[
https://issues.apache.org/jira/browse/BOOKKEEPER-799?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14220801#comment-14220801
]
Ivan Kelly commented on BOOKKEEPER-799:
---------------------------------------
https://reviews.apache.org/r/28326/
> Distribution schedule coverage sets don't take gaps in response lists into
> account when writequorum > ackquorum
> ---------------------------------------------------------------------------------------------------------------
>
> Key: BOOKKEEPER-799
> URL: https://issues.apache.org/jira/browse/BOOKKEEPER-799
> Project: Bookkeeper
> Issue Type: Bug
> Reporter: Ivan Kelly
> Assignee: Ivan Kelly
> Priority: Blocker
> Fix For: 4.4.0, 4.2.4
>
> Attachments:
> 0001-BOOKKEEPER-799-Distribution-schedule-coverage-sets-d.patch
>
>
> The algorithm now be used to check if all quorums are being covered when
> sending a read lac or fencing message is broken when writeQuorum >= ackQuorum.
> The purpose of the algorithm is to tell us when we have heard a response from
> enough nodes, that an ack quorum could not possibly have been formed without
> at least one of the nodes that we have heard responses from.
> The current algorithm works when writeQuorum == ackQuorum, as we consider all
> quorums covered if the first |ackQuorum| nodes in the writeQuorum are
> covered. However, this doesn't work in the case that it's the middle node in
> the quorum that we have heard.
> Take the example, e:4, w:3, a:2, and we've heard from node 0, and node 2. In
> this case, it is possible for the write quorum, 1,2,3 to get an ack quorum if
> 1 and 3 response.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
