[
https://issues.apache.org/jira/browse/FLINK-27804?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17542772#comment-17542772
]
Yang Wang commented on FLINK-27804:
-----------------------------------
After we suspended a FlinkDeployment with 1.15, the JobManager is still serving
for the REST API so that we could fetch the savepoint.
What I mean is it is just in the same situation with observing the savepoint
with UPGRADING state. Right?
I do not find any potential issues we could get an incorrect
savepoint/checkpoint.
> Do not observe cluster/job mid upgrade
> --------------------------------------
>
> Key: FLINK-27804
> URL: https://issues.apache.org/jira/browse/FLINK-27804
> Project: Flink
> Issue Type: Improvement
> Components: Kubernetes Operator
> Affects Versions: kubernetes-operator-1.0.0
> Reporter: Gyula Fora
> Assignee: Gyula Fora
> Priority: Critical
> Labels: pull-request-available
>
> Seems like in some weird cornercases when we observe the FINISHED job
> (stopped with savepoint) during an upgrade the recorded last snapshot is
> incorrect (still need to investigate if this is due to a Flink problem or
> what) This can lead to upgrade errors.
> This can be avoided by simply skipping the observe step when the
> reconciliation status is UPGRADING because at that point we actually know
> that the job was already shut down and state recorded correctly in the
> savepoint info.
--
This message was sent by Atlassian Jira
(v8.20.7#820007)
