[
https://issues.apache.org/jira/browse/PIG-5024?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15469771#comment-15469771
]
Xianda Ke commented on PIG-5024:
--------------------------------
To enable broadcast mechanism, add add a broadcast physical operator.
BroadcastConverter just broadcast the predecessor RDD and save the broadcast
variable to a map, which can be referenced by other function/closures.
Now, RDDConverter.convert() will take three parameters: predecessor RDDs,
broadcasted variables map and a PhysicalOperator
> add a physical operator to broadcast small RDDs
> -----------------------------------------------
>
> Key: PIG-5024
> URL: https://issues.apache.org/jira/browse/PIG-5024
> Project: Pig
> Issue Type: Sub-task
> Components: spark
> Reporter: Xianda Ke
> Assignee: Xianda Ke
> Fix For: spark-branch
>
>
> Currently, when optimize some kinds of JOIN, the indexed or sampling files
> are saved into HDFS. By setting the replication to a larger number, it serves
> as distributed cache.
> Spark's broadcast mechanism is suitable for this. It seems that we can add a
> physical operator to broadcast small RDDs.
> This will benefit the optimization of some specialized Joins, such as Skewed
> Join, Replicated Join and so on.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
