Exploring Kubernetes

With cloud native and microservice architectures gaining traction, Kubernetes (k8s) has become the standard tool for managing deployments. But what is it, do I need it, and how do I most effectively get started with it? That's what this post aims to clarify.

I'm no k8s expert. I've been picking it up because I'm interested in the devops space and because I see the problem domain it solves in my daily work. I've found the best way to learn something is to simply start working with it. Even better, is to write about it, as writing reinforces what you learned. If you can't explain something clearly, then you don't really understand it.

In this series of posts we'll develop a basic expressjs server, use k8s to develop locally and deploy it to a production environment. We'll take it step-by-step. After the first post we'll have an expressjs server running and be able to deploy it via k8s to a development environment. In further posts we'll explore local development, secret management, production clusters and stateful components, like databases.

This post assumes a basic knowledge of building and working with containerized applications. Maybe you've played with them on a toy project, or maybe you're read a lot about them. Regardless you understand the basic concepts of Docker or Docker-like tools. This won't be a container tutorial.

What is K8s & What Does it Solve?

k8s is a container orchestrator. It can declaratively manage the configuration, running state, and deployments of a cluster of containerized applications. If you've ever had to tell a client that you need to bring an app down for maintenance, tried to do rolling updates, do A/B testing, scale your app, or recover from your app going down then you understand the pain k8s solves.

With k8s you define the state you want (say three instances of your app), then tell k8s that you want to deploy a new image. k8s handles terminating the existing containers, starting up the new ones, while always keeping the desired number of container instances running. k8s maintains this state, not just when you explicitly update the application, but also when a container terminates unexpectedly.

Do You Need It?

If any of the following applies, you may want to explore k8s. Your app:

  • requires close to 100% up-time
  • has many users, or you project it will
  • has users who are very active at certain times, but not others
  • requires A/B testing

Like any framework, k8s adds complexity, and there's a learning curve. If that's not appealing, then you could still containerize your app, run the image on a VM, stop it when you need to update it, pull a new image from an image repository, and restart the service.

Even if you don't need it, it may be worth exploring k8s because it's become such a ubiquitous and influential tool in devops.

Let's Go

We're going to take a quick tour of k8s basics by building a toy app. We'll build an Expressjs server, containerize it and run it in a local k8s cluster. A cluster (very simply) is a group of "pods" running at least one container.

We will need the following tools:

The App

The server we'll build is just one route. Create a new directory called hello-k8s. cd into it. Ensure you have a recent version of node installed locally, or use nvm to install an isolated version of node. The latest stable version of node will do. Once that's installed do:

$ npm init
...
$ npm install express --save

Now copy this JavaScript to a file called server.js in the root of the directory:

const process = require("process");
const express = require("express");
const app = express();

app.get("/", (req, res) => {
  res.send(`<h1>Kubernetes Expressjs Example</h2>
  <h2>Non-Secret Configuration Example</h2>
  <ul>
    <li>MY_NON_SECRET: "${process.env.MY_NON_SECRET}"</li>
    <li>MY_OTHER_NON_SECRET: "${process.env.MY_OTHER_NON_SECRET}"</li>
  </ul>
  `);
});

app.listen(3000, () => {
  console.log("Listening on http://localhost:3000");
});

Run: $ node ./server.js and go to http://localhost:3000 to ensure the app runs. The "non secret" data should render as "undefined" for now.

Containerize It

Now let's containerize our app. Create a dockerhub account if you don't have one already. This is where we will pull and push images from and to.

