[
https://issues.apache.org/jira/browse/KAFKA-12455?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17301956#comment-17301956
]
Ron Dagostino commented on KAFKA-12455:
---------------------------------------
The test is using the default value of metadata.max.age.ms=300000 (5 minutes).
When I explicitly turn it down to metadata.max.age.ms=5000 (5 seconds) the test
passes for Raft but then fails for ZK (2 unexpected group rebalances in that
case).
I increased it to 10 seconds and then the Raft configuration failed with 3
unexpected rebalances and the ZK configuration failed with 1 unexpected
rebalance.
I decreased it to a very aggressive 1 second -- and they both passed.
We have historically seen some flakiness in the ZooKeeper version of this test,
and the fact that the test suddenly failed if we set metadata.max.age.ms to 5
or 10 seconds indicates that the it is just plain luck that the test is passing
today.
Given that the current client-side code doesn't fall back to the bootstrap
brokers when it sees no brokers available, I think any test really needs to
make it *impossible* for the client to see cluster metadata with just a single
broker. Decreasing the metadata max age decreases the possibility of it
happening but doesn't make it impossible.
Another experiment was to keep metadata.max.age.ms=300000 but define
session.timeout.ms = 30000 instead of the 10000 it was setting before -- this
is longer tyan the broker roll time, and in fact this change allows both
configurations to pass.
A further experiment was to keep metadata.max.age.ms=300000 and
session.timeout.ms = 10000 but expand to 3 brokers instead of just 2. This
should fix the issue since there would never be a situation where just 1 broker
is available, and a METADATA response would always have at least 2 brokers for
the consumer to use. Both configurations pass.
> OffsetValidationTest.test_broker_rolling_bounce failing for Raft quorums
> ------------------------------------------------------------------------
>
> Key: KAFKA-12455
> URL: https://issues.apache.org/jira/browse/KAFKA-12455
> Project: Kafka
> Issue Type: Bug
> Affects Versions: 2.8.0
> Reporter: Ron Dagostino
> Assignee: Ron Dagostino
> Priority: Blocker
>
> OffsetValidationTest.test_broker_rolling_bounce in `consumer_test.py` is
> failing because the consumer group is rebalancing unexpectedly.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
