• metrics-server (pod usage via metrics.k8s.io)
• Kubernetes cluster with in-place Pod resize support (pods/resize subresource). The operator checks capability and degrades gracefully if unsupported.

helm install resize-operator oci://ghcr.io/iamhalje/charts/resize-operator -n resize-operator-system --create-namespace

On each reconcile loop the operator:

• selects candidate Pods (namespace selector + pod label selector)
• fetches PodMetrics from metrics.k8s.io
• computes desired requests based on the policy
• executes pods/resize (dry-run + apply)
• writes annotations and emits Resized events

%%{init: {'flowchart': {'nodeSpacing': 60, 'rankSpacing': 80}}}%%
flowchart TB

  subgraph METRICS["Metrics API"]
    MS[metrics-server]
  end

  subgraph OPERATOR["Resize Operator"]
    SEL[Select candidate Pods]
    CALC[Compute desired requests]
    REC[Write annotations<br/>and emit Resized events]
  end

  subgraph CONTROL_PLANE["Kubernetes Control Plane"]
    APIS[kube-apiserver]
    SCHED[scheduler]
  end

  subgraph NODE["Node"]
    KUBELET[kubelet]
  end

  SEL -->|fetch PodMetrics| MS
  MS -->|PodMetrics| CALC
  CALC -->|pods/resize<br/>dry-run + apply| APIS
  APIS --> KUBELET
  KUBELET -->|resource update| SCHED
  APIS --> REC

The operator computes new requests based on the current resource usage from metrics.k8s.io, using headroomPercent.

• Target: target = usage * (1 + headroomPercent/100)
• Rounding: CPU is rounded to 50m steps, memory to 64Mi steps
• Small changes are skipped based on:
  • relative thresholds upPercent / downPercent
  • absolute minimum change minChangeCPU / minChangeMemory
• Bounds: computed requests are clamped to configured min/max
• Limits behavior is controlled by limitsMode:
  • Unchanged: keep limits as-is
  • EqualRequests: set limits equal to requests
• Anti-thrash:
  • stabilizationWindow: the desired target must stay stable before apply
  • cooldown: skip pods shortly after a resize

• Grafana dashboard: docs/grafana/resize-operator-dashboard.json

• CRD reference: docs/crd.md