七牛云部署 CSI 并测试模型
|
196
|
0
|
未分类

1186 字
|
32 分钟

一、部署 CSI 和镜像准备

1.1 部署命令

kubectl apply -f https://github.com/qiniu/kubernetes-csi-driver/releases/download/v0.2.0/kodo-plugin.yaml

1.2 kodo-plugin.yaml 配置文件（国内镜像版本）

---
apiVersion: storage.k8s.io/v1
kind: CSIDriver
metadata:
  name: kodoplugin.storage.qiniu.com
spec:
  attachRequired: false
  podInfoOnMount: true

---
apiVersion: apps/v1
kind: DaemonSet
metadata:
  name: kodo-csi-plugin
  namespace: test-s3
spec:
  selector:
    matchLabels:
      app: kodo-csi-plugin
  template:
    metadata:
      labels:
        app: kodo-csi-plugin
    spec:
      serviceAccount: sa.kodoplugin.storage.qiniu.com
      tolerations:
        - operator: Exists
      nodeSelector:
        kubernetes.io/os: linux
      priorityClassName: system-node-critical
      hostNetwork: true
      hostPID: true
      hostIPC: true
      containers:
        - name: csi-driver-registrar
          image: k8s-gcr.m.daocloud.io/sig-storage/csi-node-driver-registrar:v2.5.0
          imagePullPolicy: IfNotPresent
          args:
            - "--v=5"
            - "--csi-address=/var/lib/kubelet/csi-plugins/kodoplugin.storage.qiniu.com/csi.sock"
            - "--kubelet-registration-path=/var/lib/kubelet/csi-plugins/kodoplugin.storage.qiniu.com/csi.sock"
            - "--plugin-registration-path=/registration"
          volumeMounts:
            - name: kubelet-dir
              mountPath: /var/lib/kubelet/
            - name: registration-dir
              mountPath: /registration
          livenessProbe:
            exec:
              command:
                - /csi-node-driver-registrar
                - --plugin-registration-path=/registration
                - --kubelet-registration-path=/var/lib/kubelet/csi-plugins/kodoplugin.storage.qiniu.com/csi.sock
                - --mode=kubelet-registration-probe
            initialDelaySeconds: 30
            timeoutSeconds: 15

        - name: kodo-plugin
          securityContext:
            privileged: true
            capabilities:
              add: ["SYS_ADMIN"]
            allowPrivilegeEscalation: true
          image: kodoproduct/csi-plugin.storage.qiniu.com:v0.2.0
          imagePullPolicy: IfNotPresent
          command: ["/bin/bash"]
          args:
            - "-c"
            - "/entrypoint.sh --endpoint=$(CSI_ENDPOINT) --v=2 --nodeid=$(KUBE_NODE_NAME) --driver=kodo --health-port=11261"
          env:
            - name: KUBE_NODE_NAME
              valueFrom:
                fieldRef:
                  apiVersion: v1
                  fieldPath: spec.nodeName
            - name: CSI_ENDPOINT
              value: unix:///var/lib/kubelet/csi-plugins/kodoplugin.storage.qiniu.com/csi.sock
          livenessProbe:
            httpGet:
              path: /health
              port: health
              scheme: HTTP
            initialDelaySeconds: 60
            periodSeconds: 60
            timeoutSeconds: 35
            failureThreshold: 35
          ports:
            - name: health
              containerPort: 11261
              protocol: TCP
          volumeMounts:
            - name: kubelet-dir
              mountPath: /var/lib/kubelet/
              mountPropagation: Bidirectional
            - name: host-log
              mountPath: /var/log/qiniu/
            - name: bin-dir
              mountPath: /host/usr/local/bin/
            - name: systemd-dir
              mountPath: /host/etc/systemd/system/
            - name: socket-dir
              mountPath: /var/lib/qiniu/
              mountPropagation: Bidirectional
            - name: host-mount
              mountPath: /mnt
              mountPropagation: Bidirectional
            - name: host-dev
              mountPath: /dev
            - name: host-sys
              mountPath: /sys
            - name: host-proc
              mountPath: /host/proc
      volumes:
        - name: registration-dir
          hostPath:
            path: /var/lib/kubelet/plugins_registry
            type: DirectoryOrCreate
        - name: socket-dir
          hostPath:
            path: /var/lib/qiniu/
            type: DirectoryOrCreate
        - name: kubelet-dir
          hostPath:
            path: /var/lib/kubelet
            type: Directory
        - name: host-log
          hostPath:
            path: /var/log/qiniu/
            type: DirectoryOrCreate
        - name: bin-dir
          hostPath:
            path: /usr/local/bin/
            type: DirectoryOrCreate
        - name: systemd-dir
          hostPath:
            path: /etc/systemd/system/
            type: DirectoryOrCreate
        - name: host-mount
          hostPath:
            path: /mnt
            type: Directory
        - name: host-dev
          hostPath:
            path: /dev
            type: Directory
        - name: host-sys
          hostPath:
            path: /sys
            type: Directory
        - name: host-proc
          hostPath:
            path: /proc
            type: Directory
  updateStrategy:
    rollingUpdate:
      maxUnavailable: 10%
    type: RollingUpdate

