[jira] [Commented] (FLINK-7883) Make savepoints atomic with respect to state and side effects

["Steven Zhen Wu (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/FLINK-7883?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16697931#comment-16697931 ] 

Steven Zhen Wu commented on FLINK-7883:
---------------------------------------

We would love to see this happening. it is the "graceful" shutdown that we need to reduce/minimize duplicates if we are going to enable aggressive/frequent rescale events. otherwise, we are going to see frequent and significant duplicates.

> Make savepoints atomic with respect to state and side effects
> -------------------------------------------------------------
>
>                 Key: FLINK-7883
>                 URL: https://issues.apache.org/jira/browse/FLINK-7883
>             Project: Flink
>          Issue Type: Improvement
>          Components: DataStream API, Kafka Connector, State Backends, Checkpointing
>    Affects Versions: 1.3.2, 1.4.0
>            Reporter: Antoine Philippot
>            Priority: Major
>
> For a cancel with savepoint command, the JobManager trigger the cancel call once the savepoint is finished, but during the savepoint execution, kafka source continue to poll new messages which will not be part of the savepoint and will be replayed on the next application start.
> A solution could be to stop fetching the source stream task before triggering the savepoint.
> I suggest to add an interface {{StoppableFetchingSourceFunction}} with a method {{stopFetching}} that existant SourceFunction implementations could implement.
> We can add a {{stopFetchingSource}} property in 
>  {{CheckpointOptions}} class to pass the desired behaviour from {{JobManager.handleMessage(CancelJobWithSavepoint)}} to {{SourceStreamTask.triggerCheckpoint}}



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)