2016-02-10 00:12:00 +01:00
|
|
|
// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
|
2017-07-16 01:03:42 +02:00
|
|
|
// This source code is licensed under both the GPLv2 (found in the
|
|
|
|
// COPYING file in the root directory) and Apache 2.0 License
|
|
|
|
// (found in the LICENSE.Apache file in the root directory).
|
2013-10-16 23:59:46 +02:00
|
|
|
//
|
2013-12-11 04:03:13 +01:00
|
|
|
#include "db/memtable_list.h"
|
2012-11-29 01:42:36 +01:00
|
|
|
|
2014-10-28 19:54:33 +01:00
|
|
|
#ifndef __STDC_FORMAT_MACROS
|
|
|
|
#define __STDC_FORMAT_MACROS
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#include <inttypes.h>
|
2017-10-24 00:22:05 +02:00
|
|
|
#include <limits>
|
2018-10-16 04:59:20 +02:00
|
|
|
#include <queue>
|
2012-10-19 23:00:53 +02:00
|
|
|
#include <string>
|
Skip deleted WALs during recovery
Summary:
This patch record min log number to keep to the manifest while flushing SST files to ignore them and any WAL older than them during recovery. This is to avoid scenarios when we have a gap between the WAL files are fed to the recovery procedure. The gap could happen by for example out-of-order WAL deletion. Such gap could cause problems in 2PC recovery where the prepared and commit entry are placed into two separate WAL and gap in the WALs could result into not processing the WAL with the commit entry and hence breaking the 2PC recovery logic.
Before the commit, for 2PC case, we determined which log number to keep in FindObsoleteFiles(). We looked at the earliest logs with outstanding prepare entries, or prepare entries whose respective commit or abort are in memtable. With the commit, the same calculation is done while we apply the SST flush. Just before installing the flush file, we precompute the earliest log file to keep after the flush finishes using the same logic (but skipping the memtables just flushed), record this information to the manifest entry for this new flushed SST file. This pre-computed value is also remembered in memory, and will later be used to determine whether a log file can be deleted. This value is unlikely to change until next flush because the commit entry will stay in memtable. (In WritePrepared, we could have removed the older log files as soon as all prepared entries are committed. It's not yet done anyway. Even if we do it, the only thing we loss with this new approach is earlier log deletion between two flushes, which does not guarantee to happen anyway because the obsolete file clean-up function is only executed after flush or compaction)
This min log number to keep is stored in the manifest using the safely-ignore customized field of AddFile entry, in order to guarantee that the DB generated using newer release can be opened by previous releases no older than 4.2.
Closes https://github.com/facebook/rocksdb/pull/3765
Differential Revision: D7747618
Pulled By: siying
fbshipit-source-id: d00c92105b4f83852e9754a1b70d6b64cb590729
2018-05-04 00:35:11 +02:00
|
|
|
#include "db/db_impl.h"
|
2012-10-19 23:00:53 +02:00
|
|
|
#include "db/memtable.h"
|
2018-11-29 00:26:56 +01:00
|
|
|
#include "db/range_tombstone_fragmenter.h"
|
2014-01-24 23:30:28 +01:00
|
|
|
#include "db/version_set.h"
|
2017-04-06 04:02:00 +02:00
|
|
|
#include "monitoring/thread_status_util.h"
|
Fix flush not being commit while writing manifest
Summary:
Fix flush not being commit while writing manifest, which is a recent bug introduced by D60075.
The issue:
# Options.max_background_flushes > 1
# Background thread A pick up a flush job, flush, then commit to manifest. (Note that mutex is released before writing manifest.)
# Background thread B pick up another flush job, flush. When it gets to `MemTableList::InstallMemtableFlushResults`, it notices another thread is commiting, so it quit.
# After the first commit, thread A doesn't double check if there are more flush result need to commit, leaving the second flush uncommitted.
Test Plan: run the test. Also verify the new test hit deadlock without the fix.
Reviewers: sdong, igor, lightmark
Reviewed By: lightmark
Subscribers: andrewkr, omegaga, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D60969
2016-07-21 19:10:41 +02:00
|
|
|
#include "rocksdb/db.h"
|
2013-08-23 17:38:13 +02:00
|
|
|
#include "rocksdb/env.h"
|
|
|
|
#include "rocksdb/iterator.h"
|
2017-02-03 01:38:40 +01:00
|
|
|
#include "table/merging_iterator.h"
|
2019-05-31 02:39:43 +02:00
|
|
|
#include "test_util/sync_point.h"
|
2012-10-19 23:00:53 +02:00
|
|
|
#include "util/coding.h"
|
2014-04-03 06:49:51 +02:00
|
|
|
#include "util/log_buffer.h"
|
2012-10-19 23:00:53 +02:00
|
|
|
|
2013-10-04 06:49:15 +02:00
|
|
|
namespace rocksdb {
|
2012-10-19 23:00:53 +02:00
|
|
|
|
|
|
|
class InternalKeyComparator;
|
|
|
|
class Mutex;
|
|
|
|
class VersionSet;
|
|
|
|
|
2015-08-19 22:32:09 +02:00
|
|
|
void MemTableListVersion::AddMemTable(MemTable* m) {
|
|
|
|
memlist_.push_front(m);
|
|
|
|
*parent_memtable_list_memory_usage_ += m->ApproximateMemoryUsage();
|
|
|
|
}
|
|
|
|
|
|
|
|
void MemTableListVersion::UnrefMemTable(autovector<MemTable*>* to_delete,
|
|
|
|
MemTable* m) {
|
|
|
|
if (m->Unref()) {
|
|
|
|
to_delete->push_back(m);
|
|
|
|
assert(*parent_memtable_list_memory_usage_ >= m->ApproximateMemoryUsage());
|
|
|
|
*parent_memtable_list_memory_usage_ -= m->ApproximateMemoryUsage();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
MemTableListVersion::MemTableListVersion(
|
|
|
|
size_t* parent_memtable_list_memory_usage, MemTableListVersion* old)
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
: max_write_buffer_number_to_maintain_(
|
2015-08-19 22:32:09 +02:00
|
|
|
old->max_write_buffer_number_to_maintain_),
|
|
|
|
parent_memtable_list_memory_usage_(parent_memtable_list_memory_usage) {
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
if (old != nullptr) {
|
|
|
|
memlist_ = old->memlist_;
|
|
|
|
for (auto& m : memlist_) {
|
|
|
|
m->Ref();
|
|
|
|
}
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
|
|
|
|
memlist_history_ = old->memlist_history_;
|
|
|
|
for (auto& m : memlist_history_) {
|
|
|
|
m->Ref();
|
|
|
|
}
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
MemTableListVersion::MemTableListVersion(
|
2015-08-19 22:32:09 +02:00
|
|
|
size_t* parent_memtable_list_memory_usage,
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
int max_write_buffer_number_to_maintain)
|
2015-08-19 22:32:09 +02:00
|
|
|
: max_write_buffer_number_to_maintain_(max_write_buffer_number_to_maintain),
|
|
|
|
parent_memtable_list_memory_usage_(parent_memtable_list_memory_usage) {}
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
void MemTableListVersion::Ref() { ++refs_; }
|
|
|
|
|
2015-08-19 22:32:09 +02:00
|
|
|
// called by superversion::clean()
|
2014-01-28 19:35:48 +01:00
|
|
|
void MemTableListVersion::Unref(autovector<MemTable*>* to_delete) {
|
2014-01-25 22:50:30 +01:00
|
|
|
assert(refs_ >= 1);
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
--refs_;
|
|
|
|
if (refs_ == 0) {
|
|
|
|
// if to_delete is equal to nullptr it means we're confident
|
|
|
|
// that refs_ will not be zero
|
|
|
|
assert(to_delete != nullptr);
|
|
|
|
for (const auto& m : memlist_) {
|
2015-08-19 22:32:09 +02:00
|
|
|
UnrefMemTable(to_delete, m);
|
2013-11-25 20:55:36 +01:00
|
|
|
}
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
for (const auto& m : memlist_history_) {
|
2015-08-19 22:32:09 +02:00
|
|
|
UnrefMemTable(to_delete, m);
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
}
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
delete this;
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
int MemTableList::NumNotFlushed() const {
|
|
|
|
int size = static_cast<int>(current_->memlist_.size());
|
|
|
|
assert(num_flush_not_started_ <= size);
|
|
|
|
return size;
|
|
|
|
}
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
int MemTableList::NumFlushed() const {
|
|
|
|
return static_cast<int>(current_->memlist_history_.size());
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
// Search all the memtables starting from the most recent one.
|
|
|
|
// Return the most recent value found, if any.
|
|
|
|
// Operands stores the list of merge operations to apply, so far.
|
2017-01-08 23:08:51 +01:00
|
|
|
bool MemTableListVersion::Get(const LookupKey& key, std::string* value,
|
2015-05-29 23:36:35 +02:00
|
|
|
Status* s, MergeContext* merge_context,
|
Use only "local" range tombstones during Get (#4449)
Summary:
Previously, range tombstones were accumulated from every level, which
was necessary if a range tombstone in a higher level covered a key in a lower
level. However, RangeDelAggregator::AddTombstones's complexity is based on
the number of tombstones that are currently stored in it, which is wasteful in
the Get case, where we only need to know the highest sequence number of range
tombstones that cover the key from higher levels, and compute the highest covering
sequence number at the current level. This change introduces this optimization, and
removes the use of RangeDelAggregator from the Get path.
In the benchmark results, the following command was used to initialize the database:
```
./db_bench -db=/dev/shm/5k-rts -use_existing_db=false -benchmarks=filluniquerandom -write_buffer_size=1048576 -compression_type=lz4 -target_file_size_base=1048576 -max_bytes_for_level_base=4194304 -value_size=112 -key_size=16 -block_size=4096 -level_compaction_dynamic_level_bytes=true -num=5000000 -max_background_jobs=12 -benchmark_write_rate_limit=20971520 -range_tombstone_width=100 -writes_per_range_tombstone=100 -max_num_range_tombstones=50000 -bloom_bits=8
```
...and the following command was used to measure read throughput:
```
./db_bench -db=/dev/shm/5k-rts/ -use_existing_db=true -benchmarks=readrandom -disable_auto_compactions=true -num=5000000 -reads=100000 -threads=32
```
The filluniquerandom command was only run once, and the resulting database was used
to measure read performance before and after the PR. Both binaries were compiled with
`DEBUG_LEVEL=0`.
