Access to Google Cloud Storage from CKS Pods

This tutorial demonstrates how to configure a CoreWeave Kubernetes Service (CKS) cluster to authenticate to Google Cloud Storage (GCS) using OIDC Workload Identity Federation. A Kubernetes ServiceAccount in CKS will be able to access GCS directly, without any stored credentials.

Overview

CKS issues OIDC-compliant ServiceAccount tokens to pods. These tokens can be used to authenticate to external services like Google Cloud Platform (GCP) by establishing OIDC trust. This eliminates the need for long-lived credentials and allows per-ServiceAccount scoping of access to cloud resources. Benefits of this approach:

No credentials are stored in secrets or container images.
Tokens are short-lived and rotated automatically by Kubernetes.
IAM permissions can be tightly scoped to individual ServiceAccounts.
No TLS thumbprint management is required, as with AWS.

After completing this tutorial, you’ll have the following setup:

Pod requests access: The gcs-client pod uses its mounted OIDC token
GCP validates identity: Google Cloud verifies the token against your CKS cluster’s OIDC endpoint
Impersonation granted: GCP allows the pod to impersonate the gcs-reader service account
Resource access: The pod can read from your GCS bucket using temporary credentials

Prerequisites

Before you begin, ensure you have the following:

CoreWeave requirements

CKS cluster: A CoreWeave Kubernetes Service cluster with OIDC Workload Identity enabled
Cluster access: kubectl configured to access your CKS cluster
Cluster details: Your cluster’s OIDC issuer URL (we’ll show you how to find this)

Google Cloud Platform requirements

GCP project: A Google Cloud Platform project where you’ll configure Workload Identity
GCP permissions: Your GCP account must have the following IAM roles:
- Workload Identity Pool Admin (to create pools and providers)
- Service Account Admin (to create and manage service accounts)
- Project IAM Admin (to bind service accounts to workload identities)
GCS bucket: A Google Cloud Storage bucket for testing (or permission to create one)

Command line tools

gcloud CLI: Google Cloud SDK installed and authenticated
kubectl: Kubernetes command line tool configured for your CKS cluster

GCP project information

You’ll need these values during the tutorial. Gather them beforehand:

Project ID: Your GCP project ID (for example, my-project-123)
Project Number: Your GCP project number (numeric, for example, 123456789012)

To find your project details:

# Get both project ID and number
gcloud projects describe $(gcloud config get-value project)

Verify your setup

Test that everything is configured correctly:

# Verify gcloud authentication and project ID
gcloud auth list
gcloud config get-value project

# Verify kubectl access to your CKS cluster
kubectl get nodes

# Verify you have necessary GCP permissions
gcloud iam workload-identity-pools list --location=global

If any of these commands fail, resolve the authentication or permission issues before proceeding.

Set up Kubernetes resources

Before configuring GCP, you need to create the Kubernetes namespace and ServiceAccount that will be granted access to GCS.

Create the namespace

Create a namespace called foo where your workloads will run:

kubectl create namespace foo

Create the ServiceAccount

Create a ServiceAccount called bar that your pods will use:

kubectl create serviceaccount bar --namespace foo

Verify the resources

Confirm both resources were created successfully:

# Verify the namespace exists
kubectl get namespace foo

# Verify the ServiceAccount exists
kubectl get serviceaccount bar --namespace foo

# View the ServiceAccount details (including any tokens)
kubectl describe serviceaccount bar --namespace foo

Expected output should show:

Namespace foo in Active status
ServiceAccount bar exists in the foo namespace
ServiceAccount has default token secrets (these will be replaced with projected OIDC tokens)

Understanding the mapping

These Kubernetes resources will map to GCP identities as follows:

Namespace: foo → GCP attribute attribute.k8s_ns=foo
ServiceAccount: bar → GCP attribute attribute.k8s_sa=bar

When you configure the GCP Workload Identity binding, you’ll reference this specific combination (foo/bar) to ensure only pods running with this ServiceAccount in this namespace can access your GCS resources.

You can use different namespace and ServiceAccount names, but make sure to update all the GCP commands accordingly. The tutorial uses foo/bar as an example, but in production you’d typically use more descriptive names like data-pipeline/gcs-reader.

Get your OIDC issuer URL from CKS

To use identity federation, you must know the OIDC Issuer URL of your CKS cluster. This is the base URL from which token metadata and keys are served. It’s formatted as a valid HTTPS URL, such as: https://oidc.cks.coreweave.com/id/<CLUSTER_ID>. There are a few ways to obtain it:

CKS Console
CKS API
Terraform

In the Cloud Console, navigate to the Clusters page.
Click the name of the cluster to expand the cluster details panel.
The OIDC Issuer URL is displayed in the Overview section.

