Henry Robinson updated ZOOKEEPER-569:

    Attachment: zookeeper-569.patch

Here's a patch with tests that appears to fix the issue (test fails without 
fix, test succeeds with). All tests pass for me with this patch on my laptop. 

I have replaced one kludge with another here. QuorumPeer.electionAlg is set to 
null when electionType==0 until the election is actually run. This causes 
problems if you want to retrieve the electionAlg object via getElectionAlg() 
beforehand for tests. 

I've set it up so that makeLEStrategy always creates a new LeaderElection if 
electionType == 0, but also that createElectionAlgorithm sets electionAlg=new 
LeaderElection(this) instead of null, so that as long as startLeaderElection 
has been called, getElectionAlg() won't return null.

I've checked to see if this will cause any obvious problems for the call sites 
of getElectionAlg and couldn't find anything that expected null. It seems more 
consistent to me this way. The question I have is over why LeaderElection needs 
re-instantiating each time when FLE does not.

If this sounds confusing, it's because the code really is! The interaction of 
createElectionAlgorithm, startLeaderElection and makeLEStrategy is hard to 

> Failure of elected leader can lead to never-ending leader election
> ------------------------------------------------------------------
>                 Key: ZOOKEEPER-569
>                 URL: https://issues.apache.org/jira/browse/ZOOKEEPER-569
>             Project: Zookeeper
>          Issue Type: Bug
>            Reporter: Henry Robinson
>            Assignee: Henry Robinson
>             Fix For: 3.3.0
>         Attachments: zookeeper-569.patch, zookeeper-569.patch, 
> zookeeper-569.patch
> It is possible for basic LeaderElection to enter a situation where it never 
> terminates. 
> As an example, consider a three node cluster A, B and C.
> 1. In the first round, A votes for A, B votes for B and C votes for C
> 2. Since C > B > A, all nodes resolve to vote for C in the second round as 
> there is no first round winner
> 3. A, B vote for C, but C fails.
> 4. C is not elected because neither A nor B hear from it, and so votes for it 
> are discarded
> 5. A and B never reset their votes, despite not hearing from C, so continue 
> to vote for it ad infinitum. 
> Step 5 is the bug. If A and B reset their votes to themselves in the case 
> where the heard-from vote set is empty, leader election will continue.
> I do not know if this affects running ZK clusters, as it is possible that the 
> out-of-band failure detection protocols may cause leader election to be 
> restarted anyhow, but I've certainly seen this in tests. 
> I have a trivial patch which fixes it, but it needs a test (and tests for 
> race conditions are hard to write!)

This message is automatically generated by JIRA.
You can reply to this email to add a comment to the issue online.

Reply via email to