Readrandom results before PR:
```
readrandom : 4.544 micros/op 220090 ops/sec; 16.9 MB/s (63103 of 100000 found)
```
Readrandom results after PR:
```
readrandom : 11.147 micros/op 89707 ops/sec; 6.9 MB/s (63103 of 100000 found)
```
So it's actually slower right now, but this PR paves the way for future optimizations (see #4493).
----
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4449
Differential Revision: D10370575
Pulled By: abhimadan
fbshipit-source-id: 9a2e152be1ef36969055c0e9eb4beb0d96c11f4d
2018-10-24 21:29:29 +02:00
|
|
|
SequenceNumber* max_covering_tombstone_seq,
|
2017-09-11 17:58:52 +02:00
|
|
|
SequenceNumber* seq, const ReadOptions& read_opts,
|
2017-10-03 18:08:07 +02:00
|
|
|
ReadCallback* callback, bool* is_blob_index) {
|
Use only "local" range tombstones during Get (#4449)
Summary:
Previously, range tombstones were accumulated from every level, which
was necessary if a range tombstone in a higher level covered a key in a lower
level. However, RangeDelAggregator::AddTombstones's complexity is based on
the number of tombstones that are currently stored in it, which is wasteful in
the Get case, where we only need to know the highest sequence number of range
tombstones that cover the key from higher levels, and compute the highest covering
sequence number at the current level. This change introduces this optimization, and
removes the use of RangeDelAggregator from the Get path.
In the benchmark results, the following command was used to initialize the database:
```
./db_bench -db=/dev/shm/5k-rts -use_existing_db=false -benchmarks=filluniquerandom -write_buffer_size=1048576 -compression_type=lz4 -target_file_size_base=1048576 -max_bytes_for_level_base=4194304 -value_size=112 -key_size=16 -block_size=4096 -level_compaction_dynamic_level_bytes=true -num=5000000 -max_background_jobs=12 -benchmark_write_rate_limit=20971520 -range_tombstone_width=100 -writes_per_range_tombstone=100 -max_num_range_tombstones=50000 -bloom_bits=8
```
...and the following command was used to measure read throughput:
```
./db_bench -db=/dev/shm/5k-rts/ -use_existing_db=true -benchmarks=readrandom -disable_auto_compactions=true -num=5000000 -reads=100000 -threads=32
```
The filluniquerandom command was only run once, and the resulting database was used
to measure read performance before and after the PR. Both binaries were compiled with
`DEBUG_LEVEL=0`.
Readrandom results before PR:
```
readrandom : 4.544 micros/op 220090 ops/sec; 16.9 MB/s (63103 of 100000 found)
```
Readrandom results after PR:
```
readrandom : 11.147 micros/op 89707 ops/sec; 6.9 MB/s (63103 of 100000 found)
```
So it's actually slower right now, but this PR paves the way for future optimizations (see #4493).
----
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4449
Differential Revision: D10370575
Pulled By: abhimadan
fbshipit-source-id: 9a2e152be1ef36969055c0e9eb4beb0d96c11f4d
2018-10-24 21:29:29 +02:00
|
|
|
return GetFromList(&memlist_, key, value, s, merge_context,
|
|
|
|
max_covering_tombstone_seq, seq, read_opts, callback,
|
|
|
|
is_blob_index);
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
}
|
|
|
|
|
2017-10-18 02:24:25 +02:00
|
|
|
bool MemTableListVersion::GetFromHistory(
|
|
|
|
const LookupKey& key, std::string* value, Status* s,
|
Use only "local" range tombstones during Get (#4449)
Summary:
Previously, range tombstones were accumulated from every level, which
was necessary if a range tombstone in a higher level covered a key in a lower
level. However, RangeDelAggregator::AddTombstones's complexity is based on
the number of tombstones that are currently stored in it, which is wasteful in
the Get case, where we only need to know the highest sequence number of range
tombstones that cover the key from higher levels, and compute the highest covering
sequence number at the current level. This change introduces this optimization, and
removes the use of RangeDelAggregator from the Get path.
In the benchmark results, the following command was used to initialize the database:
```
./db_bench -db=/dev/shm/5k-rts -use_existing_db=false -benchmarks=filluniquerandom -write_buffer_size=1048576 -compression_type=lz4 -target_file_size_base=1048576 -max_bytes_for_level_base=4194304 -value_size=112 -key_size=16 -block_size=4096 -level_compaction_dynamic_level_bytes=true -num=5000000 -max_background_jobs=12 -benchmark_write_rate_limit=20971520 -range_tombstone_width=100 -writes_per_range_tombstone=100 -max_num_range_tombstones=50000 -bloom_bits=8
```
...and the following command was used to measure read throughput:
```
./db_bench -db=/dev/shm/5k-rts/ -use_existing_db=true -benchmarks=readrandom -disable_auto_compactions=true -num=5000000 -reads=100000 -threads=32
```
The filluniquerandom command was only run once, and the resulting database was used
to measure read performance before and after the PR. Both binaries were compiled with
`DEBUG_LEVEL=0`.
Readrandom results before PR:
```
readrandom : 4.544 micros/op 220090 ops/sec; 16.9 MB/s (63103 of 100000 found)
```
Readrandom results after PR:
```
readrandom : 11.147 micros/op 89707 ops/sec; 6.9 MB/s (63103 of 100000 found)
```
So it's actually slower right now, but this PR paves the way for future optimizations (see #4493).
----
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4449
Differential Revision: D10370575
Pulled By: abhimadan
fbshipit-source-id: 9a2e152be1ef36969055c0e9eb4beb0d96c11f4d
2018-10-24 21:29:29 +02:00
|
|
|
MergeContext* merge_context, SequenceNumber* max_covering_tombstone_seq,
|
2017-10-18 02:24:25 +02:00
|
|
|
SequenceNumber* seq, const ReadOptions& read_opts, bool* is_blob_index) {
|
2017-01-08 23:08:51 +01:00
|
|
|
return GetFromList(&memlist_history_, key, value, s, merge_context,
|
Use only "local" range tombstones during Get (#4449)
Summary:
Previously, range tombstones were accumulated from every level, which
was necessary if a range tombstone in a higher level covered a key in a lower
level. However, RangeDelAggregator::AddTombstones's complexity is based on
the number of tombstones that are currently stored in it, which is wasteful in
the Get case, where we only need to know the highest sequence number of range
tombstones that cover the key from higher levels, and compute the highest covering
sequence number at the current level. This change introduces this optimization, and
removes the use of RangeDelAggregator from the Get path.
In the benchmark results, the following command was used to initialize the database:
```
./db_bench -db=/dev/shm/5k-rts -use_existing_db=false -benchmarks=filluniquerandom -write_buffer_size=1048576 -compression_type=lz4 -target_file_size_base=1048576 -max_bytes_for_level_base=4194304 -value_size=112 -key_size=16 -block_size=4096 -level_compaction_dynamic_level_bytes=true -num=5000000 -max_background_jobs=12 -benchmark_write_rate_limit=20971520 -range_tombstone_width=100 -writes_per_range_tombstone=100 -max_num_range_tombstones=50000 -bloom_bits=8
```
...and the following command was used to measure read throughput:
```
./db_bench -db=/dev/shm/5k-rts/ -use_existing_db=true -benchmarks=readrandom -disable_auto_compactions=true -num=5000000 -reads=100000 -threads=32
```
The filluniquerandom command was only run once, and the resulting database was used
to measure read performance before and after the PR. Both binaries were compiled with
`DEBUG_LEVEL=0`.
Readrandom results before PR:
```
readrandom : 4.544 micros/op 220090 ops/sec; 16.9 MB/s (63103 of 100000 found)
```
Readrandom results after PR:
```
readrandom : 11.147 micros/op 89707 ops/sec; 6.9 MB/s (63103 of 100000 found)
```
So it's actually slower right now, but this PR paves the way for future optimizations (see #4493).
----
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4449
Differential Revision: D10370575
Pulled By: abhimadan
fbshipit-source-id: 9a2e152be1ef36969055c0e9eb4beb0d96c11f4d
2018-10-24 21:29:29 +02:00
|
|
|
max_covering_tombstone_seq, seq, read_opts,
|
|
|
|
nullptr /*read_callback*/, is_blob_index);
|
2015-05-29 23:36:35 +02:00
|
|
|
}
|
|
|
|
|
2017-09-11 17:58:52 +02:00
|
|
|
bool MemTableListVersion::GetFromList(
|
|
|
|
std::list<MemTable*>* list, const LookupKey& key, std::string* value,
|
Use only "local" range tombstones during Get (#4449)
Summary:
Previously, range tombstones were accumulated from every level, which
was necessary if a range tombstone in a higher level covered a key in a lower
level. However, RangeDelAggregator::AddTombstones's complexity is based on
the number of tombstones that are currently stored in it, which is wasteful in
the Get case, where we only need to know the highest sequence number of range
tombstones that cover the key from higher levels, and compute the highest covering
sequence number at the current level. This change introduces this optimization, and
removes the use of RangeDelAggregator from the Get path.