We'll create a Dockerfile at the root of the directory, which as you know, is a declarative way to define how your app's image should be built. I've added a few comments that may help you avoid common performance/security pitfalls when creating Dockerfiles:

# Use explicit version of node, not LTS
from node:14.17.3

# Create a non-priviliged user to run the process (server.js).
# The code should be owned by root, not world-writable, but run as the non-root user.
RUN useradd -ms /bin/bash appuser && mkdir /code && chown -R appuser /code

# Dependencies are less likely to change than app code
# and are slow to install so create layers relating to npm early
# and the actual app code which will change frequently later.
COPY package.json /code/
WORKDIR /code
RUN npm install

COPY server.js /code/

# Run the main process as the less priviliged user.
USER appuser
CMD ["node", "server.js"]

Let's build it: $ docker build -t {DOCKER_HUB_USERNAME}/hello-k8s:latest .

And now run it: $ docker run -d -p 3000:3000 {DOCKER_HUB_USERNAME}/hello-k8s:latest. Go to http://localhost:3000, and you should see the app running. Again, the non secrets will print out as null. Don't worry about that for now.

Now let's stop the running container. Do docker ps to get its ID, and stop it: $ docker stop {CONTAINER_ID}.

We've accomplished a few things by containerizing the app:

  1. Our environment is completely reprodicible, including the OS the app is run on. If, for example, we need to configure the OS specifically for the app (say add some C libraries), we can do that in the Dockerfile and exactly reproduce the app's environment on another developer's machine, in a CI build, or in production.
  2. This allows us to much more confidently spin up/down our app because we can quickly reproduce it dependencies from a common baseline.

Pushing the Image

Because we don't want to have to build the image on target machines, once we build the image we'll want to push the image to dockerhub. Let's now push the image you created to your account: $ docker push {DOCKER_HUB_USERNAME}/hello-k8s:latest. You may have to run: $ docker login first.

Move to k8s

We could stop here and simply build/tag our image after each change to the app, pull it down on the host machine, and run it using docker or docker compose (if you also want to run a containerized database). As I mentioned above, however, there are some major disadvantages to this workflow, namely:

  1. A lot of manual steps
  2. App downtime while you are bringing it down and spinning it up again
  3. This won't work easily if you want to scale your app to multiple machines

So, let's use k8s to spin up a few instances of our container. But first some quick k8s basics.

k8s Architecture

Let's discuss a few major k8s concepts, including k8s objects. First, a cluster. A cluster contains:

  • a set of node machines, which contain pods. Pods are the atomic unit of k8s and contain one or more images. A node could be a real machine or a VM. The pod is an abstraction for a worker machine.
  • A control plane, which is responsible for maintaining the desired state of the pods.

Generally we use kubectl to interact with our cluster, which is a CLI for communicating via the cluster's API. Setting up an on-premise k8s cluster is complex; it's usually only done when there's a specific requirement to run on premise, perhaps for security reasons. It's usually much easier to run a local cluster for development/testing/CI using a tool like Docker Desktop, minikube, kind etc. The local k8s cluster tool you choose depends on your requirements. For this post we are using minikube. In production most people will use a turn-key solution, like Google Kubernetes Engine (GKE).

k8s is modular. We work in terms of objects. For example, our app might include the following k8s objects:

  • pods
  • deployments
  • services
  • secrets
    etc.

We can create these imperatively using kubectl, or declaratively with yaml files (manifests) and simply point kubectl at the manifests. The big advantage of declarative configuration is we can commit these manifests to source control, and it becomes very easy for another developer, a CI system, or a production system to re-recreate the state of the cluster.

Creating a Deployment

We can create the pods that will run our express app directly, but because we want to deploy these pods, instead we will write a deployment manifest that describes how to deploy these pods. The pods' spec are handled by the deployment manifest, so we don't have to define the pods separately.

