# The custom external scaler itself (built by CI -> ghcr). KEDA dials it over gRPC.
# The "http" port is the separate observability server (health.go/server.go):
# /healthz, /readyz, and /metrics -- including scaler_signal_age_seconds, the
# per-decision staleness gauge (see docs/autoscaling-signal-staleness.md). It
# needs its own ServiceMonitor (below) to actually be scraped.
apiVersion: apps/v1
kind: Deployment
metadata:
  name: inference-scaler
  namespace: inference
  labels:
    app: inference-scaler
spec:
  replicas: 1
  selector:
    matchLabels:
      app: inference-scaler
  template:
    metadata:
      labels:
        app: inference-scaler
    spec:
      containers:
        - name: scaler
          image: ghcr.io/kornsour/keda-inference-scaler:latest   # produced by .github/workflows/ci.yml
          ports:
            - name: grpc
              containerPort: 6000
            - name: http
              containerPort: 8080
          readinessProbe:
            httpGet:
              path: /readyz
              port: http
            initialDelaySeconds: 5
            periodSeconds: 10
          livenessProbe:
            httpGet:
              path: /healthz
              port: http
            initialDelaySeconds: 5
            periodSeconds: 10
          resources:
            requests:
              cpu: 25m
              memory: 32Mi
            limits:
              cpu: 200m
              memory: 128Mi
---
apiVersion: v1
kind: Service
metadata:
  name: inference-scaler
  namespace: inference
spec:
  selector:
    app: inference-scaler
  ports:
    - name: grpc
      port: 6000
      targetPort: 6000
    - name: http
      port: 8080
      targetPort: 8080
---
# Scrape the scaler's own self-observability, not just the vLLM signals it
# reads. interval matches this deployment's streamPollInterval/pollingInterval
# (15s, see scaledobject-external.yaml) -- see
# docs/autoscaling-signal-staleness.md for why that pairing was chosen. This
# is the scaler's *own* process state, so a slower or faster scrape here
# doesn't feed back into the control loop the way the vLLM PodMonitor's
# interval does; it only affects how promptly scaler_signal_age_seconds and
# friends show up on a dashboard.
apiVersion: monitoring.coreos.com/v1
kind: ServiceMonitor
metadata:
  name: inference-scaler
  namespace: inference
  labels:
    release: kube-prometheus-stack   # match your Prometheus' serviceMonitorSelector
spec:
  selector:
    matchLabels:
      app: inference-scaler
  endpoints:
    - port: http
      path: /metrics
      interval: 15s