In the benchmark results, the following command was used to initialize the database:
```
./db_bench -db=/dev/shm/5k-rts -use_existing_db=false -benchmarks=filluniquerandom -write_buffer_size=1048576 -compression_type=lz4 -target_file_size_base=1048576 -max_bytes_for_level_base=4194304 -value_size=112 -key_size=16 -block_size=4096 -level_compaction_dynamic_level_bytes=true -num=5000000 -max_background_jobs=12 -benchmark_write_rate_limit=20971520 -range_tombstone_width=100 -writes_per_range_tombstone=100 -max_num_range_tombstones=50000 -bloom_bits=8
```
...and the following command was used to measure read throughput:
```
./db_bench -db=/dev/shm/5k-rts/ -use_existing_db=true -benchmarks=readrandom -disable_auto_compactions=true -num=5000000 -reads=100000 -threads=32
```
The filluniquerandom command was only run once, and the resulting database was used
to measure read performance before and after the PR. Both binaries were compiled with
`DEBUG_LEVEL=0`.
Readrandom results before PR:
```
readrandom : 4.544 micros/op 220090 ops/sec; 16.9 MB/s (63103 of 100000 found)
```
Readrandom results after PR:
```
readrandom : 11.147 micros/op 89707 ops/sec; 6.9 MB/s (63103 of 100000 found)
```
So it's actually slower right now, but this PR paves the way for future optimizations (see #4493).
----
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4449
Differential Revision: D10370575
Pulled By: abhimadan
fbshipit-source-id: 9a2e152be1ef36969055c0e9eb4beb0d96c11f4d
2018-10-24 21:29:29 +02:00
|
|
|
Status* s, MergeContext* merge_context,
|
|
|
|
SequenceNumber* max_covering_tombstone_seq, SequenceNumber* seq,
|
|
|
|
const ReadOptions& read_opts, ReadCallback* callback, bool* is_blob_index) {
|
2015-05-29 23:36:35 +02:00
|
|
|
*seq = kMaxSequenceNumber;
|
|
|
|
|
|
|
|
for (auto& memtable : *list) {
|
|
|
|
SequenceNumber current_seq = kMaxSequenceNumber;
|
|
|
|
|
Use only "local" range tombstones during Get (#4449)
Summary:
Previously, range tombstones were accumulated from every level, which
was necessary if a range tombstone in a higher level covered a key in a lower
level. However, RangeDelAggregator::AddTombstones's complexity is based on
the number of tombstones that are currently stored in it, which is wasteful in
the Get case, where we only need to know the highest sequence number of range
tombstones that cover the key from higher levels, and compute the highest covering
sequence number at the current level. This change introduces this optimization, and
removes the use of RangeDelAggregator from the Get path.
In the benchmark results, the following command was used to initialize the database:
```
./db_bench -db=/dev/shm/5k-rts -use_existing_db=false -benchmarks=filluniquerandom -write_buffer_size=1048576 -compression_type=lz4 -target_file_size_base=1048576 -max_bytes_for_level_base=4194304 -value_size=112 -key_size=16 -block_size=4096 -level_compaction_dynamic_level_bytes=true -num=5000000 -max_background_jobs=12 -benchmark_write_rate_limit=20971520 -range_tombstone_width=100 -writes_per_range_tombstone=100 -max_num_range_tombstones=50000 -bloom_bits=8
```
...and the following command was used to measure read throughput:
```
./db_bench -db=/dev/shm/5k-rts/ -use_existing_db=true -benchmarks=readrandom -disable_auto_compactions=true -num=5000000 -reads=100000 -threads=32
```
The filluniquerandom command was only run once, and the resulting database was used
to measure read performance before and after the PR. Both binaries were compiled with
`DEBUG_LEVEL=0`.
Readrandom results before PR:
```
readrandom : 4.544 micros/op 220090 ops/sec; 16.9 MB/s (63103 of 100000 found)
```
Readrandom results after PR:
```
readrandom : 11.147 micros/op 89707 ops/sec; 6.9 MB/s (63103 of 100000 found)
```
So it's actually slower right now, but this PR paves the way for future optimizations (see #4493).
----
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4449
Differential Revision: D10370575
Pulled By: abhimadan
fbshipit-source-id: 9a2e152be1ef36969055c0e9eb4beb0d96c11f4d
2018-10-24 21:29:29 +02:00
|
|
|
bool done =
|
|
|
|
memtable->Get(key, value, s, merge_context, max_covering_tombstone_seq,
|
|
|
|
¤t_seq, read_opts, callback, is_blob_index);
|
2015-05-29 23:36:35 +02:00
|
|
|
if (*seq == kMaxSequenceNumber) {
|
|
|
|
// Store the most recent sequence number of any operation on this key.
|
|
|
|
// Since we only care about the most recent change, we only need to
|
|
|
|
// return the first operation found when searching memtables in
|
|
|
|
// reverse-chronological order.
|
2017-09-11 17:58:52 +02:00
|
|
|
// current_seq would be equal to kMaxSequenceNumber if the value was to be
|
|
|
|
// skipped. This allows seq to be assigned again when the next value is
|
|
|
|
// read.
|
2015-05-29 23:36:35 +02:00
|
|
|
*seq = current_seq;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (done) {
|
Cache fragmented range tombstones in BlockBasedTableReader (#4493)
Summary:
This allows tombstone fragmenting to only be performed when the table is opened, and cached for subsequent accesses.
On the same DB used in #4449, running `readrandom` results in the following:
```
readrandom : 0.983 micros/op 1017076 ops/sec; 78.3 MB/s (63103 of 100000 found)
```
Now that Get performance in the presence of range tombstones is reasonable, I also compared the performance between a DB with range tombstones, "expanded" range tombstones (several point tombstones that cover the same keys the equivalent range tombstone would cover, a common workaround for DeleteRange), and no range tombstones. The created DBs had 5 million keys each, and DeleteRange was called at regular intervals (depending on the total number of range tombstones being written) after 4.5 million Puts. The table below summarizes the results of a `readwhilewriting` benchmark (in order to provide somewhat more realistic results):
```
Tombstones? | avg micros/op | stddev micros/op | avg ops/s | stddev ops/s
----------------- | ------------- | ---------------- | ------------ | ------------
None | 0.6186 | 0.04637 | 1,625,252.90 | 124,679.41
500 Expanded | 0.6019 | 0.03628 | 1,666,670.40 | 101,142.65
500 Unexpanded | 0.6435 | 0.03994 | 1,559,979.40 | 104,090.52
1k Expanded | 0.6034 | 0.04349 | 1,665,128.10 | 125,144.57
1k Unexpanded | 0.6261 | 0.03093 | 1,600,457.50 | 79,024.94
5k Expanded | 0.6163 | 0.05926 | 1,636,668.80 | 154,888.85
5k Unexpanded | 0.6402 | 0.04002 | 1,567,804.70 | 100,965.55
10k Expanded | 0.6036 | 0.05105 | 1,667,237.70 | 142,830.36
10k Unexpanded | 0.6128 | 0.02598 | 1,634,633.40 | 72,161.82
25k Expanded | 0.6198 | 0.04542 | 1,620,980.50 | 116,662.93
25k Unexpanded | 0.5478 | 0.0362 | 1,833,059.10 | 121,233.81
50k Expanded | 0.5104 | 0.04347 | 1,973,107.90 | 184,073.49
50k Unexpanded | 0.4528 | 0.03387 | 2,219,034.50 | 170,984.32
```
After a large enough quantity of range tombstones are written, range tombstone Gets can become faster than reading from an equivalent DB with several point tombstones.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4493
Differential Revision: D10842844
Pulled By: abhimadan
fbshipit-source-id: a7d44534f8120e6aabb65779d26c6b9df954c509
2018-10-26 04:25:00 +02:00
|
|
|
assert(*seq != kMaxSequenceNumber || s->IsNotFound());
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
return true;
|
|
|
|
}
|
2016-11-04 19:53:38 +01:00
|
|
|
if (!done && !s->ok() && !s->IsMergeInProgress() && !s->IsNotFound()) {
|
|
|
|
return false;
|
|
|
|
}
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
}
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2016-11-04 19:53:38 +01:00
|
|
|
Status MemTableListVersion::AddRangeTombstoneIterators(
|
2018-03-05 22:08:17 +01:00
|
|
|
const ReadOptions& read_opts, Arena* /*arena*/,
|
2018-12-18 02:26:56 +01:00
|
|
|
RangeDelAggregator* range_del_agg) {
|
2016-11-04 19:53:38 +01:00
|
|
|
assert(range_del_agg != nullptr);
|
|
|
|
for (auto& m : memlist_) {
|
2018-11-29 00:26:56 +01:00
|
|
|
// Using kMaxSequenceNumber is OK because these are immutable memtables.
|
|
|
|
std::unique_ptr<FragmentedRangeTombstoneIterator> range_del_iter(
|
|
|
|
m->NewRangeTombstoneIterator(read_opts,
|
|
|
|
kMaxSequenceNumber /* read_seq */));
|
|
|
|
range_del_agg->AddTombstones(std::move(range_del_iter));
|
2017-05-31 22:43:25 +02:00
|
|
|
}
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
2015-10-13 00:06:38 +02:00
|
|
|
void MemTableListVersion::AddIterators(
|
|
|
|
const ReadOptions& options, std::vector<InternalIterator*>* iterator_list,
|
|
|
|
Arena* arena) {
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
for (auto& m : memlist_) {
|
2014-09-05 02:40:41 +02:00
|
|
|
iterator_list->push_back(m->NewIterator(options, arena));
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
In DB::NewIterator(), try to allocate the whole iterator tree in an arena
Summary:
In this patch, try to allocate the whole iterator tree starting from DBIter from an arena
1. ArenaWrappedDBIter is created when serves as the entry point of an iterator tree, with an arena in it.
2. Add an option to create iterator from arena for following iterators: DBIter, MergingIterator, MemtableIterator, all mem table's iterators, all table reader's iterators and two level iterator.
3. MergeIteratorBuilder is created to incrementally build the tree of internal iterators. It is passed to mem table list and version set and add iterators to it.
Limitations:
(1) Only DB::NewIterator() without tailing uses the arena. Other cases, including readonly DB and compactions are still from malloc
(2) Two level iterator itself is allocated in arena, but not iterators inside it.
