P
perfscale

Core

Kubernetes

Two ways to run perfscale in a cluster: one-shot or scheduled Jobs for the CLI itself (the Helm chart below), and the perfscaled agent as a DaemonSet (it joins your fleet and picks up platform tasks).

The perfscale chart (Jobs and CronJobs)

The Perfscale/charts repo ships the CLI as a chart — a one-shot Job by default, a CronJob when schedule is set:

helm repo add perfscale https://raw.githubusercontent.com/Perfscale/charts/main/
helm install smoke perfscale/perfscale                                    # one-shot Job
helm install nightly perfscale/perfscale --set schedule="0 3 * * *"       # CronJob

The scenario comes from an inline ConfigMap (--set-file scenario.test=my.test.yaml) or an existing one (--set scenario.existingConfigMap=…). The default scenario is a self-contained SQLite smoke test, so helm install works out of the box. Engine flavors (image.flavor: k6|jmeter|locust|full), resource limits and extra CLI args are plain values — see the chart's values.yaml.

Libraries on the chart (RFC 005)

Enable cache.enabled to get a library-cache PVC mounted at /var/lib/perfscale/cache (PERFSCALE_CACHE_DIR): pre-populate it with perfscale install (from a one-time Job on the same claim) and every run reads libraries from the shared cache. With the cache disabled, libraries ride the scenario volume like in Docker, or are burned into a standalone binary.

One-shot runs as plain Jobs

No Helm needed — the engine images work as plain Jobs. Mount the scenario (ConfigMap for small tests, a PVC otherwise) and run:

apiVersion: batch/v1
kind: Job
metadata:
  name: perfscale-smoke
spec:
  template:
    spec:
      restartPolicy: Never
      containers:
        - name: perfscale
          image: ghcr.io/perfscale/perfscale:latest
          args: ["run", "-f", "/work/test.yaml", "-c", "/work/config.yaml"]
          volumeMounts:
            - name: scenario
              mountPath: /work
      volumes:
        - name: scenario
          configMap:
            name: perfscale-smoke

Libraries work the same as in Docker: local components ride the mounted volume (paths resolve relative to the declaring file); remote ones need a shared install cache (PERFSCALE_CACHE_DIR on a PVC) populated by perfscale install, or a burned standalone binary with libraries embedded.

The perfscaled agent (DaemonSet)

The platform agent runs one pod per node as a DaemonSet and joins your fleet. Its Helm chart is parked until the agent ships a container image — track Perfscale/charts#1; the chart source from before the park is in the repo's git history (git show 9718911:perfscaled/values.yaml).

Agents enforce a fail-closed fleet policy on libraries: declared by tasks:

  • PERFSCALE_LIBRARY_CAPABILITIES — the fleet ceiling (fs, clock, net:<host> globs). Default: none — libraries run, but with zero capabilities. A task granting more is rejected with a validation error, never silently downgraded.
  • PERFSCALE_LIBRARY_ALLOW_DIGESTS — optional sha256 allowlist of artifacts permitted on the fleet at all.
  • Remote libraries are never fetched at run time: they are installed by digest (perfscaled library install <url> --sha256 … or POST /api/v1/libraries/install) into a content-addressed cache, and the cached bytes are re-hashed when a task is accepted.