Page MenuHomePhabricator

Upgrade old indices in the CirrusSearch opensearch clusters
Open, Needs TriagePublic

Description

As part of the migration of the CirrusSearch cluster to opensearch 2.x @bking encountered old version indices that are not compatible with opensearch 2.x.
These indices must be re-created with the current 1.x version before we can safely upgrade to 2.x.
The list of index version is at https://docs.google.com/spreadsheets/d/1MYH3SScPgwPUoQv8gTTywOegTcrj-W33uA0Bw1gD3yg

Besides indices owned by search we see:

This task about coordinating efforts to migrate such indices.

  • Cirrus wiki indices can simply be reindexed (or dropped in the case of stale indices)
  • Cirrus metastore (might no longer be queried at search time once https://gerrit.wikimedia.org/r/c/mediawiki/extensions/CirrusSearch/+/1273690 is shipped and could be deleted/recreated from scratch using cirrus maint scripts)
  • System indices (.tasks/.ltr) have to handled separately
  • ttmserver indices will have to be reindexed using MW maint script (possibly requiring switching traffic, IIRC reindexing ttmserver can take quite some time)
  • toolhub: unclear about the procedure here (perhaps quick enough to reindex that it's fine to have a small downtime?)

Event Timeline

dcausse renamed this task from Upgrade old indices CirrusSearch opensearch cluster to Upgrade old indices in the CirrusSearch opensearch clusters.

toolhub: unclear about the procedure here (perhaps quick enough to reindex that it's fine to have a small downtime?)

Downtime during a rebuild is what has been discussed in the past for Toolhub index rebuilds. The dataset is relatively tiny. There was some day dreaming years ago about running an {elastic|open}search node as a Kubernetes pod with a persistent volume claim as the backend as a way to separate from the Cirrus clusters. I never heard if things advanced to a point where that would be possible. It would likely need the project to move out of the wikikube cluster as part of that change too. Toolhub borrowed infrastructure from all over to get into prod.

toolhub: unclear about the procedure here (perhaps quick enough to reindex that it's fine to have a small downtime?)

Downtime during a rebuild is what has been discussed in the past for Toolhub index rebuilds. The dataset is relatively tiny. There was some day dreaming years ago about running an {elastic|open}search node as a Kubernetes pod with a persistent volume claim as the backend as a way to separate from the Cirrus clusters. I never heard if things advanced to a point where that would be possible. It would likely need the project to move out of the wikikube cluster as part of that change too. Toolhub borrowed infrastructure from all over to get into prod.

It's quite recent but the dse-k8s cluster can now host opensearch clusters, ipoid and the recent semanticsearch experiment have been using this, some doc at https://wikitech.wikimedia.org/wiki/Data_Platform/Systems/OpenSearch-on-K8s.

Mentioned in SAL (#wikimedia-cloud) [2026-04-21T16:31:33Z] <dcausse> closing the toolhub_tools index in the cirrussearch opensearch@codfw (T423993)

@bd808 you may be recalling this Google doc which we started about 2 years ago.

It took quite awhile, but we did finally get the platform launched earlier this year. I'm happy to get together with you and sketch out a migration plan for toolhub at your convenience. Feel free to to read through the doc David linked above and ping me here or in Slack/IRC if you have questions. I'll create a ticket for the migration later this week and CC you if that works.

Mentioned in SAL (#wikimedia-cloud) [2026-04-24T08:36:55Z] <dcausse> remove the toolhub_tools index from the cirrussearch opensearch cluster in codfw (T423993)

Doni108 set Due Date to Mon, May 4, 12:00 AM.Mon, May 4, 4:03 PM
Doni108 removed subscribers: bd808, Nikerabbit, Aklapper and 2 others.
dcausse removed Due Date which was set to Mon, May 4, 12:00 AM.Mon, May 4, 4:34 PM
dcausse edited subscribers, added: bd808, Nikerabbit, Aklapper and 2 others; removed: Doni108.