Paired-Bloom-Filter: Balancing rounding to batches between the bottom-most level and other levels

[noamhaham]

No description provided.

[udi-speedb]

@noamhaham - Please run make format before pushing (and please do so before pushing the next commit with the updates)

[udi-speedb]

Please cleanup (remove the comment lines with "dead" code)

[udi-speedb]

Assuming you intend num_blocks_ to be a double in the calculation, please use
static_cast<double>(num_blocks_)

What I mean is to replace the "+0.0" with a static_cast

[udi-speedb]

Same here
you are using std::ceil as if it were a template which it isn't (std::ceil<size_t>)

Also, the function is called RoundDownUsableSpace and you may set rv to a value > available_size which I think is problematic (and definitely misleading).

[udi-speedb]

There is still the issue I have pointed out in the previous cycle:

Also, the function is called RoundDownUsableSpace and you may set rv to a value > available_size which I think is problematic (and definitely misleading).

[udi-speedb]

Please remove this comment if not applicable

[udi-speedb]

Indentation.
Did you run make format before committting?

[udi-speedb]

Indentation

[udi-speedb]

There is still the issue I have pointed out in the previous cycle:

Also, the function is called RoundDownUsableSpace and you may set rv to a value > available_size which I think is problematic (and definitely misleading).

[udi-speedb]

@noamhaham

• Did the branch pass 'make check' (all of the unit tests)?
• Did you run filter bench with the paired bloom filter and various bpk-s, including small ones?
• I think you should test it using db_bench with compaction running and multiple levels (and our filter obviously) to verify it runs without any issues.
• Before testing performance, QA (@Yuval-Ariel ) should probably db-stress test it, using our filter and when there are multiple levels (as the change / improvement deals with behaving differently in the bottomest level relative to higher levels)

[udi-speedb]

@noamhaham - I am approving the code of the change itself. However, please see my comments / suggestions / questions. Thanks

[noamhaham]

@udi-speedb, @erez-speedb, @Yuval-Ariel

1. Make check was completed successfully.
2. I ran filter bench with 10 bits per key and with 20 bits per key
3. We could add this kind of test to the performance tests (read while write)
4. I agree about the db-stress
5. @erez-speedb for performance tests on HDD please use the branch: paired_bloom_filter_balanced_rounding_for_tests

[Yuval-Ariel]

@udi-speedb , hey we'd like to get this into the upcoming version. can you fix the build?

[udi-speedb]

@udi-speedb , hey we'd like to get this into the upcoming version. can you fix the build?

@Yuval-Ariel The commit has no changes in the file that seems to cacuse the issue (util/build_version.cc.in)?

[udi-speedb]

I have now cherry-picked the branch's commit on the latest origin/main and I hope that would solve it.

[udi-speedb]

@Yuval-Ariel Did the trick. Please merge.

[udi-speedb]

@Yuval-Ariel
I see that the black box test failed on the following:

