[jira] [Commented] (SPARK-10797) RDD's coalesce should not write out the temporary key

["Zoltán Zvara (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/SPARK-10797?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14906397#comment-14906397 ] 

Zoltán Zvara commented on SPARK-10797:
--------------------------------------

I have prepared a solution for this, because I had to extract keys from the original values written out, to be able to visualize key-distributions. Basically what I do is, that I tell the {{ShuffleHandle}} setup at {{ShuffledRDD}}'s {{ShuffleDependency}} to write and read only values (basically the objects), so the {{DiskBlockObjectWriter}} will write out objects, instead of keys and values for each record.

I'm willing to raise a pull request against this, if you guys think this is a good approach.

> RDD's coalesce should not write out the temporary key
> -----------------------------------------------------
>
>                 Key: SPARK-10797
>                 URL: https://issues.apache.org/jira/browse/SPARK-10797
>             Project: Spark
>          Issue Type: Improvement
>          Components: Spark Core
>            Reporter: Zoltán Zvara
>
> It seems that {{RDD.coalesce}} will unnecessarily write out (to shuffle files) temporary keys used on the shuffle code path. Consider the following code:
> {code:title=RDD.scala|borderStyle=solid}
> if (shuffle) {
>       /** Distributes elements evenly across output partitions, starting from a random partition. */
>       val distributePartition = (index: Int, items: Iterator[T]) => {
>         var position = (new Random(index)).nextInt(numPartitions)
>         items.map { t =>
>           // Note that the hash code of the key will just be the key itself. The HashPartitioner
>           // will mod it with the number of total partitions.
>           position = position + 1
>           (position, t)
>         }
>       } : Iterator[(Int, T)]
>       // include a shuffle step so that our upstream tasks are still distributed
>       new CoalescedRDD(
>         new ShuffledRDD[Int, T, T](mapPartitionsWithIndex(distributePartition),
>         new HashPartitioner(numPartitions)),
>         numPartitions).values
>     } else {
> {code}
> {{ShuffledRDD}} will hash using {{position}}s as keys as in the {{distributePartition}} function. After the bucket has been chosen by the sorter {{ExternalSorter}} or {{BypassMergeSortShuffleWriter}}, the {{DiskBlockObjectWriter}} write out both the (temporary) key and value to the spacified partition. On the next stage, after reading we take only the values with {{PairRDDFunctions}}.
> This certainly has a performance impact, as we unnecessarily write/read {{Int}}s and transform the data.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org