Summary:
The file tree structure in Version is prebuilt and the range of each file is known.
On the Get() code path, we do binary search in FindFile() by comparing
target key with each file's largest key and also check the range for each L0 file.
With some pre-calculated knowledge, each key comparision that has been done can serve
as a hint to narrow down further searches:
(1) If a key falls within a L0 file's range, we can safely skip the next
file if its range does not overlap with the current one.
(2) If a key falls within a file's range in level L0 - Ln-1, we should only
need to binary search in the next level for files that overlap with the current one.
(1) will be able to skip some files depending one the key distribution.
(2) can greatly reduce the range of binary search, especially for bottom
levels, given that one file most likely only overlaps with N files from
the level below (where N is max_bytes_for_level_multiplier). So on level
L, we will only look at ~N files instead of N^L files.
Some inital results: measured with 500M key DB, when write is light (10k/s = 1.2M/s), this
improves QPS ~7% on top of blocked bloom. When write is heavier (80k/s =
9.6M/s), it gives us ~13% improvement.
Test Plan: make all check
Reviewers: haobo, igor, dhruba, sdong, yhchiang
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D17205
Summary:
D17961 has two bugs:
(1) two level iterator fails to populate FileMetaData.table_reader, causing performance regression.
(2) table cache handle the !status.ok() case in the wrong place, causing seg fault which shouldn't happen.
Test Plan: make all check
Reviewers: ljin, igor, haobo
Reviewed By: ljin
CC: yhchiang, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D17991
Summary:
One of our profilings shows that Version::Get() sometimes is slow when getting pointer of user comparators or other global objects. In this patch:
(1) we keep pointers of immutable objects in Version to avoid accesses them though option objects or cfd objects
(2) table_reader is directly cached in FileMetaData so that table cache don't have to go through handle first to fetch it
(3) If level 0 has less than 3 files, skip the filtering logic based on SST tables' key range. Smallest and largest key are stored in separated memory locations, which has potential cache misses
Test Plan: make all check
Reviewers: haobo, ljin
Reviewed By: haobo
CC: igor, yhchiang, nkg-, leveldb
Differential Revision: https://reviews.facebook.net/D17739
Summary:
Introducing RocksDBLite! Removes all the non-essential features and reduces the binary size. This effort should help our adoption on mobile.
Binary size when compiling for IOS (`TARGET_OS=IOS m static_lib`) is down to 9MB from 15MB (without stripping)
Test Plan: compiles :)
Reviewers: dhruba, haobo, ljin, sdong, yhchiang
Reviewed By: yhchiang
CC: leveldb
Differential Revision: https://reviews.facebook.net/D17835
Summary:
With multiple column families, especially when manual Flush is executed, we might roll the log file, although the current log file is empty (no data has been written to the log).
After the diff, we won't create new log file if current is empty.
Next, I will write an algorithm that will flush column families that reference old log files (i.e., that weren't flushed in a while)
Test Plan: Added an unit test. Confirmed that unit test failes in master
Reviewers: dhruba, haobo, ljin, sdong
Reviewed By: ljin
CC: leveldb
Differential Revision: https://reviews.facebook.net/D17631
Summary: When opening DB in read-only mode, client can choose to only specify a subset of column families ("default" column family can't be omitted, though)
Test Plan: added a unit test in column_family_test
Reviewers: haobo, sdong, ljin, dhruba
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D17565
Summary: It is a regression valgrind bug caused by using FileMetaData as index block handle. One of the fields of FileMetaData is not initialized after being contructed and copied, but I'm not able to find which one. Also, I realized that it's not a good idea to use FileMetaData as in TwoLevelIterator::InitDataBlock(), a copied FileMetaData can be compared with the one in version set byte by byte, but the refs can be changed. Also comparing such a large structure is slightly more expensive. Use a simpler structure instead
Test Plan:
Run the failing valgrind test (Harness.RandomizedLongDB)
make all check
Reviewers: igor, haobo, ljin
Reviewed By: igor
CC: yhchiang, leveldb
Differential Revision: https://reviews.facebook.net/D17409
Summary: Talked to <insert internal project name> folks and they found it really scary that they won't be able to roll back once they upgrade to 2.8. We should fix this.
