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:
- Implement Prepare method
- Rewrite performance tests in cuckoo_table_reader_test to write new file only if one doesn't already exist.
- Add performance tests for batch lookup along with prefetching.
Test Plan:
./cuckoo_table_reader_test --enable_perf
Results (We get better results if we used int64 comparator instead of string comparator (TBD in future diffs)):
With 100000000 items and hash table ratio 0.500000, number of hash functions used: 2.
Time taken per op is 0.208us (4.8 Mqps) with batch size of 0
With 100000000 items and hash table ratio 0.500000, number of hash functions used: 2.
Time taken per op is 0.182us (5.5 Mqps) with batch size of 10
With 100000000 items and hash table ratio 0.500000, number of hash functions used: 2.
Time taken per op is 0.161us (6.2 Mqps) with batch size of 25
With 100000000 items and hash table ratio 0.500000, number of hash functions used: 2.
Time taken per op is 0.161us (6.2 Mqps) with batch size of 50
With 100000000 items and hash table ratio 0.500000, number of hash functions used: 2.
Time taken per op is 0.163us (6.1 Mqps) with batch size of 100
With 100000000 items and hash table ratio 0.600000, number of hash functions used: 3.
Time taken per op is 0.252us (4.0 Mqps) with batch size of 0
With 100000000 items and hash table ratio 0.600000, number of hash functions used: 3.
Time taken per op is 0.192us (5.2 Mqps) with batch size of 10
With 100000000 items and hash table ratio 0.600000, number of hash functions used: 3.
Time taken per op is 0.195us (5.1 Mqps) with batch size of 25
With 100000000 items and hash table ratio 0.600000, number of hash functions used: 3.
Time taken per op is 0.191us (5.2 Mqps) with batch size of 50
With 100000000 items and hash table ratio 0.600000, number of hash functions used: 3.
Time taken per op is 0.194us (5.1 Mqps) with batch size of 100
With 100000000 items and hash table ratio 0.750000, number of hash functions used: 3.
Time taken per op is 0.228us (4.4 Mqps) with batch size of 0
With 100000000 items and hash table ratio 0.750000, number of hash functions used: 3.
Time taken per op is 0.185us (5.4 Mqps) with batch size of 10
With 100000000 items and hash table ratio 0.750000, number of hash functions used: 3.
Time taken per op is 0.186us (5.4 Mqps) with batch size of 25
With 100000000 items and hash table ratio 0.750000, number of hash functions used: 3.
Time taken per op is 0.189us (5.3 Mqps) with batch size of 50
With 100000000 items and hash table ratio 0.750000, number of hash functions used: 3.
Time taken per op is 0.188us (5.3 Mqps) with batch size of 100
With 100000000 items and hash table ratio 0.900000, number of hash functions used: 3.
Time taken per op is 0.325us (3.1 Mqps) with batch size of 0
With 100000000 items and hash table ratio 0.900000, number of hash functions used: 3.
Time taken per op is 0.196us (5.1 Mqps) with batch size of 10
With 100000000 items and hash table ratio 0.900000, number of hash functions used: 3.
Time taken per op is 0.199us (5.0 Mqps) with batch size of 25
With 100000000 items and hash table ratio 0.900000, number of hash functions used: 3.
Time taken per op is 0.196us (5.1 Mqps) with batch size of 50
With 100000000 items and hash table ratio 0.900000, number of hash functions used: 3.
Time taken per op is 0.209us (4.8 Mqps) with batch size of 100
Reviewers: sdong, yhchiang, igor, ljin
Reviewed By: ljin
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22167
Summary:
Currently, PlainTable must use mmap_reads. When PlainTable is used but
allow_mmap_reads is not set, rocksdb will fail in flush.
This diff improve Options sanitization and add MmapReadRequired() to
TableFactory.
Test Plan:
export ROCKSDB_TESTS=PlainTableOptionsSanitizeTest
make db_test -j32
./db_test
Reviewers: sdong, ljin
Reviewed By: ljin
Subscribers: you, leveldb
Differential Revision: https://reviews.facebook.net/D21939
Summary:
ManifestDumpCommand::DoCommand was allocating a VersionSet and never
freeing it.
Test Plan: make
Reviewers: igor
Reviewed By: igor
Differential Revision: https://reviews.facebook.net/D22221
Summary: A previous change triggered a change by mistake: DestroyDB() will keep info logs under DB directory. Revert the unintended change.
Test Plan: Add a unit test case to verify it.
Reviewers: ljin, yhchiang, igor
Reviewed By: igor
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22209
Summary: We need to start compression at level 1, while OptimizeForLevelComapaction() only sets up rocksdb to start compressing at level 2. I also adjusted some other things.
Test Plan: compiles
Reviewers: yinwang
Reviewed By: yinwang
Differential Revision: https://reviews.facebook.net/D22203
Summary: Make table_reader_bench cover all the three table formats.
Test Plan: Run it using three options
Reviewers: radheshyamb, ljin
Reviewed By: ljin
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22137
Summary: I was checking some functions in coding.h and coding.cc when I noticed these unused functions. Let's remove them.
Test Plan: compiles
Reviewers: sdong, ljin, yhchiang, dhruba
Reviewed By: dhruba
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22077
Summary:
Adding num_column_families flag. Adding support for column families in DoWrite and ReadRandom methods.
[Igor, please let me know if this approach sounds good. I shall add it to other methods too.]
Test Plan: Ran fillseq on 1M keys and 10 Column families and ran readrandom.
Reviewers: sdong, yhchiang, igor, ljin
Reviewed By: ljin
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D21387
Summary:
Add WriteBatchWithIndex so that a user can query data out of a WriteBatch, to support MongoDB's read-its-own-write.
WriteBatchWithIndex uses a skiplist to store the binary index. The index stores the offset of the entry in the write batch. When searching for a key, the key for the entry is read by read the entry from the write batch from the offset.
Define a new iterator class for querying data out of WriteBatchWithIndex. A user can create an iterator of the write batch for one column family, seek to a key and keep calling Next() to see next entries.
I will add more unit tests if people are OK about this API.
Test Plan:
make all check
Add unit tests.
Reviewers: yhchiang, igor, MarkCallaghan, ljin
Reviewed By: ljin
Subscribers: dhruba, leveldb, xjin
Differential Revision: https://reviews.facebook.net/D21381
Igor Canadi