[
https://issues.apache.org/jira/browse/FLINK-18828?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17186566#comment-17186566
]
Till Rohrmann commented on FLINK-18828:
---------------------------------------
I believe the question here is whether one considers if a job reaches the
{{FAILED}} state a failure of Flink or not. If {{FAILED}} is a valid outcome of
a job execution, then the return code should be zero. One could argue then that
users should configure their restart strategies to always restart if they don't
want to reach a {{FAILED}} state.
If one considers the per-job deployment mode the vehicle to run a Flink job,
then I can also see that a {{FAILED}} job state can be considered a Flink
failure and, hence, one should terminate with a non-zero exit code.
Somewhat related to this question is what is causing the job to fail. If it is
a user code fault, then I would be more inclined to say that {{FAILED}} is a
valid terminal state with a zero exit code because Flink cannot do anything
about it. If on the other hand, Flink is causing the job to reach a {{FAILED}}
state (e.g. if it does not manage to acquire enough resources, the timeouts are
too tight to run successfully on the infrastructure, a fatal error occurs,
etc.), then I can see that Flink should terminate with a non-zero exit code
indicating that it failed.
> Terminate jobmanager process with zero exit code to avoid unexpected
> restarting by K8s
> --------------------------------------------------------------------------------------
>
> Key: FLINK-18828
> URL: https://issues.apache.org/jira/browse/FLINK-18828
> Project: Flink
> Issue Type: Improvement
> Components: Runtime / Coordination
> Affects Versions: 1.10.1, 1.12.0, 1.11.1
> Reporter: Yang Wang
> Priority: Major
> Fix For: 1.12.0, 1.11.2, 1.10.3
>
>
> Currently, Flink jobmanager process terminates with a non-zero exit code if
> the job reaches the {{ApplicationStatus.FAILED}}. It is not ideal in K8s
> deployment, since non-zero exit code will cause unexpected restarting. Also
> from a framework's perspective, a FAILED job does not mean that Flink has
> failed and, hence, the return code could still be 0.
> > Note:
> This is a special case for standalone K8s deployment. For
> standalone/Yarn/Mesos/native K8s, terminating with non-zero exit code is
> harmless. And a non-zero exit code could help to check the job result quickly.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
