Building a Kubernetes PetSet Cluster for your MongoDB ReplicaSet

Some observations when working with Kubernetes 1.3 introduced PetSets to build MongoDB.

Here’s my PetSet YAML:

# mypetset.yml
apiVersion: v1
kind: Service
metadata:
  annotations:
    service.alpha.kubernetes.io/tolerate-unready-endpoints: "true"
  name: a-mongo
  labels:
    name: a-mongo
    tier: database
    type: mongo
spec:
  ports:
  - port: 27017
    name: mongo
  type: NodePort
  selector:
    tier: database
    name: a-mongo
---
apiVersion: apps/v1alpha1
kind: PetSet
metadata:
  name: a-mongo
spec:
  serviceName: "a-mongo"
  replicas: 3
  template:
    metadata:
      labels:
        name: a-mongo
        tier: database
      annotations:
        pod.alpha.kubernetes.io/initialized: "true"
    spec:
      terminationGracePeriodSeconds: 10
      containers:
      - name: mongo
        image: mongo:3.0.12
        command:
          - mongod
          - "--replSet"
          - rs0
        imagePullPolicy: Always
        ports:
        - containerPort: 27017
          name: mongo
        resources:
          requests:
            cpu: "0.1m"
            memory: "64Mi"
          limits:
            cpu: "1"
            memory: "1Gi"
        volumeMounts:
        - name: data
          mountPath: /data/db
  volumeClaimTemplates:
  - metadata:
      name: data
      annotations:
        volume.alpha.kubernetes.io/storage-class: fo
    spec:
      accessModes: [ "ReadWriteOnce" ]
      resources:
        requests:
          storage: 10Gi

If you’re on AWS, or GCE, your mongo instances are backed by persistent storage (ElasticBlockStore, GCEDisk) that will ensure that even if one of your pets die, when its revived, it will reuse the same storage.

So with your config_file, lets start initializing your petset.

$ kubectl create -f mypetset.yml

Give it a while and you should see 3 pods appear.

Note: In PetSets, pods are started one after another so if one of your pets fails startup, any other pets in the order that hasn’t been started will not be started.

$ kubectl get pods

a-mongo-0                       1/1       Running   0          22m
a-mongo-1                       1/1       Running   0          21m
a-mongo-2                       1/1       Running   0          21m

Ok looks like we’re done! Let’s initiate the cluster.

First, lets get all the available IP of our cluster.

$ kubectl describe pods a-mongo | grep IP

IP:             10.244.6.3
IP:             10.244.7.3
IP:             10.244.6.4

Awesome. Now keep this somewhere handy.

$ kubectl exec -it a-mongo-0 /bin/bash

[email protected] $ mongo
> rs.initiate()

One of the biggest caveat in setting up your mongo replicaset on PetSets is that you need to perform an extra step before adding your other cluster. We want to do this before we add members so that we don’t run into an issue where we lose our PRIMARY node due to configuration.

> cfg = rs.conf()
> cfg.members[0].host = "10.244.6.3:27017" # replace this with the cluster IP of your a-mongo-0 container.
> rs.reconfig(cfg)

Now you can safely add your members.

Note: So what happened? Apparently MongoDB ReplicaSets do not play nice if one of its members is inconsistent in its naming. ie. If you’re using IP addresses, all your cluster members should be using it. Reference

rs0:PRIMARY> rs.add("10.244.7.3")
rs0:PRIMARY> rs.add("10.244.6.4")

If you do a rs.status(), you should see the other 2 members listed as SECONDARY.

And now you have a full cluster!

References

Google Groups - Could not find member to sync from

Mongo Secondaries Stuck at Startup State