[
https://issues.apache.org/jira/browse/SPARK-17020?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15417301#comment-15417301
]
Roi Reshef commented on SPARK-17020:
------------------------------------
Nevertheless, any attempt to repartition the resulting RDD also end with having
(almost) all of its partitions stay on the same node. I made it transform into
a ShuffledRDD via PairRDDFunctions, set a HashPartitioner with 140 partitions,
and yet, I got the same data-distribution as in the screenshot I attached.
So I guess there's something very wrong with referring to a *DataFrame.rdd*
without materializing it beforehand. What and why is beyond my understanding,
currently.
> Materialization of RDD via DataFrame.rdd forces a poor re-distribution of data
> ------------------------------------------------------------------------------
>
> Key: SPARK-17020
> URL: https://issues.apache.org/jira/browse/SPARK-17020
> Project: Spark
> Issue Type: Bug
> Components: Spark Core, SQL
> Affects Versions: 1.6.1, 1.6.2, 2.0.0
> Reporter: Roi Reshef
> Attachments: dataframe_cache.PNG, rdd_cache.PNG
>
>
> Calling DataFrame's lazy val .rdd results with a new RDD with a poor
> distribution of partitions across the cluster. Moreover, any attempt to
> repartition this RDD further will fail.
> Attached are a screenshot of the original DataFrame on cache and the
> resulting RDD on cache.
--
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
