Representing soft deletes in the tests

[squeedee]

I am testing with ReconcilerTestCase and would like to emulate what happens when a childset deletes children that have finalizers (of their own). In such a case, I'd like the ResourceManager's Delete to receive the correct response from the fakeClient AND for the resources to remain in the FakeClient stubs, with deletionTimestamps added. This way, later reconcilers can query the same subresources to decide if the top level's resource should keep it's own finalizer.

Is this possible?

Alternatively, If I were to wrap a childset with WithFinalizer, it would be nice if the finalizer was not cleared if child deletion was itself a soft deletion.

[scothis]

I'd like the ResourceManager's Delete to receive the correct response from the fakeClient AND for the resources to remain in the FakeClient stubs, with deletionTimestamps added.

Looking at the fake client code in controller-runtime, it appears this is mostly already happening. The gotcha is that the passed in object isn't updated to reflect the server side value. It will need to be fetched again to see the deletion timestamp. The real client appears to behave the same way. You may see different behavior in a running controller if the resource is managed by an informer. In that case you'll need to use the APIReader to talk directly to the API server instead of the Client.

https://github.com/kubernetes-sigs/controller-runtime/blob/v0.18.2/pkg/client/fake/client.go#L1054-L1060

Logically for most cases a deleted resource with a finalizer should be treated as if it doesn't exist. The exceptions being if you are responsible for clearing the finalizer, and a new resource with the same name cannot be created.

[squeedee]

The manager instance is responsible for ensuring the object finishes it's finalization before it takes away the resource's controller's means of accessing it (a ClusterRole)...

The manager is not responsible for clearing the resource's finalizer - it's a child resource with a seperater manager.

The behavior you describe is not happening, instead:

1. A childset subreconciler set's the resource for deletion (it desires no children, triggering RMs deletion mechanism)
2. In a subsequent subreconciler, we reload the resource to see if they were properly deleted before we proceed to delete the aforementioned ClusterRole.

The code is working in reality, but not in the test. I'll dig deeper on monday and get back to you.

P.s. Apologies for all the edits, I realized i failed to make sense repeatedly. Hope I've described it clearly now.

[squeedee]

You're right it does work (when I remember to add the finalizer to the GivenObject)...

Now to my desire for Marker error on child/childset when it deletes resources but they're not removed? This would let us wrap a childset with a WithFinalizer that isn't removed until children are gone. WDYT?

[scothis]

Are you trying to do the equivalent of an owner reference with blockOwnerDeletion without an owner reference?

[squeedee]

Very much so.. Namespaces are in our way.

[scothis]

I think this will give you the necessary hook to plug in your desired behavior.

reconcilerio/runtime#525