---
apiVersion: v1
kind: ServiceAccount
metadata:
  name: sa.kodoplugin.storage.qiniu.com
  namespace: test-s3

---
apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRole
metadata:
  name: role.kodoplugin.storage.qiniu.com
rules:
  - apiGroups: [""]
    resources: ["persistentvolumes", "endpoints", "configmaps"]
    verbs: ["get", "list", "watch", "create", "delete", "update"]
  - apiGroups: [""]
    resources: ["persistentvolumeclaims", "nodes"]
    verbs: ["get", "list", "watch", "update"]
  - apiGroups: [""]
    resources: ["events"]
    verbs: ["get", "list", "watch", "create", "update", "patch"]
  - apiGroups: [""]
    resources: ["secrets", "namespaces"]
    verbs: ["get", "list"]
  - apiGroups: [""]
    resources: ["nodes", "pods"]
    verbs: ["get", "list", "watch", "update"]
  - apiGroups: ["storage.k8s.io"]
    resources: ["volumeattachments", "volumeattachments/status"]
    verbs: ["get", "list", "watch", "update", "patch"]
  - apiGroups: ["storage.k8s.io"]
    resources: ["storageclasses"]
    verbs: ["get", "list", "watch"]
  - apiGroups: ["coordination.k8s.io"]
    resources: ["leases"]
    verbs: ["get", "list", "watch", "create", "update", "patch"]

---
apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRoleBinding
metadata:
  name: binding.kodoplugin.storage.qiniu.com
subjects:
  - kind: ServiceAccount
    name: sa.kodoplugin.storage.qiniu.com
    namespace: test-s3
roleRef:
  kind: ClusterRole
  name: role.kodoplugin.storage.qiniu.com
  apiGroup: rbac.authorization.k8s.io

---
apiVersion: apps/v1
kind: Deployment
metadata:
  name: kodo-provisioner
  namespace: test-s3
spec:
  selector:
    matchLabels:
      app: kodo-provisioner
  replicas: 2
  template:
    metadata:
      labels:
        app: kodo-provisioner
    spec:
      serviceAccount: sa.kodoplugin.storage.qiniu.com
      tolerations:
        - operator: Exists
      nodeSelector:
        kubernetes.io/os: linux
      affinity:
        nodeAffinity:
          preferredDuringSchedulingIgnoredDuringExecution:
            - weight: 1
              preference:
                matchExpressions:
                  - key: node-role.kubernetes.io/master
                    operator: Exists
      priorityClassName: system-node-critical
      hostNetwork: true
      containers:
        - name: external-kodo-provisioner
          securityContext:
            privileged: true
          image: gcr.m.daocloud.io/k8s-staging-sig-storage/csi-provisioner:canary
          args:
            - "--csi-address=$(ADDRESS)"
            - "--volume-name-prefix=kodo"
            - "--timeout=150s"
            - "--leader-election=true"
            - "--retry-interval-start=500ms"
            - "--v=5"
          env:
            - name: ADDRESS
              value: /var/lib/kubelet/csi-plugins/kodoplugin.storage.qiniu.com/csi.sock
          imagePullPolicy: IfNotPresent
          volumeMounts:
            - name: kubelet-dir
              mountPath: /var/lib/kubelet/
              mountPropagation: Bidirectional
      volumes:
        - name: kubelet-dir
          hostPath:
            path: /var/lib/kubelet
            type: Directory

二、源码打包镜像

2.1 目录说明

镜像目录：

  • 10.10.207.16: /data/luozhihong/s3/CSI/kubernetes-csi-driver-0.2.0
  • 10.10.207.18: /root/luozhihong/kubernetes-csi-driver-0.2.0

