[ZOOKEEPER-710] permanent ZSESSIONMOVED error after client app reconnects to zookeeper cluster

• Type: Bug
• Status: Closed
• Priority: Blocker
• Resolution: Fixed
• Affects Version/s: 3.2.2
• Fix Version/s: 3.3.0
• Component/s: server

Environment

debian lenny; ia64; xen virtualization

• Hadoop Flags: Reviewed
• Assignee: Patrick D. Hunt
• Reporter: Lukasz Osipiuk
• Created: 18/Mar/10 11:39
• Updated: 09/Oct/23 11:22
• Resolved: 19/Mar/10 17:19

Description

Originally problem was described on Users mailing list starting with this post.
Below I restate it in more organized form.

We occasionally (few times a day) observe that our client application disconnects from Zookeeper cluster.
Application is written in C++ and we are using libzookeeper_mt library. In version 3.2.2.

The disconnects we are observing are probably related to some problems with our network infrastructure - we are observing periods with great packet loss between machines in our DC.

Sometimes after client application (i.e. zookeeper library) reconnects to zookeeper cluster we are observing that all subsequent requests return ZSESSIONMOVED error. Restarting client app helps - we always pass 0 as clientid to zookeeper_init function so old session is not reused.

On 16-03-2010 we observed few occurences of problem. Example ones:

• 22:08; client IP 10.1.112.60 (app1); sessionID 0x22767e1c9630000
• 14:21; client IP 10.1.112.61 (app2); sessionID 0x324dcc1ba580085

I attach logs of cluster and application nodes (only stuff concerining zookeeper):

• zookeeper-node1.log.2010-03-16.gz - logs of zookeepr cluster node 1 10.1.112.62
• zookeeper-node2.log.2010-03-16.gz - logs of zookeepr cluster node 2 10.1.112.63
• zookeeper-node3.log.2010-03-16.gz - logs of zookeepr cluster node 3 10.1.112.64
• app1.log.2010-03-16.gz - application logs of app1 10.1.112.60
• app2.log.2010-03-16.gz - application logs of app2 10.1.112.61

I also made some analysis of case at 22:08:

• Network glitch which resulted in problem occurred at about 22:08.
• From what I see since 17:48 node2 was the leader and it did not
  change later yesterday.
• Client was connected to node2 since 17:50
• At around 22:09 client tried to connect to every node (1,2,3).
  Connections to node1 and node3 were closed
  with exception "Exception causing close of session 0x22767e1c9630000
  due to java.io.IOException: Read error".
  Connection to node2 stood alive.
• All subsequent operations were refused with ZSESSIONMOVED error.
  Error visible both on client and on server side.