[
https://issues.apache.org/jira/browse/IGNITE-5385?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16063853#comment-16063853
]
Alexei Scherbakov commented on IGNITE-5385:
-------------------------------------------
Implemented procedure when a new coordinator will fetch ready assignments from
over nodes and finish stale exchanges (if any) with the same assignement.
Working on tests.
> Get rid of discovery custom message on exchange completion
> ----------------------------------------------------------
>
> Key: IGNITE-5385
> URL: https://issues.apache.org/jira/browse/IGNITE-5385
> Project: Ignite
> Issue Type: Improvement
> Components: cache
> Affects Versions: 2.0
> Reporter: Yakov Zhdanov
> Assignee: Alexei Scherbakov
> Priority: Blocker
> Labels: performance
> Fix For: 2.2
>
>
> Currently if late affinity assignment is on we send full partition map as a
> custom message to make sure all nodes get it. With greater number of nodes
> and caches this can cause significant slowdowns.
> We suggest to move sending to communication. In this case scenario with
> coordinator failure requires special handling, since in this case some nodes
> may receive full map, complete exchange and proceed with cache operations,
> while others may not received full map yet. In this case full map should be
> resend from new coordinator - it should be recalculated if none has received
> one from former coordinator or should be requested from one of the lucky
> receivers to get forwarded to other nodes.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
