이선민 created KAFKA-21522:
---------------------------

             Summary: GlobalStreamThread runs the uncaught exception handler 
and its shutdown with the interrupt flag set
                 Key: KAFKA-21522
                 URL: https://issues.apache.org/jira/browse/KAFKA-21522
             Project: Kafka
          Issue Type: Bug
          Components: streams
            Reporter: 이선민
            Assignee: 이선민


Follow-up of KAFKA-21070 (https://github.com/apache/kafka/pull/23549), as 
discussed there with [~lucasbru].

The PR for KAFKA-21070 (not merged yet) changes {{StreamThread.run()}} to clear 
the interrupt flag before the uncaught exception handler and 
{{completeShutdown}} run, and to restore it when the shutdown is complete. 
{{GlobalStreamThread}} has its own {{run()}} and does not do this.

If the global thread is interrupted while it is running, the interrupt surfaces 
as an {{InterruptException}}, for example from the global consumer's 
{{poll()}}. The constructor of {{InterruptException}} sets the flag again. 
{{GlobalStreamThread.run()}} catches the exception, calls 
{{streamsUncaughtExceptionHandler.accept(e)}} and then closes the state 
consumer in the {{finally}} block, all with the interrupt flag still set:





The user handler and the state listener run on an interrupted thread.



The work the handler triggers runs on the interrupted global thread. With a 
handler that returns {{SHUTDOWN_APPLICATION}} and a single stream thread, 
{{addStreamThread}} is called from there. In KAFKA-21070 the same call on an 
interrupted stream thread failed with the streams protocol, in 
{{ConsumerNetworkThread.start}} on {{initializationLatch.await()}}.



{{stateConsumer.close(wipeStateStore)}} runs interrupted, so closing the global 
consumer can be cut short. The state stores are flushed and closed on the 
interrupted thread as well.

This is from reading the code on trunk. I have not reproduced it for the global 
thread.

Proposed fix: handle it in {{GlobalStreamThread.run()}} the same way as the PR 
does in {{StreamThread.run()}}: record and clear the flag in the {{catch}} 
blocks before the handler runs and once more at the start of the {{finally}} 
block, and restore it when the shutdown is complete.



--
This message was sent by Atlassian Jira
(v10.3.26#10030026)

Reply via email to