[
https://issues.apache.org/jira/browse/ZOOKEEPER-275?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Flavio Paiva Junqueira updated ZOOKEEPER-275:
---------------------------------------------
Attachment: ZOOKEEPER-275.patch
Patch for the problems described. Includes a unit test for the first case. The
second case is difficult and I still don't have a good idea of how to write a
unit test for it. In particular, I haven't been able to crash and restart a
peer in a unit test because when I kill the listener of QuorumCnxManager and
try to create another instance, it complains that the port is in use. I tried
using setReuseAddress(true) before binding, but it still doesn't work.
> Bug in FastLeaderElection
> -------------------------
>
> Key: ZOOKEEPER-275
> URL: https://issues.apache.org/jira/browse/ZOOKEEPER-275
> Project: Zookeeper
> Issue Type: Bug
> Components: leaderElection
> Reporter: Flavio Paiva Junqueira
> Assignee: Flavio Paiva Junqueira
> Attachments: ZOOKEEPER-275.patch
>
>
> I found an execution in which leader election does not make progress. Here is
> the problematic scenario:
> - We have an ensemble of 3 servers, and we start only 2;
> - We let them elect a leader, and then crash the one with lowest id, say S_1
> (call the other S_2);
> - We restart the crashed server.
> Upon restarting S_1, S_2 has its logical clock more advanced, and S_1 has its
> logical clock set to 1. Once S_1 receives a notification from S_2, it notices
> that it is in the wrong round and it advances its logical clock to the same
> value as S_1. Now, the problem comes exactly in this point because in the
> current code S_1 resets its vote to its initial vote (its own id and zxid).
> Since S_2 has already notified S_1, it won't do it again, and we are stuck.
> The patch I'm submitting fixes this problem by setting the vote of S_1 to the
> one received if it satisfies the total order predicate ("received zxid" is
> higher or "received zxid is the same and received id is higher").
> Related to this problem, I noticed that by trying to avoid unnecessary
> notification duplicates, there could be scenarios in which a server fails
> before electing a leader and restarts before leader election succeeds. This
> could happen, for example, when there isn't enough servers available and one
> available crashes and restarts. I fixed this problem in the attached patch by
> allowing a server to send a new batch of notifications if there is at least
> one outgoing queue of pending notifications empty. This is ok because we
> space out consecutive batches of notifications.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
