Summary:
Based on my recent findings (posted in our internal group), if we use fallocate without KEEP_SIZE flag, we get superior performance of fdatasync() in append-only workloads.
This diff provides an option for user to not use KEEP_SIZE flag, thus optimizing his sync performance by up to 2x-3x.
At one point we also just called posix_fallocate instead of fallocate, which isn't very fast: http://code.woboq.org/userspace/glibc/sysdeps/posix/posix_fallocate.c.html (tl;dr it manually writes out zero bytes to allocate storage). This diff also fixes that, by first calling fallocate and then posix_fallocate if fallocate is not supported.
Test Plan: make check
Reviewers: dhruba, sdong, haobo, ljin
Reviewed By: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16761
Summary:
When the manifest is getting rolled the following happens:
1) manifest_file_number_ is assigned to a new manifest number (even though the old one is still current)
2) mutex is unlocked
3) SetCurrentFile() creates temporary file manifest_file_number_.dbtmp
4) SetCurrentFile() renames manifest_file_number_.dbtmp to CURRENT
5) mutex is locked
If FindObsoleteFiles happens between (3) and (4) it will:
1) Delete manifest_file_number_.dbtmp (because it's not in pending_outputs_)
2) Delete old manifest (because the manifest_file_number_ already points to a new one)
I introduce the concept of prev_manifest_file_number_ that will avoid the race condition.
However, we should discuss the future of MANIFEST file rolling. We found some race conditions with it last week and who knows how many more are there. Nobody is using it in production because we don't trust the implementation. Should we even support it?
Test Plan: make check
Reviewers: ljin, dhruba, haobo, sdong
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16929
Summary:
With D16767, there is a case compaction tasks are scheduled infinitely:
(1) no flush thread is configured and more than 1 compaction threads
(2) a flush is going on by one compaction hread
(3) the state of SST files is in the state that versions_->current()->NeedsCompaction() will generate a false positive (return true actually there is no work to be done)
In that case, a infinite loop will be formed.
This patch would fix it.
Test Plan: make all check
Reviewers: haobo, igor, ljin
Reviewed By: igor
CC: dhruba, yhchiang, leveldb
Differential Revision: https://reviews.facebook.net/D16863
Summary:
This is is based on https://reviews.facebook.net/D15027. It's not finished but I would like to give a prototype to avoid arena over-allocation while making better use of the already allocated memory blocks.
Instead of check approximate memtable size, we will take a deeper look at the arena, which incorporate essential idea that @sdong suggests: flush when arena has allocated its last and the last is "almost full"
Test Plan: N/A
Reviewers: haobo, sdong
Reviewed By: sdong
CC: leveldb, sdong
Differential Revision: https://reviews.facebook.net/D15051
Summary:
@igor pointed out that there is a potential data race because of the way we use the newly introduced LogBuffer. After "bg_compaction_scheduled_--" or "bg_flush_scheduled_--", they can both become 0. As soon as the lock is released after that, DBImpl's deconstructor can go ahead and deconstruct all the states inside DB, including the info_log object hold in a shared pointer of the options object it keeps. At that point it is not safe anymore to continue using the info logger to write the delayed logs.
With the patch, lock is released temporarily for log buffer to be flushed before "bg_compaction_scheduled_--" or "bg_flush_scheduled_--". In order to make sure we don't miss any pending flush or compaction, a new flag bg_schedule_needed_ is added, which is set to be true if there is a pending flush or compaction but not scheduled because of the max thread limit. If the flag is set to be true, the scheduling function will be called before compaction or flush thread finishes.
Thanks @igor for this finding!
Test Plan: make all check
Reviewers: haobo, igor
Reviewed By: haobo
CC: dhruba, ljin, yhchiang, igor, leveldb
Differential Revision: https://reviews.facebook.net/D16767
Summary: To temp fix the log buffer flushing. Flush the buffer inside the lock. Clean the trunk before we find an eventual fix.
Test Plan: make all check
Reviewers: haobo, igor
Reviewed By: igor
CC: ljin, leveldb, yhchiang
Differential Revision: https://reviews.facebook.net/D16791
Summary: Having code after SignalAll has already caused 2 bugs. Let's make sure this doesn't happen again.
