[
https://issues.apache.org/jira/browse/SPARK-2818?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Sean Owen resolved SPARK-2818.
------------------------------
Resolution: Won't Fix
> Improve joinning RDDs that transformed from the same parent RDD
> ---------------------------------------------------------------
>
> Key: SPARK-2818
> URL: https://issues.apache.org/jira/browse/SPARK-2818
> Project: Spark
> Issue Type: Improvement
> Components: Spark Core
> Reporter: Lu Lu
>
> if the joinning RDDs are originating from a same cached RDD, the DAGScheduler
> will submit redundant stages to compute and cache the common parent.
> For example:
> {code}
> val edges = sc.textFile(...).cache()
> val bigSrc = edges.groupByKey().filter(...)
> val reversed = edges.map(edge => (edge._2, edge._1))
> val bigDst = reversed.groupByKey().filter(...)
> bigSrc.join(bigDst).count
> {code}
> The final count action will trigger two stages both to compute the edges RDD.
> It will result to two performance problems:
> (1) if the resources are sufficient, these two stages will be running
> concurrently and read the same HDFS file at the same time.
> (2) if the two stages run one by one, the tasks of the latter stage can read
> the cached blocks of the edges RDD immediately. But it cannot achieve
> data-locality for these tasks because that the block location information are
> unavailable when submiting the stages.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
