[
https://issues.apache.org/jira/browse/SPARK-50332?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Dongjoon Hyun updated SPARK-50332:
----------------------------------
Priority: Major (was: Blocker)
> migrate kafka consumer offset information in spark to new MSK cluster
> ---------------------------------------------------------------------
>
> Key: SPARK-50332
> URL: https://issues.apache.org/jira/browse/SPARK-50332
> Project: Spark
> Issue Type: Sub-task
> Components: Connect, Tests
> Affects Versions: 4.0.0, 3.5.2
> Reporter: Ramakrishna
> Priority: Major
> Labels: pull-request-available
>
> I have spark job that reads messages from kafka , using
>
>
>
> {{```spark.readStream(.....)```}}
>
> method. This is deployed in AWS as sparkapp and it reads data from MSK kafka.
> Recently the devOPs decided to migrate to a different MSK cluster for cost
> reduction.
> There were some go Lang services which use their kafka offsets maintained by
> kafka server itself and the all the consumer offsets were automaticallly
> migrated to new MSK cluster since it is being managed by kafka itself.
> However for spark jobs we have structured streaming which has checkpointing
> which is being managed by spark job itself. So when this spark job is being
> made to listen to new kafka URLs , they fail with missing offsets error. To
> resolve this we need to set failOnDataLoss = false , or delete checkpoint, in
> either case we have data loss and state gets lost, if the job has some custom
> state logic .
> I have 2 requirements
> # Is there a way that kafka offsets are migrated automatically for spark
> consumers too whenever there is a migration of MSK, assuming that spark job
> has structured streaming.
> 2. Is there a way that we can delete checkpoint, so that we can remove
> the offsets tracked by checkpoint, in order to reprocess with "earliest"
> offset, however without losing the Spark state .
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