Test Plan: no test
Reviewers: sdong, dhruba, haobo
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16785
Summary: as title. also made info log output of file deletion a bit more descriptive.
Test Plan: make check; db_bench and look at LOG output
Reviewers: igor
Reviewed By: igor
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16731
Summary:
(1) Fix SanitizeOptions() to also check HashLinkList. The current
dynamic case just happens to work because the 2 classes have the same
layout.
(2) Do not delete SliceTransform object in HashSkipListFactory and
HashLinkListFactory destructor. Reason: SanitizeOptions() enforces
prefix_extractor and SliceTransform to be the same object when
Hash**Factory is used. This makes the behavior strange: when
Hash**Factory is used, prefix_extractor will be released by RocksDB. If
other memtable factory is used, prefix_extractor should be released by
user.
Test Plan: db_bench && make asan_check
Reviewers: haobo, igor, sdong
Reviewed By: igor
CC: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D16587
Summary: KSVObsolete is no longer nullptr and needs to be checked explicitly. Also did some minor code cleanup and added a stat counter to track superversion cleanups incurred in the foreground.
Test Plan: make check
Reviewers: ljin
Reviewed By: ljin
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16701
Summary: Moved LogBuffer class to an internal header. Removed some unneccesary indirection. Enabled log buffer for BackgroundCallFlush. Forced log buffer flush right after Unlock to improve time ordering of info log.
Test Plan: make check; db_bench compare LOG output
Reviewers: sdong
Reviewed By: sdong
CC: leveldb, igor
Differential Revision: https://reviews.facebook.net/D16707
Summary:
Add a check at the end of GetImpl to release SuperVersion if it becomes
obsolete. Also do Scrape() inside InstallSuperVersion so it happens more
frequent.
Test Plan:
make all check
running asan_check now
Reviewers: igor, haobo, sdong, dhruba
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16641
Summary: Now while the background thread is picking compactions, it writes out multiple info_logs, especially for universal compaction, which introduces a chance of waiting log writing in mutex, which is bad. To remove this risk, write all those info logs to a buffer and flush it after releasing the mutex.
Test Plan:
make all check
check the log lines while running some tests that trigger compactions.
Reviewers: haobo, igor, dhruba
Reviewed By: dhruba
CC: i.am.jin.lei, dhruba, yhchiang, leveldb, nkg-
Differential Revision: https://reviews.facebook.net/D16515
Summary:
Two changes:
1. DeletionState is only constructed when cleaning up is needed
2. Fix the bug of deletion state construction bug. A change was made in a previous patch: https://reviews.facebook.net/rROCKSDB774ed89c2405ee058086b099cbc8b29e243739cc#71a34e2e However, it somehow got lost when merging
Test Plan: make all check
Reviewers: kailiu, haobo, igor
Reviewed By: igor
CC: igor, dhruba, i.am.jin.lei, yhchiang, leveldb
Differential Revision: https://reviews.facebook.net/D16233
Summary:
This diff does two things:
(1) Log::Reader does not report a corruption when the last record in a log or manifest file is truncated (meaning that log writer died in the middle of the write). Inherited the code from LevelDB: https://code.google.com/p/leveldb/source/detail?r=269fc6ca9416129248db5ca57050cd5d39d177c8#
(2) Turn off mmap writes for all writes to log and manifest files
(2) is necessary because if we use mmap writes, the last record is not truncated, but is actually filled with zeros, making checksum fail. It is hard to recover from checksum failing.
Test Plan:
Added unit tests from LevelDB
Actually recovered a "corrupted" MANIFEST file.
Reviewers: dhruba, haobo
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16119
Summary:
Add an optional input parameter ReadOptions to DB::GetUpdateSince(),
which allows the verification of checksums to be disabled by setting
ReadOptions::verify_checksums to false.
Test Plan: Tests are done off-line and will not be included in the regular unit test.
Reviewers: igor
Reviewed By: igor
CC: leveldb, xjin, dhruba
Differential Revision: https://reviews.facebook.net/D16305
Summary: as title
Test Plan:
asan_check
will post results later
Reviewers: haobo, igor, dhruba, sdong
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16257
Summary:
This will also help with avoiding the deadlock. If a flush failed and we're waiting for a memtable to be flushed, we should schedule a new flush and hope a new one succeedes.