Test Plan: make check
Reviewers: haobo, ljin
Reviewed By: ljin
CC: leveldb
Differential Revision: https://reviews.facebook.net/D17343
Summary:
Previously, we used to sort all files by BySmallestFirst comparator and then re-sort level0 files in the Finalize() (recently moved to end of SaveTo).
In this diff, I chose the correct comparator at the beginning and sort the files correctly in Builder::SaveTo.
I also added a verification that all files are sorted correctly in CheckConsistency()
NOTE: This diff depends on D17037
Test Plan: make check. Will also run db_stress
Reviewers: dhruba, haobo, sdong, ljin
Reviewed By: ljin
CC: leveldb
Differential Revision: https://reviews.facebook.net/D17049
Summary: We don't preallocate MANIFEST file, even though we have an option for that. This diff preallocates manifest file every time we create it
Test Plan: make check
Reviewers: dhruba, haobo
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D17163
Summary:
This is the part that was not finished when doing the Options.max_num_files=-1 feature. For iterating non level0 SST files (which was done using two level iterator), table cache is not bypassed. With this patch, the leftover feature is done.
Test Plan: make all check; change Options.max_num_files=-1 in one of the tests to cover the codes.
Reviewers: haobo, igor, dhruba, ljin, yhchiang
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D17001
Summary:
* PartialMerge api now takes a list of operands instead of two operands.
* Add min_pertial_merge_operands to Options, indicating the minimum
number of operands to trigger partial merge.
* This diff is based on Schalk's previous diff (D14601), but it also
includes necessary changes such as updating the pure C api for
partial merge.
Test Plan:
* make check all
* develop tests for cases where partial merge takes more than two
operands.
TODOs (from Schalk):
* Add test with min_partial_merge_operands > 2.
* Perform benchmarks to measure the performance improvements (can probably
use results of task #2837810.)
* Add description of problem to doc/index.html.
* Change wiki pages to reflect the interface changes.
Reviewers: haobo, igor, vamsi
Reviewed By: haobo
CC: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D16815
Summary:
As it turns out, we need the call to ComputeCompactionScore (previously: Finalize) in CompactionPicker.
The issue caused a deadlock in db_stress: http://ci-builds.fb.com/job/rocksdb_crashtest/290/console
The last two lines before a deadlock were:
2014/03/18-22:43:41.481029 7facafbee700 (Original Log Time 2014/03/18-22:43:41.480989) Compaction nothing to do
2014/03/18-22:43:41.481041 7faccf7fc700 wait for fewer level0 files...
"Compaction nothing to do" and other thread waiting for fewer level0 files. Hm hm.
I moved the pre-sorting to SaveTo, which should fix both the original and the new issue.
Test Plan: make check for now, will run db_stress in jenkins
Reviewers: dhruba, haobo, sdong
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D17037
Summary:
This is a 500ns operation while the whole Get() call takes only a few
micro!
Test Plan: ran db_bench, for a DB with 50M keys, QPS jumps from 5.2M/s to 7.2M/s
Reviewers: haobo, igor, dhruba
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D17007
Summary:
Finalize re-sorts (read: mutates) the files_ in Version* and it is called by CompactionPicker during normal runtime. At the same time, this same Version* lives in the SuperVersion* and is accessed without the mutex in GetImpl() code path.
Mutating the files_ in one thread and reading the same files_ in another thread is a bad idea. It caused this issue: http://ci-builds.fb.com/job/rocksdb_crashtest/285/console
Long-term, we need to be more careful with method contracts and clearly document what state can be mutated when. Now that we are much faster because we don't lock in GetImpl(), we keep running into data races that were not a problem before when we were slower. db_stress has been very helpful in detecting those.
Short-term, I removed Finalize() from CompactionPicker.
Note: I believe this is an issue in current 2.7 version running in production.
Test Plan:
make check
Will also run db_stress to see if issue is gone
Reviewers: sdong, ljin, dhruba, haobo
Reviewed By: sdong
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16983
Summary:
There is a chance that an old MANIFEST is corrupted in 2.7 but just not noticed.
This check would fail them. Change it to log instead of returning a
Corruption status.
