[
https://issues.apache.org/jira/browse/FLINK-26916?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17514555#comment-17514555
]
David Morávek commented on FLINK-26916:
---------------------------------------
Hi, I think this whole thing boils down to how the LAST_STATE upgrade is
implemented. Few observations:
- The operator is not shutting down the cluster properly. It simply deletes the
underlying k8s resources.`FlinkUtils#deleteCluster(java.lang.String,
java.lang.String, io.fabric8.kubernetes.client.KubernetesClient, boolean)`
- The whole implementation seems more or less like a JM failover scenario.
Basically from the Flink standpoint JM disappears for no obvious reason, which
leaves job in a "SUSPENDED" state. This also implies that all the HA data must
remain untouched so Flink can restore to the previous state.
- This code path is not designed for application upgrades. Upgrade should be
always done by a new job submission.
- ClusterId shouldn't be reused. This guarantee is needed for example for
implementing reliable cluster shutdown. (see JobId semantics section of
https://cwiki.apache.org/confluence/pages/viewpage.action?pageId=195726435)
Could this be addressed by using native savepoints instead?
> The Operator ignores job related changes (jar, parallelism) during last-state
> upgrades
> --------------------------------------------------------------------------------------
>
> Key: FLINK-26916
> URL: https://issues.apache.org/jira/browse/FLINK-26916
> Project: Flink
> Issue Type: Bug
> Components: Kubernetes Operator
> Affects Versions: kubernetes-operator-0.1.0, kubernetes-operator-1.0.0
> Reporter: Matyas Orhidi
> Assignee: Yang Wang
> Priority: Blocker
> Labels: pull-request-available
> Fix For: kubernetes-operator-0.1.0
>
>
> RC: The old jobgraph is being reused when resuming
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
