resources:
  requests:
    cpu: "2000m"        # 保障基础转发能力，建议≥2核
    memory: "4Gi"       # 缓冲区、转码缓存所需内存
    nvidia.com/gpu: 1   # 若启用硬件转码（可选）
  limits:
    cpu: "4000m"        # 允许突发抢占，防止噪声邻居
    memory: "8Gi"

topologySpreadConstraints:
  - maxSkew: 1
    topologyKey: topology.kubernetes.io/zone
    whenUnsatisfiable: DoNotSchedule
    labelSelector:
      matchLabels:
        app: media-server

apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: media-server-hpa
  namespace: meeting
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: media-server
  minReplicas: 3
  maxReplicas: 100
  behavior:
    scaleDown:
      stabilizationWindowSeconds: 600   # 缩容冷却10分钟，防抖
      policies:
        - type: Percent
          value: 10
          periodSeconds: 60
    scaleUp:
      stabilizationWindowSeconds: 60    # 扩容响应1分钟
      policies:
        - type: Percent
          value: 50
          periodSeconds: 30
        - type: Pods
          value: 10
          periodSeconds: 30
      selectPolicy: Max
  metrics:
    # 指标1：并发流数（核心业务指标）
    - type: Pods
      pods:
        metric:
          name: media_active_streams
        target:
          type: AverageValue
          averageValue: "200"
    # 指标2：带宽利用率（网络瓶颈保护）
    - type: Pods
      pods:
        metric:
          name: media_bandwidth_usage_mbps
        target:
          type: AverageValue
          averageValue: "800"
    # 指标3：CPU负载（兜底资源指标）
    - type: Resource
      resource:
        name: cpu
        target:
          type: Utilization
          averageUtilization: 70

apiVersion: autoscaling.k8s.io/v1
kind: VerticalPodAutoscaler
metadata:
  name: media-server-vpa
spec:
  targetRef:
    kind: Deployment
    name: media-server
  updatePolicy:
    updateMode: "Initial"  # 仅在Pod创建时应用建议，避免运行中重启抖动
  resourcePolicy:
    containerPolicies:
      - containerName: media-server
        controlledResources: ["cpu", "memory"]
        minAllowed:
          cpu: "1000m"
          memory: "2Gi"
        maxAllowed:
          cpu: "8000m"
          memory: "16Gi"

livenessProbe:
  httpGet:
    path: /healthz
    port: 8080
  initialDelaySeconds: 15
  periodSeconds: 10
readinessProbe:
  httpGet:
    path: /ready
    port: 8080
  initialDelaySeconds: 5
  periodSeconds: 5
  failureThreshold: 3
# 关键：启动探针保护长启动应用
startupProbe:
  httpGet:
    path: /started
    port: 8080
  failureThreshold: 30
  periodSeconds: 10

apiVersion: karpenter.sh/v1beta1
kind: NodePool
metadata:
  name: media-pool
spec:
  template:
    spec:
      requirements:
        - key: karpenter.sh/capacity-type
          operator: In
          values: ["spot", "on-demand"]
        - key: node.kubernetes.io/instance-type
          operator: In
          values: ["ecs.g7.2xlarge", "ecs.c7.2xlarge"] # 高网络PPS规格
        - key: topology.kubernetes.io/zone
          operator: In
          values: ["cn-hangzhou-i", "cn-hangzhou-j"]
      nodeClassRef:
        name: media-nodeclass
  limits:
    cpu: "2000"  # 集群总CPU上限，防止失控
  disruption:
    consolidationPolicy: WhenEmptyOrUnderutilized
    consolidateAfter: 300s
---
apiVersion: karpenter.k8s.aws/v1beta1
kind: EC2NodeClass
metadata:
  name: media-nodeclass
spec:
  amiFamily: AL2
  subnetSelectorTerms:
    - tags:
        Name: "media-subnet-*"
  securityGroupSelectorTerms:
    - tags:
        Name: "media-sg"
  tags:
    Environment: production
    Team: media-infra

apiVersion: v1
kind: Service
metadata:
  name: media-server-lb
  annotations:
    service.beta.kubernetes.io/aws-load-balancer-type: "nlb"
    service.beta.kubernetes.io/aws-load-balancer-cross-zone-load-balancing-enabled: "true"
spec:
  type: LoadBalancer
  externalTrafficPolicy: Local
  ports:
    - name: rtp
      port: 10000
      targetPort: 10000
      protocol: UDP
    - name: rtcp
      port: 10001
      targetPort: 10001
      protocol: UDP
  selector:
    app: media-server

