workload-resizer

A Kubernetes controller for GKE that uses the in-place pod resize subresource (pods/resize, GA in K8s 1.35) to dynamically adjust pod resource requests when pods land on node types whose performance characteristics differ from the type the workload was originally calibrated for.

When a Deployment is sized for one machine family (say n2d) but the scheduler places its pods on a more powerful one (n4, ~25% more CPU per core), the original requests over-provision capacity. workload-resizer watches scheduled pods, reads the assigned node's cloud.google.com/machine-family label (configurable via --node-type-label), and patches the pod's requests in place using a configurable performance-unit matrix — without restarting the container.

Status

Pre-v1. The v1 design and scope is settled (see AGENTS.md); the suite covers unit, envtest, and Kind + KWOK e2e layers.

Quick start

# 1. Create the namespace + apply the ConfigMap — edit nodeTypes for
#    your cluster first! Doing this *before* installing the controller
#    avoids a brief CrashLoopBackOff (the controller fails fast on
#    startup if its required ConfigMap doesn't exist yet).
curl -fsSLO https://github.com/gke-demos/workload-resizer/releases/latest/download/config.yaml
$EDITOR config.yaml
kubectl apply -f config.yaml

# 2. Install the controller (RBAC, Deployment).
kubectl apply -f https://github.com/gke-demos/workload-resizer/releases/latest/download/install.yaml

# 3. (Optional) Apply a sample workload to see the resize happen.
kubectl apply -f https://raw.githubusercontent.com/gke-demos/workload-resizer/main/config/samples/deployment.yaml

# Watch the pod get resized:
kubectl get pod -l app=sample-workload -w \
  -o custom-columns=NAME:.metadata.name,\
NODE:.spec.nodeName,\
CPU:.spec.containers[0].resources.requests.cpu,\
APPLIED:.metadata.annotations.workload-resizer\\.io/applied-instance-type

See the install page for prerequisites, how to inventory your cluster's instance types, picking performance units, and upgrade/uninstall.

How it works

For each pod that lands on a node, the controller:

1. Reads the assigned node's cloud.google.com/machine-family label (configurable via --node-type-label).
2. Looks up the per-type performance unit in the global config (a ConfigMap).
3. Computes desired = clamp(original × baselinePerf / nodePerf, bounds) per container.
4. Patches the pod via /resize, mirroring the change into limits when present to preserve QoS class.
5. Records the original requests and the applied instance type as annotations on the pod, for idempotent reconciliation across controller restarts.

Full architecture and the design rationale are in AGENTS.md. The published docs site lives at https://gke-demos.github.io/workload-resizer/.

Development

dev/tools/ci             # run every CI check locally (same as GitHub Actions)
dev/tools/fix-go-format  # auto-fix formatting
make test                # envtest suite only (fast)
make test-e2e            # full Kind + KWOK e2e (~3 min)

See dev/README.md for the dev tooling layout and CONTRIBUTING.md for the contribution workflow (DCO sign-off required).

License

Apache License 2.0.