KEDA autoscaling

make keda-demo-up    # un-pause the ScaledObject (KEDA starts driving replicas)
make keda-demo-down  # re-pause (KEDA stops; return to manual make vllm-up/down)

# Restore the cluster first: needs a live GPU + Prometheus.
make vllm-up                 # 1 warm replica
make keda-demo-up            # un-pause autoscaling

kubectl -n serving get scaledobject raw-vllm                 # READY=True, ACTIVE flips on load
kubectl -n serving get hpa keda-hpa-raw-vllm -w              # KEDA-generated HPA; watch TARGETS

# Drive queue depth (concurrent requests > what 1 replica drains):
make bench                   # or: hey/vegeta against the gateway with high concurrency

# Observe: vllm:num_requests_waiting climbs → HPA desired rises (capped at 1 here by quota).
kubectl -n serving describe scaledobject raw-vllm | sed -n '/Triggers/,$p'

make keda-demo-down          # re-pause
make vllm-down               # release the GPU node ($0 idle)

# Sim backends must be running:
kubectl -n inference scale deploy/vllm-sim --replicas=1

# Land sim metrics in cluster Prometheus:
kubectl -n inference apply -f - <<'EOF'
apiVersion: monitoring.coreos.com/v1
kind: ServiceMonitor
metadata:
  name: vllm-sim
  namespace: inference
spec:
  selector:
    matchLabels:
      app: vllm-sim
  endpoints:
    - port: http
      path: /metrics
      interval: 15s
EOF
# (requires a Service selecting app=vllm-sim on port `http`; add one if the demo lacks it.)

# Autoscale the sim on queue depth:
kubectl apply -f - <<'EOF'
apiVersion: keda.sh/v1alpha1
kind: ScaledObject
metadata:
  name: vllm-sim
  namespace: inference
spec:
  scaleTargetRef:
    name: vllm-sim
  pollingInterval: 15
  cooldownPeriod: 60
  minReplicaCount: 1
  maxReplicaCount: 5
  triggers:
    - type: prometheus
      metadata:
        serverAddress: http://prometheus-operated.monitoring.svc:9090
        metricName: vllm_num_requests_waiting
        query: sum(vllm:num_requests_waiting{namespace="inference"})
        threshold: "5"
EOF

# Drive load through the gateway, watch the sim scale 1→N→min:
kubectl -n inference get hpa keda-hpa-vllm-sim -w

# Cleanup:
kubectl -n inference delete scaledobject vllm-sim
kubectl -n inference delete servicemonitor vllm-sim