When trying to run Zookeeper 3.5.4 cluster on Kubernetes, I found out that in some circumstances Zookeeper node stop listening on leader election port. This cause unavailability of ZK cluster.
Zookeeper deployed as StatefulSet in term of Kubernetes and has following dynamic configuration:
Bind address contains DNS name which generated by Kubernetes for each StatefulSet pod.
These DNS names will become resolvable after container start, but with some delay. That delay cause stopping of leader election port listener in QuorumCnxManager.Listener class.
Error happens in QuorumCnxManager.Listener "run" method, it tries to bind leader election port to hostname which not resolvable at this moment. Retry count is hard-coded and equals to 3(with backoff of 1 sec).
Zookeeper server log contains following errors:
2019-03-17 07:56:04,844 [myid:1] - WARN [QuorumPeer[myid=1](plain=/0.0.0.0:2181)(secure=disabled):QuorumPeer@1230] - Unexpected exception
java.net.SocketException: Unresolved address
2019-03-17 07:56:04,844 [myid:1] - WARN [QuorumPeer[myid=1](plain=/0.0.0.0:2181)(secure=disabled):QuorumPeer@1261] - PeerState set to LOOKING
2019-03-17 07:56:04,845 [myid:1] - INFO [QuorumPeer[myid=1](plain=/0.0.0.0:2181)(secure=disabled):QuorumPeer@1136] - LOOKING
2019-03-17 07:56:04,845 [myid:1] - INFO [QuorumPeer[myid=1](plain=/0.0.0.0:2181)(secure=disabled):FastLeaderElection@893] - New election. My id = 1, proposed zxid=0x0
2019-03-17 07:56:04,846 [myid:1] - INFO [WorkerReceiver[myid=1]:FastLeaderElection@687] - Notification: 2 (message format version), 1 (n.leader), 0x0 (n.zxid), 0xf (n.round), LOOKING (n.state), 1 (n.sid), 0x0 (n.peerEPoch), LOOKING (my state)0 (n.config version)
2019-03-17 07:56:04,979 [myid:1] - INFO [zookeeper-0.zookeeper:2183:QuorumCnxManager$Listener@892] - Leaving listener
2019-03-17 07:56:04,979 [myid:1] - ERROR [zookeeper-0.zookeeper:2183:QuorumCnxManager$Listener@894] - As I'm leaving the listener thread, I won't be able to participate in leader election any longer: zookeeper-0.zookeeper:2183
This error happens on most nodes on cluster start and Zookeeper is unable to form quorum. This will leave cluster in unusable state.
As I can see, error present on branches 3.4 and 3.5.
I think, this error can be fixed by configurable number of retries(instead of hard-coded value of 3).
Other way to fix this is removing of max retries at all. Currently, ZK server only stop leader election listener and continue to serve on other ports. Maybe, if leader election halts, we should abort process.