_batches_snapshots=1 --use_direct_io_for_flush_and_compaction=0 --use_direct_reads=0 --use_full_merge_v1=0 --use_merge=0 --use_multiget=1 --verify_checksum=1 --verify_checksum_one_in=1000000 --verify_db_one_in=100000 --write_buffer_size=4194304 --write_dbid_to_manifest=1 --writepercent=35 ``` Then repeat the following multiple times, e.g. 100 after compiling with tsan. ``` $./db_test2 --gtest_filter=DBTest2.SwitchMemtableRaceWithNewManifest ``` Pull Request resolved: https://github.com/facebook/rocksdb/pull/6249 Differential Revision: D19235077 Pulled By: riversand963 fbshipit-source-id: 79467b52f48739ce7c27e440caa2447a40653173
2020-01-07 05:08:24 +01:00
thread.join();
}
TEST_F(DBTest2, SameSmallestInSameLevel) {
// This test validates fractional casacading logic when several files at one
// one level only contains the same user key.
Options options = CurrentOptions();
options.merge_operator = MergeOperators::CreateStringAppendOperator();
DestroyAndReopen(options);
ASSERT_OK(Put("key", "1"));
ASSERT_OK(Put("key", "2"));
ASSERT_OK(db_->Merge(WriteOptions(), "key", "3"));
ASSERT_OK(db_->Merge(WriteOptions(), "key", "4"));
ASSERT_OK(Flush());
CompactRangeOptions cro;
cro.change_level = true;
cro.target_level = 2;
ASSERT_OK(dbfull()->CompactRange(cro, db_->DefaultColumnFamily(), nullptr,
nullptr));
ASSERT_OK(db_->Merge(WriteOptions(), "key", "5"));
Flush();
ASSERT_OK(db_->Merge(WriteOptions(), "key", "6"));
Flush();
ASSERT_OK(db_->Merge(WriteOptions(), "key", "7"));
Flush();
ASSERT_OK(db_->Merge(WriteOptions(), "key", "8"));
Flush();
dbfull()->TEST_WaitForCompact(true);
#ifndef ROCKSDB_LITE
ASSERT_EQ("0,4,1", FilesPerLevel());
#endif // ROCKSDB_LITE
ASSERT_EQ("2,3,4,5,6,7,8", Get("key"));
}
TEST_F(DBTest2, FileConsistencyCheckInOpen) {
Put("foo", "bar");
Flush();
SyncPoint::GetInstance()->SetCallBack(
"VersionBuilder::CheckConsistencyBeforeReturn", [&](void* arg) {
Status* ret_s = static_cast<Status*>(arg);
*ret_s = Status::Corruption("fcc");
});
SyncPoint::GetInstance()->EnableProcessing();
Options options = CurrentOptions();
options.force_consistency_checks = true;
ASSERT_NOK(TryReopen(options));
SyncPoint::GetInstance()->DisableProcessing();
}
TEST_F(DBTest2, BlockBasedTablePrefixIndexSeekForPrev) {
// create a DB with block prefix index
BlockBasedTableOptions table_options;
Options options = CurrentOptions();
table_options.block_size = 300;
table_options.index_type = BlockBasedTableOptions::kHashSearch;
table_options.index_shortening =
BlockBasedTableOptions::IndexShorteningMode::kNoShortening;
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
options.prefix_extractor.reset(NewFixedPrefixTransform(1));
Reopen(options);
Random rnd(301);
std::string large_value = rnd.RandomString(500);
ASSERT_OK(Put("a1", large_value));
ASSERT_OK(Put("x1", large_value));
ASSERT_OK(Put("y1", large_value));
Flush();
{
std::unique_ptr<Iterator> iterator(db_->NewIterator(ReadOptions()));
iterator->SeekForPrev("x3");
ASSERT_TRUE(iterator->Valid());
ASSERT_EQ("x1", iterator->key().ToString());
iterator->SeekForPrev("a3");
ASSERT_TRUE(iterator->Valid());
ASSERT_EQ("a1", iterator->key().ToString());
iterator->SeekForPrev("y3");
ASSERT_TRUE(iterator->Valid());
ASSERT_EQ("y1", iterator->key().ToString());
// Query more than one non-existing prefix to cover the case both
// of empty hash bucket and hash bucket conflict.
iterator->SeekForPrev("b1");
// Result should be not valid or "a1".