部署目录：

  • /data/luozhihong/s3

说明： 由于官方镜像系统架构不一致,需要拉取代码手动打包镜像

2.2 构建步骤

Step 1: 拉取代码

# 拉取代码(需要 tag 0.2.0)
git clone https://github.com/qiniu/kubernetes-csi-driver.git
cd kubernetes-csi-driver-0.2.0

Step 2: 准备系统命令

# 创建目录
mkdir -p docker/amd64

# 测试 rclone 是否正常
rclone version

# 拷贝系统命令到容器
cp /usr/local/bin/rclone docker/amd64/
cp /usr/local/bin/kodofs docker/amd64/

Step 3: 如果 rclone 不正常,下载并替换

# 下载 rclone
wget https://downloads.rclone.org/v1.64.0/rclone-v1.64.0-linux-amd64.zip
unzip rclone-v1.64.0-linux-amd64.zip

# 替换 rclone 命令
cp rclone-v1.64.0-linux-amd64/rclone /usr/local/bin/

2.3 Dockerfile 配置(国内镜像版本)

# 1. 构建阶段
FROM docker.m.daocloud.io/library/golang:1.21-alpine3.18 AS build-env

ARG TARGETOS
ARG TARGETARCH

COPY . /app
WORKDIR /app

# 安装依赖
RUN apk add --no-cache git make

# 设置国内代理并编译
RUN export GOPROXY=https://goproxy.cn,direct && \
    GOOS=$TARGETOS GOARCH=$TARGETARCH make build

# 2. 最终镜像
FROM docker.m.daocloud.io/library/alpine:3.18

ARG TARGETOS
ARG TARGETARCH
ARG PLUGIN_FILENAME=plugin.storage.qiniu.com
ARG CONNECTOR_FILENAME=connector.${PLUGIN_FILENAME}

# 拷贝编译结果
COPY --from=build-env /app/plugin/${PLUGIN_FILENAME} /usr/local/bin/${PLUGIN_FILENAME}
COPY --from=build-env /app/connector/${CONNECTOR_FILENAME} /usr/local/bin/${CONNECTOR_FILENAME}

# 拷贝额外文件
COPY docker/${TARGETARCH}/kodofs /usr/local/bin/kodofs
COPY docker/${TARGETARCH}/rclone /usr/local/bin/rclone
COPY docker/kodo-csi-connector.service /csiplugin-connector.service
COPY docker/entrypoint.sh /entrypoint.sh

# 赋予执行权限
RUN chmod +x /usr/local/bin/kodofs \
    /usr/local/bin/rclone \
    /usr/local/bin/${PLUGIN_FILENAME} \
    /usr/local/bin/${CONNECTOR_FILENAME} \
    /entrypoint.sh

# 安装运行依赖
RUN apk add --no-cache util-linux ca-certificates bash

ENTRYPOINT ["/entrypoint.sh"]

2.4 构建和导入镜像

# 构建镜像
docker build -t kodoproduct/csi-plugin.storage.qiniu.com:v0.2.0 .

# 导出镜像
docker save -o csi.tar kodoproduct/csi-plugin.storage.qiniu.com:v0.2.0

# 导入到 containerd
ctr -n k8s.io images import csi.tar

2.5 部署 CSI

kubectl apply -f kodo-plugin.yaml

三、使用指定的存储空间

3.1 创建 Secret

s3-secret.yaml:

apiVersion: v1
kind: Secret
metadata:
  name: kodo-csi-pv-secret
  namespace: test-s3
type: Opaque
stringData:
  accesskey: "AK"
  secretkey: "SK"
  bucketname: "ruantong-test"
  ucendpoint: "http://uc.qiniuapi.com"
  storageclass: "STANDARD"
  region: "z2"
  subdir: ""
  s3forcepathstyle: "false"

kubectl apply -f s3-secret.yaml

3.2 创建 PV 和 PVC

s3-pv-pvc.yaml:

---
# PV - 持久卷
apiVersion: v1
kind: PersistentVolume
metadata:
  name: kodo-csi-pv
  labels:
    kodo-pvname: kodo-csi-pv
spec:
  capacity:
    storage: 10Gi
  accessModes:
    - ReadWriteMany
  persistentVolumeReclaimPolicy: Retain
  csi:
    driver: kodoplugin.storage.qiniu.com
    volumeHandle: kodo-csi-pv
    volumeAttributes:
      uploadcutoff: "209715200"
      uploadchunksize: "5242880"
      uploadconcurrency: "4"
      vfscachemode: "off"
      dircacheduration: "5m0s"
      buffersize: "16777216"
      vfsreadahead: "0"
      vfscachemaxage: "1h0m0s"
      vfscachepollinterval: "1m0s"
      vfswriteback: "5s"
      vfsreadchunksize: "134217728"
      nochecksum: "no"
      nomodtime: "no"
      noseek: "no"
      readonly: "no"
      transfers: "4"
    nodePublishSecretRef:
      name: kodo-csi-pv-secret
      namespace: test-s3

