[
https://issues.apache.org/jira/browse/SPARK-13723?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15183272#comment-15183272
]
Thomas Graves commented on SPARK-13723:
---------------------------------------
see some discussion on https://github.com/apache/spark/pull/11528
> YARN - Change behavior of --num-executors when
> spark.dynamicAllocation.enabled true
> -----------------------------------------------------------------------------------
>
> Key: SPARK-13723
> URL: https://issues.apache.org/jira/browse/SPARK-13723
> Project: Spark
> Issue Type: Improvement
> Components: YARN
> Affects Versions: 2.0.0
> Reporter: Thomas Graves
>
> I think we should change the behavior when --num-executors is specified when
> dynamic allocation is enabled. Currently if --num-executors is specified
> dynamic allocation is disabled and it just uses a static number of executors.
> I would rather see the default behavior changed in the 2.x line. If dynamic
> allocation config is on then num-executors goes to max and initial # of
> executors. I think this would allow users to easily cap their usage and would
> still allow it to free up executors. It would also allow users doing ML start
> out with a # of executors and if they are actually caching the data the
> executors wouldn't be freed up. So you would get very similar behavior to if
> dynamic allocation was off.
> Part of the reason for this is when using a static number if generally wastes
> resources, especially with people doing adhoc things with spark-shell. It
> also has a big affect when people are doing MapReduce/ETL type work loads.
> The problem is that people are used to specifying num-executors so if we turn
> it on by default in a cluster config its just overridden.
> We should also update the spark-submit --help description for --num-executors
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
