[
https://issues.apache.org/jira/browse/SPARK-9197?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14633734#comment-14633734
]
Sean Owen commented on SPARK-9197:
----------------------------------
Yes, that must necessarily happen, or else in general you couldn't decommission
executors at all. Is this just a question? I think there are some discussions
about whether this should be taken into account (i.e. should pick an executor
with no cached partitions if possible). I don't think it's worth spending the
complexity and I/O to save cached partitions given that they can be re-created
if needed -- and they may not ever be needed again.
> Cached RDD partitions are lost when executors are dynamically deallocated
> -------------------------------------------------------------------------
>
> Key: SPARK-9197
> URL: https://issues.apache.org/jira/browse/SPARK-9197
> Project: Spark
> Issue Type: Bug
> Components: YARN
> Affects Versions: 1.4.1
> Reporter: Ryan Williams
>
> Currently, dynamic allocation cleans up executors that have not run any tasks
> for a certain amount of time.
> However, this often leads to cached RDD partitions being lost.
> Should dynamic allocation leave executors alone that have cached partitions?
> Should this be configurable?
> Is there any interest in code that would shuffle cached partitions around in
> preparation for executor-deallocation, to avoid this? Such logic could be
> useful in general for maintaining persisted RDDs across executor churn.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
