#!/bin/bash
# 用 Prometheus 常数 25 走一遍扩容，然后改回真实的 vLLM 排队指标。
# 不向 vLLM 发推理请求。结束时 ScaledObject 回到仓库里的那份配置。
set -u
NS=dify
ORIG=/tmp/keda-sim-orig.yaml
FAKE=/tmp/keda-sim-fake.yaml
BACK=/tmp/keda-sim-back.yaml

cat > "$FAKE" <<'EOF'
apiVersion: keda.sh/v1alpha1
kind: ScaledObject
metadata:
  name: dify-worker
  namespace: dify
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: worker
  minReplicaCount: 1
  maxReplicaCount: 3
  pollingInterval: 15
  cooldownPeriod: 60
  advanced:
    horizontalPodAutoscalerConfig:
      behavior:
        scaleUp:
          stabilizationWindowSeconds: 0
          policies:
            - type: Pods
              value: 3
              periodSeconds: 15
        scaleDown:
          stabilizationWindowSeconds: 0
          policies:
            - type: Percent
              value: 100
              periodSeconds: 15
  triggers:
    - type: prometheus
      metadata:
        serverAddress: http://prometheus.observability.svc.cluster.local:9090
        metricName: vllm_num_requests_waiting
        query: vector(25)
        threshold: "10"
        activationThreshold: "10"
EOF

cat > "$BACK" <<'EOF'
apiVersion: keda.sh/v1alpha1
kind: ScaledObject
metadata:
  name: dify-worker
  namespace: dify
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: worker
  minReplicaCount: 1
  maxReplicaCount: 3
  pollingInterval: 15
  cooldownPeriod: 60
  advanced:
    horizontalPodAutoscalerConfig:
      behavior:
        scaleUp:
          stabilizationWindowSeconds: 0
          policies:
            - type: Pods
              value: 3
              periodSeconds: 15
        scaleDown:
          stabilizationWindowSeconds: 0
          policies:
            - type: Percent
              value: 100
              periodSeconds: 15
  triggers:
    - type: prometheus
      metadata:
        serverAddress: http://prometheus.observability.svc.cluster.local:9090
        metricName: vllm_num_requests_waiting
        query: 'sum(vllm:num_requests_waiting{model_name="qwen3.6-27b"})'
        threshold: "10"
        activationThreshold: "10"
EOF

finish() {
  echo
  echo "========== 改回真实排队指标 =========="
  kubectl apply -f "$BACK"
  for i in $(seq 1 24); do
    spec=$(kubectl -n "$NS" get deploy worker -o jsonpath='{.spec.replicas}')
    ready=$(kubectl -n "$NS" get deploy worker -o jsonpath='{.status.readyReplicas}')
    echo "down t=${i} spec=${spec:-empty} ready=${ready:-empty}"
    if [ "${spec:-}" = "1" ] && [ "${ready:-}" = "1" ]; then
      break
    fi
    sleep 10
  done
  kubectl apply -f "$ORIG"
  rm -f "$FAKE" "$BACK" "$ORIG"
  echo
  echo "========== 结束后 =========="
  kubectl -n "$NS" get scaledobject dify-worker
  kubectl -n "$NS" get hpa keda-hpa-dify-worker
  kubectl -n "$NS" get pods -l app=worker -o wide
  echo -n "vLLM /health "
  curl -s -o /dev/null -w '%{http_code}\n' --max-time 5 http://127.0.0.1:8000/health || echo fail
}
trap finish EXIT

echo "========== 模拟前 =========="
kubectl -n "$NS" get scaledobject dify-worker
kubectl -n "$NS" get hpa keda-hpa-dify-worker
kubectl -n "$NS" get pods -l app=worker -o wide
echo -n "vLLM /health "
curl -s -o /dev/null -w '%{http_code}\n' --max-time 5 http://127.0.0.1:8000/health || echo fail

echo
echo "========== 把指标临时改成 25 =========="
kubectl apply -f "$FAKE"

up_ok=0
for i in $(seq 1 24); do
  spec=$(kubectl -n "$NS" get deploy worker -o jsonpath='{.spec.replicas}')
  ready=$(kubectl -n "$NS" get deploy worker -o jsonpath='{.status.readyReplicas}')
  active=$(kubectl -n "$NS" get scaledobject dify-worker -o jsonpath='{.status.conditions[?(@.type=="Active")].status}')
  echo "up t=${i} spec=${spec:-empty} ready=${ready:-empty} active=${active:-empty}"
  kubectl -n "$NS" get hpa keda-hpa-dify-worker
  if [ "${ready:-}" = "3" ]; then
    up_ok=1
    break
  fi
  sleep 10
done

echo
echo "========== 扩容后的 Pod =========="
kubectl -n "$NS" get pods -l app=worker -o wide
kubectl -n "$NS" describe scaledobject dify-worker | sed -n '/Conditions:/,/Events:/p'

if [ "$up_ok" != "1" ]; then
  echo "worker 没有在预期时间内变成 3 个就绪副本"
fi