[
https://issues.apache.org/jira/browse/SOLR-11417?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16185552#comment-16185552
]
Mano Kovacs commented on SOLR-11417:
------------------------------------
Thank you [[email protected]] for the followup.
Based on your comment, I see to things to change and either would prevent this
scenario:
bq. We can always read stale data, so we have to make sure still operate
correctly in the face of that.
The {{ZkController.getLeader()}} is currently double checks the leader, but
only in ZK. If we would also make sure if the leader confirms that he is the
leader, we could make it more robust.
bq. I guess the key question is, why are we stuck in recovering? What happens
so that we cannot be leader when the ephemeral goes away? Do we overwrite the
last active stat as not ACTIVE? In that case, it seems we should try and
prevent that when a replica simply has connection problems with the leader vs
failing a sync or something.
Yes, the last active state becomes RECOVERING, so it won't participate anymore.
We might want to restart the process if the failure is due to the leader.
I will evaluate both directions and prepare a patch with at least one.
> Crashed leader's hanging emphemral will make restarting followers stuck in
> recovering
> -------------------------------------------------------------------------------------
>
> Key: SOLR-11417
> URL: https://issues.apache.org/jira/browse/SOLR-11417
> Project: Solr
> Issue Type: Bug
> Security Level: Public(Default Security Level. Issues are Public)
> Affects Versions: 6.3
> Reporter: Mano Kovacs
> Attachments: SOLR-11417.png
>
>
> If replicas are starting up after leader crash and within the ZK session
> timeout, replicas
> * will lose leader election due to hanging ephemerals
> * will read stale data from ZK about current leader
> * will fail recovery and stuck in recovering state
> If leader is down permanently (eg. hardware failure) and all replicas are
> affected, shard will not come up (see also SOLR-7065).
> Tested on 6.3. See attached image for details.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
