Pre-scale before
the spike
ScalePilot is a cloud-agnostic Kubernetes autoscaling operator. It trains ARIMA and Holt-Winters models on your Prometheus history and patches HPA minReplicas minutes before traffic arrives while enforcing FinOps cost budgets and multi-cluster spillover.
helm install scalepilot charts/scalepilot --namespace scalepilot-system --create-namespaceThree things Kubernetes HPA can't do
ScalePilot adds proactive intelligence, geographic reach, and cost discipline to your existing autoscaling stack.
Train ARIMA or Holt-Winters models on your Prometheus history. ScalePilot pre-scales HPA minReplicas minutes before traffic spikes arrive, so your pods are warm and ready — not catching up.
- ARIMA(p,d,q) with Yule-Walker estimation
- Holt-Winters triple exponential smoothing
- 95% confidence intervals for conservative scaling
- SLO-aware ScaleUpGuard gate
When a queue depth or load metric exceeds your threshold, ScalePilot automatically spills workloads to overflow clusters — ordered by priority, capped by maxCapacity, with a cooldown to prevent thrashing.
- Metric-driven spillover (any PromQL query)
- Priority-ordered overflow cluster selection
- Automatic health-checking every 30s
- Server-side apply for idempotent cross-cluster writes
Set a monthly cost ceiling per namespace. ScalePilot polls AWS, GCP, or Azure billing APIs, computes real spend, and enforces Delay / Downgrade / Block actions automatically when you breach the limit.
- AWS Cost Explorer, GCP Billing, Azure Cost Management
- Millidollar precision avoids float rounding
- Slack & PagerDuty breach notifications
- Warning threshold before hard ceiling
Pre-scale before the spike, not after
HPA sees a CPU spike and adds pods — but those pods take 2–5 minutes to become ready. ScalePilot predicts the spike and pre-warms pods before it hits.
Collect History
ForecastPolicy queries Prometheus for up to 7 days of metric history (request rate, queue depth, CPU — any PromQL scalar).
Train Model
A background goroutine trains ARIMA(p,d,q) via Yule-Walker estimation or Holt-Winters triple exponential smoothing. The model is cached in a ConfigMap.
Predict & Pre-Scale
Every reconcile cycle loads the cached model, predicts the next leadTimeMinutes, and patches HPA minReplicas — pods start warming up before traffic arrives.
Govern & Enforce
ClusterScaleProfile enforces blackout windows and surge limits. ScalingBudget checks cloud billing and blocks, delays, or downgrades scaling if costs exceed the ceiling.
How ScalePilot compares
ScalePilot is the only open-source operator that combines proactive forecasting, geographic federation, and FinOps enforcement in a single controller.
| Capability | HPA | KEDA | Kubecost | Admiralty | ScalePilot |
|---|---|---|---|---|---|
| Reactive scaling (metric-based) | ✓ | ✓ | — | — | ✓ |
| Predictive scaling (time-series forecast) | — | — | — | — | ✓ |
| Multi-cluster workload spillover | — | — | — | Placement-based | Metric-driven |
| Cost observability | — | — | ✓ | — | ✓ |
| Cost enforcement (block/delay scaling) | — | — | — | — | ✓ |
| Blackout windows (cron-based) | — | — | — | — | ✓ |
| Per-team scaling governance | — | — | — | — | ✓ |
Ready to stop reacting to spikes?
Install ScalePilot in under 5 minutes with Helm and start forecasting your first workload.