[
https://issues.apache.org/jira/browse/SPARK-6011?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Sean Owen updated SPARK-6011:
-----------------------------
Component/s: (was: Spark Core)
YARN
Target Version/s: (was: 1.3.1)
Fix Version/s: (was: 1.3.1)
So from the PR discussion, I don't believe the proposed change can proceed. I
am not sure it gets at the underlying issue here either, which is a concern
that has been raised, for example, in
https://issues.apache.org/jira/browse/SPARK-5836 . I do think it's worth
tracking a) at least documenting this, and Marcelo's suggestion that maybe the
block manager can later do more proactive cleanup.
Have you used {{spark.cleaner.ttl}} by the way? I'm not sure if it helps in
this case.
I'd like to focus the discussion in one place so would prefer to resolve this
as a duplicate and merge into SPARK-5836
> Out of disk space due to Spark not deleting shuffle files of lost executors
> ---------------------------------------------------------------------------
>
> Key: SPARK-6011
> URL: https://issues.apache.org/jira/browse/SPARK-6011
> Project: Spark
> Issue Type: Bug
> Components: YARN
> Affects Versions: 1.2.1
> Environment: Running Spark in Yarn-Client mode
> Reporter: pankaj arora
>
> If Executors gets lost abruptly spark does not delete its shuffle files till
> application ends.
> Ours is long running application which is serving requests received through
> REST APIs and if any of the executor gets lost shuffle files are not deleted
> and that leads to local disk going out of space.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