if (iterator->Valid()) {
ASSERT_EQ("a1", iterator->key().ToString());
}
iterator->SeekForPrev("c1");
// Result should be not valid or "a1".
if (iterator->Valid()) {
ASSERT_EQ("a1", iterator->key().ToString());
}
iterator->SeekForPrev("d1");
// Result should be not valid or "a1".
if (iterator->Valid()) {
ASSERT_EQ("a1", iterator->key().ToString());
}
iterator->SeekForPrev("y3");
ASSERT_TRUE(iterator->Valid());
ASSERT_EQ("y1", iterator->key().ToString());
}
}
TEST_F(DBTest2, PartitionedIndexPrefetchFailure) {
Options options = last_options_;
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566) Summary: This PR does a few things: 1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation. 2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed: - The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated - The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory). 3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10). I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged. Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently. Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :) Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566 Reviewed By: zhichao-cao Differential Revision: D24408980 Pulled By: jay-zhuang fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
2020-10-27 18:31:34 +01:00
options.env = env_;
options.max_open_files = 20;
BlockBasedTableOptions bbto;
bbto.index_type = BlockBasedTableOptions::IndexType::kTwoLevelIndexSearch;
bbto.metadata_block_size = 128;
bbto.block_size = 128;
bbto.block_cache = NewLRUCache(16777216);
bbto.cache_index_and_filter_blocks = true;
options.table_factory.reset(NewBlockBasedTableFactory(bbto));
DestroyAndReopen(options);
// Force no table cache so every read will preload the SST file.
dbfull()->TEST_table_cache()->SetCapacity(0);
bbto.block_cache->SetCapacity(0);
Random rnd(301);
for (int i = 0; i < 4096; i++) {
ASSERT_OK(Put(Key(i), rnd.RandomString(32)));
}
ASSERT_OK(Flush());
// Try different random failures in table open for 300 times.
for (int i = 0; i < 300; i++) {
env_->num_reads_fails_ = 0;
env_->rand_reads_fail_odd_ = 8;
std::string value;
Status s = dbfull()->Get(ReadOptions(), Key(1), &value);
if (env_->num_reads_fails_ > 0) {
ASSERT_NOK(s);
} else {
ASSERT_OK(s);
}
}
env_->rand_reads_fail_odd_ = 0;
}
TEST_F(DBTest2, ChangePrefixExtractor) {
for (bool use_partitioned_filter : {true, false}) {
// create a DB with block prefix index
BlockBasedTableOptions table_options;
Options options = CurrentOptions();
// Sometimes filter is checked based on upper bound. Assert counters
// for that case. Otherwise, only check data correctness.
#ifndef ROCKSDB_LITE
bool expect_filter_check = !use_partitioned_filter;
#else
bool expect_filter_check = false;
#endif
table_options.partition_filters = use_partitioned_filter;
if (use_partitioned_filter) {
table_options.index_type =
BlockBasedTableOptions::IndexType::kTwoLevelIndexSearch;
}
table_options.filter_policy.reset(NewBloomFilterPolicy(10, false));
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
options.statistics = CreateDBStatistics();
options.prefix_extractor.reset(NewFixedPrefixTransform(2));
DestroyAndReopen(options);
Random rnd(301);
ASSERT_OK(Put("aa", ""));
ASSERT_OK(Put("xb", ""));
ASSERT_OK(Put("xx1", ""));
ASSERT_OK(Put("xz1", ""));
ASSERT_OK(Put("zz", ""));
Flush();
// After reopening DB with prefix size 2 => 1, prefix extractor
// won't take effective unless it won't change results based
// on upper bound and seek key.
options.prefix_extractor.reset(NewFixedPrefixTransform(1));
Reopen(options);
{
std::unique_ptr<Iterator> iterator(db_->NewIterator(ReadOptions()));
iterator->Seek("xa");
ASSERT_TRUE(iterator->Valid());
ASSERT_EQ("xb", iterator->key().ToString());
// It's a bug that the counter BLOOM_FILTER_PREFIX_CHECKED is not
// correct in this case. So don't check counters in this case.
