[SPARK-38320] (flat)MapGroupsWithState can timeout groups which just received inputs in the same microbatch

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 3.2.1
• Fix Version/s: 3.3.0, 3.2.2
• Component/s: Structured Streaming
• Labels: correctness releasenotes
• Assignee: Alex Balikov
• Reporter: Alex Balikov
• Shepherd: Jungtaek Lim
• Created: 24/Feb/22 17:51
• Updated: 05/Apr/22 06:29
• Resolved: 13/Mar/22 05:12

Description

We have identified an issue where the RocksDB state store iterator will not pick up store updates made after its creation. As a result of this, the timeoutProcessorIter in

https://github.com/apache/spark/blob/master/sql/core/src/main/scala/org/apache/spark/sql/execution/streaming/FlatMapGroupsWithStateExec.scala

will not pick up state changes made during newDataProcessorIter input processing. The user observed behavior is that a group state may receive input records and also be called with timeout in the same micro batch. This contradics the public documentation for GroupState -

https://spark.apache.org/docs/latest/api/scala/org/apache/spark/sql/streaming/GroupState.html

• The timeout is reset every time the function is called on a group, that is, when the group has new data, or the group has timed out. So the user has to set the timeout duration every time the function is called, otherwise, there will not be any timeout set.

Issue Links

• links to: [Github] Pull Request #35810 (alex-balikov)