[
https://issues.apache.org/jira/browse/BEAM-12229?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17333208#comment-17333208
]
Sam Whittle commented on BEAM-12229:
------------------------------------
Ugh, I see the issue. I modified how tokens were validated after testing the
change and broke it for any fused stages with more than 1 step.
The unit test is at a lower level than that misuse so it didn't catch it.
I put together PR https://github.com/apache/beam/pull/14649 but was unable to
modify StreamingDataflowWorkerTest to reproduce the issue yet due to it's
complexity. I'd like to do that before submitting.
Maybe it is best to rollback the original PR in the interim though.
The bug should only affect the cache effectiveness not corrupt data AFAICT.
Your combiner state could be different because the combiner does local
combining differently before applying if it is cached.
> WindmillStateCache has a 0% hit rate in 2.29
> --------------------------------------------
>
> Key: BEAM-12229
> URL: https://issues.apache.org/jira/browse/BEAM-12229
> Project: Beam
> Issue Type: Bug
> Components: runner-dataflow
> Affects Versions: 2.29.0
> Reporter: Steve Niemitz
> Assignee: Reuven Lax
> Priority: P1
> Fix For: 2.30.0
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> After upgrading to 2.29, I noticed that our jobs have a 0% state cache hit
> rate. I see a very high eviction rate from the cache as well (it used to be
> ~0, now its ~100,000+ evictions / second).
> We never were on 2.28, so I can't say if it worked there, but it did work on
> 2.27.
>
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