Test Plan: make all check
Reviewers: ljin, haobo
Reviewed By: haobo
Subscribers: leveldb, dhruba, yhchiang, igor
Differential Revision: https://reviews.facebook.net/D18513
2014-06-03 01:38:00 +02:00
|
|
|
void MemTableListVersion::AddIterators(
|
|
|
|
const ReadOptions& options, MergeIteratorBuilder* merge_iter_builder) {
|
|
|
|
for (auto& m : memlist_) {
|
|
|
|
merge_iter_builder->AddIterator(
|
|
|
|
m->NewIterator(options, merge_iter_builder->GetArena()));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-04-23 02:17:33 +02:00
|
|
|
uint64_t MemTableListVersion::GetTotalNumEntries() const {
|
|
|
|
uint64_t total_num = 0;
|
|
|
|
for (auto& m : memlist_) {
|
2015-03-19 00:11:02 +01:00
|
|
|
total_num += m->num_entries();
|
|
|
|
}
|
|
|
|
return total_num;
|
|
|
|
}
|
|
|
|
|
2017-02-06 23:42:38 +01:00
|
|
|
MemTable::MemTableStats MemTableListVersion::ApproximateStats(
|
|
|
|
const Slice& start_ikey, const Slice& end_ikey) {
|
|
|
|
MemTable::MemTableStats total_stats = {0, 0};
|
2015-06-13 03:04:30 +02:00
|
|
|
for (auto& m : memlist_) {
|
2017-02-06 23:42:38 +01:00
|
|
|
auto mStats = m->ApproximateStats(start_ikey, end_ikey);
|
|
|
|
total_stats.size += mStats.size;
|
|
|
|
total_stats.count += mStats.count;
|
2015-06-13 03:04:30 +02:00
|
|
|
}
|
2017-02-06 23:42:38 +01:00
|
|
|
return total_stats;
|
2015-06-13 03:04:30 +02:00
|
|
|
}
|
|
|
|
|
2015-03-19 00:11:02 +01:00
|
|
|
uint64_t MemTableListVersion::GetTotalNumDeletes() const {
|
|
|
|
uint64_t total_num = 0;
|
|
|
|
for (auto& m : memlist_) {
|
|
|
|
total_num += m->num_deletes();
|
2014-04-23 02:17:33 +02:00
|
|
|
}
|
|
|
|
return total_num;
|
|
|
|
}
|
|
|
|
|
2015-05-29 23:36:35 +02:00
|
|
|
SequenceNumber MemTableListVersion::GetEarliestSequenceNumber(
|
|
|
|
bool include_history) const {
|
|
|
|
if (include_history && !memlist_history_.empty()) {
|
|
|
|
return memlist_history_.back()->GetEarliestSequenceNumber();
|
|
|
|
} else if (!memlist_.empty()) {
|
|
|
|
return memlist_.back()->GetEarliestSequenceNumber();
|
|
|
|
} else {
|
|
|
|
return kMaxSequenceNumber;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-01-27 02:40:43 +01:00
|
|
|
// caller is responsible for referencing m
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
void MemTableListVersion::Add(MemTable* m, autovector<MemTable*>* to_delete) {
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
assert(refs_ == 1); // only when refs_ == 1 is MemTableListVersion mutable
|
2015-08-19 22:32:09 +02:00
|
|
|
AddMemTable(m);
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
|
|
|
|
TrimHistory(to_delete);
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
}
|
|
|
|
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
// Removes m from list of memtables not flushed. Caller should NOT Unref m.
|
|
|
|
void MemTableListVersion::Remove(MemTable* m,
|
|
|
|
autovector<MemTable*>* to_delete) {
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
assert(refs_ == 1); // only when refs_ == 1 is MemTableListVersion mutable
|
|
|
|
memlist_.remove(m);
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
|
2018-08-24 02:02:06 +02:00
|
|
|
m->MarkFlushed();
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
if (max_write_buffer_number_to_maintain_ > 0) {
|
|
|
|
memlist_history_.push_front(m);
|
|
|
|
TrimHistory(to_delete);
|
|
|
|
} else {
|
2015-08-19 22:32:09 +02:00
|
|
|
UnrefMemTable(to_delete, m);
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Make sure we don't use up too much space in history
|
|
|
|
void MemTableListVersion::TrimHistory(autovector<MemTable*>* to_delete) {
|
|
|
|
while (memlist_.size() + memlist_history_.size() >
|
|
|
|
static_cast<size_t>(max_write_buffer_number_to_maintain_) &&
|
|
|
|
!memlist_history_.empty()) {
|
|
|
|
MemTable* x = memlist_history_.back();
|
|
|
|
memlist_history_.pop_back();
|
|
|
|
|
2015-08-19 22:32:09 +02:00
|
|
|
UnrefMemTable(to_delete, x);
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
}
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
// Returns true if there is at least one memtable on which flush has
|
|
|
|
// not yet started.
|
2014-03-18 20:37:42 +01:00
|
|
|
bool MemTableList::IsFlushPending() const {
|
2018-10-16 04:59:20 +02:00
|
|
|
if ((flush_requested_ && num_flush_not_started_ > 0) ||
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
(num_flush_not_started_ >= min_write_buffer_number_to_merge_)) {
|
2014-10-27 22:50:21 +01:00
|
|
|
assert(imm_flush_needed.load(std::memory_order_relaxed));
|
2012-10-19 23:00:53 +02:00
|
|
|
return true;
|
|
|
|
}
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2013-06-11 23:23:58 +02:00
|
|
|
// Returns the memtables that need to be flushed.
|
2018-10-16 04:59:20 +02:00
|
|
|
void MemTableList::PickMemtablesToFlush(const uint64_t* max_memtable_id,
|
|
|
|
autovector<MemTable*>* ret) {
|
2015-03-13 18:45:40 +01:00
|
|
|
AutoThreadOperationStageUpdater stage_updater(
|
|
|
|
ThreadStatus::STAGE_PICK_MEMTABLES_TO_FLUSH);
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
const auto& memlist = current_->memlist_;
|
2019-05-11 02:53:41 +02:00
|
|
|
bool atomic_flush = false;
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
for (auto it = memlist.rbegin(); it != memlist.rend(); ++it) {
|
2012-10-19 23:00:53 +02:00
|
|
|
MemTable* m = *it;
|
2019-05-11 02:53:41 +02:00
|
|
|
if (!atomic_flush && m->atomic_flush_seqno_ != kMaxSequenceNumber) {
|
|
|
|
atomic_flush = true;
|
|
|
|
}
|
2018-10-16 04:59:20 +02:00
|
|
|
if (max_memtable_id != nullptr && m->GetID() > *max_memtable_id) {
|
|
|
|
break;
|
|
|
|
}
|
2012-10-19 23:00:53 +02:00
|
|
|
if (!m->flush_in_progress_) {
|
|
|
|
assert(!m->flush_completed_);
|
2012-11-29 01:42:36 +01:00
|
|
|
num_flush_not_started_--;
|
2012-10-19 23:00:53 +02:00
|
|
|
if (num_flush_not_started_ == 0) {
|
2014-10-27 22:50:21 +01:00
|
|
|
imm_flush_needed.store(false, std::memory_order_release);
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
2014-01-28 19:35:48 +01:00
|
|
|
m->flush_in_progress_ = true; // flushing will start very soon
|
2013-06-11 23:23:58 +02:00
|
|
|
ret->push_back(m);
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
|
|
|
}
|
2019-05-11 02:53:41 +02:00
|
|
|
if (!atomic_flush || num_flush_not_started_ == 0) {
|
|
|
|
flush_requested_ = false; // start-flush request is complete
|
|
|
|
}
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
|
|
|
|
2014-02-12 20:42:54 +01:00
|
|
|
void MemTableList::RollbackMemtableFlush(const autovector<MemTable*>& mems,
|
2018-03-05 22:08:17 +01:00
|
|
|
uint64_t /*file_number*/) {
|
2015-03-13 18:45:40 +01:00
|
|
|
AutoThreadOperationStageUpdater stage_updater(
|
|
|
|
ThreadStatus::STAGE_MEMTABLE_ROLLBACK);
|
2014-02-12 20:42:54 +01:00
|
|
|
assert(!mems.empty());
|
|
|
|
|
|
|
|
// If the flush was not successful, then just reset state.
|
2015-04-25 11:14:27 +02:00
|
|
|
// Maybe a succeeding attempt to flush will be successful.
|
2014-02-12 20:42:54 +01:00
|
|
|
for (MemTable* m : mems) {
|
|
|
|
assert(m->flush_in_progress_);
|
|
|
|
assert(m->file_number_ == 0);
|
|
|
|
|
|
|
|
m->flush_in_progress_ = false;
|
|
|
|
m->flush_completed_ = false;
|
|
|
|
m->edit_.Clear();
|
|
|
|
num_flush_not_started_++;
|
|
|
|
}
|
2014-10-27 22:50:21 +01:00
|
|
|
imm_flush_needed.store(true, std::memory_order_release);
|
2014-02-12 20:42:54 +01:00
|
|
|
}
|
|
|
|
|
2018-10-06 00:37:45 +02:00
|
|
|
// Try record a successful flush in the manifest file. It might just return
|
|
|
|
// Status::OK letting a concurrent flush to do actual the recording..
|
|
|
|
Status MemTableList::TryInstallMemtableFlushResults(
|
2014-10-02 01:19:16 +02:00
|
|
|
ColumnFamilyData* cfd, const MutableCFOptions& mutable_cf_options,
|
Skip deleted WALs during recovery
Summary:
This patch record min log number to keep to the manifest while flushing SST files to ignore them and any WAL older than them during recovery. This is to avoid scenarios when we have a gap between the WAL files are fed to the recovery procedure. The gap could happen by for example out-of-order WAL deletion. Such gap could cause problems in 2PC recovery where the prepared and commit entry are placed into two separate WAL and gap in the WALs could result into not processing the WAL with the commit entry and hence breaking the 2PC recovery logic.
