[
https://issues.apache.org/jira/browse/SPARK-27025?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16783110#comment-16783110
]
Erik van Oosten commented on SPARK-27025:
-----------------------------------------
Thanks Sean, that is very useful.
In my use case the entire data set is too big for the driver, but I can easily
fit 1/10th of it. So even with as little as 20 partitions, 2 partitions on the
driver would be fine.
In the use case there are 2 joins, and a groupby/count so this is probably a
wide transformation. So it seems that the cache/count/toLocalIterator/unpersist
approach is applicable.
The ergonomics of this approach are way worse, so I don't agree that it is
'better' to do this in application code.
> Speed up toLocalIterator
> ------------------------
>
> Key: SPARK-27025
> URL: https://issues.apache.org/jira/browse/SPARK-27025
> Project: Spark
> Issue Type: Wish
> Components: Spark Core
> Affects Versions: 2.3.3
> Reporter: Erik van Oosten
> Priority: Major
>
> Method {{toLocalIterator}} fetches the partitions to the driver one by one.
> However, as far as I can see, any required computation for the
> yet-to-be-fetched-partitions is not kicked off until it is fetched.
> Effectively only one partition is being computed at the same time.
> Desired behavior: immediately start calculation of all partitions while
> retaining the download-a-partition at a time behavior.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
