Running go test ./... fails with:

--- FAIL: TestRoundTripper (5.01s)
    openshift_test.go:27: Get "https://image-registry.openshift-image-registry.svc.cluster.local:5000/v2/": Post "https://127.0.0.1:54405/api/v1/namespaces/default/pods": dial tcp 127.0.0.1:54405: connect: connection refused

This appears to be because of the implementation of the new IsOpenShift function which assumes any error other than NotFound means the cluster is OpenShift:

func IsOpenShift() bool {
  checkOpenShiftOnce.Do(func() {
    client, err := k8s.NewKubernetesClientset()
    if err != nil {
      fmt.Print("error in NewKubernetesClientset")
      isOpenShift = false
      return
    }
    _, err = client.CoreV1().Services("openshift-image-registry").Get(context.TODO(), "image-registry", metav1.GetOptions{})
    if k8sErrors.IsNotFound(err) {
      isOpenShift = false
      return
    }
    // FIXME(lkingland): This seems to ignoire legitimate errors other than NotFound
    isOpenShift = true
  })
  return isOpenShift
}

As Evan mentioned in the PR in which this check was added, there might be a better way to detect this:

Hmm -- this will return false if the user doesn't have services get permission on the openshift-image-registry namespace.

I'm not sure if there's another / better way to detect this, either from server headers or by probing the registry and looking at its header / challenge. I think go-containerregistry uses a GET /v2/ HTTP /1.1 probe of the registry to figure out what and whether auth is required.

Originally posted by @evankanderson in #825 (comment)

The error appears to be a simple connection refused (no cluster is running locally, as one is not needed to run unit tests).

If this is expected behavior, and the test depends on a cluster being available, perhaps tagging it as an integration test would to the trick?

//go:build integration
// +build integration