[
https://issues.apache.org/jira/browse/FLINK-14480?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16957142#comment-16957142
]
Piotr Nowojski edited comment on FLINK-14480 at 10/22/19 2:53 PM:
------------------------------------------------------------------
> It looks that the expected behavior from connector is to report exception in
> close()
I don't think that we should impose such contract, as it's easy to be violated.
I talked briefly with [~sewen] and probably this check of swallowing/ignoring
the exception should be changed, to swallow the exceptions only if the task is
being canceled ({{StreamTask#isCanceled()}}).
was (Author: pnowojski):
> It looks that the expected behavior from connector is to report exception in
> close()
I don't think that we should impose such contract, as it's easy to be violated.
I talked briefly with [~sewen] and probably this check of swallowing/ignoring
the exception should be changed, to swallow the exceptions only if the task is
being canceled ({{StreamTask#isCanceled()}}.
> Sort out exception supression during snapshotting for non-running tasks.
> ------------------------------------------------------------------------
>
> Key: FLINK-14480
> URL: https://issues.apache.org/jira/browse/FLINK-14480
> Project: Flink
> Issue Type: Bug
> Components: Runtime / Checkpointing
> Affects Versions: 1.10.0
> Reporter: Arvid Heise
> Priority: Major
>
> Following-up on FLINK-14370.
> As [~becket_qin] wrote: "The cause [of the unstable test] was that when all
> the records are processed before a snapshot was taken, the records that could
> not be sent out trigger the snapshot to fail. That snapshot failure will not
> cause the job to exit. However, all the records in the KafkaProducer are
> already expired after the snapshot failure. So when the producer closes,
> there will be no more exception thrown. Thus the job finished successfully.
> It looks that the expected behavior from connector is to report exception in
> {{close()}} method as long as there was a record that could not be sent. On
> the other hand, exception thrown from
> {{CheckpointedFunction.snapshotState()}} might be ignored. Not sure if this
> is reasonable, but this expectation is not super clear from the connector
> implementation's perspective.
> In terms of immediate fix, [[email protected]] proposes to always throw
> exception as long as there has been a record sending failure. I agree that is
> the right fix per current expected behavior on the connectors."
> [~kkl0u], [~pnowojski] any other idea on how to resolve it?
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
