I am trying to set up a hadoop single node on kubernetes. The odd thing is that, when i login into the pod via kubectl exec -it <pod> /bin/bash i can happily access e.g. the name node on port 9000.

root@hadoop-5dcf94b54d-7fgfq:/hadoop/hadoop-2.8.5# telnet localhost 9000
Trying 127.0.0.1...
Connected to localhost.
Escape character is '^]'.

I can also bin/hdfs dfs -put files and such, so the cluster seems to be working fine. I can also access the ui via kubectl port-forward <podname> 50070:50070 and i see a data node up and running. So the cluster (setup is 'pseudo-distributed' as described here.) seems to be working fine.

However, when i want to access my service via kubernetes dns, i get a Connection refused.

telnet hadoop.aca534.svc.cluster.local  9000
Trying 10.32.89.21...
telnet: Unable to connect to remote host: Connection refused

What is the difference when accessing a port via k8s-dns?

The port must be open, i also can see that hadoop name node is listening on 9000.

lsof -i :9000
COMMAND  PID USER   FD   TYPE    DEVICE SIZE/OFF NODE NAME
java    2518 root  227u  IPv4 144574393      0t0  TCP localhost:9000 (LISTEN)
java    2518 root  237u  IPv4 144586825      0t0  TCP localhost:9000->localhost:58480 (ESTABLISHED)
java    2660 root  384u  IPv4 144584032      0t0  TCP localhost:58480->localhost:9000 (ESTABLISHED)

For complete reference here is my kubernetes yml service and deployment specification.

apiVersion: extensions/v1beta1
kind: Deployment
metadata:
  labels:
    service: hadoop
  name: hadoop
spec:
  selector:
    matchLabels:
      service: hadoop
  replicas: 1
  template:
    metadata:
      labels:
        service: hadoop
        run: hadoop
        track: stable
    spec:
      containers:
      - name: hadoop
        image: falcowinkler/hadoop:2.8.5
        imagePullPolicy: Never
        ports:
          # HDFS Ports
          - containerPort: 50010
          - containerPort: 50020
          - containerPort: 50070
          - containerPort: 50075
          - containerPort: 50090
          - containerPort: 8020
          - containerPort: 9000
          # Map Reduce Ports
          - containerPort: 19888
          # YARN Ports
          - containerPort: 8030
          - containerPort: 8031
          - containerPort: 8032
          - containerPort: 8033
          - containerPort: 8040
          - containerPort: 8042
          - containerPort: 8088
          - containerPort: 22
          # Other Ports
          - containerPort: 49707
          - containerPort: 2122
---
apiVersion: v1
kind: Service
metadata:
  labels:
    service: hadoop
  name: hadoop
spec:
  ports:
  - name: hadoop
    port: 9000
  - name: ssh
    port: 22
  - name: hadoop-ui
    port: 50070
  selector:
    service: hadoop
  type: ClusterIP

What is the difference when accessing a port via k8s-dns?

When you call a Pod IP address, you directly connect to a pod, not to the service.

When you call to DNS name of your service, it resolves to a Service IP address, which forward your request to actual pods using Selectors as a filter to find a destination, so it is 2 different ways of how to access pods.

Also, you cal call Service IP address directly instead of using DNS, it will works the same way. Moreover, Service IP address, unlike Pod IPs, is static, so you can use it all the time if you want.

For in-cluster communication you are using ClusterIP service mode, which is default and you set it, so everything is OK here.

Current endpoints where your service forwards requests you can get by kubectl get service $servicename -o wide in an "endpoint" column.

What about your current problems with connection, I can recommend you:

• Check endpoint of your service (there should be one or more IP addresses of pods),

• Set targetPort parameter for each of service ports, e.g:

  apiVersion: v1
  kind: Service
  metadata:
    labels:
      service: hadoop
    name: hadoop
  spec:
    ports:
    - name: hadoop
      port: 9000
      targetPort: 9000 # here is
    - name: ssh
      port: 22
      targetPort: 22 # here is
    - name: hadoop-ui
      port: 50070
      targetPort: 50070 # here is
    selector:
      service: hadoop
    type: ClusterIP
  

P.S. Here is a nice topic with explanation about how Service works. Also, you can check official documentation.