`mcmanager` does not enqueue requests for reconcilers when provider is `nil`

[tobschli]

What happened:

When trying to use a multi-cluster manager with the provider being nil, no requests are enqueued for the local manager.

How to reproduce it (as minimally and precisely as possible):

I modified the apiexport provider snippet to set the provider to nil and let it run against a kind Cluster ¹

In this cluster, there were ConfigMaps, like kube-root-ca.crt in default namespace.

When running the controller against the kind Cluster, it successfully starts, but no requests get enqueued:

❯ KUBECONFIG=path/to/kind/kubeconfig go run cmd/test/main.go 

2025-07-11T15:19:20+02:00       INFO    entrypoint      Setting up manager
2025-07-11T15:19:20+02:00       INFO    entrypoint      Starting manager
2025-07-11T15:19:20+02:00       INFO    controller-runtime.metrics      Starting metrics server
2025-07-11T15:19:20+02:00       INFO    Starting Controller     {"controller": "kcp-configmap-controller", "controllerGroup": "", "controllerKind": "ConfigMap"}
2025-07-11T15:19:20+02:00       INFO    Starting workers        {"controller": "kcp-configmap-controller", "controllerGroup": "", "controllerKind": "ConfigMap", "worker count": 1}
2025-07-11T15:19:20+02:00       INFO    controller-runtime.metrics      Serving metrics server  {"bindAddress": ":8080", "secure": false}

Environment:

• sigs.k8s.io/controller-runtime:v0.20.4
• sigs.k8s.io/multicluster-runtime:v0.20.4-alpha.7
• Kubernetes version: v1.32.2

Footnotes

1. Gist ↩

[xrstf]

I am not sure I understand the problem. You purposefully gave an empty value for the provider and now you are surprised that the runtime is not engaging with any workspace/cluster and therefore never calls your reconcile function? At first glance, of course this won't work. Without the provider the runtime has no idea about the clusters it needs to engage with. It's a miracle the program doesn't crash at some point ^^

If you want to connect to a single kind cluster, you don't need multicluster-runtime. If you want to connect to many kind clusters, you still need a "kind provider". Providers are always required and are the heart of multicluster-runtime.

Or maybe I am not understanding the nuance? Did you expect MCR to silently fallback to regular controller-runtime behaviour?

Apparently that is the intended usecase.

[embik]

Did you expect MCR to silently fallback to regular controller-runtime behaviour?

He does because that's the intended (and advertised, in our talks) behaviour. So this bug is valid. nil provider should fall back to single cluster mode, essentially.

[xrstf]

Can you still reproduce this? I checked with the kubeconfigs example and modified it to take a nil provider. I then ran it against a vanilla kind cluster and it immediately saw and reconciled all ConfigMaps in the kind cluster.

[tobschli]

Honestly, I have not touched this since July, so I would need to check again.

[skitt]

/sig multicluster