[
https://issues.apache.org/jira/browse/FLINK-29856?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17638418#comment-17638418
]
Mason Chen commented on FLINK-29856:
------------------------------------
{quote}Without notifyCheckpointComplete, it isn't clear what the benefit of an
intermediate Savepoint is (need to re-read the FLIP).
{quote}
Okay, in my re-reading, it makes sense that Savepoint is an option to take a
snapshot of the Flink job where user has full control over the lifecycle
management. However, why stop with Savepoint supports notifyCheckpointComplete
and intermediate Savepoint does not make fully make sense to me, since the stop
with Savepoint still "commits side effects".
> Triggering savepoint does not trigger operator notifyCheckpointComplete
> -----------------------------------------------------------------------
>
> Key: FLINK-29856
> URL: https://issues.apache.org/jira/browse/FLINK-29856
> Project: Flink
> Issue Type: Improvement
> Components: Runtime / Checkpointing
> Affects Versions: 1.16.0
> Reporter: Mason Chen
> Priority: Major
>
> When I trigger a savepoint with the Flink K8s operator, I verified for two
> sources (KafkaSource and MultiClusterKafkaSource) do not invoke
> notifyCheckpointComplete. This is easily reproducible in a simple pipeline
> (e.g. KafkaSource -> print). In this case, the savepoint is complete and
> successful, which is verified by the Flink Checkpoint UI tab and the
> jobmanager logs. e.g. `
> Triggering checkpoint 3 (type=SavepointType\{name='Savepoint',
> postCheckpointAction=NONE, formatType=CANONICAL})`
>
> However, when the checkpoint occurs via the interval, I do see the sources
> checkpointing properly and expected logs in the output.
> After the ticket was initially filed, I also checked with other stateful UDFs
> and observed the same behavior.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
