[
https://issues.apache.org/jira/browse/FLINK-32422?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17738107#comment-17738107
]
Matthias Pohl edited comment on FLINK-32422 at 6/28/23 1:05 PM:
----------------------------------------------------------------
I thought about it once more: The revoke leadership event should be immediately
handled by the {{LeaderElectionService}} because it means that the contender is
deregistered and wouldn't be able to handle any outstanding leadership events
anymore, anyway. In this way, we have to accept that a revoke event can be sent
to the contender twice.
It's implemented in the same way in the {{DefaultLeaderElectionService}} (see
[DefaultLeaderElectionService:233|https://github.com/apache/flink/blob/caa5f181598658403d081a0d8b733330c70ec51c/flink-runtime/src/main/java/org/apache/flink/runtime/leaderelection/DefaultLeaderElectionService.java#L233]).
I'm downgrading this issue to Minor. I will add a test case in
{{EmbeddedLeaderServiceTest}}.
was (Author: mapohl):
I thought about it once more: The revoke leadership event should be immediately
handled by the {{LeaderElectionService}} because it means that the contender is
deregistered and wouldn't be able to handle any outstanding leadership events
anymore, anyway. In this way, we have to accept that a revoke event can be sent
to the contender twice.
It's implemented in the same way in the {{DefaultLeaderElectionService}} (see
[DefaultLeaderElectionService:233|https://github.com/apache/flink/blob/caa5f181598658403d081a0d8b733330c70ec51c/flink-runtime/src/main/java/org/apache/flink/runtime/leaderelection/DefaultLeaderElectionService.java#L233]).
I'm downgrading this issue to Major. I will add a test case in
{{EmbeddedLeaderServiceTest}}.
> EmbeddedLeaderService doesn't handle the leader events properly in edge cases
> -----------------------------------------------------------------------------
>
> Key: FLINK-32422
> URL: https://issues.apache.org/jira/browse/FLINK-32422
> Project: Flink
> Issue Type: Sub-task
> Components: Runtime / Coordination
> Affects Versions: 1.18.0
> Reporter: Matthias Pohl
> Priority: Minor
>
> The leadership is granted when registering the first contender. This sets the
> leadership flag within the EmbeddedLeaderService (see
> [EmbeddedLeaderService:312ff|https://github.com/apache/flink/blob/033aca7566a0a561410b3c0e1ae8dca856cd26ce/flink-runtime/src/main/java/org/apache/flink/runtime/highavailability/nonha/embedded/EmbeddedLeaderService.java#L312]:
> the grantLeadershipCall is triggered afterwards informing the contender
> about its leadership). In the meantime, close can be called on the contender
> which deregisters the contender again calling revoke on the contender without
> having been able to gain the leadership.
> This issue was introduced by FLINK-30765.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
