[ 
https://issues.apache.org/jira/browse/FLINK-2805?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14945116#comment-14945116
 ] 

ASF GitHub Bot commented on FLINK-2805:
---------------------------------------

Github user uce commented on the pull request:

    https://github.com/apache/flink/pull/1227#issuecomment-145876037
  
    - It is very tightly coupled in in case of recovery. We won't loose this, 
but I agree to introduce an interface to hide this (with a no-op implementation 
for standalone mode and the filesystem implementation in case of recovery). I 
fully agree that it's not a good way to do it, but this is blocking #1153. Is 
that OK?
    
    - There is no check for the provided path. It's the same behaviour as the 
configured checkpoint directory `STATE_BACKEND_FS_DIR`. I think the check would 
have to be performed on the task managers. A check for "file://" does not 
suffice, because this can also be a DFS. I would open a issue for this and then 
fix it for the `STATE_BACKEND_FS_DIR` as well.
    
    - I'm still undecided about the shutdown hook. It was removed to exactly 
prevent SIGTERM from removing all recovery data, but the actor postStop will be 
called after SIGTERM as well and then it will remove it anyways. I think the 
only advantage (which doesn't justify it) is for tests.


> Make user jars available for all job managers to recover
> --------------------------------------------------------
>
>                 Key: FLINK-2805
>                 URL: https://issues.apache.org/jira/browse/FLINK-2805
>             Project: Flink
>          Issue Type: Bug
>          Components: BlobManager, JobManager
>            Reporter: Ufuk Celebi
>            Assignee: Ufuk Celebi
>
> This is a bug in https://github.com/apache/flink/pull/1153.
> In case of multiple job managers, the user jars need to be accessible by all 
> job managers (including those who arrive later).
> Since #1153 requires the file state backend to be configured, the simplest 
> solution is to make the blob server aware of the configured recovery mode and 
> put/get/delete the user jars from the file state backend as well.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to