# Kubelet 配置关键参数
cpuManagerPolicy: "static"           # 独占整核，避免CFS调度抖动
cpuManagerReconcilePeriod: "10s"
topologyManagerPolicy: "single-numa-node"  # 严格单NUMA节点分配
topologyManagerScope: "container"
kubeReserved:
  cpu: "1000m"
  memory: "2Gi"
  hugepages-1Gi: "2Gi"               # 预留大页内存供DPDK/媒体引擎使用
systemReserved:
  cpu: "500m"
  memory: "1Gi"

resources:
  requests:
    cpu: "4"           # 必须为整数，static策略下仅支持整核独占
    memory: "8Gi"
    hugepages-1Gi: "2Gi"  # 显式申请大页
  limits:
    cpu: "4"
    memory: "8Gi"
    hugepages-1Gi: "2Gi"

apiVersion: "k8s.cni.cncf.io/v1"
kind: NetworkAttachmentDefinition
metadata:
  name: sriov-media-net
  namespace: meeting
spec:
  config: '{
    "cniVersion": "0.3.1",
    "type": "sriov",
    "name": "sriov-media-net",
    "ipam": {"type": "host-local", "subnet": "10.200.0.0/16"},
    "deviceType": "vfio-pci",  # DPDK模式用 vfio-pci，内核模式用 netdevice
    "vlan": 100,
    "spoofchk": "off",
    "trust": "on"
  }'
---
# Pod 注解引用
annotations:
  k8s.v1.cni.cncf.io/networks: '[
    {"name": "sriov-media-net", "interface": "eth1", "ips": ["10.200.1.5/16"]},
    {"name": "calico", "interface": "eth0", "default-route": ["10.0.0.1"]}
  ]'

sequenceDiagram
    participant K8s as Kubernetes
    participant Media as Media Pod
    participant Signal as Signaling Server
    participant Client as Client SDK

    K8s->>Media: 发送 SIGTERM (preStop Hook 触发)
    Media->>Signal: HTTP POST /admin/drain {node_id, grace_period: 60s}
    Signal->>Signal: 标记节点状态=DRAINING, 停止新分流
    Signal->>Client: 推送 SessionUpdate {new_media_node, migrate_token}
    Client->>NewMedia: 建立新ICE/DTLS连接 (并行)
    Client->>Media: 发送 BYE / 切换流
    Media->>Signal: 确认会话数归零
    Signal->>Media: 返回 200 OK (允许销毁)
    Media->>K8s: 进程退出

#!/bin/bash
set -e
GRACE_PERIOD=${TERMINATION_GRACE_PERIOD:-60}
SIGNALING_URL="http://signaling:8080"
NODE_ID=$(hostname)

echo "[$(date)] Initiating drain for $NODE_ID, grace period: ${GRACE_PERIOD}s"

# 1. 通知信令层进入排空状态
curl -s -X POST "${SIGNALING_URL}/admin/nodes/${NODE_ID}/drain" 
  -H "Content-Type: application/json" 
  -d "{"grace_period": ${GRACE_PERIOD}}" || true

# 2. 轮询等待会话清零 (或超时)
SECONDS=0
while [ $SECONDS -lt $GRACE_PERIOD ]; do
  ACTIVE=$(curl -s "${SIGNALING_URL}/admin/nodes/${NODE_ID}/stats" | jq -r '.active_sessions // 0')
  if [ "$ACTIVE" -eq 0 ]; then
    echo "[$(date)] All sessions migrated. Safe to terminate."
    exit 0
  fi
  echo "[$(date)] Waiting for ${ACTIVE} sessions to migrate... (${SECONDS}s elapsed)"
  sleep 5
done

echo "[$(date)] Grace period exceeded. Forcing shutdown with ${ACTIVE} sessions remaining."
exit 1 # 非0退出码触发K8s强制Kill，但此时已尽力迁移

lifecycle:
  preStop:
    exec:
      command: ["/bin/bash", "/usr/local/bin/graceful-shutdown.sh"]
terminationGracePeriodSeconds: 90  # 必须 > 脚本最大执行时间

apiVersion: scheduling.k8s.io/v1
kind: PriorityClass
metadata:
  name: media-critical
value: 1000000
globalDefault: false
preemptionPolicy: PreemptLowerPriority
description: "Core media forwarding, never preempted"
---
apiVersion: scheduling.k8s.io/v1
kind: PriorityClass
metadata:
  name: media-batch
value: 10000
globalDefault: false
preemptionPolicy: Never
description: "Recording/Transcoding batch jobs, can be preempted"
---
apiVersion: scheduling.k8s.io/v1
kind: PriorityClass
metadata:
  name: spot-best-effort
value: 100
globalDefault: true
preemptionPolicy: Never
description: "Stateless auxiliary services on Spot"

