[ZOOKEEPER-4541] Ephemeral znode owned by closed session visible in 1 of 3 servers

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Duplicate
• Component/s: quorum, server
• Labels: pull-request-available
• Assignee: Unassigned
• Reporter: Håkon Hallingstad
• Created: 11/May/22 13:12
• Updated: 21/Sep/24 02:21
• Resolved: 21/Sep/24 02:21

Description

We have a ZooKeeper 3.7.0 ensemble with servers 1-3. Using zkCli.sh we saw the following znode on server 1:

stat /vespa/host-status-service/hosted-vespa:zone-config-servers/lock2/_c_be5a2484-da16-45a6-9a98-4657924040e3-lock-0000024982
cZxid = 0xa240000381f
ctime = Tue May 10 17:17:27 UTC 2022
mZxid = 0xa240000381f
mtime = Tue May 10 17:17:27 UTC 2022
pZxid = 0xa240000381f
cversion = 0
dataVersion = 0
aclVersion = 0
ephemeralOwner = 0x20006d5d6a10000
dataLength = 14
numChildren = 0

This znode was absent on server 2 and 3. A delete on the node failed everywhere.

delete /vespa/host-status-service/hosted-vespa:zone-config-servers/lock2/_c_be5a2484-da16-45a6-9a98-4657924040e3-lock-0000024982
Node does not exist: /vespa/host-status-service/hosted-vespa:zone-config-servers/lock2/_c_be5a2484-da16-45a6-9a98-4657924040e3-lock-0000024982

This makes sense as a mutable operation goes to the leader, which was server 3 and where the node is absent. Restarting server 1 did not fix the issue. Stopping server 1, removing the zookeeper database and version-2 directory, and starting the server fixed the issue.

Session 0x20006d5d6a10000 was created by a ZooKeeper client and initially connected to server 2. The client later closed the session at around the same time as the then-leader server 2 was stopped:

2022-05-10 17:17:28.141 I - cfg2 Submitting global closeSession request for session 0x20006d5d6a10000 (ZooKeeperServer)
2022-05-10 17:17:28.145 I - cfg2 Session: 0x20006d5d6a10000 closed (ZooKeeper)

That both were closed/shutdown at the same time is something we do on our side. Perhaps there is a race in the handling of closing of sessions and the shutdown of the leader?

We did a dump of server 1-3, and server 1 had two sessions that did not exist in server 2 and 3, and there was one ephemeral node reported on 1 that was not reported on server 2 and 3. The ephemeral owner matched one of the two extraneous sessions.  The dump from server 1 with the extra entries:

Global Sessions(8):
...
0x20006d5d6a10000 120000ms
0x2004360ecca0000 120000ms
...
Sessions with Ephemerals (4):
0x20006d5d6a10000:
/vespa/host-status-service/hosted-vespa:zone-config-servers/lock2/_c_be5a2484-da16-45a6-9a98-4657924040e3-lock-0000024982

Issue Links

• is superceded by: ZOOKEEPER-4785 Txn loss due to race condition in Learner.syncWithLeader() during DIFF sync
• is superceded by: ZOOKEEPER-4712 Follower.shutdown() and Observer.shutdown() do not correctly shutdown the syncProcessor, which may lead to data inconsistency
• is superceded by: ZOOKEEPER-4394 Learner.syncWithLeader got NullPointerException
• links to: GitHub Pull Request #1925