[ZOOKEEPER-3466] ZK cluster converges, but does not properly handle client connections (new in 3.5.5)

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 3.5.5

Environment

Linux

• Assignee: Unassigned
• Reporter: Dr. Jan-Philip Gehrcke
• Created: 23/Jul/19 16:48
• Updated: 26/Aug/20 17:16

Description

Hey, we explore switching from ZooKeeper 3.4.14 to ZooKeeper 3.5.5 in https://github.com/dcos/dcos.

DC/OS coordinates ZooKeeper via Exhibitor. We are not changing anything w.r.t. Exhibitor for now, and are hoping that we can use ZooKeeper 3.5.5 as a drop-in replacement for 3.4.14. This seems to work fine when Exhibitor uses a so-called static ensemble where the individual ZooKeeper instances are known a priori.

When Exhibitor however discovers individual ZooKeeper instances ("dynamic" back-end) then I think we observe a regression where ZooKeeper 3.5.5 can get into the following bad state (often, but not always):

1. three ZooKeeper instances find each other, leader election takes place (expected)
2. leader election succeeds: two followers, one leader (expected)
3. all three ZK instances respond IAMOK to RUOK  (expected)
4. all three ZK instances respond to SRVR (one says "Mode: leader", the other two say "Mode: follower")  (expected)
5. all three ZK instances respond to MNTR and show plausible output (expected)
6. Unexpected: any ZooKeeper client trying to connect to any of the three nodes observes a "connection timeout", whereas notably this is not a TCP connect() timeout. The TCP connect() succeeds, but then ZK does not seem to send the expected byte sequence to the TCP connection, and the ZK client waits for it via recv() until it hits a timeout condition. Examples for two different clients:
   1. In Kazoo we specifically hit Connection time-out: socket time-out during read
      generated here: https://github.com/python-zk/kazoo/blob/88b657a0977161f3815657878ba48f82a97a3846/kazoo/protocol/connection.py#L249
   2. In zkCli we see  Client session timed out, have not heard from server in 15003ms for sessionid 0x0, closing socket connection and attempting reconnect (org.apache.zookeeper.ClientCnxn:main-SendThread(localhost:2181))
7. This state is stable, it will last forever (well, at least for multiple hours and we didn't test longer than that).
8. In our system the ZooKeeper clients are crash-looping. They retry. What I have observed is that while they retry the ZK ensemble accumulates outstanding requests, here shown from MNTR output (emphasis mine): 
   zk_packets_received 2008
   zk_packets_sent 127
   zk_num_alive_connections 18
   zk_outstanding_requests 1880
9. The leader emits log lines confirming session timeout, example:
   [myid:3] INFO [SessionTracker:ZooKeeperServer@398] - Expiring session 0x2000642b18f0020, timeout of 10000ms exceeded [myid:3] INFO [SessionTracker:QuorumZooKeeperServer@157] - Submitting global closeSession request for session 0x2000642b18f0020
10. In this state, restarting any one of the two ZK followers results in the same state (clients don't get data from ZK upon connect).
11. In this state, restarting the ZK leader, and therefore triggering a leader re-election, almost immediately results in all clients being able to connect to all ZK instances successfully.

Issue Links

• is related to: ZOOKEEPER-3828 zookeeper clients gets connection timeout when the leader node is restarted
• relates to: ZOOKEEPER-3920 Zookeeper clients timeout after leader change due to 0.0.0.0 address when in docker environment