Let's start up minikube, which creates a local cluster:

$ minikube start
😄  minikube v1.22.0 on Darwin 11.2.3
✨  Using the docker driver based on existing profile
👍  Starting control plane node minikube in cluster minikube
🚜  Pulling base image ...
🤷  docker "minikube" container is missing, will recreate.
🔥  Creating docker container (CPUs=2, Memory=1987MB) ...
🐳  Preparing Kubernetes v1.21.2 on Docker 20.10.7 ...
🔎  Verifying Kubernetes components...
    ▪ Using image kubernetesui/dashboard:v2.1.0
    ▪ Using image gcr.io/k8s-minikube/storage-provisioner:v5
    ▪ Using image kubernetesui/metrics-scraper:v1.0.4
🌟  Enabled addons: storage-provisioner, default-storageclass, dashboard
🏄  Done! kubectl is now configured to use "minikube" cluster and "default" namespace by default

Ensure the k8s context is set to minikube. This ensures we are using our local cluster created by docker-desktop:

$ kubectl config get-contexts
CURRENT   NAME             CLUSTER          AUTHINFO         NAMESPACE
          docker-desktop   docker-desktop   docker-desktop   
          kind-mycluster   kind-mycluster   kind-mycluster   
*         minikube         minikube         minikube         default

Your terminal output will likely be different than mine. If minikube is not set as your context, then do:

$ kubectl config use-context minikube
Switched to context "minikube".

Now create a manifests directory (just to keep things organized) and create manifests/web-deployment.yaml:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: web-deployment
spec:
  replicas: 2
  selector:
    matchLabels:
      app: web-pod
  template:
    metadata:
      labels:
        app: web-pod
    spec:
      containers:
      - name: web
        image: kahunacohen/hello-k8s:latest
        ports:
        - containerPort: 3000 # The port the containers are running on internally.
          protocol: TCP

Use kubectl to create the deployment using the manifest::

$ kubectl create -f manifests/web-deployment.yaml`
deployment.apps/web-deployment created

In essence the deployment object is a wrapper for two lower level objects, pods and ReplicaSets, which you could create separately. In our case the deployment object we created implicitly creates two pods and a ReplicaSet. The ReplicaSet's job is to maintain a set of replica pods.

Now that we created our deployment, we can list the pods:

$ kubectl get pods
web-deployment-5bb9d846b6-m5nvk   1/1     Running     0          2m23s
web-deployment-5bb9d846b6-rv2kt   1/1     Running     0          2m23s

This tells us there are two pods, which are owned by web-deployment. If we inspect one of them we should see our container (among other details):

$ kubectl inspect pods web-deployment-5bb9d846b6-m5nvk
...
Containers:
  web:
    Container ID:   docker://c5e24583fa6942f8d4f791281bbf756d06add13d55f52115adb2f659b7ad48b6
    Image:          kahunacohen/hello-k8s
    Image ID:       docker-pullable://kahunacohen/hello-kube@sha256:25b2e36992c67eddb6604748c668873e750463517debeea42c870cedac182cf1
    Port:           3000/TCP
...

We can exec into the pod, just like we would using docker (don't forget to replace the pod name with one of the pod names you got from doing kubectl get pods):

$ kubectl exec --stdin --tty web-deployment-5bb9d846b6-m5nvk -- /bin/bash
root@web-deployment-5bb9d846b6-m5nvk:/code# ls /code/
node_modules  package-lock.json  package.json  server.js

Creating a Service

Great, now how do we view our app? Currently the nodes are exposed inside the cluster at ephemeral IPs. We need a service object, which is a k8s object that abstracts exposing a set of pods on the host network. Create manifests/web-service.yaml:

apiVersion: v1
kind: Service
metadata:
  name: web-service
spec:
  type: NodePort
  selector:
    app: web-pod
  ports:
    - port: 80
      targetPort: 3000
      protocol: TCP
      name: http

and

$ kubectl create -f manifests/web-service.yaml

and finally:

$ kubectl get services
NAME          TYPE        CLUSTER-IP      EXTERNAL-IP   PORT(S)        AGE
kubernetes    ClusterIP   10.96.0.1       <none>        443/TCP        10d
web-service   NodePort    10.107.232.55   <none>        80:30543/TCP   76s

You can find a deeper explanation of the service manifest here. The NodePort type is one of several options including:

  1. ClusterIP: useful for exposing pods internally within the cluster.
  2. LoadBalancer: useful in a production environment, such as AWS or GCE.

