Summary: It only covers Open() with default column family right now
Test Plan: make release
Reviewers: igor, yhchiang, sdong
Reviewed By: sdong
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22467
Summary:
Before this diff, whenever we Write to non-existing column family, Write() would fail.
This diff adds an option to not fail a Write() when WriteBatch points to non-existing column family. MongoDB said this would be useful for them, since they might have a transaction updating an index that was dropped by another thread. This way, they don't have to worry about checking if all indexes are alive on every write. They don't care if they lose writes to dropped index.
Test Plan: added a small unit test
Reviewers: sdong, yhchiang, ljin
Reviewed By: ljin
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22143
Summary:
1. assert db->Put to be true in db_stress
2. begin column family with name "1".
Test Plan: 1. ./db_stress
Reviewers: ljin, yhchiang, dhruba, sdong, igor
Reviewed By: sdong, igor
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22659
Summary: 1. db/db_impl.cc:2324 (DBImpl::BackgroundCompaction) should not raise bg_error_ when column family is dropped during compaction.
Test Plan: 1. db_stress
Reviewers: ljin, yhchiang, dhruba, igor, sdong
Reviewed By: igor
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22653
This eliminates the need to remember to call PERF_TIMER_STOP when a section has
been timed. This allows more useful design with the perf timers and enables
possible return value optimizations. Simplistic example:
class Foo {
public:
Foo(int v) : m_v(v);
private:
int m_v;
}
Foo makeFrobbedFoo(int *errno)
{
*errno = 0;
return Foo();
}
Foo bar(int *errno)
{
PERF_TIMER_GUARD(some_timer);
return makeFrobbedFoo(errno);
}
int main(int argc, char[] argv)
{
Foo f;
int errno;
f = bar(&errno);
if (errno)
return -1;
return 0;
}
After bar() is called, perf_context.some_timer would be incremented as if
Stop(&perf_context.some_timer) was called at the end, and the compiler is still
able to produce optimizations on the return value from makeFrobbedFoo() through
to main().
Summary: gcc on our dev boxes is not happy about __attribute__((unused))
Test Plan: compiles now
Reviewers: sdong, ljin
Reviewed By: ljin
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22707
Summary: We need to set contbuild for this :)
Test Plan: compiles
Reviewers: sdong, yhchiang, ljin
Reviewed By: ljin
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22701
Summary:
I have an application configured with 16 background threads. Write rates are high. L0->L1 compactions is very slow and it limits the concurrency of the system. While it's happening, other 15 threads are idle. However, when there is a need of a flush, that one thread busy with L0->L1 is doing flush, instead of any other 15 threads that are just sitting there.
This diff prevents that. If there are threads that are idle, we don't let flush preempt compaction.
Test Plan: Will run stress test
Reviewers: ljin, sdong, yhchiang
Reviewed By: sdong, yhchiang
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D22299
Summary:
BlockBasedTable sst file size can grow to a large size when universal
compaction is used. When index block exceeds 2G, pread seems to fail and
return truncated data and causes "trucated block" error. I tried to use
```
#define _FILE_OFFSET_BITS 64
```
But the problem still persists. Splitting a big write/read into smaller
batches seems to solve the problem.
Test Plan:
successfully compacted a case with resulting sst file at ~90G (2.1G
index block size)
Reviewers: yhchiang, igor, sdong
Reviewed By: sdong
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22569
Summary:
Use inlined hash functions instead of function pointer. Make number of buckets a power of two and use bitwise and instead of mod.
After these changes, we get almost 50% improvement in performance.
Results:
With 120000000 items, utilization is 89.41%, number of hash functions: 2.
Time taken per op is 0.231us (4.3 Mqps) with batch size of 0
Time taken per op is 0.229us (4.4 Mqps) with batch size of 0
Time taken per op is 0.185us (5.4 Mqps) with batch size of 0
With 120000000 items, utilization is 89.41%, number of hash functions: 2.
Time taken per op is 0.108us (9.3 Mqps) with batch size of 10
Time taken per op is 0.100us (10.0 Mqps) with batch size of 10
Time taken per op is 0.103us (9.7 Mqps) with batch size of 10
With 120000000 items, utilization is 89.41%, number of hash functions: 2.
Time taken per op is 0.101us (9.9 Mqps) with batch size of 25
Time taken per op is 0.098us (10.2 Mqps) with batch size of 25
Time taken per op is 0.097us (10.3 Mqps) with batch size of 25
With 120000000 items, utilization is 89.41%, number of hash functions: 2.
