[FLINK-33693] Force aligned barrier logic doesn't work when the aligned checkpoint timeout is enabled

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 1.18.0, 1.17.2
• Fix Version/s: 1.19.0, 1.18.1, 1.17.3
• Component/s: Runtime / Checkpointing
• Labels: pull-request-available
• Assignee: Rui Fan
• Reporter: Rui Fan
• Created: 29/Nov/23 12:45
• Updated: 07/Dec/23 14:59
• Resolved: 07/Dec/23 14:59

Description

FLINK-21945 will convert the unaligned checkpoint barrier and timeoutable aligned checkpoint barrier to force aligned barrier. However, Force aligned barrier logic doesn't work for  timeoutable aligned checkpoint barrier.

 

Bug at org.apache.flink.streaming.runtime.io.RecordWriterOutput#broadcastEvent, we will call withUnalignedUnsupported to transmit the unaligned checkpoint barrier to FORCED_ALIGNED barrier when the shuffle type cannot use unaligned checkpoint.

However, the if has one condition: isPriorityEvent. When  aligned checkpoint timeout is enabled, flink will emit one timeoutable barrier. It isn't PriorityEvent.

 

But timeoutable barrier should call withUnalignedUnsupported here.

Issue Links

• links to: GitHub Pull Request #23850
• links to: GitHub Pull Request #23888
• links to: GitHub Pull Request #23889