GKE Workload Scaling

This reference covers scaling workloads on GKE. The golden path enables VPA, OPTIMIZE_UTILIZATION autoscaling profile, and Node Auto Provisioning by default.

MCP Tools: get_k8s_resource, describe_k8s_resource, apply_k8s_manifest, patch_k8s_resource, get_cluster, update_cluster, update_node_pool

Golden Path Scaling Defaults

Setting	Golden Path Value	Notes
autoscaling.autoscalingProfile	OPTIMIZE_UTILIZATION	Aggressive scale-down for cost savings
verticalPodAutoscaling.enabled	true	VPA recommendations available
autoscaling.enableNodeAutoprovisioning	true	NAP creates node pools on demand
GPU resource limits (T4, A100)	1000000000 each	NAP can provision GPU nodes

Scaling Mechanisms

1. Manual Scaling

kubectl-only — no MCP equivalent for kubectl scale. Use kubectl directly.

kubectl scale deployment <DEPLOYMENT> --replicas=<N> -n <NAMESPACE>

2. Horizontal Pod Autoscaling (HPA)

Scales the number of pods based on metrics.

Quick setup (kubectl-only — no MCP equivalent for kubectl autoscale):

kubectl autoscale deployment <DEPLOYMENT> --cpu-percent=50 --min=1 --max=10

Manifest approach (recommended — use MCP apply_k8s_manifest):

See assets/hpa-example.yaml for a template.

apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: <DEPLOYMENT>-hpa
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: <DEPLOYMENT>
  minReplicas: 1
  maxReplicas: 10
  metrics:
  - type: Resource
    resource:
      name: cpu
      target:
        type: Utilization
        averageUtilization: 50

3. Vertical Pod Autoscaling (VPA)

Adjusts CPU and memory requests to match actual usage. Enabled by default on golden path.

Update modes:

• Off — recommendations only (safest, start here)
• Initial — sets resources only at pod creation
• Auto — restarts pods to apply new resource values
• InPlaceOrRecreate — updates resources without restart when possible (GKE 1.34+)

Create VPA in recommendation mode:

apiVersion: autoscaling.k8s.io/v1
kind: VerticalPodAutoscaler
metadata:
  name: <DEPLOYMENT>-vpa
spec:
  targetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: <DEPLOYMENT>
  updatePolicy:
    updateMode: "Off"

Read recommendations (prefer MCP describe_k8s_resource):

# MCP (preferred)
describe_k8s_resource(parent="...", resourceType="verticalpodautoscaler", name="<DEPLOYMENT>-vpa", namespace="<NAMESPACE>")

# kubectl fallback
kubectl get vpa <DEPLOYMENT>-vpa -o jsonpath='{.status.recommendation}'

See assets/vpa-example.yaml for a full template.

4. Cluster Autoscaler / Node Auto Provisioning (NAP)

On Autopilot (golden path), node scaling is fully managed. NAP automatically creates and sizes node pools based on workload demands.

For Standard clusters:

# Enable cluster autoscaler on a node pool
gcloud container clusters update <CLUSTER_NAME> --region <REGION> \
  --enable-autoscaling --node-pool <POOL_NAME> \
  --min-nodes <MIN> --max-nodes <MAX> \
  --quiet

# Enable NAP

**Autoscaling profiles:**

| Profile                | Behavior                             | Golden Path? |
| ---------------------- | ------------------------------------ | ------------ |
| `BALANCED`             | Default GKE; conservative scale-down | No           |
| `OPTIMIZE_UTILIZATION` | Aggressive scale-down; lower idle    | **Yes**      |
:                        : resources                            :              :

## Best Practices

1.  **Define resource requests**: HPA and VPA rely on accurate requests. Always
    set them.
2.  **Avoid metric conflicts**: Do not use HPA and VPA on the same metric.
    Typical pattern: HPA on CPU, VPA on memory.
3.  **Pod Disruption Budgets**: Define PDBs for all production workloads to
    ensure availability during scaling events.
4.  **HPA stabilization**: HPA has a default 5-minute stabilization window. Tune
    `behavior` for faster response if needed.
5.  **VPA "Auto" caution**: Auto mode restarts pods. Ensure your app handles
    SIGTERM gracefully. VPA requires at least 2 replicas for evictions by
    default.
6.  **Use ComputeClasses**: For workload-specific node targeting (Spot fallback,
    GPU, specific machine families), use ComputeClasses instead of node
    selectors.

## Rightsizing Workflow

1.  Deploy VPA in `Off` mode for 24+ hours
2.  Read recommendations: `kubectl describe vpa <NAME>`
3.  Compare `target` values against current `requests`
4.  Apply with 20% buffer: `new_request = target * 1.2`
5.  Use patch format to update Deployment

Condition                     | Recommendation                       | Risk
----------------------------- | ------------------------------------ | ------
CPU request >5x P95 actual    | Reduce to `P95 * 1.2`                | Medium
Memory request >3x P95 actual | Reduce to `P95 * 1.2`                | Medium
CPU request >2x P95 actual    | Rightsizing with 20% buffer          | Low
No resource limits set        | Add limits to prevent noisy-neighbor | Low