跳到主要内容

RuntimeFilter 的等待时间调整

概述

实际生产场景会遇到因为 RuntimeFilter 等待时间不合理,引起的性能问题的情况。RuntimeFilter 是一种查询优化技术,它通过运行时生成过滤条件,从而避免了对无关数据的扫描。这种优化方式能够大幅减少 I/O 操作和计算量,进而加速查询执行。下节介绍几种常见的案例帮助数据倾斜场景下的调优。

案例:RuntimeFilter 等待时间过短

参考下面 Profile 的信息:

OLAP_SCAN_OPERATOR (id=22. nereids_id=1764. table name = test_doris(test_doris)):(ExecTime: 62.870ms)
- RuntimeFilters: : RuntimeFilter: (id = 6, type = minmax, need_local_merge: true, is_broadcast: false, build_bf_cardinality: false, RuntimeFilter: (id = 7, type = in_or_bloomfilter, need_local_merge: true, is_broadcast: false, build_bf_cardinality: false,
- PushDownPredicates: []
- KeyRanges: ScanKeys:ScanKey=[null(-9223372036854775808) : 9223372036854775807]
- TabletIds: [1732763414173, 1732763414187, 1732763414201, 1732763414215]
- UseSpecificThreadToken: False
- AcquireRuntimeFilterTime: 969ns
- BlocksProduced: 1.829K (1829)
- CloseTime: 0ns
- ExecTime: 62.870ms
- InitTime: 75.703us
- KeyRangesNum: 0
- MaxScannerThreadNum: 32
- MemoryUsage:
- PeakMemoryUsage: 0.00
- NumScanners: 32
- OpenTime: 19.276ms
- ProcessConjunctTime: 30.360us
- ProjectionTime: 0ns
- RowsProduced: 7.433056M (7433056)
- RowsRead: 0
- RuntimeFilterInfo:
- ScannerWorkerWaitTime: 0ns
- TabletNum: 4
- TotalReadThroughput: 0
- WaitForDependency[OLAP_SCAN_OPERATOR_DEPENDENCY]Time: 0ns
- WaitForRuntimeFilter: 1000ms
RuntimeFilter: (id = 6, type = minmax):
- Info: [IsPushDown = false, RuntimeFilterState = NOT_READY, HasRemoteTarget = true, HasLocalTarget = false, Ignored = false]
RuntimeFilter: (id = 7, type = in_or_bloomfilter):
- Info: [IsPushDown = false, RuntimeFilterState = NOT_READY, HasRemoteTarget = true, HasLocalTarget = false, Ignored = false]

从 Profile 中可以看到:WaitForRuntimeFilter: 1000ms。这里RuntimeFilter 等待了 1000ms,但是这个 ScanOperator 并没有等到对应的 RuntimeFilter,RuntimeFilterState = NOT_READY

 RuntimeFilter: (id = 6, type = minmax):
- Info: [IsPushDown = false, RuntimeFilterState = NOT_READY, HasRemoteTarget = true, HasLocalTarget = false, Ignored = false]
RuntimeFilter: (id = 7, type = in_or_bloomfilter):
- Info: [IsPushDown = false, RuntimeFilterState = NOT_READY, HasRemoteTarget = true, HasLocalTarget = false, Ignored = false]

所以这里对应的 RuntimeFilter 的 id 6 和 7 都没有等到。通过 Profile 定位到生成 RuntimeFilter 的 Join,发现 Join 耗时

  HASH_JOIN_OPERATOR (id=26 , nereids_id=37948):
- PlanInfo
- join op: RIGHT OUTER JOIN(PARTITIONED)[]
- equal join conjunct: (id = ID)
- runtime filters: RF006[min_max] <- ID(6418/8192/1048576), RF007[in_or_bloom] <- ID(6418/8192/1048576)
- cardinality=6,418
- vec output tuple id: 27
- output tuple id: 27
- vIntermediate tuple ids: 25
- hash output slot ids: 396 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 447
- projections: USER_ID
- project output tuple id: 27
- BlocksProduced: sum 1, avg 1, max 1, min 1
- CloseTime: avg 10.111us, max 10.111us, min 10.111us
- ExecTime: avg 364.497us, max 364.497us, min 364.497us
- InitTime: avg 26.653us, max 26.653us, min 26.653us
- MemoryUsage: sum , avg , max , min
- PeakMemoryUsage: sum 0.00 , avg 0.00 , max 0.00 , min 0.00
- ProbeKeyArena: sum 0.00 , avg 0.00 , max 0.00 , min 0.00
- OpenTime: avg 45.985us, max 45.985us, min 45.985us
- ProbeRows: sum 0, avg 0, max 0, min 0
- ProjectionTime: avg 211.930us, max 211.930us, min 211.930us
- RowsProduced: sum 1, avg 1, max 1, min 1
- WaitForDependency[HASH_JOIN_OPERATOR_DEPENDENCY]Time: avg 1sec780ms, max 1sec780ms, min 1sec780ms

可以看到这个 Join 耗时大概是1sec780ms,所以 RuntimeFilter 在 1s 内并没有等到。于是调整 RuntimeFilter 的等待时间:

set runtime_filter_wait_time_ms = 3000;

调整之后,查询耗时从 5s 降低到 2s。

总结

RuntimeFilter 的等待时间需要根据场景定义,Doris 在进行一些自适应的优化改造。通过 EXPLAIN 和 PROFILE 工具观察执行瓶颈,定位对应问题,通过 SQL 的 HINT 修改 RuntimeFilter 等待时间,规避对应问题对性能的影响。