[
https://issues.apache.org/jira/browse/SOLR-13815?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16947933#comment-16947933
]
Yonik Seeley commented on SOLR-13815:
-------------------------------------
OK, I've verified what the race is (or rather what this test case hits).
At the start of the distributed add (DistributedZkUpdateProcessor.setupRequest)
the shard states are active/construction/construction (for
parent-shard/sub-shard1/sub-shard2) and this parent shard receiving the request
correctly thinks it is the leader. Then later on for the same update in
doDistribAdd(), the cluster state is re-retrieved and the shard states are now
inactive/active/active. getSubShardLeaders() thus returns null because it is
only looking for shards with states of CONSTRUCTION or RECOVERY.
We could retrieve the cluster state once per request, and that would probably
be a huge help. Still... I don't think zookeeper can update multiple znodes at
the same time, so we might still have a very small window where we see
something like inactive/construction/construction. I'm not sure what the
behavior of the current code would be in that case.
> Live split can lose data
> ------------------------
>
> Key: SOLR-13815
> URL: https://issues.apache.org/jira/browse/SOLR-13815
> Project: Solr
> Issue Type: Bug
> Security Level: Public(Default Security Level. Issues are Public)
> Reporter: Yonik Seeley
> Priority: Major
> Attachments: fail.191004_053129, fail.191004_093307
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> This issue is to investigate potential data loss during a "live" split (i.e.
> split happens while updates are flowing)
> This was discovered during the shared storage work which was based on a
> non-release branch_8x sometime before 8.3, hence the first steps are to try
> and reproduce on the master branch without any shared storage changes.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
