[
https://issues.apache.org/jira/browse/SPARK-11120?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14960169#comment-14960169
]
Ryan Williams commented on SPARK-11120:
---------------------------------------
Without dynamic allocation, you are allowed [twice the number of executors]
failures, which seems reasonable.
With dynamic allocation, {{spark.executor.instances}} doesn't get set, and so
you are allowed {{math.max(0 * 2, 3)}} failures, no matter how many executors
your job has as its min, initial, and max settings.
> maxNumExecutorFailures defaults to 3 under dynamic allocation
> -------------------------------------------------------------
>
> Key: SPARK-11120
> URL: https://issues.apache.org/jira/browse/SPARK-11120
> Project: Spark
> Issue Type: Bug
> Components: Spark Core
> Affects Versions: 1.5.1
> Reporter: Ryan Williams
> Priority: Minor
>
> With dynamic allocation, the {{spark.executor.instances}} config is 0,
> meaning [this
> line|https://github.com/apache/spark/blob/4ace4f8a9c91beb21a0077e12b75637a4560a542/yarn/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scala#L66-L68]
> ends up with {{maxNumExecutorFailures}} equal to {{3}}, which for me has
> resulted in large dynamicAllocation jobs with hundreds of executors dying due
> to one bad node serially failing executors that are allocated on it.
> I think that using {{spark.dynamicAllocation.maxExecutors}} would make most
> sense in this case; I frequently run shells that vary between 1 and 1000
> executors, so using {{s.dA.minExecutors}} or {{s.dA.initialExecutors}} would
> still leave me with a value that is lower than makes sense.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
