[
https://issues.apache.org/jira/browse/BEAM-649?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15798856#comment-15798856
]
Jean-Baptiste Onofré commented on BEAM-649:
-------------------------------------------
Basically, in the {{TransformTranslator}}, if we add caching on the {{ParDo}}
evaluator, it works fine:
{code}
context.putDataset(transform,
new BoundedDataset<>(inRDD.mapPartitions(new DoFnFunction<>(accum,
doFn,
context.getRuntimeContext(), sideInputs,
windowingStrategy)).cache()));
{code}
However, this result with "over-caching": any RDD/DStream is cached even if
it's used only one time. Moreover, over-caching could be an issue for the Spark
optimizer.
So, we have to do something smarter: basically, indicate to the {{ParDo}}
evaluator if it has to cache or not the RDD/DStream.
One way to determine if caching is require is to traverse the pipeline. We can
hook in the {{TranslationModeDetector}} pipeline visitor to determine caching
and flag for the evaluator.
> Pipeline "actions" should use foreachRDD via ParDo.
> ---------------------------------------------------
>
> Key: BEAM-649
> URL: https://issues.apache.org/jira/browse/BEAM-649
> Project: Beam
> Issue Type: Improvement
> Components: runner-spark
> Reporter: Amit Sela
> Assignee: Jean-Baptiste Onofré
>
> Spark will execute a pipeline ONLY if it's triggered by an action (batch) /
> output operation (streaming) -
> http://spark.apache.org/docs/1.6.2/streaming-programming-guide.html#output-operations-on-dstreams.
> Currently, such actions in Beam are mostly implemented via ParDo, and
> translated by the runner as a Map transformation (via mapPartitions).
> The runner overcomes this by "forcing" actions on untranslated leaves.
> While this is OK, it would be better in some cases, e.g., Sinks, to apply the
> same ParDo translation but with foreach/foreachRDD instead of
> foreachPartition/mapPartitions.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
