Kubernetes AI 推理服务部署与优化实战

3. TensorFlow Serving 部署

3.1 准备模型数据

模型文件需要持久化存储，这里使用 PVC 挂载本地或云盘。

# 创建模型目录并下载示例
mkdir -p models/mnist/1
wget -O models/mnist/1/saved_model.pb https://storage.googleapis.com/download.tensorflow.org/models/official/20181001_resnet/savedmodels/resnet_v2_fp32_savedmodel_NHWC_jpg.tar.gz

# 创建持久卷声明
kubectl create -f - <<EOF
apiVersion: v1
kind: PersistentVolumeClaim
metadata:
  name: model-pvc
  namespace: default
spec:
  accessModes:
  - ReadWriteOnce
  resources:
    requests:
      storage: 10Gi
EOF

3.2 部署 TensorFlow Serving

配置 Deployment 和 Service 后，即可对外提供服务。

deployment.yaml

apiVersion: apps/v1
kind: Deployment
metadata:
  name: tf-serving
  namespace: default
spec:
  replicas: 2
  selector:
    matchLabels:
      app: tf-serving
  template:
    metadata:
      labels:
        app: tf-serving
    spec:
      containers:
      - name: tf-serving
        image: tensorflow/serving:latest
        ports:
        - containerPort: 8500
        - containerPort: 8501
        env:
        - name: MODEL_NAME
          value: mnist
        volumeMounts:
        - name: model-volume
          mountPath: /models
      volumes:
      - name: model-volume
        persistentVolumeClaim:
          claimName: model-pvc

service.yaml

apiVersion: v1
kind: Service
metadata:
  name: tf-serving
  namespace: default
spec:
  selector:
    app: tf-serving
  ports:
  - port: 8501
    targetPort: 8501
  type: LoadBalancer

部署完成后，可以通过 curl 测试推理接口：

MODEL_SERVICE=$(kubectl get svc tf-serving -o jsonpath='{.status.loadBalancer.ingress[0].ip}')
curl -d '{"instances": [[[0.0 for _ in range(28)] for _ in range(28)]]}' -X POST http://$MODEL_SERVICE:8501/v1/models/mnist:predict

4. Triton Inference Server 部署

Triton 的优势在于支持多种框架混合部署，且内置了动态批处理等高级功能。

deployment.yaml

apiVersion: apps/v1
kind: Deployment
metadata:
  name: triton-server
  namespace: default
spec:
  replicas: 2
  selector:
    matchLabels:
      app: triton-server
  template:
    metadata:
      labels:
        app: triton-server
    spec:
      containers:
      - name: triton-server
        image: nvcr.io/nvidia/tritonserver:23.08-py3
        ports:
        - containerPort: 8000
        - containerPort: 8001
        - containerPort: 8002
        resources:
          limits:
            nvidia.com/gpu: 1
          requests:
            nvidia.com/gpu: 1
        volumeMounts:
        - name: model-volume
          mountPath: /models
      volumes:
      - name: model-volume
        persistentVolumeClaim:
          claimName: model-pvc

service.yaml

apiVersion: v1
kind: Service
metadata:
  name: triton-server
  namespace: default
spec:
  selector:
    app: triton-server
  ports:
  - port: 8000
    targetPort: 8000
  - port: 8001
    targetPort: 8001
  - port: 8002
    targetPort: 8002
  type: LoadBalancer

部署后检查状态：kubectl get pods -l app=triton-server。

5. 性能优化

5.1 模型优化

提升推理速度的基础工作：

1. 模型量化：将 FP32 转为 INT8 或 FP16，减少显存占用并加速计算。
2. 模型剪枝：移除冗余神经元，降低计算量。
3. 模型蒸馏：用大模型指导小模型训练，平衡精度与速度。

5.2 推理服务优化

开启批处理能显著提升吞吐量，同时限制 GPU 内存增长策略。

apiVersion: apps/v1
kind: Deployment
metadata:
  name: tf-serving-batched
  namespace: default
spec:
  replicas: 2
  selector:
    matchLabels:
      app: tf-serving-batched
  template:
    metadata:
      labels:
        app: tf-serving-batched
    spec:
      containers:
      - name: tf-serving
        image: tensorflow/serving:latest
        ports:
        - containerPort: 8501
        env:
        - name: MODEL_NAME
          value: mnist
        - name: TF_FORCE_GPU_ALLOW_GROWTH
          value: "true"
        - name: BATCH_SIZE
          value: "32"
        resources:
          limits:
            nvidia.com/gpu: 1
          requests:
            nvidia.com/gpu: 1

