[
https://issues.apache.org/jira/browse/FLINK-30765?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17679155#comment-17679155
]
Matthias Pohl commented on FLINK-30765:
---------------------------------------
FLINK-30759 clears things up for the
{{DispatcherRunnerLeaderElectionLifecycleManager}} and
{{DefaultDispatcherRunner}} shutdown.
Currently, the manager stops the {{LeaderElectionService}}. Closing the
{{DefaultDispatcherRunner}} happens in parallel which causes a raise condition
when {{revokeLeadership}} is called by {{LeaderElectionService.stop()}}. Both
will trigger {{DefaultDispatcherRunner.stopDispatcherLeaderProcess()}} to be
called. Merging both classes will make the shutdown process clearer.
> LeaderElectionService.stop should always call revokeLeadership
> --------------------------------------------------------------
>
> Key: FLINK-30765
> URL: https://issues.apache.org/jira/browse/FLINK-30765
> Project: Flink
> Issue Type: Sub-task
> Components: Runtime / Coordination
> Affects Versions: 1.17.0
> Reporter: Matthias Pohl
> Assignee: Matthias Pohl
> Priority: Major
>
> The current implementations of {{LeaderElectionService}} do not implement the
> {{stop()}} call consistently. Some (e.g.
> [StandaloneLeaderElectionService|https://github.com/apache/flink/blob/c6997c97c575d334679915c328792b8a3067cfb5/flink-runtime/src/main/java/org/apache/flink/runtime/leaderelection/StandaloneLeaderElectionService.java#L53]
> call revoke on the {{LeaderContender)}} whereas others don't (e.g.
> [DefaultLeaderElectionService|https://github.com/apache/flink/blob/6e1caa390882996bf2d602951b54e4bb2d9c90dc/flink-runtime/src/main/java/org/apache/flink/runtime/leaderelection/DefaultLeaderElectionService.java#L96]).
> The
> [MultipleComponentLeaderElectionService|https://github.com/apache/flink/blob/0290715a57b8d243586ab747b0cd2416c8081012/flink-runtime/src/main/java/org/apache/flink/runtime/leaderelection/DefaultMultipleComponentLeaderElectionService.java#L166]
> does call revoke on the {{LeaderContender}} instances, though.
> We should align this behavior and specify it in the LeaderElectionService
> contract before going ahead with refactoring the interfaces (FLIP-285).
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
