You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@spark.apache.org by "Apache Spark (Jira)" <ji...@apache.org> on 2022/02/14 10:54:00 UTC
[jira] [Commented] (SPARK-38124) Revive HashClusteredDistribution and apply to stream-stream join
[ https://issues.apache.org/jira/browse/SPARK-38124?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17491933#comment-17491933 ]
Apache Spark commented on SPARK-38124:
--------------------------------------
User 'HeartSaVioR' has created a pull request for this issue:
https://github.com/apache/spark/pull/35512
> Revive HashClusteredDistribution and apply to stream-stream join
> ----------------------------------------------------------------
>
> Key: SPARK-38124
> URL: https://issues.apache.org/jira/browse/SPARK-38124
> Project: Spark
> Issue Type: Bug
> Components: SQL
> Affects Versions: 3.3.0
> Reporter: Jungtaek Lim
> Assignee: Jungtaek Lim
> Priority: Blocker
> Fix For: 3.3.0
>
>
> SPARK-35703 removed HashClusteredDistribution and replaced its usages with ClusteredDistribution.
> While this works great for non stateful operators, we still need to have a separate requirement of distribution for stateful operator, because the requirement of ClusteredDistribution is too relaxed while the requirement of physical partitioning on stateful operator is quite strict.
> In most cases, stateful operators must require child distribution as HashClusteredDistribution, with below major assumptions:
> # HashClusteredDistribution creates HashPartitioning and we will never ever change it for the future.
> # We will never ever change the implementation of {{partitionIdExpression}} in HashPartitioning for the future, so that Partitioner will behave consistently across Spark versions.
> # No partitioning except HashPartitioning can satisfy HashClusteredDistribution.
>
> We should revive HashClusteredDistribution (with probably renaming specifically with stateful operator) and apply the distribution to the all stateful operators.
> SPARK-35703 only touched stream-stream join, which means stream-stream join hasn't been broken in actual releases. Let's aim the partial revert of SPARK-35703 in this ticket, and have another ticket to deal with other stateful operators, which have been broken for their introduction (2.2+).
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org