Use the CoreWeave Cloud API to query the cluster configuration and extract the serviceAccountIssuer value. Otherwise, you can call the CKS API to get the CKS OIDC config:

Get the CKS OIDC config via the API

curl -s -X GET https://api.coreweave.com/v1beta1/cks/clusters/{cluster-id} \
     -H "Content-Type: application/json" \
     -H "Authorization: Bearer {API_ACCESS_TOKEN}" \
     | jq -r '.cluster.oidc.issuerUrl'

If you’re using CoreWeave’s Terraform provider, you can export the OIDC issuer as an output during cluster provisioning.The CKS OIDC config is an attribute you can read from your Terraform provider.Ensure the issuer endpoint exposes a .well-known/openid-configuration path and a valid JWKS endpoint.

Create a workload identity pool in GCP

Create a pool to represent trusted external identities (your CKS pods):

$ gcloud iam workload-identity-pools create k8s-pool \
  --location="global" \
  --display-name="CKS Pool"

Confirm the pool was created successfully:
```
$ gcloud iam workload-identity-pools describe k8s-pool --location=global
```
Expected output should show:
- state: ACTIVE
- name: projects/PROJECT_NUMBER/locations/global/workloadIdentityPools/k8s-pool
If this fails, check that you have the Workload Identity Pool Admin role and are authenticated to the correct GCP project.

Create an OIDC provider in the pool

This setup allows GCP to map Kubernetes tokens to identities based on their namespace and ServiceAccount name:

Configure the provider to trust your CKS OIDC issuer and extract identity information from tokens. Replace <REGION> and <CLUSTER_ID> with your actual values.

$ gcloud iam workload-identity-pools providers create-oidc k8s-provider \
  --location="global" \
  --workload-identity-pool="k8s-pool" \
  --display-name="CKS OIDC Provider" \
  --issuer-uri="https://oidc.cks.coreweave.com/id/<CLUSTER_ID>" \
  --attribute-mapping="google.subject=assertion.sub,attribute.k8s_ns=assertion.kubernetes.io/serviceaccount/namespace,attribute.k8s_sa=assertion.kubernetes.io/serviceaccount/name"

Check that the provider was configured correctly:
```
$ gcloud iam workload-identity-pools providers describe k8s-provider \
  --location=global \
  --workload-identity-pool=k8s-pool
```
Expected output should include:
- state: ACTIVE
- Your CKS OIDC issuer URL in the issuerUri field
- The attribute mapping you configured

Create a Google Cloud Service Account (GSA) and grant access

Create the service account that will be impersonated by CKS workloads:

$ gcloud iam service-accounts create gcs-reader \
  --display-name="CKS GCS Reader"

Grant it permission to read from GCS. Replace <PROJECT_ID> with your actual project ID.

$ gcloud projects add-iam-policy-binding <PROJECT_ID> \
  --member="serviceAccount:gcs-reader@<PROJECT_ID>.iam.gserviceaccount.com" \
  --role="roles/storage.objectViewer"

Bind the CKS service account to the GSA

Authorize the Kubernetes ServiceAccount bar in namespace foo to impersonate the GSA via the identity pool. Replace <PROJECT_NUMBER> and <PROJECT_ID> with your actual values.

$ gcloud iam service-accounts add-iam-policy-binding gcs-reader@<PROJECT_ID>.iam.gserviceaccount.com \
  --role="roles/iam.workloadIdentityUser" \
  --member="principalSet://iam.googleapis.com/projects/<PROJECT_NUMBER>/locations/global/workloadIdentityPools/k8s-pool/attribute.k8s_ns/foo/attribute.k8s_sa/bar"

Check that the binding was created correctly. Replace <PROJECT_ID> with your actual project ID.
```
$ gcloud iam service-accounts get-iam-policy gcs-reader@<PROJECT_ID>.iam.gserviceaccount.com
```
Expected output should include a binding with:
- role: roles/iam.workloadIdentityUser
- members containing your principalSet://iam.googleapis.com/projects/... entry

Only tokens issued to the foo/bar ServiceAccount will be permitted to impersonate the gcs-reader account.

Prepare a test GCS bucket

Before testing the authentication, create a GCS bucket or use an existing one:

Create new test bucket
Use existing bucket

# Create a test bucket (bucket names must be globally unique)
gsutil mb gs://my-cks-test-bucket-$(date +%s)

# Add a test file
echo "Hello from CKS!" | gsutil cp - gs://my-cks-test-bucket-$(date +%s)/test.txt

If you have an existing bucket, ensure the gcs-reader service account has appropriate permissions. Replace [YOUR-BUCKET-NAME] with your actual bucket name.

# Grant the service account access to your bucket
gsutil iam ch serviceAccount:gcs-reader@[PROJECT-ID].iam.gserviceaccount.com:objectViewer gs://[YOUR-BUCKET-NAME]

If bucket creation fails, check the following:

Bucket names must be globally unique - try adding a timestamp or random suffix.
Ensure you have Storage Admin permissions in your GCP project.