if (expect_filter_check) {
ASSERT_EQ(0, TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED));
}
iterator->Seek("xz");
ASSERT_TRUE(iterator->Valid());
ASSERT_EQ("xz1", iterator->key().ToString());
if (expect_filter_check) {
ASSERT_EQ(0, TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED));
}
}
std::string ub_str = "xg9";
Slice ub(ub_str);
ReadOptions ro;
ro.iterate_upper_bound = &ub;
{
std::unique_ptr<Iterator> iterator(db_->NewIterator(ro));
// SeekForPrev() never uses prefix bloom if it is changed.
iterator->SeekForPrev("xg0");
ASSERT_TRUE(iterator->Valid());
ASSERT_EQ("xb", iterator->key().ToString());
if (expect_filter_check) {
ASSERT_EQ(0, TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED));
}
}
ub_str = "xx9";
ub = Slice(ub_str);
{
std::unique_ptr<Iterator> iterator(db_->NewIterator(ro));
iterator->Seek("x");
ASSERT_TRUE(iterator->Valid());
ASSERT_EQ("xb", iterator->key().ToString());
if (expect_filter_check) {
ASSERT_EQ(0, TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED));
}
iterator->Seek("xx0");
ASSERT_TRUE(iterator->Valid());
ASSERT_EQ("xx1", iterator->key().ToString());
if (expect_filter_check) {
ASSERT_EQ(1, TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED));
}
}
CompactRangeOptions compact_range_opts;
compact_range_opts.bottommost_level_compaction =
BottommostLevelCompaction::kForce;
ASSERT_OK(db_->CompactRange(compact_range_opts, nullptr, nullptr));
ASSERT_OK(db_->CompactRange(compact_range_opts, nullptr, nullptr));
// Re-execute similar queries after a full compaction
{
std::unique_ptr<Iterator> iterator(db_->NewIterator(ReadOptions()));
iterator->Seek("x");
ASSERT_TRUE(iterator->Valid());
ASSERT_EQ("xb", iterator->key().ToString());
if (expect_filter_check) {
ASSERT_EQ(2, TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED));
}
iterator->Seek("xg");
ASSERT_TRUE(iterator->Valid());
ASSERT_EQ("xx1", iterator->key().ToString());
if (expect_filter_check) {
ASSERT_EQ(3, TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED));
}
iterator->Seek("xz");
ASSERT_TRUE(iterator->Valid());
ASSERT_EQ("xz1", iterator->key().ToString());
if (expect_filter_check) {
ASSERT_EQ(4, TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED));
}
}
{
std::unique_ptr<Iterator> iterator(db_->NewIterator(ro));
iterator->SeekForPrev("xx0");
ASSERT_TRUE(iterator->Valid());
ASSERT_EQ("xb", iterator->key().ToString());
if (expect_filter_check) {
ASSERT_EQ(5, TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED));
}
iterator->Seek("xx0");
ASSERT_TRUE(iterator->Valid());
ASSERT_EQ("xx1", iterator->key().ToString());
if (expect_filter_check) {
ASSERT_EQ(6, TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED));
}
}
ub_str = "xg9";
ub = Slice(ub_str);
{
std::unique_ptr<Iterator> iterator(db_->NewIterator(ro));
iterator->SeekForPrev("xg0");
ASSERT_TRUE(iterator->Valid());
ASSERT_EQ("xb", iterator->key().ToString());
if (expect_filter_check) {
ASSERT_EQ(7, TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED));
}
}
}
}
TEST_F(DBTest2, BlockBasedTablePrefixGetIndexNotFound) {
// create a DB with block prefix index
BlockBasedTableOptions table_options;
Options options = CurrentOptions();
table_options.block_size = 300;
table_options.index_type = BlockBasedTableOptions::kHashSearch;
table_options.index_shortening =
BlockBasedTableOptions::IndexShorteningMode::kNoShortening;
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
options.prefix_extractor.reset(NewFixedPrefixTransform(1));
options.level0_file_num_compaction_trigger = 8;
Reopen(options);
ASSERT_OK(Put("b1", "ok"));
Flush();
// Flushing several files so that the chance that hash bucket
// is empty fo "b" in at least one of the files is high.
