Stig Rohde Døssing created KAFKA-4387:
-----------------------------------------
Summary: KafkaConsumer will enter an infinite loop if the polling
thread is interrupted, and either commitSync or committed is called
Key: KAFKA-4387
URL: https://issues.apache.org/jira/browse/KAFKA-4387
Project: Kafka
Issue Type: Bug
Components: clients
Affects Versions: 0.10.0.1
Reporter: Stig Rohde Døssing
Assignee: Stig Rohde Døssing
When the KafkaConsumer.commitSync method is called, the
ConsumerNetworkClient.poll(RequestFuture<?> future) method will be called with
a future that only finishes when the commit request completes, or the request
times out.
When the calling thread is interrupted, every call to the Selector underlying
the ConsumerNetworkClient will return immediately, while thread interrupt state
is not reset. The call to poll ends up looping until the request timeout, at
which point it drops back out to ConsumerCoordinator.commitOffsetsSync which
retries the request because TimeoutException is retriable. This repeats
indefinitely.
For the same reason as in https://issues.apache.org/jira/browse/KAFKA-4375, it
is good if the KafkaConsumer can handle interrupts in a reasonable way, rather
than having wakeup() be the only way to properly stop a consumer thread.
I think making ConsumerNetworkClient.maybeTriggerWakeup() throw a
WakeupException if the calling thread is interrupted makes sense, since an
interrupted thread won't be making progress in polling due to the way Selector
works, and KafkaConsumer users then don't have to handle wakeups and interrupts
separately.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
