Skip to content

Chart Testing & Promotion

Goal: a chart you can prove renders correctly before it ever reaches a cluster — and a versioned path from dev to prod that doesn't rely on "apply and see."

The chart testing pyramid

helm lint            — syntax/schema sanity (free, always run)
helm template        — render locally, eyeball the diff (no cluster needed)
helm unittest        — assert rendered output as unit tests (CI gate)
helm test            — in-cluster smoke test pods (post-deploy verification)
conftest (OPA)       — policy assertions on rendered manifests (Phase 6 pairs)

helm unittest — the missing layer most charts skip

helm template | less eyeballs output; helm-unittest asserts it. Install as a plugin:

helm plugin install https://github.com/helm-unittest/helm-unittest
helm unittest todo-app/

The suite in todo-app/tests/deployment_test.yaml teaches the pattern:

Assertion What it guards
equal spec.replicas: 1 default replica contract
image.tag: null → appVersion the default-tag fallback you think exists
autoscaling.enabled → notExists spec.replicas the conditional block every chart ships but nobody tests — replicas + HPA fighting is a real prod bug
probe port: tomcat + containerPort the service→container port contract
matchRegex :0\.9\.7$ the no-:latest values contract

Run it in CI (the gitlab-spring-app-flux package stage is where this plugs in): helm unittest failing = chart never packaged. That's the shift-left version of "the chart applied but the probes were wrong."

Promotion: version → dev → prod

flowchart LR
    MR[MR bumps version: in Chart.yaml] --> P[helm package → .tgz / OCI push]
    P --> D[deploy dev<br/>values-dev.yaml]
    D --> QA[deploy qa<br/>values-qa.yaml]
    QA --> PR[deploy prod<br/>values-prod.yaml + manual gate]

The rules:

  1. Chart version: is the artifact version — bump it every template change (semver); appVersion: is the app it ships. Confusing them in an interview is a tell.
  2. Same chart, per-env values — promotion = the same .tgz + different values-*.yaml, never a rebuilt chart per env.
  3. Immutable promotion — prod runs the exact artifact dev ran, pinned by chart version + image digest, not latest.
  4. OCI registries (helm push oci://...) are the modern chart repo — see solutions/helm-oci.

Deliberate break

  • Set autoscaling.enabled: true + leave replicas: 3 in values → the unittest notExists spec.replicas catches the conflict your eyes missed.
  • Deploy prod with tag: latest → the regex test fails — the contract is enforced in the suite, not in a wiki page.
  • Bump templates without bumping version: → a repackaged .tgz with the same version; caches and registries may serve the old artifact. Version bumps are not optional bookkeeping.

Verify

  • helm unittest runs green in CI on every chart change
  • You can name what each pyramid layer catches that the layer below can't
  • Promotion = same artifact + different values, stated in one sentence