ASSERT_OK(Put("a1", ""));
ASSERT_OK(Put("c1", ""));
Flush();
ASSERT_OK(Put("a2", ""));
ASSERT_OK(Put("c2", ""));
Flush();
ASSERT_OK(Put("a3", ""));
ASSERT_OK(Put("c3", ""));
Flush();
ASSERT_OK(Put("a4", ""));
ASSERT_OK(Put("c4", ""));
Flush();
ASSERT_OK(Put("a5", ""));
ASSERT_OK(Put("c5", ""));
Flush();
ASSERT_EQ("ok", Get("b1"));
}
#ifndef ROCKSDB_LITE
TEST_F(DBTest2, AutoPrefixMode1) {
// create a DB with block prefix index
BlockBasedTableOptions table_options;
Options options = CurrentOptions();
table_options.filter_policy.reset(NewBloomFilterPolicy(10, false));
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
options.prefix_extractor.reset(NewFixedPrefixTransform(1));
options.statistics = CreateDBStatistics();
Reopen(options);
Random rnd(301);
std::string large_value = rnd.RandomString(500);
ASSERT_OK(Put("a1", large_value));
ASSERT_OK(Put("x1", large_value));
ASSERT_OK(Put("y1", large_value));
Flush();
ReadOptions ro;
ro.total_order_seek = false;
ro.auto_prefix_mode = true;
{
std::unique_ptr<Iterator> iterator(db_->NewIterator(ro));
iterator->Seek("b1");
ASSERT_TRUE(iterator->Valid());
ASSERT_EQ("x1", iterator->key().ToString());
ASSERT_EQ(0, TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED));
}
std::string ub_str = "b9";
Slice ub(ub_str);
ro.iterate_upper_bound = &ub;
{
std::unique_ptr<Iterator> iterator(db_->NewIterator(ro));
iterator->Seek("b1");
ASSERT_FALSE(iterator->Valid());
ASSERT_EQ(1, TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED));
}
ub_str = "z";
ub = Slice(ub_str);
{
std::unique_ptr<Iterator> iterator(db_->NewIterator(ro));
iterator->Seek("b1");
ASSERT_TRUE(iterator->Valid());
ASSERT_EQ("x1", iterator->key().ToString());
ASSERT_EQ(1, TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED));
}
ub_str = "c";
ub = Slice(ub_str);
{
std::unique_ptr<Iterator> iterator(db_->NewIterator(ro));
iterator->Seek("b1");
ASSERT_FALSE(iterator->Valid());
ASSERT_EQ(2, TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED));
}
// The same queries without recreating iterator
{
ub_str = "b9";
ub = Slice(ub_str);
ro.iterate_upper_bound = &ub;
std::unique_ptr<Iterator> iterator(db_->NewIterator(ro));
iterator->Seek("b1");
ASSERT_FALSE(iterator->Valid());
ASSERT_EQ(3, TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED));
ub_str = "z";
ub = Slice(ub_str);
iterator->Seek("b1");
ASSERT_TRUE(iterator->Valid());
ASSERT_EQ("x1", iterator->key().ToString());
ASSERT_EQ(3, TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED));
ub_str = "c";
ub = Slice(ub_str);
iterator->Seek("b1");
ASSERT_FALSE(iterator->Valid());
ASSERT_EQ(4, TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED));
ub_str = "b9";
ub = Slice(ub_str);
ro.iterate_upper_bound = &ub;
iterator->SeekForPrev("b1");
ASSERT_TRUE(iterator->Valid());
ASSERT_EQ("a1", iterator->key().ToString());
ASSERT_EQ(4, TestGetTickerCount(options, BLOOM_FILTER_PREFIX_CHECKED));
ub_str = "zz";
ub = Slice(ub_str);
ro.iterate_upper_bound = &ub;
iterator->SeekToLast();
ASSERT_TRUE(iterator->Valid());
ASSERT_EQ("y1", iterator->key().ToString());
iterator->SeekToFirst();
ASSERT_TRUE(iterator->Valid());
ASSERT_EQ("a1", iterator->key().ToString());
}
}
Handle rename() failure in non-local FS (#8192) Summary: In a distributed environment, a file `rename()` operation can succeed on server (remote) side, but the client can somehow return non-ok status to RocksDB. Possible reasons include network partition, connection issue, etc. This happens in `rocksdb::SetCurrentFile()`, which can be called in `LogAndApply() -> ProcessManifestWrites()` if RocksDB tries to switch to a new MANIFEST. We currently always delete the new MANIFEST if an error occurs. This is problematic in distributed world. If the server-side successfully updates the CURRENT file via renaming, then a subsequent `DB::Open()` will try to look for the new MANIFEST and fail. As a fix, we can track the execution result of IO operations on the new MANIFEST. - If IO operations on the new MANIFEST