[KAFKA-3984] Broker doesn't retry reconnecting to an expired Zookeeper connection

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Duplicate
• Affects Version/s: 0.9.0.1, 0.10.1.1
• Assignee: Unassigned
• Reporter: Braedon Vickers
• Created: 22/Jul/16 05:41
• Updated: 21/Nov/17 15:02
• Resolved: 14/Aug/17 15:23

Description

We've been having issues with the network connectivity of our Kafka cluster, and this seems to be triggering an issue where the brokers stop trying to reconnect to Zookeeper, leaving us with a broken cluster even when the network has recovered.

When network issues begin we see java.net.NoRouteToHostException exceptions from org.apache.zookeeper.ClientCnxn as it attempts to re-establish the connection. If the network issue resolves itself while we are only getting these errors the broker seems to reconnect fine.

However, a lot of the time we end up with a message like this:

[2016-07-22 00:21:44,181] FATAL Could not establish session with zookeeper (kafka.server.KafkaHealthcheck)
org.I0Itec.zkclient.exception.ZkException: Unable to connect to <zookeeper hosts>
	at org.I0Itec.zkclient.ZkConnection.connect(ZkConnection.java:71)
	at org.I0Itec.zkclient.ZkClient.reconnect(ZkClient.java:1279)
...
Caused by: java.net.UnknownHostException: <zookeeper host>
	at java.net.InetAddress.getAllByName(InetAddress.java:1126)
	at java.net.InetAddress.getAllByName(InetAddress.java:1192)
	at org.apache.zookeeper.client.StaticHostProvider.<init>(StaticHostProvider.java:61)
	at org.apache.zookeeper.ZooKeeper.<init>(ZooKeeper.java:445)
...

(apologies for the partial stack traces - I'm having to try and reconstruct them from a less than ideal centralised logging setup.)

If this happens, the broker stops trying to reconnect to Zookeeper, and we have to restart it.

It looks like while the org.apache.zookeeper.Zookeeper client's state isn't Expired it will keep retrying the connection, and will recover OK when the network is back. However, once it changes to Expired (not entirely sure how that happens - based on the session timeout perhaps?) zkclient closes the existing client and attempts to create a new one. If the network is still down, the client constructor throws a java.net.UnknownHostException, zkclient calls handleSessionEstablishmentError() on KafkaHealthcheck, KafkaHealthcheck.handleSessionEstablishmentError() logs a "Fatal" error and does nothing else.

It seems like some form of retry needs to happen here, or the broker is stuck with no Zookeeper connection indefinitely.KafkaHealthcheck.handleSessionEstablishmentError() used to kill the JVM, but that was removed in https://issues.apache.org/jira/browse/KAFKA-2405. Killing the JVM would be better than doing nothing, as then your init system could restart it, allowing it to recover once the network was back.

Our cluster is running 0.9.0.1, so not sure if it affects 0.10.0.0 as well. However, it seems likely, as there doesn't seem to be any code changes in kafka or zkclient that would affect this behaviour.

Issue Links

• is duplicated by: KAFKA-5473 handle ZK session expiration properly when a new session can't be established