Flink on Kubernetes (flink-operator) 部署Flink

flink on k8s 官网

https://nightlies.apache.org/flink/flink-kubernetes-operator-docs-release-1.1/docs/try-flink-kubernetes-operator/quick-start/
我的部署脚本和官网不一样，有些地方官网不够详细

部署k8s集群

注意，按照默认配置至少有两台worker

安装helm

https://helm.sh/zh/docs/intro/install/

安装flink opreator

kubectl create -f https://github.com/jetstack/cert-manager/releases/download/v1.8.2/cert-manager.yaml
helm repo add flink-operator-repo https://downloads.apache.org/flink/flink-kubernetes-operator-1.1.0/
helm install flink-kubernetes-operator flink-operator-repo/flink-kubernetes-operator -n flink

安装完成后，资源如下

[root@k8s1 flinkinstall]# helm list -n flink
NAME NAMESPACE REVISION UPDATED STATUS CHART APP VERSION
flink-kubernetes-operator flink 1 2024-03-07 16:57:48.374299701 +0800 CST deployed flink-kubernetes-operator-1.7.0 1.7.0
[root@k8s1 flinkinstall]# kubectl get all -A
NAMESPACE NAME READY STATUS RESTARTS AGE
cert-manager pod/cert-manager-66b646d76-gkw55 1/1 Running 1 2d2h
cert-manager pod/cert-manager-cainjector-59dc9659c7-pkgrm 1/1 Running 1 2d2h
cert-manager pod/cert-manager-webhook-7f7787f7fd-wd5vv 1/1 Running 1 2d2h
flink pod/flink-kubernetes-operator-857d48ff65-45mg2 2/2 Running 6 5d20h

NAMESPACE NAME TYPE CLUSTER-IP EXTERNAL-IP PORT(S) AGE
cert-manager service/cert-manager ClusterIP 192.178.138.19 <none> 9402/TCP 2d2h
cert-manager service/cert-manager-webhook ClusterIP 192.178.130.219 <none> 443/TCP 2d2h
flink service/flink-operator-webhook-service ClusterIP 192.178.139.67 <none> 443/TCP 5d20h

NAMESPACE NAME READY UP-TO-DATE AVAILABLE AGE
cert-manager deployment.apps/cert-manager 1/1 1 1 2d2h
cert-manager deployment.apps/cert-manager-cainjector 1/1 1 1 2d2h
cert-manager deployment.apps/cert-manager-webhook 1/1 1 1 2d2h
flink deployment.apps/flink-kubernetes-operator 1/1 1 1 5d20h

NAMESPACE NAME DESIRED CURRENT READY AGE
cert-manager replicaset.apps/cert-manager-66b646d76 1 1 1 2d2h
cert-manager replicaset.apps/cert-manager-cainjector-59dc9659c7 1 1 1 2d2h
cert-manager replicaset.apps/cert-manager-webhook-7f7787f7fd 1 1 1 2d2h
flink replicaset.apps/flink-kubernetes-operator-857d48ff65 1 1 1 5d20h
此时k8s集群就可以支持我们按照flink-opreator的指定格式提交flink任务了

提交flink任务

session模式与application模式区别在于资源隔离度

session模式： jobmanager预先启动，随时准备接收flink jar，启动taskmanager，flink任务结束后jobmanager不退出，所有flink任务共享同一个jobmanager，资源隔离差，某个flink任务导致jobmanager异常，会影响到其他flink任务，小任务，不在乎异常情况可以用
application模式：每次提交flink任务才会启动一个jobmanger，flink任务结束后，jobmanager也退出，隔离效果好，生产常用
per-job模式：这个模式与application模式类似，区别在于client的运行位置，但是新版的flink已经删除了这种提交方式

这里是flink on yarn的运行模式

https://blog.csdn.net/java_creatMylief/article/details/126172793

application模式

apiVersion: flink.apache.org/v1beta1
kind: FlinkDeployment
metadata:
  name: pod-template-example
spec:
  image: flink:1.15
  flinkVersion: v1_15
  flinkConfiguration:
    taskmanager.numberOfTaskSlots: "2"
  serviceAccount: flink
  podTemplate:
    apiVersion: v1
    kind: Pod
    metadata:
      name: pod-template
    spec:
      serviceAccount: flink
      containers:
        # Do not change the main container name
        - name: flink-main-container
          env:
          volumeMounts:
            - mountPath: /flink-logs
              name: flink-logs
      initContainers:
        - name: init-nginx
          image: busybox
          command: [ 'sh','-c','wget http://192.168.33.2/phoenix-client-1.0-SNAPSHOT-jar-with-dependencies.jar -O /flink-logs/StateMachineExample1.jar' ]
          volumeMounts:
            - mountPath: /flink-logs
              name: flink-logs
      volumes:
        - name: flink-logs
          emptyDir: { }
  jobManager:
    resource:
      memory: "1024m"
      cpu: 1
  taskManager:
    resource:
      memory: "1024m"
      cpu: 1
  job:
    jarURI: local:///flink-logs/StateMachineExample1.jar
    parallelism: 1
    entryClass: org.examplexxx.test
    args: [/path/from/data,/path/to/data]
    initialSavepointPath: hdfs://flink/ckpath/xxxxx
    
kubectl apply -f ${name}.yaml

kubectl port-forward svc/basic-example-rest 8081 --address 192.168.33.81

访问 http://192.168.33.81:8081

jarURI: local:///flink-logs/StateMachineExample1.jar

此处jarURL只得是docker内部路径，且不支持远程路径（http/s3/hdfs），因此需要将jar包放到docker内部。

1、可以将flink版本和jar包打到一个镜像中。

2、可以使用pvc挂载进去。

3、使用initContainers和 containers使用相同的挂载路径，然后使用远程文件下载放到挂载路径中，containers就能获取到该jar包

此处使用第三种情况，使用initContainers变相支持远程文件地址，使用起来比较方便。

yarn-application 对比

yarn-application	k8s-application
-p (并行度)	spec.job.parallelism
-yjm (jobmanager内存)	spec.jobManager.resource.memory
-ytm (taskmanager内存)	spec.taskManager.resource.memory
-ys (taskmanger的slot槽数)	spec.flinkConfiguration.taskmanager.numberOfTaskSlots
-c (主类)	spec.job.entryClass
jar (jar包)	spec.job.jarURI
-s (恢复点启动)	spec.job.initialSavepointPath

session模式

部署session cluster

apiVersion: flink.apache.org/v1beta1
kind: FlinkDeployment
metadata:
  namespace: flink
  name: session-deployment-only
spec:
  image: flink:1.13.6
  flinkVersion: v1_13
  imagePullPolicy: IfNotPresent # 镜像拉去策略，优先本地，没有，仓库拉去
  ingress:
    template: "flink.k8s.io/{{namespace)}/{{name}}(/|$)(.*)"
    className: "Nginx"
    annotations:
      taskmanager.numberOfTaskSlots: "2"
    serviceAccount: flink
    jobManager:
      replicas: 1
      resource:
        memory: "1024m"
        cpu: 1
    taskManager:
      replicas: 1
      resource:
        memory: "1024m"
        cpu: 1

kubectl apply -f ${name}.yaml

部署cluster完成，配置svcType 后即可访问，flink web ui，此时jobManager是启动着的 taskmanager随着flink jar进行启动和停止

部署flink jar

apiVersion: flink.apache.org/v1beta1
kind: FlinkSessionJob
metadata:
  namespace: flink
  name: session-job-only
  job:
    jarUrl: sasa
    entryClass: aa
    parallelism: 1
    upgradeMode: stateless


kubectl apply -f ${name}.yaml