fail, then we know the CURRENT must point to the original MANIFEST. Therefore, it is safe to remove the new MANIFEST. - If IO operations on the new MANIFEST all succeed, but somehow we end up in the clean up code block, then we do not know whether CURRENT points to the new or old MANIFEST. (For local POSIX-compliant FS, it should still point to old MANIFEST, but it does not matter if we keep the new MANIFEST.) Therefore, we keep the new MANIFEST. - Any future `LogAndApply()` will switch to a new MANIFEST and update CURRENT. - If process reopens the db immediately after the failure, then the CURRENT file can point to either the new MANIFEST or the old one, both of which exist. Therefore, recovery can succeed and ignore the other. Pull Request resolved: https://github.com/facebook/rocksdb/pull/8192 Test Plan: make check Reviewed By: zhichao-cao Differential Revision: D27804648 Pulled By: riversand963 fbshipit-source-id: 9c16f2a5ce41bc6aadf085e48449b19ede8423e4
2021-04-20 03:10:23 +02:00
class RenameCurrentTest : public DBTestBase,
public testing::WithParamInterface<std::string> {
public:
RenameCurrentTest()
: DBTestBase("rename_current_test", /*env_do_fsync=*/true),
sync_point_(GetParam()) {}
~RenameCurrentTest() override {}
void SetUp() override {
env_->no_file_overwrite_.store(true, std::memory_order_release);
}
void TearDown() override {
env_->no_file_overwrite_.store(false, std::memory_order_release);
}
void SetupSyncPoints() {
SyncPoint::GetInstance()->DisableProcessing();
SyncPoint::GetInstance()->SetCallBack(sync_point_, [&](void* arg) {
Status* s = reinterpret_cast<Status*>(arg);
assert(s);
*s = Status::IOError("Injected IO error.");
});
}
const std::string sync_point_;
};
INSTANTIATE_TEST_CASE_P(DistributedFS, RenameCurrentTest,
::testing::Values("SetCurrentFile:BeforeRename",
"SetCurrentFile:AfterRename"));
TEST_P(RenameCurrentTest, Open) {
Destroy(last_options_);
Options options = GetDefaultOptions();
options.create_if_missing = true;
SetupSyncPoints();
SyncPoint::GetInstance()->EnableProcessing();
Status s = TryReopen(options);
ASSERT_NOK(s);
SyncPoint::GetInstance()->DisableProcessing();
Reopen(options);
}
TEST_P(RenameCurrentTest, Flush) {
Destroy(last_options_);
Options options = GetDefaultOptions();
options.max_manifest_file_size = 1;
options.create_if_missing = true;
Reopen(options);
ASSERT_OK(Put("key", "value"));
SetupSyncPoints();
SyncPoint::GetInstance()->EnableProcessing();
ASSERT_NOK(Flush());
ASSERT_NOK(Put("foo", "value"));
SyncPoint::GetInstance()->DisableProcessing();
Reopen(options);
ASSERT_EQ("value", Get("key"));
ASSERT_EQ("NOT_FOUND", Get("foo"));
}
TEST_P(RenameCurrentTest, Compaction) {
Destroy(last_options_);
Options options = GetDefaultOptions();
options.max_manifest_file_size = 1;
options.create_if_missing = true;
Reopen(options);
ASSERT_OK(Put("a", "a_value"));
ASSERT_OK(Put("c", "c_value"));
ASSERT_OK(Flush());
ASSERT_OK(Put("b", "b_value"));
ASSERT_OK(Put("d", "d_value"));
ASSERT_OK(Flush());
SetupSyncPoints();
SyncPoint::GetInstance()->EnableProcessing();
ASSERT_NOK(db_->CompactRange(CompactRangeOptions(), /*begin=*/nullptr,
/*end=*/nullptr));
ASSERT_NOK(Put("foo", "value"));
SyncPoint::GetInstance()->DisableProcessing();
Reopen(options);
ASSERT_EQ("NOT_FOUND", Get("foo"));
ASSERT_EQ("d_value", Get("d"));
}
TEST_F(DBTest2, BottommostTemperature) {
Options options = CurrentOptions();
options.bottommost_temperature = Temperature::kWarm;
options.level0_file_num_compaction_trigger = 2;
Reopen(options);
ASSERT_OK(Put("foo", "bar"));
ASSERT_OK(Put("bar", "bar"));
ASSERT_OK(Flush());
ASSERT_OK(Put("foo", "bar"));
ASSERT_OK(Put("bar", "bar"));
ASSERT_OK(Flush());
ASSERT_OK(dbfull()->TEST_WaitForCompact());
Reopen(options);
ColumnFamilyMetaData metadata;
db_->GetColumnFamilyMetaData(&metadata);
ASSERT_EQ(1, metadata.file_count);
ASSERT_EQ(Temperature::kWarm, metadata.levels[1].files[0].temperature);
}
#endif // ROCKSDB_LITE
// WAL recovery mode is WALRecoveryMode::kPointInTimeRecovery.
