[ https://issues.apache.org/jira/browse/SPARK-29438?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16949465#comment-16949465 ]
Jungtaek Lim commented on SPARK-29438: -------------------------------------- Could you please link the actual code block from Github repo? * current master branch https://github.com/apache/spark/blob/6390f02f9fba059ec5d089a68c8d758aca35c9cd/sql/core/src/main/scala/org/apache/spark/sql/execution/streaming/state/StateStore.scala#L267-L280 * current branch-2.4 branch https://github.com/apache/spark/blob/4f46e8f804cba6d845116cb7daf9b4c682e6a0f1/sql/core/src/main/scala/org/apache/spark/sql/execution/streaming/state/StateStore.scala#L267-L280 * current branch-2.3 branch https://github.com/apache/spark/blob/75cc3b2da9ee0b51ecf0f13169f2b634e36a60c4/sql/core/src/main/scala/org/apache/spark/sql/execution/streaming/state/StateStore.scala#L265-L278 using taskId in the path is completely new to me so would like to confirm. > Failed to get state store in stream-stream join > ----------------------------------------------- > > Key: SPARK-29438 > URL: https://issues.apache.org/jira/browse/SPARK-29438 > Project: Spark > Issue Type: Bug > Components: Structured Streaming > Affects Versions: 2.4.4 > Reporter: Genmao Yu > Priority: Critical > > Now, Spark use the `TaskPartitionId` to determine the StateStore path. > {code:java} > TaskPartitionId \ > StateStoreVersion --> StoreProviderId -> StateStore > StateStoreName / > {code} > In spark stages, the task partition id is determined by the number of tasks. > As we said the StateStore file path depends on the task partition id. So if > stream-stream join task partition id is changed against last batch, it will > get wrong StateStore data or fail with non-exist StateStore data. In some > corner cases, it happened. Following is a sample pseudocode: > {code:java} > val df3 = streamDf1.join(streamDf2) > val df5 = streamDf3.join(batchDf4) > val df = df3.union(df5) > df.writeStream...start() > {code} > A simplified DAG like this: > {code:java} > DataSourceV2Scan Scan Relation DataSourceV2Scan DataSourceV2Scan > (streamDf3) | (streamDf1) (streamDf2) > | | | | > Exchange(200) Exchange(200) Exchange(200) Exchange(200) > | | | | > Sort Sort | | > \ / \ / > \ / \ / > SortMergeJoin StreamingSymmetricHashJoin > \ / > \ / > \ / > Union > {code} > Stream-Steam join task Id will start from 200 to 399 as they are in the same > stage with `SortMergeJoin`. But when there is no new incoming data in > `streamDf3` in some batch, it will generate a empty LocalRelation, and then > the SortMergeJoin will be replaced with a BroadcastHashJoin. In this case, > Stream-Steam join task Id will start from 1 to 200. Finally, it will get > wrong StateStore path through TaskPartitionId, and failed with error reading > state store delta file. > {code:java} > LocalTableScan Scan Relation DataSourceV2Scan DataSourceV2Scan > | | | | > BroadcastExchange | Exchange(200) Exchange(200) > | | | | > | | | | > \ / \ / > \ / \ / > BroadcastHashJoin StreamingSymmetricHashJoin > \ / > \ / > \ / > Union > {code} > In my job, I closed the auto BroadcastJoin feature (set > spark.sql.autoBroadcastJoinThreshold=-1) to walk around this bug. We should > make the StateStore path determinate but not depends on TaskPartitionId. -- 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