Allow broker's service clusterIP customisation

[MMirelli]

Fixes #497

Motivation

This customisation is useful to configure headless vs non-headless broker's service. The default is headless broker service, i.e. a service for which kubernetes does not allocate an IP address (https://kubernetes.io/docs/concepts/services-networking/service/#type-clusterip). A headless service is a very simple type of service that doesn't seem to work well when pulsar service is exposed by pulsar-proxy via a nodeport.

Modifications

Adding 3 modes to deploy the pulsar broker service:

1. a valid IPv4 address
2. auto mode, i.e. clusterIP: ''
3. headless, i.e. clusterIP: 'None'

Verifying this change

• Run custom verification
• Make sure that the change passes the CI checks.

Custom verification

Mode 1. -- valid IPv4 address

Valid

values.yaml

broker:
  service:
    clusterIP: 255.255.255.255

Running helm diff upgrade pulsar-test ./apache/pulsar-helm-chart/charts/pulsar/ -n pulsar -f ./values.yaml, returns

+   type: ClusterIP
    ports:
    # prometheus needs to access /metrics endpoint
    - name: http
      port: 8080
    - name: "pulsar"
      port: 6650
+   clusterIP: "255.255.255.255"
    selector:
      app: pulsar
      release: pulsar-test
      component: broker

Invalid

values.yaml

broker:
  service:
    clusterIP: 255.255.255.256

Running helm diff upgrade pulsar-test ./apache/pulsar-helm-chart/charts/pulsar/ -n pulsar -f ./values.yaml, returns

+   type: ClusterIP
    ports:
    # prometheus needs to access /metrics endpoint
    - name: http
      port: 8080
    - name: "pulsar"
      port: 6650
+   clusterIP: "None"
    selector:
      app: pulsar
      release: pulsar-test
      component: broker

Mode 2. -- auto

values.yaml

broker:
  service:
    clusterIP: 'auto'

Running helm diff upgrade pulsar-test ./apache/pulsar-helm-chart/charts/pulsar/ -n pulsar -f ./values.yaml, returns

+   type: ClusterIP
    ports:
    # prometheus needs to access /metrics endpoint
    - name: http
      port: 8080
    - name: "pulsar"
      port: 6650
+   clusterIP: ""
    selector:
      app: pulsar
      release: pulsar-test
      component: broker

Mode 3. -- headless

values.yaml

broker:

Running helm diff upgrade pulsar-test ./apache/pulsar-helm-chart/charts/pulsar/ -n pulsar -f ./values.yaml, returns

+   type: ClusterIP
    ports:
    # prometheus needs to access /metrics endpoint
    - name: http
      port: 8080
    - name: "pulsar"
      port: 6650
+   clusterIP: "None"
    selector:
      app: pulsar
      release: pulsar-test
      component: broker

[MMirelli]

Just validated latest changes on my local k3d, we should be good!

[lhotari]

A headless service is a very simple type of service that doesn't seem to work well when pulsar service is exposed by pulsar-proxy via a nodeport.

@MMirelli what problem is showing up in this case?

[lhotari]

LGTM

[MMirelli]

A headless service is a very simple type of service that doesn't seem to work well when pulsar service is exposed by pulsar-proxy via a nodeport.

@MMirelli what problem is showing up in this case?

When the pulsar service is exposed by pulsar-proxy via a NodePort and the broker's service is headless, it seems that the proxy cannot find the broker as no IP address is directly assigned to the broker. A detail is that in my local cluster I am using the advertisedListeners broker config.

Anyway, I might remember the issue wrong. I'll reproduce (perhaps DEBUG logging will give us further insights) and confirm.

[MMirelli]

Actually there doesn't seem to be a problem with broker service being headless. It must have been some other configuration that broke it.

[MMirelli]

Anyway I did notice today, that using non-headless seems to have some advantages. With headless broker service, today I observed re-connections when using pulsar-client, that I didn't see in non-headless mode. It seems to be a configuration to play a bit with :)