[FLINK-3595] Kafka09 consumer thread does not interrupt when stuck in record emission

• Type: Bug
• Status: Closed
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 1.0.0
• Fix Version/s: 1.0.1, 1.1.0
• Component/s: Connectors / Kafka
• Assignee: Ufuk Celebi
• Reporter: Stephan Ewen
• Created: 09/Mar/16 13:34
• Updated: 04/Apr/16 19:28
• Resolved: 04/Apr/16 19:28

Description

When canceling a job, the Kafka 0.9 Consumer Thread may be stuck in a blocking method (output emitting) and never wakes up.

The thread as a whole cannot be simply interrupted, because of a bug in Kafka that makes the consumer freeze/hang up on interrupt.

There are two possible solutions:

• allow and call interrupt when the consumer thread is emitting elements
• destroy the output network buffer pools eagerly on canceling. The Kafka thread will then throw an exception if it is stuck in emitting elements and it will terminate, which is accepted in case the status is canceled.