TEST_F(DBTest2, PointInTimeRecoveryWithIOErrorWhileReadingWal) {
Options options = CurrentOptions();
DestroyAndReopen(options);
ASSERT_OK(Put("foo", "value0"));
Close();
SyncPoint::GetInstance()->DisableProcessing();
SyncPoint::GetInstance()->ClearAllCallBacks();
bool should_inject_error = false;
SyncPoint::GetInstance()->SetCallBack(
"DBImpl::RecoverLogFiles:BeforeReadWal",
[&](void* /*arg*/) { should_inject_error = true; });
SyncPoint::GetInstance()->SetCallBack(
"LogReader::ReadMore:AfterReadFile", [&](void* arg) {
if (should_inject_error) {
ASSERT_NE(nullptr, arg);
*reinterpret_cast<Status*>(arg) = Status::IOError("Injected IOError");
}
});
SyncPoint::GetInstance()->EnableProcessing();
options.avoid_flush_during_recovery = true;
options.wal_recovery_mode = WALRecoveryMode::kPointInTimeRecovery;
Status s = TryReopen(options);
ASSERT_TRUE(s.IsIOError());
}
TEST_F(DBTest2, PointInTimeRecoveryWithSyncFailureInCFCreation) {
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency(
{{"DBImpl::BackgroundCallFlush:Start:1",
"PointInTimeRecoveryWithSyncFailureInCFCreation:1"},
{"PointInTimeRecoveryWithSyncFailureInCFCreation:2",
"DBImpl::BackgroundCallFlush:Start:2"}});
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
CreateColumnFamilies({"test1"}, Options());
ASSERT_OK(Put("foo", "bar"));
// Creating a CF when a flush is going on, log is synced but the
// closed log file is not synced and corrupted.
port::Thread flush_thread([&]() { ASSERT_NOK(Flush()); });
TEST_SYNC_POINT("PointInTimeRecoveryWithSyncFailureInCFCreation:1");
CreateColumnFamilies({"test2"}, Options());
env_->corrupt_in_sync_ = true;
TEST_SYNC_POINT("PointInTimeRecoveryWithSyncFailureInCFCreation:2");
flush_thread.join();
env_->corrupt_in_sync_ = false;
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
// Reopening the DB should not corrupt anything
Options options = CurrentOptions();
options.wal_recovery_mode = WALRecoveryMode::kPointInTimeRecovery;
ReopenWithColumnFamilies({"default", "test1", "test2"}, options);
}
TEST_F(DBTest2, RenameDirectory) {
Options options = CurrentOptions();
DestroyAndReopen(options);
ASSERT_OK(Put("foo", "value0"));
Close();
auto old_dbname = dbname_;
auto new_dbname = dbname_ + "_2";
EXPECT_OK(env_->RenameFile(dbname_, new_dbname));
options.create_if_missing = false;
dbname_ = new_dbname;
ASSERT_OK(TryReopen(options));
ASSERT_EQ("value0", Get("foo"));
Destroy(options);
dbname_ = old_dbname;
}
} // namespace ROCKSDB_NAMESPACE
#ifdef ROCKSDB_UNITTESTS_WITH_CUSTOM_OBJECTS_FROM_STATIC_LIBS
extern "C" {
void RegisterCustomObjects(int argc, char** argv);
}
#else
void RegisterCustomObjects(int /*argc*/, char** /*argv*/) {}
#endif // !ROCKSDB_UNITTESTS_WITH_CUSTOM_OBJECTS_FROM_STATIC_LIBS
int main(int argc, char** argv) {
ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
::testing::InitGoogleTest(&argc, argv);
RegisterCustomObjects(argc, argv);
return RUN_ALL_TESTS();
}