K8S Pod定向部署到指定节点的实现全过程

# 查看集群所有节点名称
kubectl get nodes

# 为node-01添加标签：hardware=gpu
kubectl label nodes node-01 hardware=gpu

# 验证标签是否添加成功（查看node-01的Labels字段）
kubectl describe node node-01 | grep Labels

apiVersion: v1
kind: Pod
metadata:
  name: tensorflow-gpu-pod
spec:
  containers:
  - name: tensorflow-container
    image: tensorflow/tensorflow:latest-gpu  # 依赖GPU的TensorFlow镜像
    command: ["sleep", "3600"]  # 保持Pod运行1小时，便于验证
  nodeSelector:  # 定向到标签为hardware=gpu的节点
    hardware: gpu

# 部署Pod
kubectl apply -f gpu-pod.yaml

# 查看Pod调度结果（观察NODE字段是否为node-01）
kubectl get pods tensorflow-gpu-pod -o wide

kubectl label nodes node-01 hardware-  # 标签键后加“-”表示删除该标签

apiVersion: v1
kind: Pod
metadata:
  name: ssd-pod-hard
spec:
  containers:
  - name: nginx-container
    image: nginx:latest
  affinity:
    nodeAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:  # 硬亲和性
        nodeSelectorTerms:
        - matchExpressions:
          - key: zone
            operator: In  # 匹配规则：zone的值在[shanghai]中
            values:
            - shanghai
          - key: storage
            operator: Exists  # 匹配规则：节点存在storage标签（不限制值，只要有该标签即可）

apiVersion: v1
kind: Pod
metadata:
  name: web-pod-soft
spec:
  containers:
  - name: tomcat-container
    image: tomcat:9.0
  affinity:
    nodeAffinity:
      preferredDuringSchedulingIgnoredDuringExecution:  # 软亲和性
      - weight: 80  # 权重（1-100），值越大优先级越高
        preference:
          matchExpressions:
          - key: cpu
            operator: In
            values:
            - high  # 优先匹配cpu=high的节点

# 部署两个Pod
kubectl apply -f ssd-pod-hard.yaml -f web-pod-soft.yaml

# 查看调度结果：
# ssd-pod-hard仅会在zone=shanghai且有storage标签的节点上运行
# web-pod-soft优先在cpu=high节点运行，若无则在其他节点运行
kubectl get pods -o wide

# 为node-02添加污点
kubectl taint nodes node-02 gpu-only=true:NoSchedule

# 查看节点污点（观察Taints字段）
kubectl describe node node-02 | grep Taints

apiVersion: v1
kind: Pod
metadata:
  name: ai-training-pod
spec:
  containers:
  - name: pytorch-container
    image: pytorch/pytorch:latest
    command: ["sleep", "3600"]
  tolerations:  # 配置容忍，抵消节点污点
  - key: "gpu-only"
    operator: "Equal"  # 匹配规则：key和value与污点完全一致
    value: "true"
    effect: "NoSchedule"  # 匹配污点的effect

apiVersion: v1
kind: Pod
metadata:
  name: normal-pod
spec:
  containers:
  - name: alpine-container
    image: alpine:latest
    command: ["sleep", "3600"]

# 部署两个Pod
kubectl apply -f ai-training-pod.yaml -f normal-pod.yaml

# 查看调度结果：
# ai-training-pod（有容忍）会调度到node-02
# normal-pod（无容忍）会被node-02排斥，调度到其他节点
kubectl get pods -o wide

# （可选）删除节点污点，恢复节点为普通节点
kubectl taint nodes node-02 gpu-only=true:NoSchedule-