Your autoscaler is always late. It notices demand at 9:00 and starts a pod that is ready at 9:02. For two minutes you are short, every single morning.

presage forecasts what demand will be when the new pods are actually ready, and provisions for that instead. It uses TimesFM, Google's open time-series model, with no training and no data leaving your cluster.

It scales Deployments, StatefulSets, anything with a scale subresource — and Agones Fleets.

---

Does it actually work?

Here is presage against a conventional autoscaler on two weeks of simulated game-server traffic. Dashed white is the capacity actually needed; the coloured lines are what each strategy provisioned.

Both track the shape. The difference is timing — and timing is the whole game:

	average replicas	short on
conventional autoscaler	11.5	1.5% of the time
presage (TimesFM)	12.2	0.2% of the time

Spending the same money on a bigger buffer instead would leave you short 0.8% of the time — four times worse than presage for the same cost.

Numbers from make backtest-charts, run against a reproducible synthetic signal. Run it on your metrics before believing any of it — that is exactly what the tool is for, and it will tell you if forecasting buys you nothing.

Try it

helm install presage oci://ghcr.io/breezycourses/charts/presage \
  --namespace presage-system --create-namespace

Then point it at something:

apiVersion: scaling.presage.sh/v1alpha1
kind: PredictiveScaler
metadata:
  name: api
spec:
  mode: Shadow                    # watch it before you trust it
  scaleTargetRef: {apiVersion: apps/v1, kind: Deployment, name: api}
  signal:
    prometheus:
      address: http://prometheus-operated.monitoring:9090
      query: sum(rate(http_requests_total{app="api"}[5m]))
  leadTime: {source: Static, static: 90s}
  capacity: {perReplica: "250"}
  policy: {minReplicas: 2, maxReplicas: 50}

kubectl get predictivescalers

NAME   MODE     TARGET   CURRENT   RECOMMENDED   READY
api    Shadow   api      4         7             True

Every scaler starts in Shadow mode: it publishes what it would do and changes nothing. Compare it against reality for a couple of weeks, then flip to Enforce.

Three things worth knowing

It cannot starve your workload. A conventional reactive calculation runs alongside the forecast and sets a floor. A badly wrong forecast can only ever give you too many replicas, never too few.

Uncertainty buys capacity, not hesitation. Capacity is sized to a high quantile of the forecast, so when the model is unsure you get more headroom rather than a paralysed autoscaler.

A presage outage is a non-event for Agones. presage never writes Fleet replicas; it answers the FleetAutoscaler webhook and returns an error whenever it cannot answer honestly, so Agones falls straight through to a plain buffer policy.

Documentation

Getting started	install, shadow, evaluate, enforce
How it decides	forecast → replica count, and the safety properties
Backtesting	score it against your own history before adopting
Agones	Fleet autoscaling and the fallback that makes it safe
Architecture	the pieces and why they are separate
Operations	metrics, alerts, "why did it pick that number"
Comparison	versus HPA, KEDA, PredictKube, Agones Schedule
API reference	generated from the CRDs
Examples	Deployment, Agones Fleet, multi-signal

Status

Alpha. The API is v1alpha1 and will change. What it does not do yet:

• No covariates — scheduled events and releases are not fed to the model, and that is where the biggest remaining gain is.
• Never run on a production cluster. It scales a real workload on a real cluster in CI, but nobody's players depend on it yet.
• No KEDA or HPA adapter. The scale subresource covers most of that ground.

On a clean weekly curve, a plain seasonal baseline may match TimesFM — presage ships that baseline in-process and the backtest will tell you which one wins on your workload. Use the cheaper one if it ties.

Contributing

Issues and pull requests welcome. See CONTRIBUTING.md, SECURITY.md for private vulnerability reporting, and GOVERNANCE.md for how the project is run.

Licence

Apache-2.0. TimesFM is a separate Apache-2.0 project and is not vendored here; the checkpoint is pulled at runtime from a pinned revision.