***note: appending --bloom-bits (7.000000) to --filter-uri***
[2023-02-12 11:46:58.582634] Running db_stress with pid=14160: ./db_stress --acquire_snapshot_one_in=10000 --adaptive_readahead=0 --allow_concurrent_memtable_write=0 --allow_data_in_errors=True --async_io=0 --avoid_flush_during_recovery=0 --avoid_unnecessary_blocking_io=0 --backup_max_size=104857600 --backup_one_in=100000 --batch_protection_bytes_per_key=0 --block_size=4096 --bloom_bits=13 --bottommost_compression_type=zlib --bytes_per_sync=262144 --cache_index_and_filter_blocks=0 --cache_size=8388608 --cache_type=lru_cache --charge_compression_dictionary_building_buffer=0 --charge_file_metadata=0 --charge_filter_construction=0 --charge_table_reader=0 --checkpoint_one_in=1000000 --checksum_type=kXXH3 --clear_column_family_one_in=0 --compact_files_one_in=1000000 --compact_range_one_in=1000000 --compaction_pri=4 --compaction_ttl=1000 --compare_full_db_state_snapshot=1 --compression_max_dict_buffer_bytes=0 --compression_max_dict_bytes=0 --compression_parallel_threads=1 --compression_type=snappy --compression_use_zstd_dict_trainer=0 --compression_zstd_max_train_bytes=0 --continuous_verification_interval=0 --customopspercent=0 --data_block_hash_table_util_ratio=0.51 --data_block_index_type=1 --db=/tmp/rocksdb_crashtest_blackboxoogywfea --db_write_buffer_size=0 --delpercent=18 --delrangepercent=3 --destroy_db_initially=0 --detect_filter_construct_corruption=0 --disable_wal=0 --enable_compaction_filter=0 --enable_pipelined_write=0 --expected_values_dir=/tmp/rocksdb_crashtest_expected_gb2qn1ct --fail_if_options_file_error=1 --file_checksum_impl=crc32c --filter_uri= --flush_one_in=1000000 --format_version=5 --get_current_wal_file_one_in=0 --get_live_files_one_in=100000 --get_property_one_in=1000000 --get_sorted_wal_files_one_in=0 --index_block_restart_interval=8 --index_type=2 --ingest_external_file_one_in=0 --initial_auto_readahead_size=16384 --initiate_wbm_flushes=0 --iterpercent=23 --key_len_percent_dist=3,5,7,28,6,11,23,13,4 --level_compaction_dynamic_level_bytes=True --long_running_snapshots=1 --mark_for_compaction_one_file_in=10 --max_auto_readahead_size=0 --max_background_compactions=20 --max_bytes_for_level_base=10485760 --max_key=102400 --max_key_len=9 --max_manifest_file_size=1073741824 --max_write_batch_group_size_bytes=16777216 --max_write_buffer_number=3 --max_write_buffer_size_to_maintain=8388608 --memtable_prefix_bloom_size_ratio=0 --memtable_protection_bytes_per_key=1 --memtable_whole_key_filtering=0 --memtablerep=skip_list --mmap_read=0 --mock_direct_io=False --nooverwritepercent=40 --num_file_reads_for_auto_readahead=0 --num_iterations=51 --open_files=-1 --open_metadata_write_fault_one_in=0 --open_read_fault_one_in=0 --open_write_fault_one_in=16 --ops_per_thread=100000000 --optimize_filters_for_memory=1 --paranoid_file_checks=1 --partition_filters=1 --partition_pinning=3 --pause_background_one_in=1000000 --periodic_compaction_seconds=2 --prefix_size=-1 --prefixpercent=0 --prepopulate_block_cache=1 --progress_reports=0 --read_fault_one_in=0 --readahead_size=0 --readpercent=28 --recycle_log_file_num=1 --reopen=0 --ribbon_starting_level=999 --secondary_cache_fault_one_in=32 --secondary_cache_uri=compressed_secondary_cache://capacity=8388608;enable_custom_split_merge=true --seed=25961358 --set_options_one_in=10000 --snapshot_hold_ops=100000 --sst_file_manager_bytes_per_sec=104857600 --sst_file_manager_bytes_per_truncate=1048576 --stats_dump_period_sec=0 --subcompactions=3 --sync=0 --sync_fault_injection=1 --sync_wal_one_in=100000 --target_file_size_base=2097152 --target_file_size_multiplier=2 --test_batches_snapshots=0 --top_level_index_pinning=3 --unpartitioned_pinning=0 --use_direct_io_for_flush_and_compaction=1 --use_direct_reads=0 --use_dynamic_delay=0 --use_full_merge_v1=False --use_merge=0 --use_multiget=0 --user_timestamp_size=0 --value_size_mult=32 --verify_before_write=False --verify_checksum=1 --verify_checksum_one_in=1000000 --verify_db_one_in=100000 --verify_sst_unique_id_in_manifest=1 --wal_bytes_per_sync=524288 --wal_compression=zstd --write_buffer_size=134217728 --write_dbid_to_manifest=1 --writepercent=28


[2023-02-12 11:47:05.979145] WARNING: db_stress (pid=14160) ended before kill: exitcode=1

Exit Before Killing
stdout:

stderr:
WARNING: use_zstd_dict_trainer is false but zstd finalizeDictionary cannot be used because ZSTD 1.4.5+ is not linked with the binary. zstd dictionary trainer will be used.
WARNING: prefix_size is non-zero but memtablerep != prefix_hash
ASAN:DEADLYSIGNAL
=================================================================
==14160==ERROR: AddressSanitizer: FPE on unknown address 0x5[559](https://github.com/speedb-io/speedb/actions/runs/4155651486/jobs/7189154145#step:6:560)0387ce56 (pc 0x55590387ce56 bp 0x7ffd69e95220 sp 0x7ffd69e95210 T0)
    #0 0x55590387ce55 in rocksdb::FSFileState::DropRandomUnsyncedData(rocksdb::Random*) utilities/fault_injection_fs.cc:88
    #1 0x55590387d3bc in rocksdb::FaultInjectionTestFS::DropRandomUnsyncedFileData(rocksdb::Random*) utilities/fault_injection_fs.cc:880
    #2 0x5559027f7720 in rocksdb::StressTest::Open(rocksdb::SharedState*) db_stress_tool/db_stress_test_base.cc:2659
    #3 0x5559027fcabe in rocksdb::StressTest::InitDb(rocksdb::SharedState*) db_stress_tool/db_stress_test_base.cc:308
    #4 0x555902788b84 in rocksdb::RunStressTest(rocksdb::StressTest*) db_stress_tool/db_stress_driver.cc:76
    #5 0x5559026c4905 in rocksdb::db_stress_tool(int, char**) db_stress_tool/db_stress_tool.cc:366
    #6 0x7f964aeb5c86 in __libc_start_main (/lib/x86_64-linux-gnu/libc.so.6+0x21c86)
    #7 0x5559026bfca9 in _start (/__w/speedb/speedb/db_stress+0x2bfca9)

AddressSanitizer can not provide additional info.
SUMMARY: AddressSanitizer: FPE utilities/fault_injection_fs.cc:88 in rocksdb::FSFileState::DropRandomUnsyncedData(rocksdb::Random*)
==14160==ABORTING

make[1]: *** [blackbox_crash_test] Error 2
make: *** [blackbox_asan_crash_test] Error 2
crash_test.mk:65: recipe for target 'blackbox_crash_test' failed
make[1]: Leaving directory '/__w/speedb/speedb'
Makefile:1045: recipe for target 'blackbox_asan_crash_test' failed
Error: Process completed with exit code 2.

However, the failure occurred when it wasn't using the paired bloom filter IMO. Are you familiar with this failure?

[Yuval-Ariel]

no im not, i'll run it on main to see maybe its because of recent additions.