[
https://issues.apache.org/jira/browse/HDDS-13976?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18089324#comment-18089324
]
Sarveksha Yeshavantha Raju commented on HDDS-13976:
---------------------------------------------------
The biggest issue is that write pipeline is usually gone when we get to the
stuck state, causing {{getPipelineLeaderId()}} to return null.
The change would at best unblock a narrow set of cases shortly after close,
while still carrying data-loss risk, so we are not continuing with this
implementation.
> Use last pipeline leader ID to handle quasi closed stuck containers
> -------------------------------------------------------------------
>
> Key: HDDS-13976
> URL: https://issues.apache.org/jira/browse/HDDS-13976
> Project: Apache Ozone
> Issue Type: Sub-task
> Reporter: Ivan Andika
> Assignee: Sarveksha Yeshavantha Raju
> Priority: Major
>
> This is simply an idea.
> Currently, we mainly use BCSID and originNodeId to check whether we should
> close a container. However, these two information are sometimes not enough to
> ensure that the close quasi closed containers can be safely closed, thus
> causing quasi-closed stuck containers.
> It might be possible for us to use the last leader ID and use the Raft leader
> guarantees (i.e. highest committed index) to add more cases where container
> can still be closed if the origin node ID of the replicas are equal to last
> leader ID (despite the <= 3 unique origin node ID). This is because unlike
> Raft leader needs to apply the transaction before replying to the client,
> whereas Raft follower in the majority quorum only requires to commit, but not
> necessary apply it. Additionally, we might need to ensure that the leader is
> ready (i.e. it already applied all the logs until its term's start up log
> entry).
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
