Sean Quah created KAFKA-21021:
---------------------------------
Summary: test_fencing_static_consumer is flaky for consumer
protocol when duplicate static members do not start up at the same time.
Key: KAFKA-21021
URL: https://issues.apache.org/jira/browse/KAFKA-21021
Project: Kafka
Issue Type: Bug
Components: system tests
Reporter: Sean Quah
test_fencing_static_consumer creates a group with 3 static members (consumer)
and tries to join 2 additional static members (conflict_consumer) using
duplicate instance ids (num_conflict_consumers=2). Under the consumer protocol,
the original 3 members are expected to remain active and keep their
assignments, while the 2 duplicate-instance-id members are rejected with a
fatal UnreleasedInstanceIdException and terminate.
The test uses VerifiableConsumer to start the duplicate-instance-id members,
which considers them started when both nodes are simultaneously reporting a
non-Dead state. But since these consumers are destined to die within about a
second of starting, the "started" check only succeeds if both nodes happen to
be alive at the same polling instant. If one node's startup lags behind the
other's by more than ~1-second, the two nodes are never simultaneously
non-Dead, and VerifiableConsumer.start() times out after 120 seconds with
"Verifiable consumer didn't finish startup in 120 seconds," even though the
fencing itself behaved correctly.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
