Oneby Wang created KAFKA-21588:
----------------------------------

             Summary: Flaky test 
TransactionsWithTieredStoreTest#testRecoveryFromEpochOverflow
                 Key: KAFKA-21588
                 URL: https://issues.apache.org/jira/browse/KAFKA-21588
             Project: Kafka
          Issue Type: Test
            Reporter: Oneby Wang


TransactionsWithTieredStoreTest#testRecoveryFromEpochOverflow intermittently 
times out in the second producer's initTransactions\(\), before the test 
reaches its epoch rollover assertion.

On trunk revision 53f279fc5882c8644fbf378fea8929852348da6c \(macOS, Temurin 
25.0.4, Gradle 9.8.2\), repetitions 7 and 9 of 10 failed with retries disabled.

{code:bash}
./gradlew :storage:test \-\-rerun \-\-no\-parallel \-\-max\-workers=3 
  \-PmaxParallelForks=1 \-PmaxTestRetries=0 \-Pkafka.cluster.test.repeat=10 
  \-\-tests 
org.apache.kafka.tiered.storage.integration.TransactionsWithTieredStoreTest.testRecoveryFromEpochOverflow
{code}

In both failures, the helper had set the producer epoch to 32765 while the 
coordinator was still in PREPARE\_ABORT with COMPLETE\_ABORT pending. The 
completion callback carried epoch 1 and threw IllegalStateException. The second 
producer kept retrying InitProducerId until the test timed out after 60 seconds.

The first abort should complete before the helper sets the boundary epoch. A 
test\-only change waits for COMPLETE\_ABORT with no pending transition, then 
sets the epoch under the same metadata lock. With the original asynchronous 
send, the tiered and non\-tiered tests each passed 10/10 locally. In a separate 
run with a 750 ms pause before appending the first COMPLETE\_ABORT, the 
original test reproduced the same failure and the changed test passed.

The \[seven\-day trunk 
report|[https://develocity.apache.org/scans/tests?search.relativeStartTime=P7D&search.rootProjectNames=kafka&search.tags=github%2Ctrunk&search.tasks=test&search.timeZoneId=UTC&tests.container=org.apache.kafka.tiered.storage.integration.TransactionsWithTieredStoreTest&tests.test=testRecoveryFromEpochOverflow%28ClusterInstance%29%5B1%5D|https://develocity.apache.org/scans/tests?search.relativeStartTime=P7D&search.rootProjectNames=kafka&search.tags=github%2Ctrunk&search.tasks=test&search.timeZoneId=UTC&tests.container=org.apache.kafka.tiered.storage.integration.TransactionsWithTieredStoreTest&tests.test=testRecoveryFromEpochOverflow%28ClusterInstance%29%5B1%5D]\]
 showed 80 passed, 0 failed and 0 flaky for this method when checked on 20 
November 2026.





--
This message was sent by Atlassian Jira
(v10.3.26#10030026)

Reply via email to