[
https://issues.apache.org/jira/browse/BEAM-12654?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17476490#comment-17476490
]
Kenneth Knowles commented on BEAM-12654:
----------------------------------------
Yes, the assumption in Dataflow (before Beam) was that every shuffle was
checkpointing so that it always had a stable value upon replay. That is why so
many connectors have reshuffles in them that are useless on Flink and Spark. We
need to migrate them all to {{@RequiresStableInput}} and also implement that
correctly in Spark and Flink.
> FileIO can produce duplicates in output files
> ---------------------------------------------
>
> Key: BEAM-12654
> URL: https://issues.apache.org/jira/browse/BEAM-12654
> Project: Beam
> Issue Type: Bug
> Components: beam-model, io-java-files, runner-flink, runner-spark,
> sdk-java-core
> Reporter: Jozef Vilcek
> Priority: P3
>
> FileIO can produce duplicates in output files - depending on a runner.
> Concrete example for Spark when executing as batch:
> When using FileIO with specific number of shards, it will use default
> sharding function which is a round robin shard assignment with random seed.
> In multistage pipeline, data between stages are hold by shuffle service until
> downstream stage request it for further computations. If shuffle results
> computed with this seeded shard function are lost - e.g. shuffle service
> fails because of HW error - then Spark will attempt to recover data by
> computing them again from source data. As a result of a random seed sharding,
> this will assign different shard - and therefore key to the element.
> More details are discussed in this thread:
> https://lists.apache.org/thread.html/r5e91d1996479defbf5e896dca3cf237ee2d9b59396cb3c4edf619df1%40%3Cdev.beam.apache.org%3E
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
