[
https://issues.apache.org/jira/browse/ZOOKEEPER-159?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Flavio Paiva Junqueira updated ZOOKEEPER-159:
---------------------------------------------
Attachment: ZOOKEEPER-159.patch
This patch passes the unit tests and my manual tests and benchmarks.
> Cover two corner cases of leader election
> -----------------------------------------
>
> Key: ZOOKEEPER-159
> URL: https://issues.apache.org/jira/browse/ZOOKEEPER-159
> Project: Zookeeper
> Issue Type: Bug
> Components: leaderElection
> Reporter: Flavio Paiva Junqueira
> Fix For: 3.0.0
>
> Attachments: ZOOKEEPER-159.patch
>
>
> In the default implementation of leader election, there are two undesirable
> cases that need to be covered:
> 1- If there is a leader elected and this leader is supported by at least
> quorum of peers, then it can happen that one peer disconnects from the
> leader, and initiates a new leader election. As it is a new leader election,
> we increment the logical clock of this peer, and according to the current
> implementation, this peer won't accept any vote from a peer with a lower LE
> turn (corresponds to the value of the logical clock of the voting peer). The
> attached patch corrects this problem by allowing a peer to go back to a
> previous epoch in the case a majority votes for a leader, and the peer also
> receives a vote from the leader. This feature allows a peer to correct a
> false suspicion of the current leader;
> 2- If a peer advances to a new turn before others, then it may end up voting
> for a peer that either does not have the highest server id or the the highest
> zxid. The attached patch fixes this problem by resetting the vote a peer when
> it updates its logical clock upon receiving a notification with a higher turn
> value.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.