Improve ExtensionConfig reconcile performance

[killianmuldoon]

Leading on from #7985 we should ensure watch on ExtensionConfig correctly filters by extensions actually referenced in the ClusterClass. This might include adding an index of index of ExtensionConfigs -> ClusterClass to quickly find all ClusterClasses using a particular ExtensionConfig.

/area runtime-sdk

[fabriziopandini]

/triage accepted

[sbueringer]

Current state:

If we receive a create/update event on an ExtensionConfig we trigger a reconcile of all ClusterClass that have a DiscoverVariablesExtension configured. The only optimization in place is that if the ExtensionConfig has a namespaceSelector set we only do it for matching ClusterClasses.

Example:

• 10 ClusterClasses with each referencing 2 ExtensionConfigs (with overall 20 ExtensionConfigs)
• That means every resync period we get 20 reconciles for every ClusterClass (ignoring deduplication if multiple ClusterClass items are added to the queue at the same time)

I think this can lead to pretty high load on Runtime Extensions

[sbueringer]

Oh wait I think it's totally broken.

Looks like we don't map ExtensionConfig to ClusterClass, we map ExtensionConfig to ExtensionConfig (i.e. we have to return ClusterClass name/namespace not ExtensionConfig name)

https://github.com/kubernetes-sigs/cluster-api/blob/3e16abc3df71314bbaee770846beba45c0eb170d/internal/controllers/clusterclass/clusterclass_controller.go#L412C1-L412C1

(@killianmuldoon can you double check)

[killianmuldoon]

Yeah you're right. The current behaviour does trigger a reconcile - but it always returns an error after looking up the ClusterClass. Will open a bugfix now.

[fabriziopandini]

Does the reconcile queue take care of duplicate reconcile requests for the same cc?

[sbueringer]

Hard to predict. I think it depends on when/how exactly the resync for ExtensionConfigs come in and how quickly the ClusterClass is reconciled.

As soon as a ClusterClass reconcile is started the next event would mark the ClusterClass as dirty in the queue and lead to another reconcile. I think it's simpler to fix the issue then trying to estimate the exact impact (as it might vary based on performance etc.)

[killianmuldoon]

/assign