---
# PVC - 持久卷声明
apiVersion: v1
kind: PersistentVolumeClaim
metadata:
  name: kodo-pvc
  namespace: test-s3
spec:
  accessModes:
    - ReadWriteMany
  storageClassName: ''
  resources:
    requests:
      storage: 10Gi
  selector:
    matchLabels:
      kodo-pvname: kodo-csi-pv

kubectl apply -f s3-pv-pvc.yaml

四、动态创建存储空间

4.1 创建 Secret

s3-secret.yaml:

apiVersion: v1
kind: Secret
metadata:
  name: kodo-csi-sc-secret
  namespace: test-s3
type: Opaque
stringData:
  accesskey: "AK"
  secretkey: "SK"
  bucketname: "ruantong-test"
  ucendpoint: "http://uc.qiniuapi.com"
  storageclass: "STANDARD"
  region: "z2"
  subdir: ""
  s3forcepathstyle: "false"

kubectl apply -f s3-secret.yaml

4.2 创建 StorageClass

sc.yaml:

apiVersion: storage.k8s.io/v1
kind: StorageClass
metadata:
  name: kodo-csi-sc
parameters:
  uploadcutoff: "209715200"
  uploadchunksize: "5242880"
  uploadconcurrency: "4"
  vfscachemode: "off"
  dircacheduration: "5m0s"
  buffersize: "16777216"
  vfsreadahead: "0"
  vfscachemaxage: "1h0m0s"
  vfscachepollinterval: "1m0s"
  vfswriteback: "5s"
  vfsreadchunksize: "134217728"
  nochecksum: "no"
  nomodtime: "no"
  noseek: "no"
  readonly: "no"
  transfers: "4"
  csi.storage.k8s.io/provisioner-secret-name: kodo-csi-sc-secret
  csi.storage.k8s.io/provisioner-secret-namespace: test-s3
provisioner: kodoplugin.storage.qiniu.com
reclaimPolicy: Retain  # Retain:保留存储空间及数据 | Delete:删除存储空间及数据

kubectl apply -f sc.yaml

4.3 创建 PVC

pvc.yaml:

apiVersion: v1
kind: PersistentVolumeClaim
metadata:
  name: kodo-pvc
  namespace: test-s3
spec:
  accessModes:
    - ReadWriteMany
  storageClassName: kodo-csi-sc
  resources:
    requests:
      storage: 10Gi

kubectl apply -f pvc.yaml

五、部署模型

5.1 部署 BGE-M3 模型

bge-m3-s3.yaml:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: vllm-bge-m3
  namespace: test-s3
spec:
  replicas: 1
  selector:
    matchLabels:
      app: vllm-bge-m3
  template:
    metadata:
      labels:
        app: vllm-bge-m3
    spec:
      nodeSelector:
        kubernetes.io/hostname: bcm-headnode
      containers:
      - name: vllm-bge-m3
        image: bge-m3:v0.1.1
        command:
        - vllm
        - serve
        - /kodo
        - --served-model-name=bge-m3
        - --host=0.0.0.0
        - --port=5000
        - --tensor-parallel-size=1
        - --max-model-len=4096
        - --gpu-memory-utilization=0.3
        env:
        - name: CUDA_VISIBLE_DEVICES
          value: "0"
        ports:
        - containerPort: 5000
          name: http
        volumeMounts:
        - name: kodo-storage
          mountPath: /kodo
          readOnly: true
        - name: shm
          mountPath: /dev/shm
        resources:
          limits:
            memory: "8Gi"
            cpu: "4"
          requests:
            memory: "8Gi"
            cpu: "2"
      volumes:
      - name: kodo-storage
        persistentVolumeClaim:
          claimName: kodo-pvc
      - name: shm
        emptyDir:
          medium: Memory
          sizeLimit: 4Gi

