Matthias J. Sax created KAFKA-21103:
---------------------------------------
Summary: Kafka Streams might loose a thread unnoticed
Key: KAFKA-21103
URL: https://issues.apache.org/jira/browse/KAFKA-21103
Project: Kafka
Issue Type: Bug
Components: streams
Reporter: Matthias J. Sax
KS allows to register an uncaught-exception-handler, which has a REPLACE_THREAD
return code.
However, it's possible that starting up the new thread fails, and such an error
is swallowed silently, letting the user think the thread was restarted even if
it was not.
This is especially bad for single-threaded clients, which would stay in state
RUNNING and not transitioning into DEAD or ERROR.
A known workaround is to inspect IQ metadata – when the thread dies, it's
correctly removed there.
In `KafkaStreams#replaceStreamThread()` we call `addStreamsThread()` but ignore
it's returned Optional. We should check if the Optional is empty or not, and
either retry, or somehow surface the error otherwise (throwing an exception
might also work, which we would need to catch elsewhere, and re-triggering the
handler?)
Depending how we want to fix this, a KIP might be needed?
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
