[
https://issues.apache.org/jira/browse/HIVE-8913?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14242813#comment-14242813
]
Chao commented on HIVE-8913:
----------------------------
different record handler threads should process different partitions of the RDD.
Without being cached, each such thread should go through {{HadoopRDD#compute}},
and have IOContext initialized.
The issue here is, I think, a thread can be assigned to a {{ShuffleMapTask}},
which will get the input RDD through broadcast variables. And, hence, that
thread won't go through {{HadoopRDD#compute}}.
> Make SparkMapJoinResolver handle runtime skew join [Spark Branch]
> -----------------------------------------------------------------
>
> Key: HIVE-8913
> URL: https://issues.apache.org/jira/browse/HIVE-8913
> Project: Hive
> Issue Type: Improvement
> Components: Spark
> Reporter: Rui Li
> Assignee: Rui Li
> Attachments: HIVE-8913.1-spark.patch, HIVE-8913.2-spark.patch
>
>
> Sub-task of HIVE-8406.
> Now we have {{SparkMapJoinResolver}} in place. But at the moment, it doesn't
> handle the map join task created by upstream SkewJoinResolver, i.e. those
> wrapped in a ConditionalTask. We have to implement this part for runtime skew
> join to work on spark. To do so, we can borrow logic from {{MapJoinResolver}}.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