5.3 自动缩放

根据负载自动调整副本数，节省成本的同时应对流量高峰。

apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: tf-serving-hpa
  namespace: default
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: tf-serving
  minReplicas: 2
  maxReplicas: 10
  metrics:
  - type: Resource
    resource:
      name: cpu
      target:
        type: Utilization
        averageUtilization: 70
  - type: Resource
    resource:
      name: memory
      target:
        type: Utilization
        averageUtilization: 80

6. 监控与可观测性

6.1 监控配置

集成 Prometheus 可以方便地采集指标。

apiVersion: monitoring.coreos.com/v1
kind: ServiceMonitor
metadata:
  name: tf-serving-monitor
  namespace: monitoring
spec:
  selector:
    matchLabels:
      app: tf-serving
  endpoints:
  - port: 8501
    path: /v1/monitoring/prometheus
    interval: 15s

6.2 日志管理

合理设置日志级别，避免磁盘爆满，同时保留关键调试信息。

apiVersion: apps/v1
kind: Deployment
metadata:
  name: tf-serving
  namespace: default
spec:
  # ...省略部分字段...
  template:
    spec:
      containers:
      - name: tf-serving
        image: tensorflow/serving:latest
        env:
        - name: TF_CPP_MIN_LOG_LEVEL
          value: "0"
        - name: TF_ENABLE_GPU_GARBAGE_COLLECTION
          value: "true"
        args:
        - --model_name=mnist
        - --model_base_path=/models/mnist
        - --enable_batching=true
        - --batching_parameters_file=/models/batching_parameters.txt

7. 安全最佳实践

7.1 模型安全

• 模型加密：保护静态模型文件不被窃取。
• 访问控制：利用 RBAC 限制谁可以调用服务。
• 版本管理：记录模型变更历史，便于回滚。

7.2 网络安全

通过 NetworkPolicy 限制 Pod 间的通信，只允许必要的流量通过。

apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
  name: ai-inference-network-policy
  namespace: default
spec:
  podSelector:
    matchLabels:
      app: tf-serving
  policyTypes:
  - Ingress
  - Egress
  ingress:
  - from:
    - podSelector:
        matchLabels:
          app: api-gateway
    ports:
    - protocol: TCP
      port: 8501
  egress:
  - to:
    - podSelector:
        matchLabels:
          app: monitoring
    ports:
    - protocol: TCP
      port: 9090

8. 实际应用场景

8.1 多模型部署

Triton 支持在一个实例中加载多个模型，节省资源。

apiVersion: apps/v1
kind: Deployment
metadata:
  name: triton-multi-model
  namespace: default
spec:
  replicas: 2
  selector:
    matchLabels:
      app: triton-multi-model
  template:
    metadata:
      labels:
        app: triton-multi-model
    spec:
      containers:
      - name: triton-server
        image: nvcr.io/nvidia/tritonserver:23.08-py3
        ports:
        - containerPort: 8000
        - containerPort: 8001
        - containerPort: 8002
        resources:
          limits:
            nvidia.com/gpu: 1
          requests:
            nvidia.com/gpu: 1
        volumeMounts:
        - name: model-volume
          mountPath: /models
      volumes:
      - name: model-volume
        persistentVolumeClaim:
          claimName: models-pvc

8.2 A/B 测试

利用 Ingress 的 Canary 特性进行灰度发布。

apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  name: ai-inference-ingress
  namespace: default
  annotations:
    nginx.ingress.kubernetes.io/canary: "true"
    nginx.ingress.kubernetes.io/canary-weight: "20"
spec:
  rules:
  - host: inference.example.com
    http:
      paths:
      - path: /v1/models
        pathType: Prefix
        backend:
          service:
            name: tf-serving-v2
            port:
              number: 8501

9. 故障排查

9.1 常见问题解决

遇到服务不可用或性能下降时，先检查底层资源。

# 查看 GPU 使用情况
kubectl exec -it <pod-name> -- nvidia-smi

# 查看推理服务日志
kubectl logs -l app=tf-serving

# 检查模型状态
curl http://<service-ip>:8501/v1/models/mnist

# 测试推理服务
curl -d '{"instances": [[[0.0 for _ in range(28)] for _ in range(28)]]}' -X POST http://<service-ip>:8501/v1/models/mnist:predict

9.2 调试技巧

1. 启用详细日志：设置 TF_CPP_MIN_LOG_LEVEL=0 获取更多信息。
2. 使用 GPU 分析工具：配合 nvidia-smi 和 nvprof 定位瓶颈。
3. 检查网络连接：确保服务端口未被防火墙拦截。
4. 验证模型格式：确认输入数据形状与模型定义一致。

10. 总结

Kubernetes 为 AI 推理提供了弹性的基础设施。关键在于合理配置 GPU 资源、选择匹配的推理框架、实施性能优化策略，并建立完善的监控与安全体系。遵循这些实践，能够构建出既高效又可靠的 AI 服务架构。