Before the commit, for 2PC case, we determined which log number to keep in FindObsoleteFiles(). We looked at the earliest logs with outstanding prepare entries, or prepare entries whose respective commit or abort are in memtable. With the commit, the same calculation is done while we apply the SST flush. Just before installing the flush file, we precompute the earliest log file to keep after the flush finishes using the same logic (but skipping the memtables just flushed), record this information to the manifest entry for this new flushed SST file. This pre-computed value is also remembered in memory, and will later be used to determine whether a log file can be deleted. This value is unlikely to change until next flush because the commit entry will stay in memtable. (In WritePrepared, we could have removed the older log files as soon as all prepared entries are committed. It's not yet done anyway. Even if we do it, the only thing we loss with this new approach is earlier log deletion between two flushes, which does not guarantee to happen anyway because the obsolete file clean-up function is only executed after flush or compaction)
This min log number to keep is stored in the manifest using the safely-ignore customized field of AddFile entry, in order to guarantee that the DB generated using newer release can be opened by previous releases no older than 4.2.
Closes https://github.com/facebook/rocksdb/pull/3765
Differential Revision: D7747618
Pulled By: siying
fbshipit-source-id: d00c92105b4f83852e9754a1b70d6b64cb590729
2018-05-04 00:35:11 +02:00
|
|
|
const autovector<MemTable*>& mems, LogsWithPrepTracker* prep_tracker,
|
|
|
|
VersionSet* vset, InstrumentedMutex* mu, uint64_t file_number,
|
|
|
|
autovector<MemTable*>* to_delete, Directory* db_directory,
|
|
|
|
LogBuffer* log_buffer) {
|
2015-03-13 18:45:40 +01:00
|
|
|
AutoThreadOperationStageUpdater stage_updater(
|
|
|
|
ThreadStatus::STAGE_MEMTABLE_INSTALL_FLUSH_RESULTS);
|
2012-10-19 23:00:53 +02:00
|
|
|
mu->AssertHeld();
|
|
|
|
|
2018-10-06 00:37:45 +02:00
|
|
|
// Flush was successful
|
|
|
|
// Record the status on the memtable object. Either this call or a call by a
|
|
|
|
// concurrent flush thread will read the status and write it to manifest.
|
2013-10-28 06:55:46 +01:00
|
|
|
for (size_t i = 0; i < mems.size(); ++i) {
|
2013-06-11 23:23:58 +02:00
|
|
|
// All the edits are associated with the first memtable of this batch.
|
2013-10-28 07:08:54 +01:00
|
|
|
assert(i == 0 || mems[i]->GetEdits()->NumEntries() == 0);
|
2013-06-11 23:23:58 +02:00
|
|
|
|
2013-10-28 06:55:46 +01:00
|
|
|
mems[i]->flush_completed_ = true;
|
|
|
|
mems[i]->file_number_ = file_number;
|
2013-06-11 23:23:58 +02:00
|
|
|
}
|
2012-10-19 23:00:53 +02:00
|
|
|
|
2015-04-25 11:14:27 +02:00
|
|
|
// if some other thread is already committing, then return
|
2012-10-19 23:00:53 +02:00
|
|
|
Status s;
|
|
|
|
if (commit_in_progress_) {
|
2018-10-06 00:37:45 +02:00
|
|
|
TEST_SYNC_POINT("MemTableList::TryInstallMemtableFlushResults:InProgress");
|
2012-10-19 23:00:53 +02:00
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Only a single thread can be executing this piece of code
|
|
|
|
commit_in_progress_ = true;
|
|
|
|
|
Fix flush not being commit while writing manifest
Summary:
Fix flush not being commit while writing manifest, which is a recent bug introduced by D60075.
The issue:
# Options.max_background_flushes > 1
# Background thread A pick up a flush job, flush, then commit to manifest. (Note that mutex is released before writing manifest.)
# Background thread B pick up another flush job, flush. When it gets to `MemTableList::InstallMemtableFlushResults`, it notices another thread is commiting, so it quit.
# After the first commit, thread A doesn't double check if there are more flush result need to commit, leaving the second flush uncommitted.
Test Plan: run the test. Also verify the new test hit deadlock without the fix.
Reviewers: sdong, igor, lightmark
Reviewed By: lightmark
Subscribers: andrewkr, omegaga, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D60969
2016-07-21 19:10:41 +02:00
|
|
|
// Retry until all completed flushes are committed. New flushes can finish
|
|
|
|
// while the current thread is writing manifest where mutex is released.
|
|
|
|
while (s.ok()) {
|
|
|
|
auto& memlist = current_->memlist_;
|
2018-10-06 00:37:45 +02:00
|
|
|
// The back is the oldest; if flush_completed_ is not set to it, it means
|
|
|
|
// that we were assigned a more recent memtable. The memtables' flushes must
|
|
|
|
// be recorded in manifest in order. A concurrent flush thread, who is
|
|
|
|
// assigned to flush the oldest memtable, will later wake up and does all
|
|
|
|
// the pending writes to manifest, in order.
|
Fix flush not being commit while writing manifest
Summary:
Fix flush not being commit while writing manifest, which is a recent bug introduced by D60075.
The issue:
# Options.max_background_flushes > 1
# Background thread A pick up a flush job, flush, then commit to manifest. (Note that mutex is released before writing manifest.)
# Background thread B pick up another flush job, flush. When it gets to `MemTableList::InstallMemtableFlushResults`, it notices another thread is commiting, so it quit.
# After the first commit, thread A doesn't double check if there are more flush result need to commit, leaving the second flush uncommitted.
Test Plan: run the test. Also verify the new test hit deadlock without the fix.
Reviewers: sdong, igor, lightmark
Reviewed By: lightmark
Subscribers: andrewkr, omegaga, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D60969
2016-07-21 19:10:41 +02:00
|
|
|
if (memlist.empty() || !memlist.back()->flush_completed_) {
|
2012-10-19 23:00:53 +02:00
|
|
|
break;
|
|
|
|
}
|
Fix flush not being commit while writing manifest
Summary:
Fix flush not being commit while writing manifest, which is a recent bug introduced by D60075.
The issue:
# Options.max_background_flushes > 1
# Background thread A pick up a flush job, flush, then commit to manifest. (Note that mutex is released before writing manifest.)
# Background thread B pick up another flush job, flush. When it gets to `MemTableList::InstallMemtableFlushResults`, it notices another thread is commiting, so it quit.
# After the first commit, thread A doesn't double check if there are more flush result need to commit, leaving the second flush uncommitted.
Test Plan: run the test. Also verify the new test hit deadlock without the fix.
Reviewers: sdong, igor, lightmark
Reviewed By: lightmark
Subscribers: andrewkr, omegaga, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D60969
2016-07-21 19:10:41 +02:00
|
|
|
// scan all memtables from the earliest, and commit those
|
2018-10-06 00:37:45 +02:00
|
|
|
// (in that order) that have finished flushing. Memtables
|
Fix flush not being commit while writing manifest
Summary:
Fix flush not being commit while writing manifest, which is a recent bug introduced by D60075.
The issue:
# Options.max_background_flushes > 1
# Background thread A pick up a flush job, flush, then commit to manifest. (Note that mutex is released before writing manifest.)
# Background thread B pick up another flush job, flush. When it gets to `MemTableList::InstallMemtableFlushResults`, it notices another thread is commiting, so it quit.
# After the first commit, thread A doesn't double check if there are more flush result need to commit, leaving the second flush uncommitted.
Test Plan: run the test. Also verify the new test hit deadlock without the fix.
Reviewers: sdong, igor, lightmark
Reviewed By: lightmark
Subscribers: andrewkr, omegaga, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D60969
2016-07-21 19:10:41 +02:00
|
|
|
// are always committed in the order that they were created.
|
|
|
|
uint64_t batch_file_number = 0;
|
|
|
|
size_t batch_count = 0;
|
|
|
|
autovector<VersionEdit*> edit_list;
|
Skip deleted WALs during recovery
Summary:
This patch record min log number to keep to the manifest while flushing SST files to ignore them and any WAL older than them during recovery. This is to avoid scenarios when we have a gap between the WAL files are fed to the recovery procedure. The gap could happen by for example out-of-order WAL deletion. Such gap could cause problems in 2PC recovery where the prepared and commit entry are placed into two separate WAL and gap in the WALs could result into not processing the WAL with the commit entry and hence breaking the 2PC recovery logic.
Before the commit, for 2PC case, we determined which log number to keep in FindObsoleteFiles(). We looked at the earliest logs with outstanding prepare entries, or prepare entries whose respective commit or abort are in memtable. With the commit, the same calculation is done while we apply the SST flush. Just before installing the flush file, we precompute the earliest log file to keep after the flush finishes using the same logic (but skipping the memtables just flushed), record this information to the manifest entry for this new flushed SST file. This pre-computed value is also remembered in memory, and will later be used to determine whether a log file can be deleted. This value is unlikely to change until next flush because the commit entry will stay in memtable. (In WritePrepared, we could have removed the older log files as soon as all prepared entries are committed. It's not yet done anyway. Even if we do it, the only thing we loss with this new approach is earlier log deletion between two flushes, which does not guarantee to happen anyway because the obsolete file clean-up function is only executed after flush or compaction)
This min log number to keep is stored in the manifest using the safely-ignore customized field of AddFile entry, in order to guarantee that the DB generated using newer release can be opened by previous releases no older than 4.2.
Closes https://github.com/facebook/rocksdb/pull/3765
Differential Revision: D7747618
Pulled By: siying
fbshipit-source-id: d00c92105b4f83852e9754a1b70d6b64cb590729
2018-05-04 00:35:11 +02:00
|
|
|
autovector<MemTable*> memtables_to_flush;
|
Fix flush not being commit while writing manifest
Summary:
Fix flush not being commit while writing manifest, which is a recent bug introduced by D60075.
The issue:
# Options.max_background_flushes > 1
# Background thread A pick up a flush job, flush, then commit to manifest. (Note that mutex is released before writing manifest.)
# Background thread B pick up another flush job, flush. When it gets to `MemTableList::InstallMemtableFlushResults`, it notices another thread is commiting, so it quit.
