Local Testing (No GPU Required)

Test the CastSlice webhook mutation logic on a local KinD or Minikube cluster with no physical NVIDIA GPU. You'll mock GPU capacity on the node and verify that the annotation-triggered rewrite works end-to-end.

ℹ️

This guide requires CastSlice to already be installed. If you haven't done that yet, follow the Getting Started guide first.

Step 1 — Mock a Node with nvidia.com/gpu capacity

Kubernetes does not allow kubectl patch to add arbitrary extended resources to status.capacity directly because the status sub-resource is guarded by the API server. The standard workaround is to PATCH the node directly via the raw API through kubectl proxy.

⚠️

This modifies node status in your cluster. Only do this in a development or local cluster — never in production.

Find your node name and start kubectl proxy

bash
# Get the first (or only) node name
$ NODE=$(kubectl get nodes -o jsonpath='{.items[0].metadata.name}')
$ echo "Patching node: $NODE"
Patching node: kind-control-plane

# Start proxy in background, save PID
$ kubectl proxy &
$ PROXY_PID=$!
Starting to serve on 127.0.0.1:8001

Patch the node status to add fake GPU capacity

bash
$ curl -s -X PATCH \
  -H "Content-Type: application/json-patch+json" \
  "http://localhost:8001/api/v1/nodes/${NODE}/status" \
  --data '[
    {"op":"add","path":"/status/capacity/nvidia.com~1gpu","value":"4"},
    {"op":"add","path":"/status/allocatable/nvidia.com~1gpu","value":"4"}
  ]'

# Stop the proxy
$ kill $PROXY_PID

After this, the node advertises 4 fake nvidia.com/gpu resources. Kubernetes will schedule Pods requesting GPU resources against this node.

Verify fake capacity is visible

bash
$ kubectl describe node ${NODE} | grep -A 8 "Capacity:"
Capacity:
  nvidia.com/gpu:  4
  cpu:             8
  memory:          16316112Ki
  pods:            110
Allocatable:
  nvidia.com/gpu:  4

✓ Node now reports 4 × nvidia.com/gpu available

Step 2 — Apply the Test Pod

The repository includes a pre-built test Pod at docs/test-pod.yaml that requests a GPU and includes the castops.io/optimize: "true" annotation.

Create the test Pod

bash
$ kubectl apply -f docs/test-pod.yaml
pod/gpu-test-pod created

The Pod spec requests nvidia.com/gpu: 1 and has the opt-in annotation. CastSlice intercepts the CREATE admission request and rewrites it before it is persisted.

Step 3 — Verify the Mutation

Inspect the actual resource spec that was stored — this is what CastSlice wrote, not what you applied.

Check the stored resource limits

bash
$ kubectl get pod gpu-test-pod \
  -o jsonpath='{.spec.containers[0].resources}' | jq .
{
  "limits": {
    "nvidia.com/gpu-shared": "1",
    "memory": "4Gi"
  },
  "requests": {
    "memory": "4Gi"
  }
}

✓ nvidia.com/gpu is gone — nvidia.com/gpu-shared: 1 took its place

✅

Mutation confirmed. The webhook intercepted the CREATE request, detected the opt-in annotation, and rewrote the resource limits via JSON Patch before the Pod was persisted to etcd.

Troubleshooting

Pod stays Pending after apply

If the fake node doesn’t also have the nvidia.com/gpu-shared resource, the Pod won’t be schedulable (it’s waiting for a resource that doesn’t exist on any node). For purely testing webhook mutation, you can check the stored spec immediately after kubectl apply — the Pod doesn’t need to reach Running state to verify the mutation happened.

bash — check webhook logs if mutation didn't happen
$ kubectl logs -n cast-slice deploy/cast-slice

Webhook is not mutating

Check that:

The Pod (or Pod template) has castops.io/optimize: "true" in its metadata.annotations
CastSlice is Running: kubectl get pods -n cast-slice
The webhook is registered: kubectl get mutatingwebhookconfigurations
cert-manager issued the certificate: kubectl get certificate -n cast-slice

curl: (7) Failed to connect to localhost:8001

kubectl proxy may have already been killed or the port is different. Re-run kubectl proxy & and check the port it reports.