Use projected OIDC token in pod and access GCS

In your workload, configure a projected service account token with the proper audience. Create a pod YAML file called gcs-client-pod.yaml with the following content, filling in the placeholder values for [PROJECT-NUMBER], [PROJECT-ID], and [YOUR-BUCKET-NAME].

gcs-client-pod.yaml

apiVersion: v1
kind: Pod
metadata:
  name: gcs-client
spec:
  serviceAccountName: bar
  containers:
  - name: gcs
    image: google/cloud-sdk:slim
    command:
    - bash
    - -c
    - |
      gcloud iam workload-identity-pools create-cred-config \
        projects/[PROJECT-NUMBER]/locations/global/workloadIdentityPools/k8s-pool/providers/k8s-provider \
        --service-account="gcs-reader@[PROJECT-ID].iam.gserviceaccount.com" \
        --credential-source-file=/var/run/secrets/tokens/oidc-token \
        --output-file=/tmp/creds.json && \
      gcloud auth login --cred-file=/tmp/creds.json && \
      gsutil ls gs://[YOUR-BUCKET-NAME]
    volumeMounts:
    - name: oidc-token
      mountPath: /var/run/secrets/tokens
    - name: creds
      mountPath: /tmp
  volumes:
  - name: oidc-token
    projected:
      sources:
      - serviceAccountToken:
          path: oidc-token
          audience: //iam.googleapis.com/projects/[PROJECT-NUMBER]/locations/global/workloadIdentityPools/k8s-pool/providers/k8s-provider
          expirationSeconds: 3600
  - name: creds
    emptyDir: {}

This pod uses the projected token to generate GCP-compatible credentials dynamically at runtime, then uses them to read from GCS.

Deploy and test the pod

Apply the pod configuration:

$ kubectl apply -f gcs-client-pod.yaml -n foo

Wait for the pod to start and check its status:

$ kubectl get pod gcs-client -n foo
$ kubectl logs gcs-client -n foo

If the pod fails, check the following:

Error 403: Permission denied: Check that the gcs-reader service account has access to your bucket.
Invalid token: Verify your OIDC issuer URL matches your cluster’s actual endpoint.
Pod won’t start: Ensure the bar ServiceAccount exists in the foo namespace.

Debug commands:

# Check if token is being mounted
kubectl exec gcs-client -n foo -- ls -la /var/run/secrets/tokens/

# View detailed pod events
kubectl describe pod gcs-client -n foo

Clean up resources

If you’re done testing, remove the resources to avoid charges and maintain security:

Remove Kubernetes resources

# Delete the test pod
kubectl delete pod gcs-client -n foo

# Delete the ServiceAccount (optional, if you're not using it elsewhere)
kubectl delete serviceaccount bar -n foo

# Delete the namespace (optional, will remove everything in it)
kubectl delete namespace foo

Remove GCP resources

# Remove the IAM binding
gcloud iam service-accounts remove-iam-policy-binding gcs-reader@<PROJECT_ID>.iam.gserviceaccount.com \
  --role="roles/iam.workloadIdentityUser" \
  --member="principalSet://iam.googleapis.com/projects/<PROJECT_NUMBER>/locations/global/workloadIdentityPools/k8s-pool/attribute.k8s_ns/foo/attribute.k8s_sa/bar"

# Delete the Google Cloud Service Account
gcloud iam service-accounts delete gcs-reader@<PROJECT_ID>.iam.gserviceaccount.com

# Delete the OIDC provider
gcloud iam workload-identity-pools providers delete k8s-provider \
  --location=global \
  --workload-identity-pool=k8s-pool

# Delete the Workload Identity Pool
gcloud iam workload-identity-pools delete k8s-pool \
  --location=global

Remove test bucket (if created)

# Delete the test bucket and its contents
gsutil rm -r gs://[YOUR-BUCKET-NAME]

Deleting shared resourcesDeleting the Workload Identity Pool will break authentication for any other applications using it. Only delete shared resources if you’re sure they’re not needed elsewhere.

CoreWeave Kubernetes Service

Documentation Index

​Overview

​Prerequisites

​CoreWeave requirements

​Google Cloud Platform requirements

​Command line tools

​GCP project information

​Verify your setup

​Set up Kubernetes resources

​Create the namespace

​Create the ServiceAccount

​Verify the resources

​Understanding the mapping

​Get your OIDC issuer URL from CKS

​Create a workload identity pool in GCP

​Create an OIDC provider in the pool

​Create a Google Cloud Service Account (GSA) and grant access

​Bind the CKS service account to the GSA

​Prepare a test GCS bucket

​Use projected OIDC token in pod and access GCS

​Deploy and test the pod

​Clean up resources

​Remove Kubernetes resources

​Remove GCP resources

​Remove test bucket (if created)

Overview

Prerequisites

CoreWeave requirements

Google Cloud Platform requirements

Command line tools

GCP project information

Verify your setup

Set up Kubernetes resources

Create the namespace

Create the ServiceAccount

Verify the resources

Understanding the mapping

Get your OIDC issuer URL from CKS

Create a workload identity pool in GCP

Create an OIDC provider in the pool

Create a Google Cloud Service Account (GSA) and grant access

Bind the CKS service account to the GSA

Prepare a test GCS bucket

Use projected OIDC token in pod and access GCS

Deploy and test the pod

Clean up resources

Remove Kubernetes resources

Remove GCP resources

Remove test bucket (if created)