# After the first commit, thread A doesn't double check if there are more flush result need to commit, leaving the second flush uncommitted.
Test Plan: run the test. Also verify the new test hit deadlock without the fix.
Reviewers: sdong, igor, lightmark
Reviewed By: lightmark
Subscribers: andrewkr, omegaga, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D60969
2016-07-21 19:10:41 +02:00
|
|
|
// enumerate from the last (earliest) element to see how many batch finished
|
|
|
|
for (auto it = memlist.rbegin(); it != memlist.rend(); ++it) {
|
|
|
|
MemTable* m = *it;
|
|
|
|
if (!m->flush_completed_) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
if (it == memlist.rbegin() || batch_file_number != m->file_number_) {
|
|
|
|
batch_file_number = m->file_number_;
|
2017-03-16 03:22:52 +01:00
|
|
|
ROCKS_LOG_BUFFER(log_buffer,
|
|
|
|
"[%s] Level-0 commit table #%" PRIu64 " started",
|
|
|
|
cfd->GetName().c_str(), m->file_number_);
|
Fix flush not being commit while writing manifest
Summary:
Fix flush not being commit while writing manifest, which is a recent bug introduced by D60075.
The issue:
# Options.max_background_flushes > 1
# Background thread A pick up a flush job, flush, then commit to manifest. (Note that mutex is released before writing manifest.)
# Background thread B pick up another flush job, flush. When it gets to `MemTableList::InstallMemtableFlushResults`, it notices another thread is commiting, so it quit.
# After the first commit, thread A doesn't double check if there are more flush result need to commit, leaving the second flush uncommitted.
Test Plan: run the test. Also verify the new test hit deadlock without the fix.
Reviewers: sdong, igor, lightmark
Reviewed By: lightmark
Subscribers: andrewkr, omegaga, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D60969
2016-07-21 19:10:41 +02:00
|
|
|
edit_list.push_back(&m->edit_);
|
Skip deleted WALs during recovery
Summary:
This patch record min log number to keep to the manifest while flushing SST files to ignore them and any WAL older than them during recovery. This is to avoid scenarios when we have a gap between the WAL files are fed to the recovery procedure. The gap could happen by for example out-of-order WAL deletion. Such gap could cause problems in 2PC recovery where the prepared and commit entry are placed into two separate WAL and gap in the WALs could result into not processing the WAL with the commit entry and hence breaking the 2PC recovery logic.
Before the commit, for 2PC case, we determined which log number to keep in FindObsoleteFiles(). We looked at the earliest logs with outstanding prepare entries, or prepare entries whose respective commit or abort are in memtable. With the commit, the same calculation is done while we apply the SST flush. Just before installing the flush file, we precompute the earliest log file to keep after the flush finishes using the same logic (but skipping the memtables just flushed), record this information to the manifest entry for this new flushed SST file. This pre-computed value is also remembered in memory, and will later be used to determine whether a log file can be deleted. This value is unlikely to change until next flush because the commit entry will stay in memtable. (In WritePrepared, we could have removed the older log files as soon as all prepared entries are committed. It's not yet done anyway. Even if we do it, the only thing we loss with this new approach is earlier log deletion between two flushes, which does not guarantee to happen anyway because the obsolete file clean-up function is only executed after flush or compaction)
This min log number to keep is stored in the manifest using the safely-ignore customized field of AddFile entry, in order to guarantee that the DB generated using newer release can be opened by previous releases no older than 4.2.
Closes https://github.com/facebook/rocksdb/pull/3765
Differential Revision: D7747618
Pulled By: siying
fbshipit-source-id: d00c92105b4f83852e9754a1b70d6b64cb590729
2018-05-04 00:35:11 +02:00
|
|
|
memtables_to_flush.push_back(m);
|
Fix flush not being commit while writing manifest
Summary:
Fix flush not being commit while writing manifest, which is a recent bug introduced by D60075.
The issue:
# Options.max_background_flushes > 1
# Background thread A pick up a flush job, flush, then commit to manifest. (Note that mutex is released before writing manifest.)
# Background thread B pick up another flush job, flush. When it gets to `MemTableList::InstallMemtableFlushResults`, it notices another thread is commiting, so it quit.
# After the first commit, thread A doesn't double check if there are more flush result need to commit, leaving the second flush uncommitted.
Test Plan: run the test. Also verify the new test hit deadlock without the fix.
Reviewers: sdong, igor, lightmark
Reviewed By: lightmark
Subscribers: andrewkr, omegaga, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D60969
2016-07-21 19:10:41 +02:00
|
|
|
}
|
|
|
|
batch_count++;
|
2016-07-06 03:09:59 +02:00
|
|
|
}
|
2013-06-11 23:23:58 +02:00
|
|
|
|
2018-10-06 00:37:45 +02:00
|
|
|
// TODO(myabandeh): Not sure how batch_count could be 0 here.
|
Fix flush not being commit while writing manifest
Summary:
Fix flush not being commit while writing manifest, which is a recent bug introduced by D60075.
The issue:
# Options.max_background_flushes > 1
# Background thread A pick up a flush job, flush, then commit to manifest. (Note that mutex is released before writing manifest.)
# Background thread B pick up another flush job, flush. When it gets to `MemTableList::InstallMemtableFlushResults`, it notices another thread is commiting, so it quit.
# After the first commit, thread A doesn't double check if there are more flush result need to commit, leaving the second flush uncommitted.
Test Plan: run the test. Also verify the new test hit deadlock without the fix.
Reviewers: sdong, igor, lightmark
Reviewed By: lightmark
Subscribers: andrewkr, omegaga, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D60969
2016-07-21 19:10:41 +02:00
|
|
|
if (batch_count > 0) {
|
Skip deleted WALs during recovery
Summary:
This patch record min log number to keep to the manifest while flushing SST files to ignore them and any WAL older than them during recovery. This is to avoid scenarios when we have a gap between the WAL files are fed to the recovery procedure. The gap could happen by for example out-of-order WAL deletion. Such gap could cause problems in 2PC recovery where the prepared and commit entry are placed into two separate WAL and gap in the WALs could result into not processing the WAL with the commit entry and hence breaking the 2PC recovery logic.
Before the commit, for 2PC case, we determined which log number to keep in FindObsoleteFiles(). We looked at the earliest logs with outstanding prepare entries, or prepare entries whose respective commit or abort are in memtable. With the commit, the same calculation is done while we apply the SST flush. Just before installing the flush file, we precompute the earliest log file to keep after the flush finishes using the same logic (but skipping the memtables just flushed), record this information to the manifest entry for this new flushed SST file. This pre-computed value is also remembered in memory, and will later be used to determine whether a log file can be deleted. This value is unlikely to change until next flush because the commit entry will stay in memtable. (In WritePrepared, we could have removed the older log files as soon as all prepared entries are committed. It's not yet done anyway. Even if we do it, the only thing we loss with this new approach is earlier log deletion between two flushes, which does not guarantee to happen anyway because the obsolete file clean-up function is only executed after flush or compaction)
This min log number to keep is stored in the manifest using the safely-ignore customized field of AddFile entry, in order to guarantee that the DB generated using newer release can be opened by previous releases no older than 4.2.
Closes https://github.com/facebook/rocksdb/pull/3765
Differential Revision: D7747618
Pulled By: siying
fbshipit-source-id: d00c92105b4f83852e9754a1b70d6b64cb590729
2018-05-04 00:35:11 +02:00
|
|
|
if (vset->db_options()->allow_2pc) {
|
|
|
|
assert(edit_list.size() > 0);
|
|
|
|
// We piggyback the information of earliest log file to keep in the
|
|
|
|
// manifest entry for the last file flushed.
|
|
|
|
edit_list.back()->SetMinLogNumberToKeep(PrecomputeMinLogNumberToKeep(
|
|
|
|
vset, *cfd, edit_list, memtables_to_flush, prep_tracker));
|
|
|
|
}
|
|
|
|
|
Fix flush not being commit while writing manifest
Summary:
Fix flush not being commit while writing manifest, which is a recent bug introduced by D60075.
The issue:
# Options.max_background_flushes > 1
# Background thread A pick up a flush job, flush, then commit to manifest. (Note that mutex is released before writing manifest.)
# Background thread B pick up another flush job, flush. When it gets to `MemTableList::InstallMemtableFlushResults`, it notices another thread is commiting, so it quit.
# After the first commit, thread A doesn't double check if there are more flush result need to commit, leaving the second flush uncommitted.
Test Plan: run the test. Also verify the new test hit deadlock without the fix.
Reviewers: sdong, igor, lightmark
Reviewed By: lightmark
Subscribers: andrewkr, omegaga, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D60969
2016-07-21 19:10:41 +02:00
|
|
|
// this can release and reacquire the mutex.
|
|
|
|
s = vset->LogAndApply(cfd, mutable_cf_options, edit_list, mu,
|
|
|
|
db_directory);
|
|
|
|
|
|
|
|
// we will be changing the version in the next code path,
|
|
|
|
// so we better create a new one, since versions are immutable
|
|
|
|
InstallNewVersion();
|
|
|
|
|
|
|
|
// All the later memtables that have the same filenum
|
|
|
|
// are part of the same batch. They can be committed now.
|
|
|
|
uint64_t mem_id = 1; // how many memtables have been flushed.
|
2018-06-28 21:16:10 +02:00
|
|
|
|
|
|
|
// commit new state only if the column family is NOT dropped.
|
|
|
|
// The reason is as follows (refer to
|
|
|
|
// ColumnFamilyTest.FlushAndDropRaceCondition).
|
|
|
|
// If the column family is dropped, then according to LogAndApply, its
|
2018-10-06 00:37:45 +02:00
|
|
|
// corresponding flush operation is NOT written to the MANIFEST. This
|
2018-06-28 21:16:10 +02:00
|
|
|
// means the DB is not aware of the L0 files generated from the flush.
|
|
|
|
// By committing the new state, we remove the memtable from the memtable
|
|
|
|
// list. Creating an iterator on this column family will not be able to
|
|
|
|
// read full data since the memtable is removed, and the DB is not aware
|
|
|
|
// of the L0 files, causing MergingIterator unable to build child
|
|
|
|
// iterators. RocksDB contract requires that the iterator can be created
|
|
|
|
// on a dropped column family, and we must be able to
|
|
|
|
// read full data as long as column family handle is not deleted, even if
|
|
|
|
// the column family is dropped.
|
|
|
|
if (s.ok() && !cfd->IsDropped()) { // commit new state
|
Fix flush not being commit while writing manifest
Summary:
Fix flush not being commit while writing manifest, which is a recent bug introduced by D60075.
