# KEDA ScaledObject — the centerpiece, validated locally against the mock.
#
# Scales on INFERENCE-AWARE signals from Prometheus, not CPU:
#   1. queue depth  (vllm:num_requests_waiting) — requests waiting for a slot
#   2. KV-cache util (vllm:gpu_cache_usage_perc) — decode memory pressure
#
# These are the exact metric names a real vLLM emits, so this object is reused
# unchanged when the GPU InferenceService replaces the mock (see ../../k8s/).
apiVersion: keda.sh/v1alpha1
kind: ScaledObject
metadata:
  name: mock-vllm-scaler
  namespace: inference
spec:
  scaleTargetRef:
    name: mock-vllm
  minReplicaCount: 1           # min 1, not 0: KEDA can't wake a scaled-to-zero
  maxReplicaCount: 5           # Deployment on a queue metric (no pod = no metric);
                               # scale-to-zero would need an HTTP activator.
  advanced:
    horizontalPodAutoscalerConfig:
      behavior:
        scaleDown:
          # Don't thrash: wait for sustained calm before scaling back in.
          stabilizationWindowSeconds: 60
  triggers:
    - type: prometheus
      metadata:
        serverAddress: http://kube-prometheus-stack-prometheus.monitoring.svc:9090
        metricName: requests_waiting
        # Total queued across all replicas; >3 sustained = scale out.
        query: sum(vllm:num_requests_waiting)
        threshold: "3"
    - type: prometheus
      metadata:
        serverAddress: http://kube-prometheus-stack-prometheus.monitoring.svc:9090
        metricName: kv_cache_util
        # Hottest replica's KV-cache usage; past 70% we're near the concurrency
        # ceiling, so add capacity before TTFT degrades.
        query: max(vllm:gpu_cache_usage_perc)
        threshold: "0.7"