// pkg/plugins/mediatopology/mediatopology.go
package mediatopology

import (
    "context"
    "k8s.io/kubernetes/pkg/scheduler/framework"
    "k8s.io/kubernetes/pkg/scheduler/framework/plugins/names"
)

// 1. 定义插件名称与权重常量
const (
    PluginName = "MediaTopologyAffinity"
    // 网络质量权重占比 30%，需通过 ConfigMap 动态热加载
    DefaultNetworkWeight int64 = 30 
)

type MediaTopologyPlugin struct {
    handle framework.Handle
    // 注入外部网络拓扑服务客户端（gRPC/HTTP）
    topologyClient TopologyServiceClient 
    // 权重配置原子更新
    weights *atomic.Value 
}

func New(_ context.Context, _ runtime.Object, h framework.Handle) (framework.Plugin, error) {
    return &MediaTopologyPlugin{
        handle: h,
        topologyClient: NewTopologyClient(h.ClientSet()),
        weights: &atomic.Value{},
    }, nil
}

// 2. Filter 阶段：硬性拦截不满足最低网络质量的节点
func (p *MediaTopologyPlugin) Filter(ctx context.Context, state *framework.CycleState, pod *v1.Pod, nodeInfo *framework.NodeInfo) *framework.Status {
    // 仅对媒体负载生效（通过 Label/Annotation 识别）
    if !isMediaWorkload(pod) { return nil }
    
    nodeZone := nodeInfo.Node().Labels[topology.kubernetes.io/zone]
    podPreferredZone := getPreferredZoneFromAnnotation(pod) // 由上层 Controller 注入
    
    // 场景：跨 AZ 调度必须满足最低带宽阈值（如 10Gbps 可用）
    if nodeZone != podPreferredZone {
        linkQuality, err := p.topologyClient.GetLinkQuality(ctx, podPreferredZone, nodeZone)
        if err != nil || linkQuality.AvailableBandwidthGbps < 10 {
            return framework.NewStatus(framework.Unschedulable, 
                fmt.Sprintf("Cross-AZ link %s->%s insufficient bandwidth: %.2f Gbps", podPreferredZone, nodeZone, linkQuality.AvailableBandwidthGbps))
        }
    }
    return nil
}

// 3. Score 阶段：多维度精细化打分
func (p *MediaTopologyPlugin) Score(ctx context.Context, state *framework.CycleState, pod *v1.Pod, nodeName string) (int64, *framework.Status) {
    nodeInfo, _ := p.handle.SnapshotSharedLister().NodeInfos().Get(nodeName)
    weights := p.weights.Load().(WeightConfig)
    
    var totalScore int64
    
    // 维度 A: 网络链路质量评分 (0-100)
    netScore := p.scoreNetworkQuality(pod, nodeInfo.Node())
    totalScore += netScore * weights.NetworkWeight
    
    // 维度 B: 存储数据局部性评分 (0-100)
    // 查询 CSI 驱动暴露的 PV 拓扑信息 或 JuiceFS/CubeFS 元数据服务
    storageScore := p.scoreDataLocality(pod, nodeInfo.Node())
    totalScore += storageScore * weights.StorageWeight
    
    // 维度 C: 会话亲和性评分 (0-100)
    sessionScore := p.scoreSessionAffinity(pod, nodeName)
    totalScore += sessionScore * weights.SessionWeight
    
    // 归一化至框架要求的 0 - MaxNodeScore (默认 100)
    return totalScore / 100, nil
}

// 4. Permit 阶段：实现“等待网络就绪”语义
// 解决 Pod 创建快于网络控制平面下发路由/ACL 规则的竞态条件
func (p *MediaTopologyPlugin) Permit(ctx context.Context, state *framework.CycleState, pod *v1.Pod, nodeName string) (*framework.Status, time.Duration) {
    if !needWaitNetworkReady(pod) { return nil, 0 }
    
    // 向网络控制器查询目标节点 VPC/ENI/安全组 是否 Ready
    ready, timeout := p.networkController.WaitForNodeReady(ctx, nodeName, 30*time.Second)
    if !ready {
        return framework.NewStatus(framework.Wait, "Waiting for network interfaces ready"), timeout
    }
    return nil, 0
}

graph LR
    A[kube-scheduler<br/>Main Loop] -->|gRPC ScoreRequest| B(Media Scorer Service<br/>Deployment HPA)
    B --> C[(Network Topology DB<br/>Redis/Etcd/TiKV)]
    B --> D[(Storage Heatmap API<br/>CSI/Object Storage)]
    B --> E[(Cost Engine<br/>FinOps System)]
    B --> F[ML Predictor<br/>流量预测/异常检测]
    C -->|实时探测数据| G[eBPF Prober DaemonSet]

