Scheduling of GroupByKey and CombinePerKey operations

Posted by Pawel Bartoszek on
URL: http://deprecated-apache-flink-user-mailing-list-archive.369.s1.nabble.com/Scheduling-of-GroupByKey-and-CombinePerKey-operations-tp17864.html

Can I ask why some operations run only one slot? I understand that file writes should happen only one one slot but GroupByKey operation could be distributed across all slots. I am having around 20k distinct keys every minute. Is there any way to break this operator chain? 

 

I noticed that CombinePerKey operations that don't have IO related transformation are scheduled across all 32 slots.

My cluster has 32 slots across 2 task managers. Running Beam 2.2. and Flink 1.3.2

2018-01-18, 13:56:28	2018-01-18, 14:37:14	40m 45s	GroupByKey -> ParMultiDo(WriteShardedBundles) -> ParMultiDo(Anonymous) -> xxx.pipeline.output.io.file.WriteWindowToFile-SumPlaybackBitrateResult2/TextIO.Write/WriteFiles/Reshuffle/Window.Into()/Window.Assign.out -> ParMultiDo(ReifyValueTimestamp) -> ToKeyedWorkItem	149 MB	333,672	70.8 MB	19	32	00320000	RUNNING
Start Time End Time Duration Bytes received Records received Bytes sent Records sent Attempt Host Status 2018-01-18, 13:56:28 40m 45s 2.30 MB 0 2.21 MB 0 1 xxx RUNNING 2018-01-18, 13:56:28 40m 45s 2.30 MB 0 2.21 MB 0 1 xxx RUNNING 2018-01-18, 13:56:28 40m 45s 2.30 MB 0 2.21 MB 0 1 xxx RUNNING 2018-01-18, 13:56:28 40m 45s 77.5 MB 333,683 2.21 MB 20 1 xxx RUNNING 2018-01-18, 13:56:28 40m 45s 2.30 MB 0 2.21 MB 0 1 xxx RUNNING 2018-01-18, 13:56:28 40m 45s 2.30 MB 0 2.21 MB 0 1 xxx RUNNING

Thanks,
Pawel