[jira] [Commented] (SPARK-35725) Support repartition expand partitions in AQE
[ https://issues.apache.org/jira/browse/SPARK-35725?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17682475#comment-17682475 ] Penglei Shi commented on SPARK-35725: - [~ulysses] Thx! > Support repartition expand partitions in AQE > > > Key: SPARK-35725 > URL: https://issues.apache.org/jira/browse/SPARK-35725 > Project: Spark > Issue Type: Sub-task > Components: SQL >Affects Versions: 3.2.0 >Reporter: XiDuo You >Assignee: XiDuo You >Priority: Major > Fix For: 3.2.0 > > > Currently, we don't support expand partition dynamically in AQE which is not > friendly for some data skew job. > Let's say if we have a simple query: > {code:java} > SELECT * FROM table DISTRIBUTE BY col > {code} > The column of `col` is skewed, then some shuffle partitions would handle too > much data than others. > If we haven't inroduced extra shuffle, we can optimize this case by expanding > partitions in AQE. -- This message was sent by Atlassian Jira (v8.20.10#820010) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-35725) Support repartition expand partitions in AQE
[ https://issues.apache.org/jira/browse/SPARK-35725?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17682473#comment-17682473 ] XiDuo You commented on SPARK-35725: --- [~Penglei Shi] Kyuubi community has a Spark extension to support similar stuff about `finalStageConfigIsolation`, see docs [https://kyuubi.readthedocs.io/en/v1.6.1-incubating/extensions/engines/spark/rules.html.] Hope it can help you. > Support repartition expand partitions in AQE > > > Key: SPARK-35725 > URL: https://issues.apache.org/jira/browse/SPARK-35725 > Project: Spark > Issue Type: Sub-task > Components: SQL >Affects Versions: 3.2.0 >Reporter: XiDuo You >Assignee: XiDuo You >Priority: Major > Fix For: 3.2.0 > > > Currently, we don't support expand partition dynamically in AQE which is not > friendly for some data skew job. > Let's say if we have a simple query: > {code:java} > SELECT * FROM table DISTRIBUTE BY col > {code} > The column of `col` is skewed, then some shuffle partitions would handle too > much data than others. > If we haven't inroduced extra shuffle, we can optimize this case by expanding > partitions in AQE. -- This message was sent by Atlassian Jira (v8.20.10#820010) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-35725) Support repartition expand partitions in AQE
[ https://issues.apache.org/jira/browse/SPARK-35725?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17682469#comment-17682469 ] Penglei Shi commented on SPARK-35725: - [~ulysses] Yes, the final shuffle coming with Rebalance needs a target size different from the existed one for optimizing skew partitions and coalescing small partitions. Do you think it's reasonable to use a different target size depending on shuffle origin in CoalesceShufflePartitions and OptimizeSkewInRebalancePartitions? > Support repartition expand partitions in AQE > > > Key: SPARK-35725 > URL: https://issues.apache.org/jira/browse/SPARK-35725 > Project: Spark > Issue Type: Sub-task > Components: SQL >Affects Versions: 3.2.0 >Reporter: XiDuo You >Assignee: XiDuo You >Priority: Major > Fix For: 3.2.0 > > > Currently, we don't support expand partition dynamically in AQE which is not > friendly for some data skew job. > Let's say if we have a simple query: > {code:java} > SELECT * FROM table DISTRIBUTE BY col > {code} > The column of `col` is skewed, then some shuffle partitions would handle too > much data than others. > If we haven't inroduced extra shuffle, we can optimize this case by expanding > partitions in AQE. -- This message was sent by Atlassian Jira (v8.20.10#820010) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-35725) Support repartition expand partitions in AQE
[ https://issues.apache.org/jira/browse/SPARK-35725?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17682460#comment-17682460 ] XiDuo You commented on SPARK-35725: --- [~Penglei Shi] , It will cause inconsistent. The feature `rebalance` acutally depends on two rule: 1. optimize skew in rebalance, 2: coalesce shuffle partition. The key is rule2 always uses the advisory size to coalesce small partitions. so I think what you need is a config that can control both rule1 and rule2 rather than rule1 only in final shuffle for writefiles. That's why I reuse the existed advisory size config. > Support repartition expand partitions in AQE > > > Key: SPARK-35725 > URL: https://issues.apache.org/jira/browse/SPARK-35725 > Project: Spark > Issue Type: Sub-task > Components: SQL >Affects Versions: 3.2.0 >Reporter: XiDuo You >Assignee: XiDuo You >Priority: Major > Fix For: 3.2.0 > > > Currently, we don't support expand partition dynamically in AQE which is not > friendly for some data skew job. > Let's say if we have a simple query: > {code:java} > SELECT * FROM table DISTRIBUTE BY col > {code} > The column of `col` is skewed, then some shuffle partitions would handle too > much data than others. > If we haven't inroduced extra shuffle, we can optimize this case by expanding > partitions in AQE. -- This message was sent by Atlassian Jira (v8.20.10#820010) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-35725) Support repartition expand partitions in AQE
[ https://issues.apache.org/jira/browse/SPARK-35725?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17682455#comment-17682455 ] Penglei Shi commented on SPARK-35725: - Hi [~ulysses] , advisory partition size for shuffle will be used for splitting skew partitions at present, which is default 64MB. Usually we prefer a samll value such as 32/64mb for bertter performance for intermediate shuffle. But in last shuffle which is come with Rebalance, we prefer a large value such as 128/256mb so that the written files are in appropriate size. So i think we need another target size to split skew partitions in Rebalance, what's your suggestion? > Support repartition expand partitions in AQE > > > Key: SPARK-35725 > URL: https://issues.apache.org/jira/browse/SPARK-35725 > Project: Spark > Issue Type: Sub-task > Components: SQL >Affects Versions: 3.2.0 >Reporter: XiDuo You >Assignee: XiDuo You >Priority: Major > Fix For: 3.2.0 > > > Currently, we don't support expand partition dynamically in AQE which is not > friendly for some data skew job. > Let's say if we have a simple query: > {code:java} > SELECT * FROM table DISTRIBUTE BY col > {code} > The column of `col` is skewed, then some shuffle partitions would handle too > much data than others. > If we haven't inroduced extra shuffle, we can optimize this case by expanding > partitions in AQE. -- This message was sent by Atlassian Jira (v8.20.10#820010) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-35725) Support repartition expand partitions in AQE
[ https://issues.apache.org/jira/browse/SPARK-35725?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17361566#comment-17361566 ] Apache Spark commented on SPARK-35725: -- User 'ulysses-you' has created a pull request for this issue: https://github.com/apache/spark/pull/32883 > Support repartition expand partitions in AQE > > > Key: SPARK-35725 > URL: https://issues.apache.org/jira/browse/SPARK-35725 > Project: Spark > Issue Type: Sub-task > Components: SQL >Affects Versions: 3.2.0 >Reporter: XiDuo You >Priority: Major > > Currently, we don't support expand partition dynamically in AQE which is not > friendly for some data skew job. > Let's say if we have a simple query: > {code:java} > SELECT * FROM table DISTRIBUTE BY col > {code} > The column of `col` is skewed, then some shuffle partitions would handle too > much data than others. > If we haven't inroduced extra shuffle, we can optimize this case by expanding > partitions in AQE. -- This message was sent by Atlassian Jira (v8.3.4#803005) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org