deploy-gke

name	deploy-gke
description	Deploy ML service to GKE with Kustomize overlays and Workload Identity
allowed-tools	["Read","Grep","Glob","Bash(docker:*)","Bash(gcloud:*)","Bash(gsutil:*)","Bash(kubectl:*)","Bash(kustomize:*)","Bash(curl:*)"]
when_to_use	Use when deploying a service to GCP GKE cluster. Examples: 'deploy bankchurn to GKE', 'push to GCP production', 'GKE deployment'
argument-hint	<service-name> <version-tag> [environment]
arguments	["service-name","version-tag","environment"]
authorization_mode	{"dev":"AUTO","staging":"CONSULT","prod":"STOP"}

Deploy to GKE

Authorization Protocol

This skill enforces the Agent Behavior Protocol (AGENTS.md). Actions per environment:

Env	Mode	What the agent does
dev	AUTO	Execute all steps without asking
staging	CONSULT	Show the full plan (image tag, diff, namespace) and wait for a human "proceed" before kubectl apply
prod	STOP	Do NOT apply. Instruct the user to merge an approved PR and let GitHub Actions with environment: production (required_reviewers) perform the deploy

If you are in prod mode and the human insists, output:

[AGENT MODE: STOP]
Operation: Direct kubectl apply to production cluster
Reason: Prod deploys require the governed path (see ADR-002)
Waiting for: Merge to main + GitHub Environment approval

Then halt.

Pre-Flight Checklist

• Verify context: kubectl config current-context must be GKE cluster
• Docker image built and pushed to Artifact Registry
• Kustomize overlay patched with correct image tag
• Terraform applied for any new infrastructure
• Model artifact uploaded to GCS
• All tests passing in CI

Step 1: Verify Cluster Context

kubectl config current-context
# Expected: gke_{PROJECT_ID}_{REGION}_{CLUSTER_NAME}

NEVER proceed if context is wrong. Switch with:

gcloud container clusters get-credentials {CLUSTER} --region {REGION} --project {PROJECT}

Step 2: Build and Push Image

# Tag with version and SHA
export VERSION=v{X.Y.Z}
export SHA=$(git rev-parse --short HEAD)
export REGISTRY={REGION}-docker.pkg.dev/{PROJECT_ID}/{REPO}

docker build -t ${REGISTRY}/{service}:${VERSION} -t ${REGISTRY}/{service}:sha-${SHA} .
docker push ${REGISTRY}/{service}:${VERSION}
docker push ${REGISTRY}/{service}:sha-${SHA}

Step 3: Update Kustomize Overlay

# k8s/overlays/gcp-{env}/kustomization.yaml  (env = dev | staging | production)
images:
  - name: {service}-predictor
    newName: {REGION}-docker.pkg.dev/{PROJECT_ID}/{REPO}/{service}
    newTag: {VERSION}

Step 4: Apply Manifests

# Apply the overlay matching the target environment.
# Production deploys are gated by the dev → staging → prod chain (ADR-011);
# manual application here is for dev iteration or emergency only.
kubectl apply -k k8s/overlays/gcp-{env}/    # env = dev | staging | production
kubectl rollout status deployment/{service}-predictor -n {namespace} --timeout=300s

Step 5: Smoke Test

# Get service URL
export SVC_URL=$(kubectl get ingress -n {namespace} -o jsonpath='{.items[0].status.loadBalancer.ingress[0].ip}')

# Health check
curl -f http://${SVC_URL}/health
curl -f http://${SVC_URL}/ready

# Test prediction with a schema-valid scaffold payload. Add
# `-H "X-API-Key: ${API_KEY}"` when API_AUTH_ENABLED=true.
curl -X POST http://${SVC_URL}/predict \
  -H "Content-Type: application/json" \
  -d '{
    "entity_id": "deploy-smoke-001",
    "slice_values": {"smoke": "gke"},
    "feature_a": 42.0,
    "feature_b": 50000.0,
    "feature_c": "category_A"
  }'

# Metrics scrape smoke
curl -s http://${SVC_URL}/metrics | grep "_requests_total"

Step 6: Verify Monitoring

• Prometheus scraping /metrics from new pods
• Grafana dashboard showing new version
• No alert firing in AlertManager

Rollback (if needed)

kubectl rollout undo deployment/{service}-predictor -n {namespace}
kubectl rollout status deployment/{service}-predictor -n {namespace}

Workload Identity Verification

# Verify SA annotation
kubectl get serviceaccount {service}-sa -n {namespace} -o yaml | grep "iam.gke.io"

# Test GCS access from pod
kubectl exec -it {pod} -n {namespace} -- gsutil ls gs://{model-bucket}/