[jira] [Commented] (SPARK-2121) Not fully cached when there is enough memory in ALS

["sam (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/SPARK-2121?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14247148#comment-14247148 ] 

sam commented on SPARK-2121:
----------------------------

I seem to be getting this problem too, I have a job that should be caching a reasonably small data set (using _AND_DISK just in case, but it really is a small data set).  Unfortunately my executors get killed and my job re-runs around half of the job to recover the lost data.  Now it's a very expensive job, although it's a small data set it iterates over a large broadcasted variable many many times, so I really don't want it to get recomputed (difference between a 6 hour job and a 9 hour job).

Is there currently any work around? Like just increasing some of the many many configurables?

> Not fully cached when there is enough memory in ALS
> ---------------------------------------------------
>
>                 Key: SPARK-2121
>                 URL: https://issues.apache.org/jira/browse/SPARK-2121
>             Project: Spark
>          Issue Type: Bug
>          Components: Block Manager, MLlib, Spark Core
>    Affects Versions: 1.0.0
>            Reporter: Shuo Xiang
>
> While factorizing a large matrix using the latest Alternating Least Squares (ALS) in mllib, from sparkUI it looks like that spark fail to cache all the partitions of some RDD while memory is sufficient. Please find [this post](http://apache-spark-user-list.1001560.n3.nabble.com/Not-fully-cached-when-there-is-enough-memory-tt7429.html) for screenshots. This may cause subsequent job failures while executing `userOut.Count()` or `productsOut.count`.



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

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org