Test without an accelerator

Noctaya's controllers, gateway, scaler integration, and scale-to-zero lifecycle can run on a CPU-only workstation. The E2E suite uses a fake accelerator resource and a CPU vLLM stub; it does not validate drivers, device plugins, vendor runtimes, performance, or physical hardware support.

Prerequisites

Install:

• the Go version declared by go.mod;
• make;
• Docker, or Podman;
• Kind;
• kubectl; and
• Helm.

No existing Kubernetes cluster is required.

Run local checks

make test
make lint

make test covers unit, envtest, adapter-rendering, gateway, model, and stub behavior. It also regenerates and formats source, then writes cover.out; inspect the working tree afterward.

Run the scale-to-zero E2E suite

Run the External Push lifecycle:

make test-e2e

The command:

1. creates the disposable Kind cluster noctaya-test-e2e;
2. uses a dedicated kubeconfig;
3. installs KEDA 2.20.1;
4. builds and loads the operator, gateway, and stub images;
5. runs the External Push lifecycle suite; and
6. deletes the cluster and kubeconfig on exit.

The runner refuses to reuse an existing cluster with that name or overwrite its kubeconfig.

The suite verifies:

• two-replica operator placement, disruption protection, measured leader handoff, and continued reconciliation;
• idle backend scale-to-zero;
• cold activation and streaming completion;
• concurrent scale-out through 0 → 1 → 2 → 0;
• aggregate demand from two gateways during gateway replacement;
• PodDisruptionBudget enforcement during voluntary gateway eviction;
• configurable queue overflow with 429 and Retry-After;
• Secret-backed client authentication, negative paths, and in-place key rotation;
• backend Pod replacement, request recovery, and return to zero;
• backend image-pull failure reporting and recovery after correcting the runtime;
• graceful drain of an active stream;
• fast cold-start rejection with 503;
• push activation before the fallback polling interval;
• allowed and denied traffic through the opt-in ingress NetworkPolicy profile;
• KEDA External Push over mutual TLS, including rejection of a plaintext scaler client.

CI runs the same lifecycle through .github/workflows/test-e2e.yml. Upgrade and rollback remain targeted release checks; follow the contributor release checklist.

Use Podman

Select Podman for both the image build and the Kind provider:

KIND_EXPERIMENTAL_PROVIDER=podman make test-e2e CONTAINER_TOOL=podman

CPU vLLM stub

test/vllm-stub/ provides only the interfaces needed by the suite:

Endpoint	Purpose
/health	Delayed readiness for cold-start tests
/v1/chat/completions and /v1/completions	JSON and streaming SSE responses
/metrics and /control	Runtime-shaped metrics and test-time metric updates

Run its focused tests with:

go test ./test/vllm-stub/...

Physical accelerator claims must follow the hardware validation requirements.