Kubernetes Learning Notes - Part 1 - Deployment

Welcome to the first part of my Kubernetes learning notes series. In this blog post, I’m going to record my learning experience for deploying the overmind web service to a Kubernetes cluster.

The overmind web service is a simple and contrived microservice for managing zerglings. Don’t worry if you’re not a starcraft fan. The details of the web service doesn’t really matter. The web service has a couple of endpoints, some of which involve talking to a database.


First, we need to install kubectl - the Kubernetes client CLI tool. Please follow the instruction here.

Then we will need to setup a local testing cluster. For that, we use minikube - a tool that spins up a single-node kubernetes instance on your local machine via virtualbox (or other hypervisor). Please refer to the minikube README for how to set it up and run it locally.

After you downloaded the minikube binary and put it in your $PATH, run minikube start:

Starting local Kubernetes v1.6.0 cluster...
Starting VM...
SSH-ing files into VM...
Setting up certs...
Starting cluster components...
Connecting to cluster...
Setting up kubeconfig...
Kubectl is now configured to use the cluster.

Now your “cluster” is up and running and your kubectl is configured to use the minikube context:

$ kubectl get node
minikube   Ready     3d        v1.6.0

Pods, Replica Sets, Deployments

Pods, replica sets and deployments are core concepts involved in deploying a container to our cluster.


A Pod is the most basic building block of a Kubernetes. A pod represents a running process on your cluster. In container terms, a pod is one or more containers the need to work together. They containers in the same pod share the same networking namspace, which means container A can talk to container B in the same pod via localhost. In the overmind service example, we will be deploying the overmind docker container as a pod in the cluster.

Replica Set

A Replica Set is a kubernetes “controller” that ensures at any given time, the correct number of “replicas” of a pod is present in the cluster. For example, if I declare that the overmind pod has a replica of 3. The replica set controller makes sure that there are 3 overmind pods running in the cluster. If for some reason, one pod died, the controller will spin up another pod in the cluster to bring the total number of replicas to 3.


A Deployment declares the desired state for pods or replica sets. For example, you can declare that the deployment should use the rolling upgrade strategy so we can have zero-downtime deployment.

Let’s start by creating a deployment manifest for our overmind service.

Deployment Manifest

Let’s start writing a deployment manifest for our service.

Save the following as overmind-deployment.yaml.

apiVersion: extensions/v1beta1
kind: Deployment
    name: overmind
    replicas: 3
                app: overmind
                tier: web
                - name: overmind
                  image: kevinjqiu/overmind:1
                      - name: OVERMIND_HTTP_ADDR
                        value: ""

The yaml file here describes the Deployment object. The deployment object declares that we want 3 replicas of the overmind pod, and the pod definition is encoded below in the template section. The schema for template is the same as the schema for pod.

Here we declare that each pod will include a single container using the kevinjqiu/overmind:1 image. We also specify the environment variable for our pod which defines the bind address for our service.

Start Deployment

After we have the deployment manifest, let’s start a deployment by using the kubectl command:

$ kubectl apply -f overmind-deployment.yaml
deployment "overmind" create d

Let’s verify the deployment object is correctly created:

$ kubectl get deployment
overmind   3         3         3            3           1m

The deployment manifest will also create another two types of objects as discussed above: pods and replica sets. Let’s verify that:

$ kubectl get pods
NAME                       READY     STATUS    RESTARTS   AGE
overmind-581439581-h3f3f   1/1       Running   0          3m
overmind-581439581-hbq0g   1/1       Running   0          3m
overmind-581439581-szkk5   1/1       Running   0          3m

$ kubectl get rs
NAME                 DESIRED   CURRENT   READY     AGE
overmind-581439581   3         3         3         4m

So far so good. As you can see, we have 3 replicas of the pod running in the cluster. Each pod does have its own ip address. They can be shown using kubectl get pods -o wide command:

$ kubectl get pods -o wide
NAME                       READY     STATUS    RESTARTS   AGE       IP           NODE
overmind-581439581-h3f3f   1/1       Running   0          6m   minikube
overmind-581439581-hbq0g   1/1       Running   0          6m   minikube
overmind-581439581-szkk5   1/1       Running   0          6m   minikube

From your host, there isn’t a route to get to those ip addresses but they are reachable on the hosts in the cluster.

To prove that, let’s ssh into the minikube instance:

minikube ssh

Try curling one of these ip addresses:

$ curl

So there it is. Our pod is up and running correctly.


Our pods may have been deployed correctly in the cluster, but there are a couple of problems: * the pods are not load balanced * if we setup load balancer externally, and if pods die and the replica set (controller) spawns a new pod, how are we to have the load balancer update its backends?

This is where service comes in. A service is an abstraction which defines the access to a logical group of pods. Defining a service will give the pods a unified (virtual) IP address, distribute the load and watches for pod changes so it can add/remove pods from its available backends.

Service Manifest

Let’s go ahead and make a service manifest. Create a file overmind-service.yaml:

apiVersion: v1
kind: Service
    name: overmind
        app: overmind
        tier: web
        app: overmind
        tier: web
    type: NodePort
        - port: 8080

A service uses selector to collect the pods and load balance them. In the deployment manifest, we declare that the overmind pods have the labels app=overmind and tier=web. Here, we use this selector in the selector section of the spec to declare that the service should select these pods.

We also specify type: NodePort. This defines the way we “publish” our service. NodePort is a simple mechanism that each node in the cluster will proxy the same port into the service. The other allowed type is LoadBalancer which allows you to integrate Kubernetes with a Cloud Provider and setup load balancer in the cloud (e.g., ELB).

Create the Service

Let’s apply the manifest and test it out.

$ kubectl apply -f overmind-service.yaml
service "overmind" created

Verify that the service exists:

$ kubectl get svc -o wide
kubernetes     <none>        443/TCP          6h        <none>
overmind    <nodes>       8080:30674/TCP   8m        app=overmind,tier=web

Keep in mind that these ip addresses are still cluster-wide. There isn’t a route to those ip addresses that you can curl with. To allow traffic from outside of the container, we need to setup Ingress but that’s a topic for later. Fortunately, minikube has a convenient way for us to allow traffic into the container. Use the following command:

$ minikube service --url overmind

Now we can use this url to access the overmind service deployed on Kubernetes:

$ curl $(minikube service --url overmind)/_health


In this blog post, we’ve gone through the basics of how to deploy to a Kubernetes cluster using deployments with pods and replica sets. We also used service to abstract the access of these pods.

You may notice in the above curl output that it says the brain of overmind is “damaged”. This is due to the fact that we haven’t deployed the “brain” - CouchDB to the cluster yet. Deploying CouchDB (a stateful service) involves container volumes, persistent volumes and claims, stateful sets and a little bit of secret management. Stay tuned for the next blog post where I’ll be writing about my learning experience with these topics. Ciao!

comments powered by Disqus