The issue:
# Options.max_background_flushes > 1
# Background thread A pick up a flush job, flush, then commit to manifest. (Note that mutex is released before writing manifest.)
# Background thread B pick up another flush job, flush. When it gets to `MemTableList::InstallMemtableFlushResults`, it notices another thread is commiting, so it quit.
# After the first commit, thread A doesn't double check if there are more flush result need to commit, leaving the second flush uncommitted.
Test Plan: run the test. Also verify the new test hit deadlock without the fix.
Reviewers: sdong, igor, lightmark
Reviewed By: lightmark
Subscribers: andrewkr, omegaga, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D60969
2016-07-21 19:10:41 +02:00
|
|
|
while (batch_count-- > 0) {
|
|
|
|
MemTable* m = current_->memlist_.back();
|
2017-03-16 03:22:52 +01:00
|
|
|
ROCKS_LOG_BUFFER(log_buffer, "[%s] Level-0 commit table #%" PRIu64
|
|
|
|
": memtable #%" PRIu64 " done",
|
|
|
|
cfd->GetName().c_str(), m->file_number_, mem_id);
|
Fix flush not being commit while writing manifest
Summary:
Fix flush not being commit while writing manifest, which is a recent bug introduced by D60075.
The issue:
# Options.max_background_flushes > 1
# Background thread A pick up a flush job, flush, then commit to manifest. (Note that mutex is released before writing manifest.)
# Background thread B pick up another flush job, flush. When it gets to `MemTableList::InstallMemtableFlushResults`, it notices another thread is commiting, so it quit.
# After the first commit, thread A doesn't double check if there are more flush result need to commit, leaving the second flush uncommitted.
Test Plan: run the test. Also verify the new test hit deadlock without the fix.
Reviewers: sdong, igor, lightmark
Reviewed By: lightmark
Subscribers: andrewkr, omegaga, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D60969
2016-07-21 19:10:41 +02:00
|
|
|
assert(m->file_number_ > 0);
|
|
|
|
current_->Remove(m, to_delete);
|
|
|
|
++mem_id;
|
|
|
|
}
|
|
|
|
} else {
|
2019-05-15 22:14:18 +02:00
|
|
|
for (auto it = current_->memlist_.rbegin(); batch_count-- > 0; ++it) {
|
Fix flush not being commit while writing manifest
Summary:
Fix flush not being commit while writing manifest, which is a recent bug introduced by D60075.
The issue:
# Options.max_background_flushes > 1
# Background thread A pick up a flush job, flush, then commit to manifest. (Note that mutex is released before writing manifest.)
# Background thread B pick up another flush job, flush. When it gets to `MemTableList::InstallMemtableFlushResults`, it notices another thread is commiting, so it quit.
# After the first commit, thread A doesn't double check if there are more flush result need to commit, leaving the second flush uncommitted.
Test Plan: run the test. Also verify the new test hit deadlock without the fix.
Reviewers: sdong, igor, lightmark
Reviewed By: lightmark
Subscribers: andrewkr, omegaga, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D60969
2016-07-21 19:10:41 +02:00
|
|
|
MemTable* m = *it;
|
|
|
|
// commit failed. setup state so that we can flush again.
|
2017-03-16 03:22:52 +01:00
|
|
|
ROCKS_LOG_BUFFER(log_buffer, "Level-0 commit table #%" PRIu64
|
|
|
|
": memtable #%" PRIu64 " failed",
|
|
|
|
m->file_number_, mem_id);
|
Fix flush not being commit while writing manifest
Summary:
Fix flush not being commit while writing manifest, which is a recent bug introduced by D60075.
The issue:
# Options.max_background_flushes > 1
# Background thread A pick up a flush job, flush, then commit to manifest. (Note that mutex is released before writing manifest.)
# Background thread B pick up another flush job, flush. When it gets to `MemTableList::InstallMemtableFlushResults`, it notices another thread is commiting, so it quit.
# After the first commit, thread A doesn't double check if there are more flush result need to commit, leaving the second flush uncommitted.
Test Plan: run the test. Also verify the new test hit deadlock without the fix.
Reviewers: sdong, igor, lightmark
Reviewed By: lightmark
Subscribers: andrewkr, omegaga, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D60969
2016-07-21 19:10:41 +02:00
|
|
|
m->flush_completed_ = false;
|
|
|
|
m->flush_in_progress_ = false;
|
|
|
|
m->edit_.Clear();
|
|
|
|
num_flush_not_started_++;
|
|
|
|
m->file_number_ = 0;
|
|
|
|
imm_flush_needed.store(true, std::memory_order_release);
|
|
|
|
++mem_id;
|
|
|
|
}
|
2013-06-11 23:23:58 +02:00
|
|
|
}
|
2016-07-06 03:09:59 +02:00
|
|
|
}
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
|
|
|
commit_in_progress_ = false;
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2012-11-29 01:42:36 +01:00
|
|
|
// New memtables are inserted at the front of the list.
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
void MemTableList::Add(MemTable* m, autovector<MemTable*>* to_delete) {
|
|
|
|
assert(static_cast<int>(current_->memlist_.size()) >= num_flush_not_started_);
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
InstallNewVersion();
|
2014-01-27 02:40:43 +01:00
|
|
|
// this method is used to move mutable memtable into an immutable list.
|
|
|
|
// since mutable memtable is already refcounted by the DBImpl,
|
|
|
|
// and when moving to the imutable list we don't unref it,
|
|
|
|
// we don't have to ref the memtable here. we just take over the
|
|
|
|
// reference from the DBImpl.
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
current_->Add(m, to_delete);
|
2013-08-23 08:10:02 +02:00
|
|
|
m->MarkImmutable();
|
2012-10-19 23:00:53 +02:00
|
|
|
num_flush_not_started_++;
|
|
|
|
if (num_flush_not_started_ == 1) {
|
2014-10-27 22:50:21 +01:00
|
|
|
imm_flush_needed.store(true, std::memory_order_release);
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Returns an estimate of the number of bytes of data in use.
|
2015-08-19 22:32:09 +02:00
|
|
|
size_t MemTableList::ApproximateUnflushedMemTablesMemoryUsage() {
|
2014-11-06 20:14:28 +01:00
|
|
|
size_t total_size = 0;
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
for (auto& memtable : current_->memlist_) {
|
2014-11-06 20:14:28 +01:00
|
|
|
total_size += memtable->ApproximateMemoryUsage();
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
2014-11-06 20:14:28 +01:00
|
|
|
return total_size;
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
|
|
|
|
2015-08-19 22:32:09 +02:00
|
|
|
size_t MemTableList::ApproximateMemoryUsage() { return current_memory_usage_; }
|
|
|
|
|
2017-10-24 00:22:05 +02:00
|
|
|
uint64_t MemTableList::ApproximateOldestKeyTime() const {
|
|
|
|
if (!current_->memlist_.empty()) {
|
|
|
|
return current_->memlist_.back()->ApproximateOldestKeyTime();
|
|
|
|
}
|
|
|
|
return std::numeric_limits<uint64_t>::max();
|
|
|
|
}
|
|
|
|
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
void MemTableList::InstallNewVersion() {
|
|
|
|
if (current_->refs_ == 1) {
|
|
|
|
// we're the only one using the version, just keep using it
|
|
|
|
} else {
|
|
|
|
// somebody else holds the current version, we need to create new one
|
|
|
|
MemTableListVersion* version = current_;
|
2015-08-19 22:32:09 +02:00
|
|
|
current_ = new MemTableListVersion(¤t_memory_usage_, current_);
|
2014-01-25 22:50:30 +01:00
|
|
|
current_->Ref();
|
MemTableListVersion
Summary:
MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful.
This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team.
I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet())
Test Plan: `make check` works. I will also do `make valgrind_check` before commit
Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 23:52:08 +01:00
|
|
|
version->Unref();
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Skip deleted WALs during recovery
Summary:
This patch record min log number to keep to the manifest while flushing SST files to ignore them and any WAL older than them during recovery. This is to avoid scenarios when we have a gap between the WAL files are fed to the recovery procedure. The gap could happen by for example out-of-order WAL deletion. Such gap could cause problems in 2PC recovery where the prepared and commit entry are placed into two separate WAL and gap in the WALs could result into not processing the WAL with the commit entry and hence breaking the 2PC recovery logic.
Before the commit, for 2PC case, we determined which log number to keep in FindObsoleteFiles(). We looked at the earliest logs with outstanding prepare entries, or prepare entries whose respective commit or abort are in memtable. With the commit, the same calculation is done while we apply the SST flush. Just before installing the flush file, we precompute the earliest log file to keep after the flush finishes using the same logic (but skipping the memtables just flushed), record this information to the manifest entry for this new flushed SST file. This pre-computed value is also remembered in memory, and will later be used to determine whether a log file can be deleted. This value is unlikely to change until next flush because the commit entry will stay in memtable. (In WritePrepared, we could have removed the older log files as soon as all prepared entries are committed. It's not yet done anyway. Even if we do it, the only thing we loss with this new approach is earlier log deletion between two flushes, which does not guarantee to happen anyway because the obsolete file clean-up function is only executed after flush or compaction)
This min log number to keep is stored in the manifest using the safely-ignore customized field of AddFile entry, in order to guarantee that the DB generated using newer release can be opened by previous releases no older than 4.2.
