[jira] [Commented] (SPARK-19353) Support binary I/O in PipedRDD

["Sergei Lebedev (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/SPARK-19353?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15852498#comment-15852498 ] 

Sergei Lebedev commented on SPARK-19353:
----------------------------------------

For reference: we have a fully backward-compatible [implementation|https://github.com/criteo-forks/spark/pull/26] of binary PipedRDD in our GitHub fork.

> Support binary I/O in PipedRDD
> ------------------------------
>
>                 Key: SPARK-19353
>                 URL: https://issues.apache.org/jira/browse/SPARK-19353
>             Project: Spark
>          Issue Type: Improvement
>            Reporter: Sergei Lebedev
>            Priority: Minor
>
> The current design of RDD.pipe is very restrictive. 
> It is line-based, each element of the input RDD [gets serialized|https://github.com/apache/spark/blob/master/core/src/main/scala/org/apache/spark/rdd/PipedRDD.scala#L143] into one or more lines. Similarly for the output of the child process, one line corresponds to a single element of the output RDD. 
> It allows to customize the output format via {{printRDDElement}}, but not the input format.
> It is not designed for extensibility. The only way to get a "BinaryPipedRDD" is to copy/paste most of it and change the relevant parts.
> These limitations have been discussed on [SO|http://stackoverflow.com/questions/27986830/how-to-pipe-binary-data-in-apache-spark] and the mailing list, but alas no issue has been created.
> A possible solution to at least the first two limitations is to factor out the format into a separate object (or objects). For instance, {{InputWriter}} and {{OutputReader}}, following Hadoop streaming API. 
> {code}
> trait InputWriter[T] {
>     def write(os: OutputStream, elem: T)
> }
> trait OutputReader[T] {
>     def read(is: InputStream): T
> }
> {code}
> The default configuration would be to write and read in line-based format, but the users will also be able to selectively swap those to the appropriate implementations.



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)

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