[
https://issues.apache.org/jira/browse/SOLR-13239?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18122093#comment-18122093
]
Nick Shanin commented on SOLR-13239:
------------------------------------
🤖 AI text below 🤖 (posted on behalf of Nick Shanin)
I have been looking at this ticket and have a draft patch on a branch
(solr-13239-submit) that stops ZkStateReader.getCurrentCollections() from
reporting collections whose state has not been published yet, but I am holding
it rather than opening a PR.
On review, the approach resolves lazy collection references synchronously while
the collection-set notification lock is held, so a slow or failing ZooKeeper
read could delay or break notifications for unrelated collections. That
trade-off needs a redesign, most likely resolving the state outside the lock or
recording only validated state as it arrives, and I do not want to propose the
current shape.
The branch is kept for reference. Happy to pick this back up or hand it over if
someone has a preferred direction.
> CollectionStateWatcher reports new collections before they really exist
> -----------------------------------------------------------------------
>
> Key: SOLR-13239
> URL: https://issues.apache.org/jira/browse/SOLR-13239
> Project: Solr
> Issue Type: Bug
> Components: clients - java
> Affects Versions: 7.6
> Reporter: Hendrik Haddorp
> Priority: Major
>
> A CollectionStateWatcher registered via
> org.apache.solr.common.cloud.ZkStateReader.registerCloudCollectionsListener
> gets invoked as soon as the CloudSolrClient detects a new collection. This is
> based on having a watch on the /collections znode. When the
> CollectionStateWatcher tries to read out information about the new collection
> via zkStateReader.getClusterState() there is a good chance that no
> DocCollection can be found. The reason for that is that a DocCollection is
> based on the state.json below the collection. As this znode is below the
> collection znode it does need to be created a bit later. So there is a race
> condition.
> One can run into the same problem if one tries to register a
> CollectionStateWatcher via ZkStateReader.registerCollectionStateWatcher
> straight after a new collection is found. The watcher is then being invoked
> with the DocCollection set to null as it also can not find the DocCollection
> object. Null does however indicate that the collection was being deleted.
> see also the mail thread about this:
> https://www.mail-archive.com/[email protected]&q=subject:%22Re%5C%3A+CloudSolrClient+getDocCollection%22&o=newest&f=1
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
