Eric Chang created KAFKA-21489:
----------------------------------
Summary: CreatePartitions uses an inflated replication factor
while a partition of the topic is being reassigned
Key: KAFKA-21489
URL: https://issues.apache.org/jira/browse/KAFKA-21489
Project: Kafka
Issue Type: Bug
Components: controller
Affects Versions: 4.4.0, 4.3.0, 4.2.0, 4.1.0, 4.0.0
Reporter: Eric Chang
Assignee: Eric Chang
In KRaft, {{ReplicationControlManager.createPartitions}} takes the replication
factor from the length of an existing partition's {{replicas}} array. During a
reassignment, this array holds both the target replicas and the replicas being
removed. If that partition is being reassigned, new partitions get more
replicas than the topic should have.
Steps to reproduce (unit test in {{ReplicationControlManagerTest}}):
# Create a topic with one partition and replicas [0, 1].
# Start a reassignment of partition 0 to [2, 3] and do not complete it. The
replicas of partition 0 become [2, 3, 0, 1].
# Call {{createPartitions}} to add one partition without a manual assignment.
The new partition gets 4 replicas instead of 2.
# Call {{createPartitions}} with a manual assignment of 2 replicas. The request
fails with {{INVALID_REPLICA_ASSIGNMENT}}.
Background:
In ZooKeeper mode, {{ZkAdminManager}} rejected CreatePartitions with
{{REASSIGNMENT_IN_PROGRESS}} while the topic was being reassigned. KIP-195
lists this error, and KAFKA-9038 limited the check to the topic in the request.
The KRaft controller does not have this check.
Possible fixes:
* Reject CreatePartitions with {{REASSIGNMENT_IN_PROGRESS}} when any partition
of the topic is being reassigned. This matches the ZooKeeper behavior.
* Compute the replication factor without the removing replicas
({{replicas.length - removingReplicas.length}}).
Related: KAFKA-21252
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)