Skip to content
GreenKube

Introduction

GreenKube is an open-source optimization loop for Kubernetes clusters run with GitOps. It collects optimization recommendations from the tools you already run, ranks them by value and risk, delivers the best ones as reviewable pull requests to your GitOps repository, and then measures what each merged change actually saved: in money, in CO₂e, and in released capacity (CPU cores and memory freed).

sources ─► recommendation ─► plan (diff + digest) ─► pull request ─► merge
▲ │
│ Argo CD / Flux sync
│ │
ledger ◄── measurement vs counterfactual ◄── apply detection ◄─────────┘

Teams running Kubernetes do not lack optimization advice. The hard part comes after the advice:

  • Trust. VPA, Goldilocks, KRR and cloud recommenders use different algorithms and windows, and they disagree. Which one is right for this workload?
  • Safe application. With GitOps, the cluster mirrors Git. Resizing pods in place creates drift that Argo CD or Flux reverts, or that has to be hidden with ignoreDifferences.
  • Proof. Most tools stop at potential savings. Few measure what happened after the change, and fewer link the result to the commit that caused it.
StageWhat GreenKube does
AggregateStarts from GreenKube’s own measurements and reads the other sources of advice in your cluster (VPA and Goldilocks, Robusta KRR reports and anything sent to the push API) into one signal model, provenance kept.
DecideGreenKube’s measured value decides each size; the other sources raise the confidence when they agree and lower it when they disagree (one vote per engine, contested advice labelled and never approved automatically). A source alone never creates a recommendation. Each recommendation carries evidence, a projected impact, a risk level, a confidence score and a risk-adjusted expected value.
DeliverFinds the workload’s Git source through Argo CD or Flux, renders the change in an isolated sandbox, and opens a pull request on GitHub, GitLab or Gitea once a signed-in user approves the exact diff (its digest).
VerifyWaits for Argo CD or Flux to sync the merged revision, then detects that the new configuration is live.
MeasureCompares the workload after the change with a counterfactual built from the 7 days before it, with a 90% confidence interval, confounder checks and health gates, and compares the result with the projection.
RecordWrites every outcome to an append-only, hash-chained ledger linked to the pull request, merge commit, synced revision and measurement.

Git is the only write path

No admission webhooks, no in-place mutation, no drift. Every change, including a rollback, is a pull request.

Projected is never realized

Projected, measured allocated and realized (billed) savings are labelled and never added together. Released capacity is never converted to money.

Fail closed, explain why

When a change cannot be automated safely, GreenKube refuses and tells you why and what would fix it.

Evidence travels with the change

The pull request carries the usage evidence, impact, risk, rendered diff and what will be verified after the merge.

  • Platform and SRE teams who run Argo CD or Flux and want optimization changes to go through the same review as any other change.
  • FinOps practitioners who need measured savings, linked to commits, instead of a list of potential savings.
  • Sustainability teams who want CO₂e next to cost on every change, with documented methods and limits.

It runs in the cluster it optimizes, on-premises and on clouds (AWS, GCP, Azure, OVHcloud, Scaleway, kubeadm, k3s, bare metal).

  • It never changes a workload directly in the cluster. Without GitOps, manual mode gives you the kubectl steps and measures the change once you mark it applied.
  • It does not claim billed savings it has not measured. Today it measures allocated savings (what the changed workloads stop being charged) and shows the expected share that reaches the bill. Pool-level measurement of billed savings is on the roadmap.
  • It recommends what it can deliver and measure: CPU and memory requests and limits (decreases and increases) and deletions of orphaned LoadBalancers and PersistentVolumes. Advisory-only analyzers (zombie pods, idle namespaces, off-peak scaling, node and NodePool advice) were removed in this release.