Flavio Paiva Junqueira updated ZOOKEEPER-362:

    Attachment: ZOOKEEPER-362.patch

This patch fixes the problem in the description. More concretely, it does the 

1- It synchronizes QuorumCnxManager::connectOne so that there are no competing 
connections to the same server;
2- It doesn't remove an existing connection in 
QuorumCnxManager::receiveConnection when winning the challenge;
3- it eliminates the second definition of "ss" in QuorumCnxManager::Listener. 
This was a pretty silly bug (my fault of course);
4- It adds a deadline to semapahores in FLENewEpochTest so that it doesn't wait 
5- If thread 0 finishes before thread 1, then thread 1 initiates a new round 
after waiting for 1s. This is what happens in a real deployment as a follower 
gives up on its elected leader if the elected leader takes too long to 
acknowledge its leadership. As we don't run the follower/leader part of the 
code in this test, moving to the next round doesn't happen automatically.

> Issues with FLENewEpochTest
> ---------------------------
>                 Key: ZOOKEEPER-362
>                 URL: https://issues.apache.org/jira/browse/ZOOKEEPER-362
>             Project: Zookeeper
>          Issue Type: Bug
>    Affects Versions: 3.1.1
>            Reporter: Flavio Paiva Junqueira
>             Fix For: 3.2.0
>         Attachments: ZOOKEEPER-362.patch
> I have been able to identify two reasons that cause FLENewEpochTest to fail:
> 1- There is a race condition that is triggered when two peers try to 
> establish a connection to each other for leader election. Basically, if they 
> start roughly at the same time, the server with highest id will try to open 
> two connections. The two competing connections will lead to one notification 
> message to be lost. This message happens to be critical for this two process 
> scenario; 
> 2- The code to shut down a peer is not working well with the unit tests. For 
> this particular unit test, we need to be able to shut down a peer completely 
> to check the situation the test tries to reproduce. However, it seems that in 
> some runs timing causes the other peers to believe it is still alive, and end 
> up electing it. This peer, however, eventually shuts down and leader election 
> fails.

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