[
https://issues.apache.org/jira/browse/FLINK-26409?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17776480#comment-17776480
]
Tzu-Li (Gordon) Tai commented on FLINK-26409:
---------------------------------------------
This test has indeed historically been quite flaky.
If I remember correctly, the test was modified to test against multi-broker
clusters due to it being flaky in the past, but apparently that didn't fully
resolve things.
A few data points:
* I double checked if the KafkaSource is doing anything special with the Kafka
consumer in order for it to handle partition leader switches, and it isn't -
leader switching is internal implementation details of the Kafka consumer
client.
* We already have an e2e exactly-once test with artificial failures within a
Flink job.
All in all, it sounds like a good call to remove this flaky test as it's not
really adding much value for test coverage.
> Remove meaningless Kafka connector test case
> KafkaConsumerTestBase.runBrokerFailureTest
> ---------------------------------------------------------------------------------------
>
> Key: FLINK-26409
> URL: https://issues.apache.org/jira/browse/FLINK-26409
> Project: Flink
> Issue Type: Bug
> Components: Connectors / Kafka
> Reporter: Qingsheng Ren
> Priority: Major
>
> {{KafkaConsumerTestBase#runBrokerFailureTest}} is actually validating the
> functionality of Kafka broker and consumer (change partition leader when the
> ex-leader is down and let consumer switch to the new one without exception)
> instead of KafkaSource / FlinkKafkaConsumer, so we could consider removing
> this case.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
