Usage Guide

Once PPA is installed, you can enable proactive scaling for your applications by defining a Model resource and configuring KEDA to utilize it.

1. Define a Predictive Model

The Model Custom Resource triggers the automated pipeline that fetches historical metrics and trains a Prophet model for your workload.

apiVersion: ppa.io/v1alpha1
kind: Model
metadata:
  name: my-app-cpu-prediction
  namespace: my-apps
spec:
  prometheusUrl: "http://prometheus-operated.monitoring:9090"
  queryType: "cpu" # Options: "cpu" | "memory"
  workload:
    type: "deployment"
    name: "my-app"

Automation Workflow

When you apply this resource, the following sequence occurs automatically: 1. Orchestration: Kro detects the new Model. 2. Initial Training: An Argo Workflow trains the baseline model. 3. Scheduled Training: A CronWorkflow ensures the model is retrained hourly. 4. Inference Service: A KServe InferenceService is deployed to serve the latest model.

2. Configure KEDA ScaledObject

Link your workload to the PPA External Scaler by configuring KEDA to trigger based on the model's predictions.

apiVersion: keda.sh/v1alpha1
kind: ScaledObject
metadata:
  name: my-app-scaler
  namespace: my-apps
spec:
  scaleTargetRef:
    name: my-app
  minReplicaCount: 1
  maxReplicaCount: 10
  triggers:
    - type: external-push
      metadata:
        scalerAddress: ppa-scaler.ppa:9090
        modelName: my-app-cpu-prediction
        horizon: "5"       # Minutes to look ahead
        targetValue: "0.8"  # Target utilization (e.g., 80%)

3. Monitoring

Verify the lifecycle of your predictive models through the command line:

# Check training workflow status
argo list -n my-apps

# Check model availability and inference URL
kubectl get models my-app-cpu-prediction -o yaml