[ 
https://issues.apache.org/jira/browse/KAFKA-12366?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17306708#comment-17306708
 ] 

John Roesler commented on KAFKA-12366:
--------------------------------------

Thanks for the report [~vcrfxia] , I have reviewed the benchmark results 
offline, and I believe this was caused by the task idling improvements from 
KIP-695.

[https://cwiki.apache.org/confluence/display/KAFKA/KIP-695%3A+Further+Improve+Kafka+Streams+Timestamp+Synchronization]

That feature was actually pulled from 2.8, so I adjusted the report to target 
3.0, just to avoid confusion.

Also, the benchmark in question shows performance for the last few release 
branches to be about 68k +/- 7k and for trunk to be around 62K +/- 5K, which is 
certainly a drop, but still in the ballpark.

In contrast, the change is a significant improvement to the semantics of that 
join. The reason it's slightly slower now is that it is joining the stream and 
table records in the correct order, which fixes a pretty bad past behavior of 
producing many missed join results (i.e., dropping stream records because 
there's no table record to join with, but only because we didn't ingest the 
table record yet). We were able to compute those results faster, but only 
because we were missing a lot of the output one would expect.

Since the perf is still pretty close, and because it seems like the slight drop 
is well worth the improved results, I'll go ahead and close this. Thanks for 
raising it for review!

> Performance regression in stream-table joins on trunk
> -----------------------------------------------------
>
>                 Key: KAFKA-12366
>                 URL: https://issues.apache.org/jira/browse/KAFKA-12366
>             Project: Kafka
>          Issue Type: Bug
>          Components: streams
>            Reporter: Victoria Xia
>            Assignee: John Roesler
>            Priority: Blocker
>             Fix For: 3.0.0
>
>
> Stream-table join benchmarks have revealed a significant performance 
> regression on trunk as compared to the latest release version. We should 
> investigate as a blocker prior to the 2.8 release.



--
This message was sent by Atlassian Jira
(v8.3.4#803005)

Reply via email to