Time taken per op is 0.100us (10.0 Mqps) with batch size of 50
Time taken per op is 0.097us (10.3 Mqps) with batch size of 50
Time taken per op is 0.097us (10.3 Mqps) with batch size of 50
With 120000000 items, utilization is 89.41%, number of hash functions: 2.
Time taken per op is 0.102us (9.8 Mqps) with batch size of 100
Time taken per op is 0.098us (10.2 Mqps) with batch size of 100
Time taken per op is 0.115us (8.7 Mqps) with batch size of 100
With 100000000 items, utilization is 74.51%, number of hash functions: 2.
Time taken per op is 0.201us (5.0 Mqps) with batch size of 0
Time taken per op is 0.155us (6.5 Mqps) with batch size of 0
Time taken per op is 0.152us (6.6 Mqps) with batch size of 0
With 100000000 items, utilization is 74.51%, number of hash functions: 2.
Time taken per op is 0.089us (11.3 Mqps) with batch size of 10
Time taken per op is 0.084us (11.9 Mqps) with batch size of 10
Time taken per op is 0.086us (11.6 Mqps) with batch size of 10
With 100000000 items, utilization is 74.51%, number of hash functions: 2.
Time taken per op is 0.087us (11.5 Mqps) with batch size of 25
Time taken per op is 0.085us (11.7 Mqps) with batch size of 25
Time taken per op is 0.093us (10.8 Mqps) with batch size of 25
With 100000000 items, utilization is 74.51%, number of hash functions: 2.
Time taken per op is 0.094us (10.6 Mqps) with batch size of 50
Time taken per op is 0.094us (10.7 Mqps) with batch size of 50
Time taken per op is 0.093us (10.8 Mqps) with batch size of 50
With 100000000 items, utilization is 74.51%, number of hash functions: 2.
Time taken per op is 0.092us (10.9 Mqps) with batch size of 100
Time taken per op is 0.089us (11.2 Mqps) with batch size of 100
Time taken per op is 0.088us (11.3 Mqps) with batch size of 100
With 80000000 items, utilization is 59.60%, number of hash functions: 2.
Time taken per op is 0.154us (6.5 Mqps) with batch size of 0
Time taken per op is 0.168us (6.0 Mqps) with batch size of 0
Time taken per op is 0.190us (5.3 Mqps) with batch size of 0
With 80000000 items, utilization is 59.60%, number of hash functions: 2.
Time taken per op is 0.081us (12.4 Mqps) with batch size of 10
Time taken per op is 0.077us (13.0 Mqps) with batch size of 10
Time taken per op is 0.083us (12.1 Mqps) with batch size of 10
With 80000000 items, utilization is 59.60%, number of hash functions: 2.
Time taken per op is 0.077us (13.0 Mqps) with batch size of 25
Time taken per op is 0.073us (13.7 Mqps) with batch size of 25
Time taken per op is 0.073us (13.7 Mqps) with batch size of 25
With 80000000 items, utilization is 59.60%, number of hash functions: 2.
Time taken per op is 0.076us (13.1 Mqps) with batch size of 50
Time taken per op is 0.072us (13.8 Mqps) with batch size of 50
Time taken per op is 0.072us (13.8 Mqps) with batch size of 50
With 80000000 items, utilization is 59.60%, number of hash functions: 2.
Time taken per op is 0.077us (13.0 Mqps) with batch size of 100
Time taken per op is 0.074us (13.6 Mqps) with batch size of 100
Time taken per op is 0.073us (13.6 Mqps) with batch size of 100
With 70000000 items, utilization is 52.15%, number of hash functions: 2.
Time taken per op is 0.190us (5.3 Mqps) with batch size of 0
Time taken per op is 0.186us (5.4 Mqps) with batch size of 0
Time taken per op is 0.184us (5.4 Mqps) with batch size of 0
With 70000000 items, utilization is 52.15%, number of hash functions: 2.
Time taken per op is 0.079us (12.7 Mqps) with batch size of 10
Time taken per op is 0.070us (14.2 Mqps) with batch size of 10
Time taken per op is 0.072us (14.0 Mqps) with batch size of 10
With 70000000 items, utilization is 52.15%, number of hash functions: 2.
