Stephan Ewen resolved FLINK-3594.
       Resolution: Fixed
    Fix Version/s: 1.1.0

Fixed via 34110fefc6e49bcbe1ae8de090fa43a5f6ef5fec

> StreamTask may fail when checkpoint is concurrent to regular termination
> ------------------------------------------------------------------------
>                 Key: FLINK-3594
>                 URL: https://issues.apache.org/jira/browse/FLINK-3594
>             Project: Flink
>          Issue Type: Bug
>            Reporter: Chesnay Schepler
>            Assignee: Stephan Ewen
>            Priority: Critical
>              Labels: test-stability
>             Fix For: 1.1.0
> Some tests in the KafkaConsumerTestBase rely on throwing a SuccessException 
> to stop the streaming job if the test condition is fulfilled.
> The job then fails, and it is checked whether the cause was a 
> SuccessException. if so, the test is marked as a success, otherwise as a 
> failure.
> However, should this exception be thrown while a checkpoint is being 
> triggered, the exception that stop the job is not the SuccessException, but a 
> CancelTaskException.
> This should affect every test that uses the SuccessException.
> observed here: https://travis-ci.org/apache/flink/jobs/114523189
> The problem is that the exception causes the StreamTask to enter the finally 
> block inside invoke(), which sets isRunning to false. Within 
> triggerCheckpoint() isRunning is then checked for being false, and if so a 
> CancelTaskException is thrown.
> This seems like an issue of the runtime; i observed other tests failing, 
> without giving a good cause since the CancelTaskException masks it.
> I was wondering whether triggerCheckpoint() could return false instead of 
> throwing an exception, and simply assume that an exception will be thrown 
> within invoke().

This message was sent by Atlassian JIRA

Reply via email to