[ https://issues.apache.org/jira/browse/ZOOKEEPER-975?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12983331#action_12983331 ]
Flavio Junqueira commented on ZOOKEEPER-975: -------------------------------------------- Thanks for bringing this up, Vishal. This is not a new observation, although I can't remember if we discussed it in a jira or not. In general, I'm lukewarm about this change. It is certainly not an issue to avoid the server going into LEADING before it goes correctly into LOOKING, but I'm not entirely comfortable with manipulating the queues of notifications. Being able to have two servers concurrently thinking they are leading is a situation supported by our protocols, and such a modification would be an optimization to avoid the unnecessary LEADING step. Regarding application recovery time, we don't have a load balance scheme at this point, which could be quite useful, so bringing a new follower up does not guarantee that clients will move their sessions to the new follower. Note that this situation occurs only if there is an ensemble running and a server joins or recovers. > new peer goes in LEADING state even if ensemble is online > --------------------------------------------------------- > > Key: ZOOKEEPER-975 > URL: https://issues.apache.org/jira/browse/ZOOKEEPER-975 > Project: ZooKeeper > Issue Type: Bug > Affects Versions: 3.3.2 > Reporter: Vishal K > Fix For: 3.4.0 > > > Scenario: > 1. 2 of the 3 ZK nodes are online > 2. Third node is attempting to join > 3. Third node unnecessarily goes in "LEADING" state > 4. Then third goes back to LOOKING (no majority of followers) and finally > goes to FOLLOWING state. > While going through the logs I noticed that a peer C that is trying to > join an already formed cluster goes in LEADING state. This is because > QuorumCnxManager of A and B sends the entire history of notification > messages to C. C receives the notification messages that were > exchanged between A and B when they were forming the cluster. > In FastLeaderElection.lookForLeader(), due to the following piece of > code, C quits lookForLeader assuming that it is supposed to lead. > 740 //If have received from all nodes, then > terminate > 741 if ((self.getVotingView().size() == > recvset.size()) && > 742 > (self.getQuorumVerifier().getWeight(proposedLeader) != 0)){ > 743 self.setPeerState((proposedLeader == > self.getId()) ? > 744 ServerState.LEADING: > learningState()); > 745 leaveInstance(); > 746 return new Vote(proposedLeader, > proposedZxid); > 747 > 748 } else if (termPredicate(recvset, > This can cause: > 1. C to unnecessarily go in LEADING state and wait for tickTime * initLimit > and then restart the FLE. > 2. C waits for 200 ms (finalizeWait) and then considers whatever > notifications it has received to make a decision. C could potentially > decide to follow an old leader, fail to connect to the leader, and > then restart FLE. See code below. > 752 if (termPredicate(recvset, > 753 new Vote(proposedLeader, proposedZxid, > 754 logicalclock))) { > 755 > 756 // Verify if there is any change in the > proposed leader > 757 while((n = recvqueue.poll(finalizeWait, > 758 TimeUnit.MILLISECONDS)) != null){ > 759 if(totalOrderPredicate(n.leader, > n.zxid, > 760 proposedLeader, > proposedZxid)){ > 761 recvqueue.put(n); > 762 break; > 763 } > 764 } > In general, this does not affect correctness of FLE since C will > eventually go back to FOLLOWING state (A and B won't vote for > C). However, this delays C from joining the cluster. This can in turn > affect recovery time of an application. > Proposal: A and B should send only the latest notification (most > recent) instead of the entire history. Does this sound reasonable? -- This message is automatically generated by JIRA. - You can reply to this email to add a comment to the issue online.