# ConfigMap: scheduler-config (通过 VolumeMount 热加载)
apiVersion: v1
kind: ConfigMap
metadata:
  name: scheduler-config
  namespace: kube-system
data:
  plugin-config.yaml: |
    apiVersion: kubescheduler.config.k8s.io/v1
    kind: KubeSchedulerConfiguration
    profiles:
    - schedulerName: media-scheduler
      pluginConfig:
      - name: MediaTopologyAffinity
        args:
          # 运维/控制器可动态修改此处权重，无需重启调度器
          networkWeight: 30
          storageWeight: 25
          costWeight: 10
          # 熔断开关
          circuitBreaker:
            enabled: true
            level: "L1" # L0/L1/L2/L3

# Node 标签示例 (由 CSI Controller 自动维护)
labels:
  topology.kubernetes.io/zone: cn-hangzhou-a
  storage.juicefs.io/metadata-zone: cn-hangzhou-a      # 元数据引擎所在 AZ
  storage.juicefs.io/hot-data-tier: "ssd"              # 本地热数据介质类型
  storage.juicefs.io/cache-capacity-gb: "2048"         # 本地缓存容量
  storage.juicefs.io/cache-hit-rate: "0.92"            # 近 5min 缓存命中率
annotations:
  storage.juicefs.io/volume-topology: '{"vol-001": "cn-hangzhou-a", "vol-002": "cn-hangzhou-b"}' # 该节点可直连的卷分布

// 调度器 Score 阶段伪代码
func (p *MediaTopologyPlugin) scoreCost(pod *v1.Pod, node *v1.Node) int64 {
    budget := p.getBudgetToken(pod.Namespace) // 剩余预算 (GB)
    estimatedCrossAZTraffic := p.estimateTraffic(pod, node) // 基于历史画像预测
    
    if estimatedCrossAZTraffic == 0 { return 100 } // 同 AZ，满分
    
    // 预算充足：轻微惩罚；预算耗尽：重度惩罚/硬拦截
    ratio := float64(estimatedCrossAZTraffic) / float64(budget)
    if ratio > 1.0 { return -1000 } // 触发 Filter 硬拦截更合理
    return int64(100 * (1 - ratio*ratio)) // 二次曲线惩罚
}

{
  "timestamp": "2024-05-20T10:00:00.123Z",
  "event": "PodScheduled",
  "pod": "live-transcode-7b9d4-fzx2k",
  "namespace": "media-prod",
  "targetNode": "cn-hangzhou-a-wkr-001",
  "scheduler": "media-scheduler",
  "plugins": {
    "MediaTopologyAffinity": {
      "phase": "Score",
      "durationMs": 12,
      "details": {
        "candidateNodes": 45,
        "scores": [
          {"node": "cn-hangzhou-a-wkr-001", "total": 92, "breakdown": {"network": 95, "storage": 90, "session": 85, "cost": 98}},
          {"node": "cn-hangzhou-b-wkr-003", "total": 68, "breakdown": {"network": 40, "storage": 30, "session": 20, "cost": 95}}
        ],
        "selectedReason": "Highest composite score: Data locality (PV in Zone A) + Low cross-AZ cost"
      }
    },
    "NodeResourcesFit": {"score": 88, "breakdown": {"cpu": 90, "memory": 85, "nvidia.com/gpu": 100}}
  },
  "circuitBreakerLevel": "L0"
}

# CRD: MediaSchedulingIntent
apiVersion: scheduling.media.io/v1alpha1
kind: MediaSchedulingIntent
metadata:
  name: live-core-intent
spec:
  workloadSelector:
    matchLabels:
      app: live-transcode
  objectives:
    - name: LatencyOptimization
      priority: Critical
      spec:
        maxP99LatencyMs: 50          # 端到端调度决策延迟
        maxNetworkRttMs: 2           # 目标节点间网络 RTT
    - name: CostControl
      priority: High
      spec:
        maxCrossAZTrafficRatio: 0.1  # 跨 AZ 流量占比 < 10%
        budgetRef: "monthly-bandwidth-budget"
    - name: HAResilience
      priority: High
      spec:
        minAZSpread: 2               # 至少分布在 2 个 AZ
        maxPodsPerAZRatio: 0.6       # 单 AZ 不超过 60%
  constraints:
    - type: Hardware
      spec:
        requiredGPU: "nvidia.com/gpu: t4/v100/a100"
        minVRAM: 16Gi

# kube-scheduler 启动命令关键参数
--leader-elect=true 
--profiling=false 
--feature-gates=SchedulingGates=true,PodSchedulingReadiness=true 
--config=/etc/kubernetes/scheduler-config.yaml 
--v=4 
--log-format=json