Jiangjie Qin created KAFKA-5453:
-----------------------------------
Summary: Controller may miss requests sent to the broker when zk
session timeout happens.
Key: KAFKA-5453
URL: https://issues.apache.org/jira/browse/KAFKA-5453
Project: Kafka
Issue Type: Bug
Reporter: Jiangjie Qin
The issue I encountered was the following:
1. Partition reassignment was in progress, one replica of a partition is being
reassigned from broker 1 to broker 2.
2. Controller received an ISR change notification which indicates broker 2 has
caught up.
3. Controller was sending StopReplicaRequest to broker 1.
4. Broker 1 zk session timeout occurs. Controller removed broker 1 from the
cluster and cleaned up the queue. i.e. the StopReplicaRequest was removed from
the ControllerChannelManager.
5. Broker 1 reconnected to zk and act as if it is still a follower replica of
the partition.
6. Broker 1 will always receive exception from the leader because it is not in
the replica list.
Not sure what is the correct fix here. It seems that broke 1 in this case
should ask the controller for the latest replica assignment.
There are two related bugs:
1. when a {{NotAssignedReplicaException}} is thrown from
{{Partition.updateReplicaLogReadResult()}}, the other partitions in the same
request will failed to update the fetch timestamp and offset and thus also drop
out of the ISR.
2. The {{NotAssignedReplicaException}} was not properly returned to the
replicas, instead, a UnknownServerException is returned.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
