[SOLR-17153] CloudSolrClient should not throw "Collection not found" with an out-dated ClusterState

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Incomplete
• Fix Version/s: 9.6
• Component/s: SolrJ
• Labels: pull-request-available
• Assignee: Unassigned
• Reporter: David Smiley
• Created: 06/Feb/24 21:42
• Updated: Yesterday 17:11
• Resolved: Yesterday 02:19

Description

(Present status: reverted / not accomplished.  New issues needed to continue)

Today, CloudSolrClient will locally fail if it's asked to send a request to a collection that it thinks does not exist due to its local ClusterState view being out-of-date.  We shouldn't fail!  And most SolrCloud tests should then remove their waitForState calls that follow collection creation!  Other stale state matters are out-of-scope.

Proposal: CloudSolrClient shouldn't try and be too smart.  Always route a request to Solr (any node); don't presume its state is up-to-date.  Maybe, after a response is received, it can check if its state has been updated and if not then explicitly get a new state.  Or not if that's too complicated.

Issue Links

• causes: SOLR-17275 Major performance regression of CloudSolrClient in Solr 9.6.0 when using aliases
• is related to: SOLR-17605 Umbrella: CloudSolrClient: HTTP ClusterStateProvider
• links to: Dev list
• links to: GitHub Pull Request #2363
• links to: Github Pull Request #2363
• links to: GitHub Pull Request #2392
• links to: GitHub Pull Request #2393