If paranoid_checks = false, Wait() will still hang on ENOSPC, but at least it will automatically continue when the space frees up. Current behavior both hangs and deadlocks.
Also, I renamed some 'compaction' to 'flush'. 'compaction' was leveldb way of saying things.
Test Plan: make check
Reviewers: dhruba, haobo, ljin
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16281
Summary:
More info here: https://github.com/facebook/rocksdb/issues/89
If flush fails because of ENOSPC, we have a deadlock problem. This is a quick fix that will continue the normal operation when user deletes the file and frees up the space on the device.
We need to address the issue more broadly with bg_error_ cleanup.
Test Plan: make check
Reviewers: dhruba, haobo, ljin
Reviewed By: ljin
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16275
Summary: Currently, the first transaction log file ignore bytes_per_sync and other storage-related options. It is not consistent. Fix it.
Test Plan: make all check. See the options set in GDB.
Reviewers: haobo, kailiu
Reviewed By: haobo
CC: igor, ljin, yhchiang, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D16215
Summary: Provide a public API for users to access the table properties for each SSTable.
Test Plan: Added a unit tests to test the function correctness under differnet conditions.
Reviewers: haobo, dhruba, sdong
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16083
Summary:
1. Add some more implementation-aware tests for PlainTable
2. move from a hard-coded one index per 16 rows in one prefix to a configurable number. Also, make hash table ratio = 0 means binary search only. Also fixes some divide 0 risks.
3. Explicitly support total order (only use binary search)
4. some code cleaning up.
Test Plan: make all check
Reviewers: haobo, kailiu
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16023
Summary: Added a bit more information to compaction context, requested by internal team at FB.
Test Plan: Modified CompactionFilter test to make sure is_manual_compaction is properly set.
Reviewers: haobo
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16095
Summary: Clean up IOErrors so that it only indicates errors talking to device.
Test Plan: make all check
Reviewers: igor, haobo, dhruba, emayanke
Reviewed By: igor
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15831
Summary: This covers existing table files before DB open happens and avoids contention on table cache
Test Plan: db_test
Reviewers: haobo, sdong, igor, dhruba
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16089
Summary:
Use super_version insider NewIterator to avoid Ref() each component
separately under mutex
The new added bench shows NewIterator QPS increases from 515K to 719K
No meaningful improvement for multiget I guess due to its relatively small
cost comparing to 90 keys fetch in the test.
Test Plan: unit test and db_bench
Reviewers: igor, sdong
Reviewed By: igor
CC: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D15609
Summary: In PlainTable, use one single byte to represent 8 bytes of internal bytes, if seqID = 0 and it is value type (which should be common for bottom most files). It is to save 7 bytes for uncompressed cases.
Test Plan: make all check
Reviewers: haobo, dhruba, kailiu
Reviewed By: haobo
CC: igor, leveldb
Differential Revision: https://reviews.facebook.net/D15489
Summary:
VersionSet::next_file_number_ is always assumed to be strictly greater than VersionSet::log_number_. In our new recovery code, we artificially set log_number_ to be (log_number + 1), so that once we flush, we don't recover from the same log file again (this is important because of merge operator non-idempotence)
When we set VersionSet::log_number_ to (log_number + 1), we also have to mark that file number used, such that next_file_number_ is increased to a legal level. Otherwise, VersionSet might assert.
This has not be a problem so far because here's what happens:
1. assume next_file_number is 5, we're recovering log_number 10
2. in DBImpl::Recover() we call MarkFileNumberUsed with 10. This will set VersionSet::next_file_number_ to 11.
3. If there are some updates, we will call WriteTable0ForRecovery(), which will use file number 11 as a new table file and advance VersionSet::next_file_number_ to 12.
4. When we LogAndApply() with log_number 11, assertion is true: assert(11 <= 12);
However, this was a lucky occurrence. Even though this diff doesn't cause a bug, I think the issue is important to fix.
Test Plan: In column families I have different recovery logic and this code path asserted. When adding MarkFileNumberUsed(log_number + 1) assert is gone.
Reviewers: dhruba, kailiu
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15783
Summary: I didn't figure out the reason why the feature of zeroing out earlier sequence ID is disabled in universal compaction. I do see bottommost_level is set correctly. It should simply work if we remove the constraint of universal compaction.
