[FLINK-23043] Wrong master address displayed when run application in native K8s with -Dkubernetes.rest-service.exposed.type=NodePort

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Duplicate
• Affects Version/s: 1.13.1
• Component/s: Deployment / Kubernetes
• Labels: pull-request-available
• Assignee: Shang Yuanchun
• Reporter: Shang Yuanchun
• Created: 20/Jun/21 05:20
• Updated: 19/Aug/21 15:22
• Resolved: 19/Aug/21 15:21

Description

When running with Native Kubernetes in application mode, and Fabric8 Kubernetes client is configured by Service account, if you specify -Dkubernetes.rest-service.exposed.type=NodePort, the IP address in JobManager Web Interface url displayed is not master node, but instead it is the ClusterIP of service kubernetes.default.svc.

Steps to reproduce:

1.Start a Kubernetes cluster
2.Bind ClusterRole edit to service account: default (for simplicity here)

3.Run kubectl run flink --image flink:latest sleep inf

4.Into container of step 3, run 

./bin/flink run-application --target kubernetes-application -Dkubernetes.cluster-id=my-application-cluster -Dkubernetes.container.image=flink:latest -Dkubernetes.rest-service.exposed.type=NodePort local:///opt/flink/examples/streaming/TopSpeedWindowing.jar

Here it will say that JobManager Web interface is at http://10.96.0.1:32333 

But actually 10.96.0.1 here is the environment variable of KUBERNETES_SERVICE_HOST, and it is the ClusterIP of service kubernetes.default.svc, not the address of master node, so http://10.96.0.1:32333 is not accessiable.

Issue Links

• duplicates: FLINK-23507 The node IP obtained in NodePort mode is a VIP
• links to: GitHub Pull Request #16209