Closes https://github.com/facebook/rocksdb/pull/3765
Differential Revision: D7747618
Pulled By: siying
fbshipit-source-id: d00c92105b4f83852e9754a1b70d6b64cb590729
2018-05-04 00:35:11 +02:00
|
|
|
uint64_t MemTableList::PrecomputeMinLogContainingPrepSection(
|
|
|
|
const autovector<MemTable*>& memtables_to_flush) {
|
2016-04-18 20:11:51 +02:00
|
|
|
uint64_t min_log = 0;
|
|
|
|
|
|
|
|
for (auto& m : current_->memlist_) {
|
Skip deleted WALs during recovery
Summary:
This patch record min log number to keep to the manifest while flushing SST files to ignore them and any WAL older than them during recovery. This is to avoid scenarios when we have a gap between the WAL files are fed to the recovery procedure. The gap could happen by for example out-of-order WAL deletion. Such gap could cause problems in 2PC recovery where the prepared and commit entry are placed into two separate WAL and gap in the WALs could result into not processing the WAL with the commit entry and hence breaking the 2PC recovery logic.
Before the commit, for 2PC case, we determined which log number to keep in FindObsoleteFiles(). We looked at the earliest logs with outstanding prepare entries, or prepare entries whose respective commit or abort are in memtable. With the commit, the same calculation is done while we apply the SST flush. Just before installing the flush file, we precompute the earliest log file to keep after the flush finishes using the same logic (but skipping the memtables just flushed), record this information to the manifest entry for this new flushed SST file. This pre-computed value is also remembered in memory, and will later be used to determine whether a log file can be deleted. This value is unlikely to change until next flush because the commit entry will stay in memtable. (In WritePrepared, we could have removed the older log files as soon as all prepared entries are committed. It's not yet done anyway. Even if we do it, the only thing we loss with this new approach is earlier log deletion between two flushes, which does not guarantee to happen anyway because the obsolete file clean-up function is only executed after flush or compaction)
This min log number to keep is stored in the manifest using the safely-ignore customized field of AddFile entry, in order to guarantee that the DB generated using newer release can be opened by previous releases no older than 4.2.
Closes https://github.com/facebook/rocksdb/pull/3765
Differential Revision: D7747618
Pulled By: siying
fbshipit-source-id: d00c92105b4f83852e9754a1b70d6b64cb590729
2018-05-04 00:35:11 +02:00
|
|
|
// Assume the list is very short, we can live with O(m*n). We can optimize
|
|
|
|
// if the performance has some problem.
|
|
|
|
bool should_skip = false;
|
|
|
|
for (MemTable* m_to_flush : memtables_to_flush) {
|
|
|
|
if (m == m_to_flush) {
|
|
|
|
should_skip = true;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (should_skip) {
|
2016-04-18 20:11:51 +02:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
auto log = m->GetMinLogContainingPrepSection();
|
|
|
|
|
|
|
|
if (log > 0 && (min_log == 0 || log < min_log)) {
|
|
|
|
min_log = log;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return min_log;
|
|
|
|
}
|
|
|
|
|
2019-01-04 05:53:52 +01:00
|
|
|
// Commit a successful atomic flush in the manifest file.
|
|
|
|
Status InstallMemtableAtomicFlushResults(
|
|
|
|
const autovector<MemTableList*>* imm_lists,
|
|
|
|
const autovector<ColumnFamilyData*>& cfds,
|
|
|
|
const autovector<const MutableCFOptions*>& mutable_cf_options_list,
|
|
|
|
const autovector<const autovector<MemTable*>*>& mems_list, VersionSet* vset,
|
2019-01-31 23:28:53 +01:00
|
|
|
InstrumentedMutex* mu, const autovector<FileMetaData*>& file_metas,
|
2019-01-04 05:53:52 +01:00
|
|
|
autovector<MemTable*>* to_delete, Directory* db_directory,
|
|
|
|
LogBuffer* log_buffer) {
|
|
|
|
AutoThreadOperationStageUpdater stage_updater(
|
|
|
|
ThreadStatus::STAGE_MEMTABLE_INSTALL_FLUSH_RESULTS);
|
|
|
|
mu->AssertHeld();
|
|
|
|
|
|
|
|
size_t num = mems_list.size();
|
|
|
|
assert(cfds.size() == num);
|
|
|
|
if (imm_lists != nullptr) {
|
|
|
|
assert(imm_lists->size() == num);
|
|
|
|
}
|
|
|
|
for (size_t k = 0; k != num; ++k) {
|
|
|
|
#ifndef NDEBUG
|
|
|
|
const auto* imm =
|
|
|
|
(imm_lists == nullptr) ? cfds[k]->imm() : imm_lists->at(k);
|
|
|
|
if (!mems_list[k]->empty()) {
|
|
|
|
assert((*mems_list[k])[0]->GetID() == imm->GetEarliestMemTableID());
|
|
|
|
}
|
|
|
|
#endif
|
2019-01-31 23:28:53 +01:00
|
|
|
assert(nullptr != file_metas[k]);
|
2019-01-04 05:53:52 +01:00
|
|
|
for (size_t i = 0; i != mems_list[k]->size(); ++i) {
|
|
|
|
assert(i == 0 || (*mems_list[k])[i]->GetEdits()->NumEntries() == 0);
|
|
|
|
(*mems_list[k])[i]->SetFlushCompleted(true);
|
2019-01-31 23:28:53 +01:00
|
|
|
(*mems_list[k])[i]->SetFileNumber(file_metas[k]->fd.GetNumber());
|
2019-01-04 05:53:52 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
Status s;
|
|
|
|
|
|
|
|
autovector<autovector<VersionEdit*>> edit_lists;
|
|
|
|
uint32_t num_entries = 0;
|
|
|
|
for (const auto mems : mems_list) {
|
|
|
|
assert(mems != nullptr);
|
|
|
|
autovector<VersionEdit*> edits;
|
|
|
|
assert(!mems->empty());
|
|
|
|
edits.emplace_back((*mems)[0]->GetEdits());
|
|
|
|
++num_entries;
|
|
|
|
edit_lists.emplace_back(edits);
|
|
|
|
}
|
2019-02-14 03:23:10 +01:00
|
|
|
// Mark the version edits as an atomic group if the number of version edits
|
|
|
|
// exceeds 1.
|
|
|
|
if (cfds.size() > 1) {
|
|
|
|
for (auto& edits : edit_lists) {
|
|
|
|
assert(edits.size() == 1);
|
|
|
|
edits[0]->MarkAtomicGroup(--num_entries);
|
|
|
|
}
|
|
|
|
assert(0 == num_entries);
|
2019-01-04 05:53:52 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
// this can release and reacquire the mutex.
|
|
|
|
s = vset->LogAndApply(cfds, mutable_cf_options_list, edit_lists, mu,
|
|
|
|
db_directory);
|
|
|
|
|
|
|
|
for (size_t k = 0; k != cfds.size(); ++k) {
|
|
|
|
auto* imm = (imm_lists == nullptr) ? cfds[k]->imm() : imm_lists->at(k);
|
|
|
|
imm->InstallNewVersion();
|
|
|
|
}
|
|
|
|
|
2019-05-20 19:37:37 +02:00
|
|
|
if (s.ok() || s.IsColumnFamilyDropped()) {
|
2019-01-04 05:53:52 +01:00
|
|
|
for (size_t i = 0; i != cfds.size(); ++i) {
|
|
|
|
if (cfds[i]->IsDropped()) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
auto* imm = (imm_lists == nullptr) ? cfds[i]->imm() : imm_lists->at(i);
|
|
|
|
for (auto m : *mems_list[i]) {
|
|
|
|
assert(m->GetFileNumber() > 0);
|
|
|
|
uint64_t mem_id = m->GetID();
|
|
|
|
ROCKS_LOG_BUFFER(log_buffer,
|
|
|
|
"[%s] Level-0 commit table #%" PRIu64
|
|
|
|
": memtable #%" PRIu64 " done",
|
|
|
|
cfds[i]->GetName().c_str(), m->GetFileNumber(),
|
|
|
|
mem_id);
|
|
|
|
imm->current_->Remove(m, to_delete);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
for (size_t i = 0; i != cfds.size(); ++i) {
|
|
|
|
auto* imm = (imm_lists == nullptr) ? cfds[i]->imm() : imm_lists->at(i);
|
|
|
|
for (auto m : *mems_list[i]) {
|
|
|
|
uint64_t mem_id = m->GetID();
|
|
|
|
ROCKS_LOG_BUFFER(log_buffer,
|
|
|
|
"[%s] Level-0 commit table #%" PRIu64
|
|
|
|
": memtable #%" PRIu64 " failed",
|
|
|
|
cfds[i]->GetName().c_str(), m->GetFileNumber(),
|
|
|
|
mem_id);
|
|
|
|
m->SetFlushCompleted(false);
|
|
|
|
m->SetFlushInProgress(false);
|
|
|
|
m->GetEdits()->Clear();
|
|
|
|
m->SetFileNumber(0);
|
|
|
|
imm->num_flush_not_started_++;
|
|
|
|
}
|
|
|
|
imm->imm_flush_needed.store(true, std::memory_order_release);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2019-05-18 04:16:51 +02:00
|
|
|
void MemTableList::RemoveOldMemTables(uint64_t log_number,
|
|
|
|
autovector<MemTable*>* to_delete) {
|
|
|
|
assert(to_delete != nullptr);
|
|
|
|
InstallNewVersion();
|
|
|
|
auto& memlist = current_->memlist_;
|
|
|
|
for (auto it = memlist.rbegin(); it != memlist.rend(); ++it) {
|
|
|
|
MemTable* mem = *it;
|
|
|
|
if (mem->GetNextLogNumber() > log_number) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
current_->Remove(mem, to_delete);
|
|
|
|
--num_flush_not_started_;
|
|
|
|
if (0 == num_flush_not_started_) {
|
|
|
|
imm_flush_needed.store(false, std::memory_order_release);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-10-04 06:49:15 +02:00
|
|
|
} // namespace rocksdb
|