[
https://issues.apache.org/jira/browse/FLINK-3190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15069684#comment-15069684
]
Ufuk Celebi commented on FLINK-3190:
------------------------------------
I agree that a fixed number of retries can be a limitation. There is a PR [1],
which addresses the hard coded restart behaviour. I didn't have a close look at
the changes, but I think that a strategy like the one you suggest can be added
easily after the PR is merged.
[1] https://github.com/apache/flink/pull/1470
> Retry rate limits for DataStream API
> ------------------------------------
>
> Key: FLINK-3190
> URL: https://issues.apache.org/jira/browse/FLINK-3190
> Project: Flink
> Issue Type: Improvement
> Reporter: Sebastian Klemke
> Priority: Minor
>
> For a long running stream processing job, absolute numbers of retries don't
> make much sense: The job will accumulate transient errors over time and will
> die eventually when thresholds are exceeded. Rate limits are better suited in
> this scenario: A job should only die, if it fails too often in a given time
> frame. To better overcome transient errors, retry delays could be used, as
> suggested in other issues.
> Absolute numbers of retries can still make sense, if failing operators don't
> make any progress at all. We can measure progress by OperatorState changes
> and by observing output, as long as the operator in question is not a sink.
> If operator state changes and/or operator produces output, we can assume it
> makes progress.
> As an example, let's say we configured a retry rate limit of 10 retries per
> hour and a non-sink operator A. If the operator fails once every 10 minutes
> and produces output between failures, it should not lead to job termination.
> But if the operator fails 11 times in an hour or does not produce output
> between 11 consecutive failures, job should be terminated.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