The NodePort type we are using is mostly used in development. It forwards requests from outside the cluster to inside and is randomly chosen between a range of ports. The important thing to note is this sequence of requests this service allows. In our case a request comes in to our docker desktop cluster at:

  1. localhost:30543 ...which forwards requests to:
  2. ...our service to an internal IP at port 80. The service selects pods as per the spec and forwards to:
  3. IP:3000, where our container is running.

This tells us that the NodePort 30543 on localhost is forwarding to port 80, where are service is. So if we go to: http://localhost:30543, we should see our express app. Except we don't! If we were using docker desktop as our local cluster, we could, but minikube is a bit more complex in that it more closely resembles a production cluster (it allows load balancers etc.). With minikube we have one more bit of redirection to see our app because we have to tunnel into its VM. Do:

$ minikube service web-service

This should tunnel in and open up a tab where the app renders.

Bring Down a Pod

Let's test the ReplicaSet that the deployment implicitly created by killing a pod and see what happens:

$ kubectl get pods
web-deployment-fdc6dddb7-9vb9s   1/1     Running       0          5h18m
web-deployment-fdc6dddb7-tnhzd   1/1     Running       0          5h18m

Indeed, two pods are running. Kill one:

$ kubectl delete pods web-deployment-fdc6dddb7-9vb9s

We can see the pod we killed is in the process of terminating, and there's a new pod already running to take its place. That's the ReplicaSet in action: maintaining the desired state of the cluster. Now a bit later we'll see the pod we killed is gone, replaced entirely by a new one:

kubectl get pods
web-deployment-fdc6dddb7-n4pf8   1/1     Running             0          2m27s
web-deployment-fdc6dddb7-tnhzd   1/1     Running             0          5h22m

Non-sensitive Run-time Data

A great place to keep non-sensitive, run-time configuration is in environment variables. Traditionally we've set them when the OS starts up, perhaps in a .env file deployed along with the app.

It's import to understand the security implications of environment variables. While setting sensitive data in environment variables may be more secure than hard-coding them in source code, it's still bad practice because they are easily leaked via logs, and/or third-party dependencies could read them and "phone home." We'll discuss managing sensitive data later in k8s, but just know that for now we are only going to store non-sensitive data in environment variables for our run-time configuration.

If you wanted to set env vars for our project, you could explicitly create a pod manifest file and set them there. A better way may be to use the ConfigMap object. This decouples your
configuration from your pods/images and allows them to be more easily re-used.

Create a web-configmap.yaml file and put it in manifests:

apiVersion: v1
data:
  MY_NON_SECRET: foo
  MY_OTHER_NON_SECRET: bar
kind: ConfigMap
metadata:
  name: web-configmap
  namespace: default

Now do:

$ kubectl create -f manifests/web-configmap.yaml 
configmap/web-configmap created
$
$ kubectl describe configmap web-configmap
Name:         web-configmap
Namespace:    default
Labels:       <none>
Annotations:  <none>

Data
====
MY_NON_SECRET:
----
foo
MY_OTHER_NON_SECRET:
----
bar
Events:  <none>

Great, so now our config data is created as just another k8s object. There are several ways to consume this data from our pods, one of which is via env vars. Let's add the envFrom block to our web-deployment.yaml to map the pod to the configmap:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: web-deployment
spec:
  replicas: 2
  selector:
    matchLabels:
      app: web-pod
  template:
    metadata:
      labels:
        app: web-pod
    spec:
      containers:
      - name: web
        image: kahunacohen/hello-k8s
        envFrom:
          - configMapRef:
              name: web-configmap
        ports:
        - containerPort: 3000
        protocol: TCP

Then do:

 kubectl replace -f manifests/web-deployment.yaml
 deployment.apps/web-deployment replaced

Here we use the replace command to update the deployment in-place. There is no service
outage because, even while k8s updates the deployment, it always maintains the desired state of two pods running. If you check the pods with kubectl get pods you can see how the state of the cluster changes until the new pods are running and the old ones are terminated.