Test Plan: make all check
Reviewers: haobo, dhruba, kailiu, igor
Reviewed By: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15423
Summary:
In DBImpl we keep track of some statistics internally and expose them via GetProperty(). This diff encapsulates all the internal statistics into a class InternalStatisics. Most of it is copy/paste.
Apart from cleaning up db_impl.cc, this diff is also necessary for Column families, since every column family should have its own CompactionStats, MakeRoomForWrite-stall stats, etc. It's much easier to keep track of it in every column family if it's nicely encapsulated in its own class.
Test Plan: make check
Reviewers: dhruba, kailiu, haobo, sdong, emayanke
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15273
Summary: as title
Test Plan: unit test
Reviewers: haobo, igor, sdong, kailiu, dhruba
Reviewed By: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15435
Summary:
@dhruba, I'm not sure where we need to sync the directory. I implemented the function in Env() and added the dir sync just after we close the newly created file in the builder.
Should I also add FsyncDir() to new files that get created by a compaction?
Test Plan: Confirmed that FsyncDir is returning Status::OK()
Reviewers: dhruba, haobo
Reviewed By: dhruba
CC: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D14751
Summary: There is no reason to have functions NeedCompaction(), MaxCompactionScore() and MaxCompactionScoreLevel() in VersionSet, since they don't access any data in VersionSet.
Test Plan: make check
Reviewers: kailiu, haobo, sdong
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15333
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
Summary: as title
Test Plan:
make all check
What else tests shall I cover?
Reviewers: igor, haobo
CC:
Differential Revision: https://reviews.facebook.net/D15339
Summary:
This diff implements a special type of iterator that doesn't create a snapshot
(can be used to read newly inserted data) and is optimized for doing sequential
reads.
TailingIterator uses current superversion number to determine whether to
invalidate its internal iterators. If the version hasn't changed, it can often
avoid doing expensive seeks over immutable structures (sst files and immutable
memtables).
Test Plan:
* new unit tests
* running LD with this patch
Reviewers: igor, dhruba, haobo, sdong, kailiu
Reviewed By: sdong
CC: leveldb, lovro, march
Differential Revision: https://reviews.facebook.net/D15285
Summary:
This diff does two things:
* Rethinks how we call Recover() with read_only option. Before, we call it with pointer to memtable where we'd like to apply those changes to. This memtable is set in db_impl_readonly.cc and it's actually DBImpl::mem_. Why don't we just apply updates to mem_ right away? It seems more intuitive.
* Changes when we apply updates to manifest. Before, the process is to recover all the logs, flush it to sst files and then do one giant commit that atomically adds all recovered sst files and sets the next log number. This works good enough, but causes some small troubles for my column family approach, since I can't have one VersionEdit apply to more than single column family[1]. The change here is to commit the files recovered from logs right away. Here is the state of the world before the change:
1. Recover log 5, add new sst files to edit
2. Recover log 7, add new sst files to edit
3. Recover log 8, add new sst files to edit
4. Commit all added sst files to manifest and mark log files 5, 7 and 8 as recoverd (via SetLogNumber(9) function)
After the change, we'll do:
1. Recover log 5, commit the new sst files and set log 5 as recovered
2. Recover log 7, commit the new sst files and set log 7 as recovered
3. Recover log 8, commit the new sst files and set log 8 as recovered
The added (small) benefit is that if we fail after (2), the new recovery will only have to recover log 8. In previous case, we'll have to restart the recovery from the beginning. The bigger benefit will be to enable easier integration of multiple column families in Recovery code path.
[1] I'm happy to dicuss this decison, but I believe this is the cleanest way to go. It also makes backward compatibility much easier. We don't have a requirement of adding multiple column families atomically.
Test Plan: make check
Reviewers: dhruba, haobo, kailiu, sdong
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15237
Summary:
This moves the use of versions_ to before the mutex is unlocked
to avoid a possible race.
Task ID: #
Blame Rev:
Test Plan:
make check
Revert Plan:
Database Impact:
Memcache Impact:
Other Notes:
EImportant:
- begin *PUBLIC* platform impact section -
Bugzilla: #
- end platform impact -
Reviewers: haobo, dhruba
Reviewed By: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15279
