Seeing flakey testing where default namespace doesn't exist

[jonathan-innis]

Currently, we are running envtest in Karpenter where we are starting up all of the binaries and then applying pods to the default namespace. Every now and then, we are seeing failures in our pod metric testing (seen here: https://github.com/kubernetes-sigs/karpenter/actions/runs/7251475446/job/19753909305?pr=885). When this happens, it always contains the same error

Pod Metrics [It] should update the pod state metrics
/home/runner/work/karpenter/karpenter/pkg/controllers/metrics/pod/suite_test.go:56

  [FAILED] Expected success, but got an error:
      <*errors.StatusError | 0xc000134aa0>: 
      namespaces "default" not found
      {
          ErrStatus: {
              TypeMeta: {Kind: "", APIVersion: ""},
              ListMeta: {
                  SelfLink: "",
                  ResourceVersion: "",
                  Continue: "",
                  RemainingItemCount: nil,
              },
              Status: "Failure",
              Message: "namespaces \"default\" not found",
              Reason: "NotFound",
              Details: {Name: "default", Group: "", Kind: "namespaces", UID: "", Causes: nil, RetryAfterSeconds: 0},
              Code: 404,
          },
      }
  In [It] at: /home/runner/work/karpenter/karpenter/pkg/controllers/metrics/pod/suite_test.go:58 @ 12/18/23 17:16:52.[71](https://github.com/kubernetes-sigs/karpenter/actions/runs/7251475446/job/19753909305?pr=885#step:5:72)9

This error indicates that the default namespace doesn't exist at the apiserver when we are applying the object; however, I would generally expect this namespace to exist once the apiserver starts up.

It's worth noting that we don't see this error in other testing where we are running envtest.Environment with the CRDs option set. I suspect that this is due to the fact that this gives the setup a little more time to populate this namespace at the apiserver before it returns.

Is there a general recommendation here? Obviously we could wait for the default namespace to exist in our own code, but it would be nice if we could be assured that it would always exist when we start testing from upstream.

[troy0820]

/kind support

[sbueringer]

It's worth noting that we don't see this error in other testing where we are running envtest.Environment with the CRDs option set. I suspect that this is due to the fact that this gives the setup a little more time to populate this namespace at the apiserver before it returns.

Probably as envtest first deploys CRDs and then waits for them to be available.

Which client call gives you this error?

[jonathan-innis]

Probably as envtest first deploys CRDs and then waits for them to be available

Yeah, this is what I was suspecting. There must be enough time in there that it's able to bring-up the default namespace.

Which client call gives you this error

It's a call to create a pod with no namespace specified.

[sbueringer]

@jonathan-innis I'm fine with adding a "waitForDefaultNamespace" call here: https://github.com/kubernetes-sigs/controller-runtime/blob/21779fbe6419e03ae1cd86a69b35d2725c6f7558/pkg/envtest/server.go#L273C1-L273C1

[jonathan-innis]

Perfect, I'll try to put-up a PR to add this in shortly 🎉 Thanks for the help!

[jonathan-innis]

@sbueringer Opened a PR with the wait mechanism. Frustratingly, I've been trying to repro this locally but no dice, so it's hard for me to tell if this is actually going to resolve our issue. I'd suspect that it will, but it's hard to say without being able to prove it.

Let me know if you have any thoughts on ways I could force this race.