Now when you go to localhost:{NodePort} you should see:

  • MY_NON_SECRET: "foo"
  • MY_OTHER_NON_SECRET: "bar"

Jobs

A very useful k8s object is jobs, specifically cronjobs, which are merely scheduled jobs.

Jobs are created like any other object in k8s. Every job is run in a new pod, and they stick around. You can customize how many pods are kept etc. in the manifest.

Let's try it. Create manifests/print-hello.yaml:

apiVersion: batch/v1
kind: CronJob
metadata:
  name: print-hello
spec:
  schedule: "*/5 * * * *"
  successfulJobsHistoryLimit: 2
  failedJobsHistoryLimit: 2
  jobTemplate:
    spec:
      template:
        spec:
          containers:
          - name: print-hello
            image: kahunacohen/hello-kube:latest
            imagePullPolicy: IfNotPresent
            command:
            - /bin/sh
            - -c
            - date; echo Hello from the Kubernetes cluster
          restartPolicy: OnFailure

This will create a pod every 5 minutes, printing out the date and the message as specified in the command block. It will only keep the two most recent pods--the rest will be destroyed.

Let's create it:

$ kubectl create -f manifests/print-hello.yaml
$
$ kubectl get jobs
NAME             COMPLETIONS   DURATION   AGE
hello-27114520   1/1           9s         80s

We can see a pod was created for this job:

$ kubectl get pods
NAME                                                READY   STATUS       RESTARTS   AGE
hello-27114520-4wnmf                      0/1        Completed   0               75s
web-deployment-65b8bccdfd-bb54g  1/1        Running       1               18h
web-deployment-65b8bccdfd-mfwx8  1/1       Running       1               18h

And we can ask the pod for logs so we can see the output:

$ kubectl logs hello-27114520-4wnmf
Wed Jul 21 12:40:09 UTC 2021
Hello from the Kubernetes cluster

Very cool!

Updating App Code

When we updated the deployment with the ConfigMap we saw that there was no down-time. k8s maintained the desired state of the pods. Once the deployment was replaced, the config variables became defined and were rendered.

But how do we actually update the app, not just the deployment? A few things trigger k8s to redeploy your actual app. The main way we trigger this is by setting a new image via kubectl's set image command. The workflow is:

  1. Make a change to your app code.
  2. Rebuild a new image, tagging it with a new version. Set the latest tag to the same image as the newly tagged image.
  3. Push both the newly tagged image and latest to an image repo.
  4. Call kubectl set image to set the deployment to the new image.

Let's see this in detail:

  1. In server.js, change the header "Kubernetes Expressjs Example" to "Kubernetes Expressjs Example 123". If you go to localhost:{NODE_PORT}, you won't see your change because the deployment is still using image tagged "latest" that it pulled from dockerhub.
  2. Now, build a new image and tag it with both your tag and "latest":
    $ docker build -t {DOCKER_HUB_USERNAME}}/hello-k8s:latest \
    -t {DOCKER_HUB_USERNAME}/hello-k8s:0.1 .
  3. Push the tags:
    docker push {DOCKER_HUB_USERNAME}/hello-k8s:latest && \
    docker push {DOCKER_HUB_USERNAME}/hello-k8s:0.1
  4. Now we need to set the existing deployment's image to the new tag: 0.1, selecting
    the container labeled "web" to replace:
$ kubectl set image deployment/web-deployment web=kahunacohen/hello-k8s:0.1
deployment.apps/web-deployment image updated

Now if we do kubectl get pods we should see that k8s is resolving the current state to the desired state, terminating the existing pods, while creating new ones with the new container. After some time, you should be able to refresh localhost:{NODE_PORT} and see your change.

The obvious problem in this development workflow is the manual steps and the time it takes between when you make a change to application code and when you can see the change. That is the topic of the next (soon to be written) post.