Predictive Autoscaling

Reactive autoscaling waits for the metric to move before it adds replicas, which costs at least one pod start-up time of headroom. Predictive autoscaling closes part of that gap: it projects the metric trend of the observation window forward and, in Auto mode, uses the projection as a floor for scale-up. The reactive decision is never replaced, and a projection never scales a workload down.

The feature is opt-in through spec.predictive. When the block is absent nothing in this page applies. Today the block is accepted and validated only: applying it writes no status.predictive and changes no replica count. The controller that consumes these fields lands in a follow-up.

Modes

mode	Behaviour
Preview	The controller computes the projection and records it in status.predictive, but the replica decision stays exactly as it is without the block. Use it to watch what the projection would have asked for before trusting it. This is the default when mode is omitted.
Auto	The projection is converted into a replica count and used as a floor: final = max(reactive, projected). It can raise a scale-up, never trigger or lower a scale-down.

spec:
  predictive:
    mode: Auto
    horizonSeconds: 120

Status

status.predictive reports the last evaluation:

Field	Meaning
mode	Effective mode of the evaluation, after defaulting to Preview.
metric	targetMetric of the source the evaluation was derived from.
observedValue	Mean of the samples used by the fit, as a decimal string in the same unit as the metric sample and targetValue.
predictedValue	Value of the fitted line at now + horizonSeconds, as a decimal string in the same unit as the metric sample and targetValue.
predictedReplicas	Replica count the projection alone asks for.
reactiveReplicas	Replica count the reactive path selected in the same round, before the predictive floor.
wouldBeReplicas	Replica count the same round would apply in Auto, after the same caps, bounds and cooldown. In Auto it equals the applied count; in Preview it shows whether Auto would change the decision.
lastUpdated	When the prediction was computed.

The block is removed when spec.predictive is removed.

Example

apiVersion: autoscaling.aibrix.ai/v1alpha1
kind: PodAutoscaler
metadata:
  name: deepseek-r1-distill-llama-8b-predictive
  namespace: default
  labels:
    app.kubernetes.io/name: aibrix
    app.kubernetes.io/managed-by: kustomize
  annotations:
    autoscaling.aibrix.ai/scale-down-cooldown-window: 3m
spec:
  scalingStrategy: KPA
  minReplicas: 1
  maxReplicas: 8
  observeWindowSeconds: 180
  panicWindowSeconds: 60
  predictive:
    mode: Preview
    horizonSeconds: 120
  metricsSources:
    - metricSourceType: pod
      protocolType: http
      port: '8000'
      path: metrics
      targetMetric: gpu_cache_usage_perc
      targetValue: '0.5'
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: deepseek-r1-distill-llama-8b

Limitations

• Predictive scaling is not supported with scalingStrategy: HPA. The HPA strategy delegates the decision to a native HorizontalPodAutoscaler resource, which has no channel to receive the projection, so the combination is rejected by the validating webhook.

• The projection uses the observeWindowSeconds window even when KPA is in panic mode. The panic window stays a reactive signal, and the projection is deliberately the slower of the two.

• The projection is linear. Traffic with a strong daily shape still benefits from schedules, which set bounds rather than a trend.