[
https://issues.apache.org/jira/browse/BEAM-10991?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17205142#comment-17205142
]
Reuven Lax commented on BEAM-10991:
-----------------------------------
pr/12980 rolls back this PR.
Just a quick browse of that PR shows some possible problems. If a timer shows
up in a bundle but is then deleted before it fires (i.e. in a processElement or
in a different timer firing), then I believe this PR will cause us to leak the
watermark hold.
Note: that the behavior prior to this PR is that the timer will fire despite
having been deleted, because it was already captured in the bundle. That
behavior is also incorrect, so I think it should be high priority to fix the
problems in that PR and resubmit it.
> Timers don't release watermark holds in dataflow on 2.24
> --------------------------------------------------------
>
> Key: BEAM-10991
> URL: https://issues.apache.org/jira/browse/BEAM-10991
> Project: Beam
> Issue Type: Bug
> Components: runner-dataflow
> Affects Versions: 2.24.0
> Reporter: Steve Niemitz
> Assignee: Kenneth Knowles
> Priority: P1
> Fix For: 2.25.0
>
> Time Spent: 50m
> Remaining Estimate: 0h
>
> We have multiple streaming pipelines (using state + timers) that, after
> upgrading to 2.24, exhibited very strange watermark behavior. The watermark
> on some stateful DoFns would advance to the end of the first window, and then
> get stuck there forever, even preventing the job from draining.
> I was able to track the problem down to
> [https://github.com/apache/beam/commit/88acc5267f759d81e9836a9db17b9e0ee521c785.|https://github.com/apache/beam/commit/88acc5267f759d81e9836a9db17b9e0ee521c785]
> After revering it, the behavior went back to normal. It looks like its
> possible in that commit that watermark holds for some timers aren't being
> cleared.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
