Skip to main content

VPA

Lumie deploys VPA in recommendation-only mode. It is there to compute resource advice, not to automatically change running workloads.

Source paths​

  • lumie-infra/observability/vpa/argocd.yaml
  • lumie-infra/observability/vpa/helm-values.yaml

Runtime contract​

  • recommender enabled
  • updater disabled
  • admission controller disabled
  • ignoreDifferences rule for CPU limits to avoid ArgoCD drift noise

What that means operationally​

  • VPA gathers history and computes recommended requests.
  • No controller in this deployment rewrites live pod specs.
  • Goldilocks is the main consumer-facing surface for those recommendations.

Failure modes​

  • Operators may assume VPA is auto-rightsizing workloads; it is not.
  • If the recommender is down, Goldilocks loses fresh data but workloads keep running unchanged.
  • If a team later enables updater or admission without revisiting repo assumptions, the operational model changes materially and this page must be updated.

Verification​

kubectl get applications.argoproj.io -n argocd vpa
kubectl get pods -n vpa
kubectl describe deploy -n vpa vpa-recommender
kubectl get vpa -A