Stig Rohde Døssing created KAFKA-3635:
-----------------------------------------
Summary: Broker leadership state can become inconsistent between
Zookeeper/MetadataCache and the ReplicaManager's Partition
Key: KAFKA-3635
URL: https://issues.apache.org/jira/browse/KAFKA-3635
Project: Kafka
Issue Type: Bug
Affects Versions: 0.9.0.1
Reporter: Stig Rohde Døssing
Priority: Critical
On a 3 broker cluster, broker 2 and 3 managed to get into a state where neither
was fully the leader. The state in Zookeeper and the metadata cache (accessed
via TopicMetadataRequest) states that broker 3 is the leader, but trying to
produce/consume messages from that broker throws
NotLeaderForPartitionException. Meanwhile, consuming from broker 2 is possible,
but a metadata request to it reports that broker 3 is the leader.
Since the leader lookup when handling fetch requests is via Partition and via
MetadataCache for metadata requests, I'm guessing the two can get into an
inconsistent state.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
