[jira] [Comment Edited] (FLINK-25849) Differentiate TaskManager sessions when registering at the JobMaster

["Till Rohrmann (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/FLINK-25849?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17483088#comment-17483088 ] 

Till Rohrmann edited comment on FLINK-25849 at 1/27/22, 12:02 PM:
------------------------------------------------------------------

I am not sure whether Akka can handle this situation properly. But if this is not a problem, then the address of the TaskExecutor should be enough for the discrimination.


was (Author: till.rohrmann):
I am not sure whether Akka can handle this situation properly.

> Differentiate TaskManager sessions when registering at the JobMaster
> --------------------------------------------------------------------
>
>                 Key: FLINK-25849
>                 URL: https://issues.apache.org/jira/browse/FLINK-25849
>             Project: Flink
>          Issue Type: Sub-task
>          Components: Runtime / Coordination
>    Affects Versions: 1.15.0
>            Reporter: Till Rohrmann
>            Assignee: Till Rohrmann
>            Priority: Major
>              Labels: pull-request-available
>
> With the introduction of configurable {{ResourceID}} for {{TaskManager}} processes, it can happen that a restarted {{TaskManager}} process will be restarted with the same {{ResourceID}}. When it now tries to register at the {{JobMaster}}, the {{JobMaster}} won't recognize it as a new instance because it only compares the {{ResourceID}}. As a consequence, the {{JobMaster}} things that this is a duplicate registration and ignores it.
> It would be better if the {{TaskManager}} would send a session id with the registration that could then be used to decide whether a new instance tries to register at the {{JobMaster}} and, therefore, the old one needs to be disconnected or whether the registration attempt is a duplicate. This would speed up the cluster reconciliation because we would not have to wait for the heartbeat timeout to occur.



--
This message was sent by Atlassian Jira
(v8.20.1#820001)