[jira] [Commented] (RATIS-666) Coalesced heartbeat in multiraft

["Jitendra Nath Pandey (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/RATIS-666?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16916012#comment-16916012 ] 

Jitendra Nath Pandey commented on RATIS-666:
--------------------------------------------

Thanks for bringing this up, [~timmylicheng].
Multiraft will be a very useful feature for Ozone. However, we should probably limit the number of raft groups that each node can be part of. For example, if each node can be part of say only 5 groups, it will have 5 heartbeats to deal with and total number of heartbeats in the cluster will still not bloat too much. 
   Ozone must try to evenly distribute the groups and avoid extreme situations where one node is leader to all as it will completely skew the performance behavior.

> Coalesced heartbeat in multiraft
> --------------------------------
>
>                 Key: RATIS-666
>                 URL: https://issues.apache.org/jira/browse/RATIS-666
>             Project: Ratis
>          Issue Type: Improvement
>          Components: raft-group
>            Reporter: Li Cheng
>            Priority: Major
>
> I'm using this issue to discuss the coalesced heartbeat plan in multi-raft. We are looking at incorporating multi-raft feature in ratis into Hadoop Ozone. So in ozone, every datanode would be in multiple raft groups or say pipelines with multi-raft, which brings:
>  # Is there any plan for coalesced heartbeat on single node? 
>  # Are we going to use gRPC to achieve coalesced heartbeat like what cockroach does? Shall we assume only Java APIs are required?
>  # Either we have coalesced heartbeat, every node would have chances to be selected as leader in each raft group. So to the extreme extend, one node, say node A, would be the leader to all raft groups. If we implement coalesced heartbeat, there would more easily push node A to be the bottleneck for future stumbling in performance. Any idea on how to avoid this extremity? Maybe do a candidate scrub?
>  # How do we plan to test the 'single node, multi raft groups' scenario? Furthermore, if we allow coalesced heartbeat configurable, how to determine when and whether to use it?
>  
> [~szetszwo] [~Sammi] [~xyao] [~waterlx]



--
This message was sent by Atlassian Jira
(v8.3.2#803003)