[
https://issues.apache.org/jira/browse/IGNITE-8828?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16805136#comment-16805136
]
Alexey Goncharuk commented on IGNITE-8828:
------------------------------------------
I think the change requires a bit more work.
* First, I think we need to add {{InitNewCoordinatorFuture}} hang handling
(see IGNITE-11626 - it is currently not even being diagnosed as a hang source)
* Second, the timeouts and minimal number of owners should be changed at
runtime similarly to baseline auto-adjust timeout
* I think we should also have an ability to kick pending nodes out of topology
via a command even if timeout is set to 0
* Additionally, it would be great to collect all debug information via the
command line utility to understand the cause of the exchange hang
> Detecting and stopping unresponsive nodes during Partition Map Exchange
> -----------------------------------------------------------------------
>
> Key: IGNITE-8828
> URL: https://issues.apache.org/jira/browse/IGNITE-8828
> Project: Ignite
> Issue Type: Improvement
> Components: general
> Reporter: Sergey Chugunov
> Assignee: Ilya Lantukh
> Priority: Major
> Labels: iep-25
> Fix For: 2.8
>
> Original Estimate: 264h
> Remaining Estimate: 264h
>
> During PME process coordinator (1) gathers local partition maps from all
> nodes and (2) sends calculated full partition map back to all nodes in the
> topology.
> However if one or more nodes fail to send local information on step 1 for any
> reason, PME process hangs blocking all operations. The only solution will be
> to manually identify and stop nodes which failed to send info to coordinator.
> This should be done by coordinator itself: in case it didn't receive in time
> local partition maps from any nodes, it should check that stopping these
> nodes won't lead to data loss and then stop them forcibly.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