requirements:
  - key: "karpenter.k8s.aws/instance-category"
    operator: In
    values: ["c", "m", "r"]  # 计算/内存/通用型均可
  - key: "karpenter.k8s.aws/instance-generation"
    operator: In
    values: ["6", "7", "8"]  # 跨代际规避库存风险
  - key: "topology.kubernetes.io/zone"
    operator: In
    values: ["cn-hangzhou-i", "cn-hangzhou-j", "cn-hangzhou-k"]
  - key: "karpenter.sh/capacity-type"
    operator: In
    values: ["spot", "on-demand"]  # 优先Spot，无库存自动回落On-Demand
limits:
  cpu: "5000"
disruption:
  consolidationPolicy: WhenEmptyOrUnderutilized
  consolidateAfter: 60s
  # 关键：设置过期时间，强制轮换老旧实例规格
  expireAfter: 720h  # 30天强制重建，规避硬件老化/安全补丁积压

# .gitlab-ci.yml 片段
chaos_validation:
  stage: validate
  image: chaosiq/chaostoolkit:latest
  script:
    - chaos run experiment.yaml --settings=env.yaml
    - |
      # 解析实验报告，阻断发布
      if [ $(jq '.status' report.json) != "completed" ]; then
        echo "Chaos experiment failed!"
        exit 1
      fi
  artifacts:
    reports:
      junit: report.xml
    when: always
  rules:
    - if: $CI_COMMIT_BRANCH == "main" || $CI_PIPELINE_SOURCE == "merge_request_event"

infra/
├── base/
│   ├── deployment.yaml       # 通用模板 (资源/探针/安全上下文)
│   ├── hpa.yaml              # HPA 模板 (参数化: {{.Values.hpa.targetStreams}})
│   ├── pdb.yaml
│   ├── servicemonitor.yaml   # Prometheus 采集规则
│   └── kustomization.yaml
├── overlays/
│   ├── dev/
│   │   ├── kustomization.yaml  # replicas=2, spotOnly=true
│   │   └── values.yaml
│   ├── staging/
│   │   ├── kustomization.yaml  # replicas=10, mixed
│   │   └── values.yaml
│   └── prod/
│       ├── kustomization.yaml  # replicas=50, multi-AZ, on-demand baseline
│       ├── values.yaml
│       └── hpa-patch.yaml      # 生产特有: 更保守的缩容窗口
├── clusters/
│   ├── cn-hangzhou/            # 集群级资源: NodePool, PriorityClass, NetworkPolicy
│   └── cn-shanghai/
└── apps/
    └── media-server/           # ArgoCD Application 定义
        ├── application.yaml
        └── project.yaml

apiVersion: argoproj.io/v1alpha1
kind: AnalysisTemplate
metadata:
  name: media-server-canary-analysis
spec:
  args:
    - name: service-name
    - name: namespace
  metrics:
    - name: success-rate
      interval: 1m
      count: 10  # 观测10分钟
      successCondition: result[0] >= 0.995
      failureLimit: 3
      provider:
        prometheus:
          address: http://prometheus.monitoring:9090
          query: |
            sum(rate(media_signaling_join_success_total{service="{{args.service-name}}",namespace="{{args.namespace}}"}[1m]))
            /
            sum(rate(media_signaling_join_total{service="{{args.service-name}}",namespace="{{args.namespace}}"}[1m]))
    - name: p99-latency
      successCondition: result[0] <= 300  # 300ms
      provider:
        prometheus:
          query: |
            histogram_quantile(0.99, sum(rate(media_ice_connection_duration_seconds_bucket{service="{{args.service-name}}",namespace="{{args.namespace}}"}[1m])) by (le))
---
# Rollout 策略引用
apiVersion: argoproj.io/v1alpha1
kind: Rollout
spec:
  strategy:
    canary:
      canaryMetadata:
        labels:
          role: canary
      stableMetadata:
        labels:
          role: stable
      steps:
        - setWeight: 10
        - pause: {duration: 5m}
        - analysis:
            templates:
              - templateName: media-server-canary-analysis
            args:
              - name: service-name
                value: media-server-canary
              - name: namespace
                value: meeting
        - setWeight: 50
        - pause: {duration: 10m}
        - analysis:
            templates:
              - templateName: media-server-canary-analysis
        - setWeight: 100