I’m using the latest version of google-kubernetes (1.22.8-gke.202) in a Kubernetes managed cluster. I also have a custom service account that has access to the "Artifact Registry Reader" scope that should grant it permission to pull private images from the repository – calling this custom-service-account.
I’ve validated that the nodes themselves have the custom-service-account service account linked to them within Compute Engine. Kubernetes is setup with a service account that is linked to the IAM service account with the same name through workload identity. However, when I try to spawn a pod that pulls from my private repo it fails indefinitely.
Events:
Type Reason Age From Message
---- ------ ---- ---- -------
Warning FailedScheduling 21m (x3 over 24m) default-scheduler 0/2 nodes are available: 2 node(s) were unschedulable.
Warning FailedScheduling 19m default-scheduler no nodes available to schedule pods
Normal NotTriggerScaleUp 18m (x25 over 24m) cluster-autoscaler pod didn't trigger scale-up: 1 node(s) had taint {reserved-pool: true}, that the pod didn't tolerate
Normal Scheduled 18m default-scheduler Successfully assigned default/test-service-a-deployment-5757fc5797-b54gx to gke-personal-XXXX--personal-XXXX--ac9a05b6-16sb
Normal Pulling 17m (x4 over 18m) kubelet Pulling image "us-central1-docker.pkg.dev/personal-XXXX/my-test-repo/my-test-repo-business-logic:latest"
Warning Failed 17m (x4 over 18m) kubelet Failed to pull image "us-central1-docker.pkg.dev/personal-XXXX/my-test-repo/my-test-repo-business-logic:latest": rpc error: code = Unknown desc = failed to pull and unpack image "us-central1-docker.pkg.dev/personal-XXXX/my-test-repo/my-test-repo-business-logic:latest": failed to resolve reference "us-central1-docker.pkg.dev/personal-XXXX/my-test-repo/my-test-repo-business-logic:latest": failed to authorize: failed to fetch anonymous token: unexpected status: 403 Forbidden
Warning Failed 17m (x4 over 18m) kubelet Error: ErrImagePull
Warning Failed 16m (x6 over 18m) kubelet Error: ImagePullBackOff
Normal BackOff 3m27s (x65 over 18m) kubelet Back-off pulling image "us-central1-docker.pkg.dev/personal-XXXX/my-test-repo/my-test-repo-business-logic:latest"
I’ve also ssh’ed into the nodes themselves and at least by default with a regular docker pull
or crictl pull
see this same error.
So, the specific questions I have:
- How is GCP injecting the service account credentials into Kubernetes/Docker worker that tries to launch the images? Is it expected that the regular docker command doesn’t seem to have these credentials?
- Do I need to manually bootstrap some additional authentication for Kubernetes aside from just inheriting the service account on the pods?
EDIT: Result of here
> gcloud container clusters describe personal-XXXX-gke --zone us-central1-a --format="value(workloadIdentityConfig.workloadPool)"
personal-XXXX.svc.id.goog
> gcloud container node-pools describe personal-XXXX-gke-node-pool --cluster personal-XXXX-gke --format="value(config.workloadMetadataConfig.mode)" --zone us-central1-a
GKE_METADATA
> kubectl describe serviceaccount --namespace default be-service-account
Name: be-service-account
Namespace: default
Labels: <none>
Annotations: iam.gke.io/gcp-service-account: [email protected]
Image pull secrets: <none>
Mountable secrets: be-service-account-token-jmss9
Tokens: be-service-account-token-jmss9
Events: <none>
> gcloud iam service-accounts get-iam-policy [email protected]
bindings:
- members:
- serviceAccount:personal-XXXX.svc.id.goog[default/be-service-account]
role: roles/iam.workloadIdentityUser
etag: BwXjqJ9DC6A=
version: 1
2
Answers
When checking for access to artifact registry, please check permission and scopes as per this documentation.
Depending on how your cluster is created, various scopes are added. https://cloud.google.com/kubernetes-engine/docs/how-to/access-scopes#create_with_sa
In my case, I created Autopilot cluster from the console (UI) and did everything you did w.r.t linking service accounts – turns out the default service account that gets applied does not get the scope cloud-platform.
I ended up re-creating the cluster with the right service account (non-default) for my autopilot nodes. https://cloud.google.com/sdk/gcloud/reference/container/clusters/create#–scopes. Most likely to use the CLI for future creations.