Time taken per op is 0.080us (12.5 Mqps) with batch size of 25
Time taken per op is 0.072us (14.0 Mqps) with batch size of 25
Time taken per op is 0.071us (14.1 Mqps) with batch size of 25
With 70000000 items, utilization is 52.15%, number of hash functions: 2.
Time taken per op is 0.082us (12.1 Mqps) with batch size of 50
Time taken per op is 0.071us (14.1 Mqps) with batch size of 50
Time taken per op is 0.073us (13.6 Mqps) with batch size of 50
With 70000000 items, utilization is 52.15%, number of hash functions: 2.
Time taken per op is 0.080us (12.5 Mqps) with batch size of 100
Time taken per op is 0.077us (13.0 Mqps) with batch size of 100
Time taken per op is 0.078us (12.8 Mqps) with batch size of 100
Test Plan:
make check all
make valgrind_check
make asan_check
Reviewers: sdong, ljin
Reviewed By: ljin
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22539
Summary:
When reading from kBlockCacheTier, ForwardIterator's internal child iterators
may end up in the incomplete state (read was unable to complete without doing
disk I/O). `ForwardIterator::status()` will correctly report that; however, the
iterator may be stuck in that state until all sub-iterators are rebuilt:
* `NeedToSeekImmutable()` may return false even if some sub-iterators are
incomplete
* one of the child iterators may be an empty iterator without any state other
that the kIncomplete status (created using `NewErrorIterator()`); seeking on
any such iterator has no effect -- we need to construct it again
Akin to rebuilding iterators after a superversion bump, this diff makes forward
iterator reset all incomplete child iterators when `Seek()` or `Next()` are
called.
Test Plan: TEST_TMPDIR=/dev/shm/rocksdbtest ROCKSDB_TESTS=TailingIterator ./db_test
Reviewers: igor, sdong, ljin
Reviewed By: ljin
Subscribers: lovro, march, leveldb
Differential Revision: https://reviews.facebook.net/D22575
Summary: It is too expensive to bump ticker to every key/vaue pair
Test Plan: make release
Reviewers: sdong, yhchiang, igor
Reviewed By: igor
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22527
Summary:
1. remove class InternalFilterPolicy in db/dbformat.h
2. Transformation from internal key to user key is done in filter_block.cc
3. This is a preparation for patch D20979
Test Plan:
make all check
valgrind ./db_test
Reviewers: igor, yhchiang, ljin, sdong
Reviewed By: sdong
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22509
Summary: This implements a cache friendly version of Cuckoo Hash in which, in case of collission, we try to insert in next few locations. The size of the neighborhood to check is taken as an input parameter in builder and stored in the table.
Test Plan:
make check all
cuckoo_table_{db,reader,builder}_test
Reviewers: sdong, ljin
Reviewed By: ljin
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22455
Summary:
Based on discussions from t4982833. This is just a short-term fix, I plan to revamp manual compaction process as part of t4982812.
Also, I think we should schedule automatic compactions at the very end of manual compactions, not when we're done with one level. I made that change as part of this diff. Let me know if you disagree.
Test Plan: make check for now
Reviewers: sdong, tnovak, yhchiang, ljin
Reviewed By: yhchiang
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22401
Summary: No __thread for ios.
Test Plan: compile works for ios now
Reviewers: ljin, dhruba
Reviewed By: dhruba
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22491
Summary: This assert makes Insert O(n^2) instead of O(n) in debug mode. Memtable insert is in the critical path. No need to assert uniqunnes of the key here, since we're adding a sequence number to it anyway.
Test Plan: none
Reviewers: sdong, ljin
Reviewed By: ljin
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22443
Summary:
- New Uint64 comparator
- Modify Reader and Builder to take custom user comparators instead of bytewise comparator
- Modify logic for choosing unused user key in builder
- Modify iterator logic in reader
- test changes
Test Plan:
cuckoo_table_{builder,reader,db}_test
make check all
Reviewers: ljin, sdong
Reviewed By: ljin
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D22377
Summary: In DBImpl::Recover method, while loading memtables, also check if memtables are empty. Use this in DBImplReadonly to determine whether to lookup memtable or not.
