You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@ignite.apache.org by "Sergey Chugunov (Jira)" <ji...@apache.org> on 2026/05/14 12:00:00 UTC

[jira] [Commented] (IGNITE-27746) MDC. Implement parallel ping of DC2's nodes with the connection recovery.

    [ https://issues.apache.org/jira/browse/IGNITE-27746?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18080883#comment-18080883 ] 

Sergey Chugunov commented on IGNITE-27746:
------------------------------------------

I have an idea of how to improve this feature.

This patch introduces a new mechanism performing connection checks in parallel to an old one based on a sequential testing on nodes.

Parallel checks mechanism may say to the sequential one: there is an alive node in a remote datacenter, keep testing nodes. But the sequential mechanism has a limit of 3 nodes at max it is allowed to check before segmenting current node from topology.

So if this alive node is farther away than 3 nodes, then current node will segment itself anyway.

We may improve it by better integration between parallel and sequential mechanisms.

> MDC. Implement parallel ping of DC2's nodes with the connection recovery.
> -------------------------------------------------------------------------
>
>                 Key: IGNITE-27746
>                 URL: https://issues.apache.org/jira/browse/IGNITE-27746
>             Project: Ignite
>          Issue Type: Sub-task
>            Reporter: Vladimir Steshin
>            Priority: Major
>              Labels: IEP-140, ise
>             Fix For: 2.19
>
>          Time Spent: 40m
>  Remaining Estimate: 0h
>
> Consider:
>  * The Multy-DC feature is on.
>  * A corner node from DC1 can't send a message to it's next node in DC2.
>  * DC2 is unavailable.
>  * No node of DC1 can connect to any node in DC2.
> To prevent sequential nodes failure in DC1 we need to extend the connection recovery mechanics. We need to know whether DC2 is completely unavailable. If so, we switch to DC/brain split but keep nodes of DC1 online. To achive this we might ping DC2's nodes from the edge node while it does the normal connection recovery under the same connection recovery timeout. If the recovery fails and no ping to DC2 is success, we consider DC1 to work separatelly from DC2.



--
This message was sent by Atlassian Jira
(v8.20.10#820010)