[
https://issues.apache.org/jira/browse/SOLR-6530?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14139170#comment-14139170
]
Alan Woodward commented on SOLR-6530:
-------------------------------------
Think I just came across another version of this:
1. A is leader, and is distributing docs to B
2. A gets a large GC pause, during which B takes over as leader
3. A wakes up again, continues to send docs from DistributedUpdateProcessor,
but now gets errors in response because B is refusing the updates (as they're
marked as FROMLEADER, and B is now the leader)
4. In DUP.doFinish(), A finds that it has a bunch of errors from B, and so it
attempts to put B into recovery, with the same result as point 4 in the issue
description
> Commits under network partition can put any node in down state by any node
> --------------------------------------------------------------------------
>
> Key: SOLR-6530
> URL: https://issues.apache.org/jira/browse/SOLR-6530
> Project: Solr
> Issue Type: Bug
> Components: SolrCloud
> Reporter: Shalin Shekhar Mangar
> Priority: Critical
> Fix For: 5.0, 6.0
>
> Attachments: SOLR-6530.patch, SOLR-6530.patch
>
>
> Commits are executed by any node in SolrCloud i.e. they're not routed via the
> leader like other updates.
> # Suppose there's 1 collection, 1 shard, 2 replicas (A and B) and A is the
> leader
> # Suppose a commit request is made to node B during a time where B cannot
> talk to A due to a partition for any reason (failing switch, heavy GC,
> whatever)
> # B fails to distribute the commit to A (times out) and asks A to recover
> # This was okay earlier because a leader just ignores recovery requests but
> with leader initiated recovery code, B puts A in the "down" state and A can
> never get out of that state.
> tl;dr; During network partitions, if enough commit/optimize requests are sent
> to the cluster, all the nodes in the cluster will eventually be marked as
> "down".
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
