Split Load Using NATS Partitions

 

In this post we will review NATS partitioning and how to use it to split load among multiple pods.

We've already reviewed the steps to setup a NATS cluster in kubernetes in this post. As part of the NATS statefulset template we have a config map, which is mounted into the NATS server container.

- name: nats
  args:
  - --config
  - /etc/nats-config/nats.conf

volumeMounts:
- mountPath: /etc/nats-config
  name: config

volumes:
- configMap:
    name: nats-config

The ConfigMap is as follows:

apiVersion: v1
kind: ConfigMap
metadata:
  name: nats-config
data:
  nats.conf: |
    {
      "cluster": {
        "name": "nats",
        "no_advertise": true,
        "port": 6222,
        "routes": [
          "nats://nats-0.nats-headless:6222",
          "nats://nats-1.nats-headless:6222"
        ]
      },
      "http_port": 8222,
      "lame_duck_duration": "30s",
      "lame_duck_grace_period": "10s",
      "pid_file": "/var/run/nats/nats.pid",
      "port": 4222,
      "server_name": $SERVER_NAME,
      "mappings": {        
        "application.*": "application.{{partition(3,1)}}.{{wildcard(1)}}"
      }
    }

The NATS partitioning is configured by the mapping section in the config. In this case the producer publish messages to the NATS queue "application-<APPLICATION_ID>".

We want to split the load among 3 pods, and hence we configure the following:

"application.*": "application.{{partition(3,1)}}.{{wildcard(1)}}"

This provides the following instruction to the NATS:

Take the value used in the first wild card "application.*" , and use it to split into 3 partitions 0,1,2. Then publish the message in the following queue "application.<PARTITION_ID>.<APPLICATION ID>".

Now we can have 3 pods, each subsribing to the prefix: "application.<PARTITION_ID>.*", and the messages are split accordingly. Notice that a specific APPLICATION ID is always assigned to the same PARTITION_ID.

Also notice that this is a static assignment, regardless of the load on each application. In case we have hightly unbalanced applications load, this might not be a suitable method.