Test Plan: make
Reviewers: haobo, igor
Reviewed By: igor
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16923
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:
In the last fix, I forgot to point to the writer when comparing edit,
which is apparently not correct.
Test Plan: still running make whitebox_crash_test
Reviewers: igor, haobo, igor2
Reviewed By: igor2
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16911
Summary:
Here is what it can cause probelm:
There is one memtable flush and one compaction. Both call LogAndApply(). If both edits are applied in the same batch with flush edit first and the compaction edit followed. LogAndApplyHelper() will assign compaction edit current VersionSet's log number(which should be smaller than the log number from flush edit). It cause log_numbers in MANIFEST to be not monotonic increasing, which violates the assume Recover() makes. What is more is after comitting to MANIFEST file, log_number_ in VersionSet is updated to the log_number from the last edit, which is the compaction one. It ends up not updating the log_number.
Test Plan:
make whitebox_crash_test
got another assertion about iter->valid(), not sure if that is related
to this.
Reviewers: igor, haobo
Reviewed By: igor
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16875
Summary:
I'm cleaning up some code preparing for the big diff review tomorrow. This is the first part of the cleanup.
Changes are mostly cosmetic. The goal is to decrease amount of code difference between columnfamilies and master branch.
This diff also fixes race condition when dropping column family.
Test Plan: Ran db_stress with variety of parameters
Reviewers: dhruba, haobo
Differential Revision: https://reviews.facebook.net/D16833
Summary: Column family should be dropped after the change has been commited
Test Plan: db stress
Reviewers: dhruba, haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16779
Summary:
If verify_checksums_in_compaction is true, compaction will verify checksums. This is default.
If it's false, compaction doesn't verify checksums. This is useful for in-memory workloads.
Test Plan: corruption_test
Reviewers: dhruba
Reviewed By: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16695
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:
Column family IDs should be unique, even if column family is dropped. To achieve this, we save max column family in manifest.
Note that the diff is still not ready. I'm only using differential to move the patch to my Mac machine.
Test Plan: added a test to column_family_test
Reviewers: dhruba, haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16581
Summary:
EncodeTo(&record) does not overwrite, it appends to it.
This means that group commit log and apply will look something like:
record1
record1record2
record1record2record3
I'm surprised this didn't show up in production, but I think the reason is that MANIFEST group commit almost never happens.
This bug turned up in column family work, where opening a database failed with "adding a same column family twice".
Test Plan: Tested the change in column family branch and observed that the problem is gone (with db_stress)
Reviewers: dhruba, haobo
Reviewed By: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16461
Summary:
Previous code is too convoluted and I must be drunk for letting
such code to be written without a second thought.
Thanks to the discussion with @sdong, I added the `Options` when
generating the flusher, thus avoiding the tricks.
Just FYI: I resisted to add Options in flush_block_policy.h since I
wanted to avoid cyclic dependencies: FlushBlockPolicy dpends on Options
and Options also depends FlushBlockPolicy... While I appreciate my
effort to prevent it, the old design turns out creating more troubles than
it tried to avoid.
Test Plan: ran ./table_test
Reviewers: sdong
Reviewed By: sdong
CC: sdong, leveldb
Differential Revision: https://reviews.facebook.net/D16503
Summary:
I though I might get away with as little changes to LogAndApply() as possible. It turns out this is not the case.
This diff introduces different behavior of LogAndApply() for three cases:
1. column family add
2. column family drop
3. no-column family manipulation
(1) and (2) don't support group commit yet.
There were a lot of problems with old version od LogAndApply, detected by db_stress. The biggest was non-atomicity of manifest writes and metadata changes (i.e. if column family add is in manifest, it also has to be in in-memory data structure).
Test Plan: db_stress
Reviewers: dhruba, haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16491
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:
This fixes few bugs with CreateColumnFamily
* We first have to LogAndApply and then call VersionSet::CreateColumnFamily. Otherwise, WriteSnapshot might be invoked, writing out column family add inside of LogAndApply, even though it's not really committed
* Fix LogAndApplyHelper() to not apply log number to column_family_data, which is in case of column family add, just a dummy (default) column family
* Create SuperVerion when creating column family
Test Plan: column_family_test
Reviewers: dhruba, haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16443
