[
https://issues.apache.org/jira/browse/BEAM-848?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15637890#comment-15637890
]
Amit Sela commented on BEAM-848:
--------------------------------
I have this patched-up but I'm waiting to find the time to actually test and
see this break before the patch, and run successfully after.
Breaking this should be easy enough if the input from the {{UnboundedSource}}
is not Kryo-serializable.
> Make post-read (unbounded) shuffle use coders instead of Kryo.
> --------------------------------------------------------------
>
> Key: BEAM-848
> URL: https://issues.apache.org/jira/browse/BEAM-848
> Project: Beam
> Issue Type: Bug
> Components: runner-spark
> Reporter: Amit Sela
> Assignee: Amit Sela
>
> The SparkRunner uses {{mapWithState}} to read and manage CheckpointMarks, and
> this stateful operation will be followed by a shuffle:
> https://github.com/apache/spark/blob/master/streaming/src/main/scala/org/apache/spark/streaming/dstream/MapWithStateDStream.scala#L159
> It would be best to use coders here:
> https://github.com/apache/incubator-beam/blob/master/runners/spark/src/main/java/org/apache/beam/runners/spark/io/SparkUnboundedSource.java#L71
> in order to get an Iterator<byte[]> and decode afterwards.
> This method is preferred for two reasons:
> 1. Known coding should be faster then Kryo.
> 2. Users are already required to explicitly use coders when authoring a
> Pipeline and the runner should avoid adding complexity of "registering"
> additional serializers.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