Test Plan:
db_test
make check all
Reviewers: sdong, yhchiang, ljin, igor
Reviewed By: ljin
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22281
Summary: also fix HISTORY.md
Test Plan: make all check
Reviewers: sdong, yhchiang, igor
Reviewed By: igor
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22437
Summary:
It was creating BlockBasedTableOptions object in a loop without calling
destroy()
Test Plan: valgrind ./c_test --leak-check=full --show-reachable=yes
Reviewers: sdong, igor
Reviewed By: igor
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22431
Summary:
Two things:
1. Use hash-based index for data column family
2. Use Get() instead of Iterator Seek() when DB is opened read-only
Test Plan: added read-only test in unit test
Reviewers: yinwang
Reviewed By: yinwang
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22323
Summary: Add a virtual function in table factory that will print table options
Test Plan: make release
Reviewers: igor, yhchiang, sdong
Reviewed By: sdong
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22149
Summary: as title
Test Plan:
tested on my mac
make rocksdbjava
make jtest
Reviewers: sdong, igor, yhchiang
Reviewed By: yhchiang
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D21963
Summary:
I will move compression related options in a separate diff since this
diff is already pretty lengthy.
I guess I will also need to change JNI accordingly :(
Test Plan: make all check
Reviewers: yhchiang, igor, sdong
Reviewed By: igor
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D21915
Summary:
I am currently working on a project that uses RocksDB. While debugging some perf issues, I came up across interesting compaction concurrency issue. Namely, I had 15 idle threads and a good comapction to do, but CompactionPicker returned "Compaction nothing to do". Here's how Internal stats looked:
2014/08/22-08:08:04.551982 7fc7fc3f5700 ------- DUMPING STATS -------
2014/08/22-08:08:04.552000 7fc7fc3f5700
** Compaction Stats [default] **
Level Files Size(MB) Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) RW-Amp W-Amp Rd(MB/s) Wr(MB/s) Rn(cnt) Rnp1(cnt) Wnp1(cnt) Wnew(cnt) Comp(sec) Comp(cnt) Avg(sec) Stall(sec) Stall(cnt) Avg(ms)
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
L0 7/5 353 1.0 0.0 0.0 0.0 2.3 2.3 0.0 0.0 0.0 9.4 0 0 0 0 247 46 5.359 8.53 1 8526.25
L1 2/2 86 1.3 2.6 1.9 0.7 2.6 1.9 2.7 1.3 24.3 24.0 39 19 71 52 109 11 9.938 0.00 0 0.00
L2 26/0 833 1.3 5.7 1.7 4.0 5.2 1.2 6.3 3.0 15.6 14.2 47 112 147 35 373 44 8.468 0.00 0 0.00
L3 12/0 505 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0 0 0 0 0 0 0.000 0.00 0 0.00
Sum 47/7 1778 0.0 8.3 3.6 4.6 10.0 5.4 8.1 4.4 11.6 14.1 86 131 218 87 728 101 7.212 8.53 1 8526.25
Int 0/0 0 0.0 2.4 0.8 1.6 2.7 1.2 11.5 6.1 12.0 13.6 20 43 63 20 203 23 8.845 0.00 0 0.00
Flush(GB): accumulative 2.266, interval 0.444
Stalls(secs): 0.000 level0_slowdown, 0.000 level0_numfiles, 8.526 memtable_compaction, 0.000 leveln_slowdown_soft, 0.000 leveln_slowdown_hard
Stalls(count): 0 level0_slowdown, 0 level0_numfiles, 1 memtable_compaction, 0 leveln_slowdown_soft, 0 leveln_slowdown_hard
** DB Stats **
Uptime(secs): 336.8 total, 60.4 interval
Cumulative writes: 61584000 writes, 6480589 batches, 9.5 writes per batch, 1.39 GB user ingest
Cumulative WAL: 0 writes, 0 syncs, 0.00 writes per sync, 0.00 GB written
Interval writes: 11235257 writes, 1175050 batches, 9.6 writes per batch, 259.9 MB user ingest
Interval WAL: 0 writes, 0 syncs, 0.00 writes per sync, 0.00 MB written
To see what happened, go here: 47b452cfcf/db/compaction_picker.cc (L430)
* The for loop started with level 1, because it has the worst score.
* PickCompactionBySize on L429 returned nullptr because all files were being compacted
* ExpandWhileOverlapping(c) returned true (because that's what it does when it gets nullptr!?)
* for loop break-ed, never trying compactions for level 2 :( :(
This bug was present at least since January. I have no idea how we didn't find this sooner.
Test Plan:
Unit testing compaction picker is hard. I tested this by running my service and observing L0->L1 and L2->L3 compactions in parallel. However, for long-term, I opened the task #4968469