---
apiVersion: v1
kind: Service
metadata:
  name: vllm-service
  namespace: test-s3
spec:
  type: NodePort
  selector:
    app: vllm-bge-m3
  ports:
  - name: http
    protocol: TCP
    port: 5000
    targetPort: 5000
    nodePort: 30501

kubectl apply -f bge-m3-s3.yaml

5.2 测试模型

curl http://localhost:30501/v1/embeddings \
  -H "Content-Type: application/json" \
  -d '{
    "model": "bge-m3",
    "input": "你好,vLLM 向量服务测试"
  }'

六、问题排查

6.1 部署 CSI 时常见问题

⚠️ 问题 1: 镜像无法拉取

解决方案: 需要切换成国内可访问的镜像

⚠️ 问题 2: kodoproduct/csi-plugin.storage.qiniu.com:v0.2.0 镜像架构不兼容

原因: DockerHub 镜像不兼容系统架构

解决方案:

  1. 拉取代码仓库的代码手动打包镜像
  2. 修改 Dockerfile 中的 FROM 为国内镜像源
  3. 检查 rclone version 是否正常
  4. 如果不正常,下载并替换 rclone: wget https://downloads.rclone.org/v1.64.0/rclone-v1.64.0-linux-amd64.zipcp rclone-v1.64.0-linux-amd64/rclone /usr/local/bin/
  5. 打镜像时:
    • 新建目录: docker/amd64
    • 拷贝主机系统的 /usr/local/bin/rclone 和 /usr/local/bin/kodofs 到 docker/amd64/
  6. docker build 后导入 ctr 或上传到 Harbor: docker save -o csi.tar kodoproduct/csi-plugin.storage.qiniu.com:v0.2.0ctr -n k8s.io images import csi.tar

6.2 部署 BGE-M3 时常见问题

⚠️ 问题 3: GPU 共享环境变量

解决方案: 添加 GPU 环境变量

env:
- name: CUDA_VISIBLE_DEVICES
  value: "0"

⚠️ 问题 4: 镜像准备

确保将 bge-m3 镜像导入 ctr 或上传到 Harbor

附录

参数说明

参数说明默认值
uploadcutoff上传切割大小209715200
uploadchunksize上传块大小5242880
uploadconcurrency上传并发数4
vfscachemodeVFS 缓存模式off
dircacheduration目录缓存时长5m0s
buffersize缓冲区大小16777216
transfers传输并发数4

常用命令

# 查看 CSI 插件状态
kubectl get pods -n test-s3

# 查看 PV/PVC 状态
kubectl get pv,pvc -n test-s3

# 查看 StorageClass
kubectl get sc

# 查看日志
kubectl logs -n test-s3 <pod-name>

# 删除资源
kubectl delete -f <yaml-file>

文档版本: v1.0
最后更新: 2024

暂无评论

发送评论 编辑评论 


				

			

						

				

					

						

							

								
							

							
						

					

				

				

					

						

							

								
							

							
						

					

				

				

					

						

							

								
							

							
							
													

					

				

			

			
			

				

					
				

			

				

											

							
							
						

																				
					
											
						

	

		
|´・ω・)ノ
ヾ(≧∇≦*)ゝ
(☆ω☆)
（╯‵□′）╯︵┴─┴
￣﹃￣
(/ω＼)
∠( ᐛ 」∠)＿
(๑•̀ㅁ•́ฅ)
→_→
୧(๑•̀⌄•́๑)૭
٩(ˊᗜˋ*)و
(ノ°ο°)ノ
(´இ皿இ｀)
⌇●﹏●⌇
(ฅ´ω`ฅ)
(╯°A°)╯︵○○○
φ(￣∇￣o)
ヾ(´･ ･｀｡)ノ"
( ง ᵒ̌皿ᵒ̌)ง⁼³₌₃
(ó﹏ò｡)
Σ(っ °Д °;)っ
( ,,´･ω･)ﾉ"(´っω･｀｡)
╮(╯▽╰)╭
o(*////▽////*)q
＞﹏＜
( ๑´•ω•) "(ㆆᴗㆆ)
😂
😀
😅
😊
🙂
🙃
😌
😍
😘
😜
😝
😏
😒
🙄
😳
😡
😔
😫
😱
😭
💩
👻
🙌
🖕
👍
👫
👬
👭
🌚
🌝
🙈
💊
😶
🙏
🍦
🍉
😣
Source: github.com/k4yt3x/flowerhd
	

	

		
颜文字
Emoji
小恐龙
花!
	


									

			

			
			
		

	






上一篇
下一篇