Niklas Semmler created FLINK-26502:
--------------------------------------
Summary: Multiple component leader election has different
close/stop behavior
Key: FLINK-26502
URL: https://issues.apache.org/jira/browse/FLINK-26502
Project: Flink
Issue Type: Bug
Components: Runtime / Coordination
Affects Versions: 1.15.0
Reporter: Niklas Semmler
The new multi component leader election driver behaves different when the
service is closed.
The
[ZooKeeperLeaderElectionDriver#close|https://github.com/apache/flink/blob/master/flink-runtime/src/main/java/org/apache/flink/runtime/leaderelection/ZooKeeperLeaderElectionDriver.java#L120]
method closes the Zookeeper connection. In contrast, the
[MultipleComponentLeaderElectionDriverAdapter#close|https://github.com/apache/flink/blob/master/flink-runtime/src/main/java/org/apache/flink/runtime/leaderelection/MultipleComponentLeaderElectionDriverAdapter.java#L55]
(also a LeaderElectionDriver object) only unregisters the event handler. It
relies on the
[ZooKeeperMultipleComponentLeaderElectionDriver#close|https://github.com/apache/flink/blob/master/flink-runtime/src/main/java/org/apache/flink/runtime/leaderelection/ZooKeeperMultipleComponentLeaderElectionDriver.java#L106]
method to close the Zookeeper connection. Currently this method is only called
when the HighAvailabilityServices are stopped. This difference in behavior
means that previously enabled tests are not working anymore FLINK-25235.
Together with [~dmvk], I propose that we align the behavior of the
MultipleComponentLeaderElectionDriverAdapter to the existing behavior. We can
do so by closing the Zookeeper connection once all components that are part of
the multiple component leader election have stopped.
PS: I use ZooKeeper here as an example, the same can be said about Kubernetes
as well.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
