gke-scaling

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

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 <DEPLOYMEN

Manifest approach (recommended — use MCP apply_k8s_manifest):

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

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:

Read recommendations (prefer MCP describe_k8s_resource):

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:

Autoscaling profiles:

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