apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: vllm-hpa
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: vllm-llm
  minReplicas: 2
  maxReplicas: 10
  metrics:
  - type: Resource
    resource:
      name: cpu
      target:
        type: Utilization
        averageUtilization: 70

apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: vllm-hpa-custom
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: vllm-llm
  minReplicas: 2
  maxReplicas: 10
  metrics:
  # GPU 利用率指标
  - type: Pods
    pods:
      metric:
        name: gpu_utilization_percent
      target:
        type: AverageValue
        averageValue: "80"
  # 请求队列长度
  - type: Pods
    pods:
      metric:
        name: vllm_queue_length
      target:
        type: AverageValue
        averageValue: "16"

apiVersion: keda.sh/v1alpha1
kind: ScaledObject
metadata:
  name: vllm-scaling
  namespace: llm-serving
spec:
  scaleTargetRef:
    name: vllm-llm
  pollingInterval: 15      # 每 15 秒检查一次指标
  cooldownPeriod: 300      # 缩容冷却 5 分钟（避免抖动）
  minReplicaCount: 2       # 始终保留 2 个副本
  maxReplicaCount: 20
  advanced:
    horizontalPodAutoscalerConfig:
      behavior:
        scaleUp:
          stabilizationWindowSeconds: 60  # 扩缩稳定窗口
          policies:
          - type: Pods
            value: 2        # 每次最多加 2 个 Pod
            periodSeconds: 60
        scaleDown:
          stabilizationWindowSeconds: 600  # 缩容更保守（10 分钟）
          policies:
          - type: Pods
            value: 1        # 每次只缩 1 个 Pod
            periodSeconds: 120
  triggers:
  # Prometheus 指标触发
  - type: prometheus
    metadata:
      serverAddress: http://prometheus.monitoring:9090
      query: avg_over_time(vllm_num_requests_waiting{namespace="llm-serving"}[2m])
      threshold: "10"       # 排队请求超过 10 个开始扩容
  # GPU 利用率触发
  - type: prometheus
    metadata:
      serverAddress: http://prometheus.monitoring:9090
      query: avg(DCGM_FI_DEV_GPU_UTIL{namespace="llm-serving"})
      threshold: "85"       # GPU 利用率超过 85% 扩容

# 自定义 Metrics Adapter 示例（Prometheus Adapter 配置）
rules:
  - seriesQuery: 'vllm_gpu_cache_usage_ratio{namespace!="",pod!=""}'
    resources:
      overrides:
        namespace: {resource: "namespace"}
        pod: {resource: "pod"}
    name:
      matches: "vllm_gpu_cache_usage_ratio"
    metricsQuery: 'avg by(pod) (rate(vllm_gpu_cache_usage_ratio[2m]))'

# Warm Pool 实现：使用 KEDA 保持最小副本 + Cluster Autoscaler 预 provision
# 策略：始终保持 minReplicaCount = 2
# 即使用户请求为 0，也保留 2 个已加载模型的 Pod

# 预测性扩容实现：KEDA Cron Scaler
  triggers:
  - type: cron
    metadata:
      timezone: Asia/Shanghai
      start: 0 9 * * *     # 每天 9:00 开始
      end: 0 10 * * *      # 每天 10:00 结束
      desiredReplicas: "8"  # 早高峰扩容到 8 个

spec:
  template:
    spec:
      terminationGracePeriodSeconds: 120  # 给 2 分钟完成优雅退出
      containers:
      - name: vllm
        lifecycle:
          preStop:
            exec:
              command:
              - /bin/sh
              - -c
              - |
                # 1. 通知 LB 停止分发流量
                sleep 10
                # 2. 等待正在处理的请求完成
                # 3. 清理 GPU 资源
                curl -X POST http://localhost:8000/shutdown