Justine Olshan created KAFKA-18137:
--------------------------------------
Summary: Unloading transaction state incorrectly removes loading
partitions
Key: KAFKA-18137
URL: https://issues.apache.org/jira/browse/KAFKA-18137
Project: Kafka
Issue Type: Task
Reporter: Justine Olshan
Assignee: Justine Olshan
When there is a become follower transition on a transaction coordinator state
partition, we intend to unload the state partition. However, we pass the new
epoch to the method that does the unloading. In that method, we create a
`TransactionPartitionAndLeaderEpoch` object comprising of the topic partition
and the epoch that we use as a key to remove the partition from loading.
However, we wouldn't ever expect to see this epoch in that map since we only
load on the leader. See the code snippet:
[https://github.com/apache/kafka/blob/d00f0ecf1a1a082c97564f4b807e7a342472b57a/core/src/main/scala/kafka/coordinator/transaction/TransactionStateManager.scala#L602]
We could have a partition load after the unloading occurs, and that partition
will be stuck storing stale state on the broker until it restarts. While this
may not immediately cause a correctness issue, we should try to properly clean
up state.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
