Azure Functions has become my standard framework to build serverless backends for Single Page Applications (SPAs). It’s easy to get started with Azure Functions, although the platform itself is really powerful and highly configurable. I really love the cloud hosting capabilities for Azure Functions, but from time to time, it’s not allowed to use the public cloud as the deployment target.

Azure Functions can easily be hosted in different environments. You can download the entire runtime and execute it on-premises (if you have to) or you can use the existing Docker image and deploy your serverless backend to Kubernetes. And that’s exactly what this post will describe, you’ll create a sample Azure Functions project including a docker container, deploy it to a Kubernetes cluster and let it automatically scale based on CPU utilization.

What do you need

In order to follow the instructions in this article, you need several things installed and configured on your system.

Once you’ve installed and configured those tools, you’re set and it’s time to get started.

Creating the Azure Functions Sample

Azure Functions Core Tools allow you to spin up new Azure Functions projects quickly. The Command Line Interface (CLI) accepts various arguments. For demonstration purpose, we’ll just use the --sample argument which will create a small Azure Function which receives a name (string) from the QueryString using an HTTP trigger and produces a new HTTP Response. As second argument we’ll use --docker which instructs CLI to generate a Dockerfile for our new Azure Functions project.

mkdir azfunction-on-k8s && cd azfunction-on-k8s

func init --sample --docker

That’s it. Our new Azure Functions project has been created, including the required Dockerfile.

Create the Docker image

I’ll use public Docker Hub for now, but I’ll also work with your ACR service running in Azure. So let’s build the docker image by leveraging the generated Dockerfile. Remember to tag your image properly – depending on your registry.

docker build -t <%YOUR_REPOSITORY%>/azfunction-on-k8s .

Docker will now instruct the service to pull all required filesystem layers and follow the instructions from the Dockerfile to build the image. Once the command has finished, you will find the new image when executing docker images, docker automatically assigns the latest tag to the image if you don’t specify a tag manually.

Push the Docker image to the registry

In order to push the docker image to your registry, you have to be authenticated. Once authenticated (by following the instructions from docker login), you can push the image to the registry.

docker login
docker push <%YOUR_REPOSITORY%>/azfunction-on-k8s

You can visit the website of your registry (eg. or the Azure Portal to verify the existence of your image.

Connect to your Kubernetes cluster

First, let’s connect to the correct Kubernetes cluster. kubectl has a concept of contexts, each unique Kubernetes cluster has a context. To examine the contexts known by your kubectl installation and set the context to use your favorite cluster.

kubectl config get-contexts

You’ll receive a list of registered contexts, where one of those is marked as current context (in the first column).

Contexts for kubectl

We will be using minikube for this post, so let’s ensure minikube is up and running and change the kubectl context to minikube.

Preparing minikube to support autoscaling

Minikube is a lightweight Kubernetes cluster for usage on your local development machine. That said, by default, not all required features are activated to use autoscaling on minikube. Minikube offers some so-called add-ons. In order to use autoscaling based on metrics, the heapster add-on needs to be enabled. You can get a list of all add-ons by executing

minikube addons list

Minikube add-ons with disabled heapster

As you can see, heapster is disabled, let’s enable it now.

minikube addons enable heapster

It’s important to enable heapster before deploying our docker image yo the cluster

Minikube add-ons: heapster enabled

Deploying to the Kubernetes cluster

Let’s deploy our docker image to the k8s cluster. Normally you would create a deployment using a simple yaml file, but for demonstration purpose, we’ll just use the kubectl command for now. For real deployments, you always want to use a proper yaml deployment definition.

kubectl run azure-function-on-kubernetes

This creates a new Kubernetes deployment named azure-function-on-kubernetes based on your docker image. It’ll expose the port 80. By specifying --requests=cpu=100m, we request 100 millicpu for our container. So it’s 0,1 actual CPU. (Further details on resource management for containers and pods can be found here).

Once your deployment has the desired state (check it by querying kubectl get deploy), we need to expose it using a service. Normally you would use a service of type LoadBalancer, but allocating external IP addresses (that’s what the LoadBalancer also has to do…) is only implemented by (private and public) cloud vendors. As an alternative, we’ll use a service of type NodePort and query the url manually from minikube.

kubectl expose deployment azure-function-on-kubernetes --type=NodePort

Wait for the service to be created using kubectl get svc -w (stop watching using CTRL+C) and grab the url.

minikube service azure-function-on-kubernetes --url

Minikube will immediately print the URL where you can access the Azure Function at. Give it a try, open a browser and navigate to the url, first you should see the Azure Function welcome page. In order to access the actual function, you’ve to append the path to the function which is /api/httpfunction/?name=<%YOUR_NAME%>. So in my case, the absolute url is

The Sample Azure Function has been invoked successfully

Enable Autoscaling

The funny part of this article is of course autoscaling. We all want to see Kubernetes taking care of our containers and pods and dynamically scale the application based on actual load. πŸ€˜πŸΌπŸš€

HorizontalPodAutoscaler is also a first citizen object in Kubernetes which can also be described in yaml for real deployments, you definitely want to create a yaml file instead of using the corresponding kubectl command.

kubectl autoscale deploy azure-function-on-kubernetes

This command tells Kubernetes to scale the number of pods up to 30. If the CPU utilization hit’s 30% new pods will automatically be deployed. Since version 1.6.0 of Kubernetes you can also configure delays for down- and upscaling. By default Kubernetes will delay downscaling by 5 minutes and upscaling by 3 minutes. To change the default just append --horizontal-pod-autoscaler-downscale-delay=10m0s and --horizontal-pod-autoscaler-upscale-delay=1m0s.

Let’s generate some load

Now that everything is in place, it’s time to generate some load. There are plenty of different ways how to achieve this. The easiest way is to just use a while command right inside of the terminal like shown below or you can follow the instruction in the official k8s docs and create a small pod, which will issue HTTP requests as long as the pod is running.

Option 1: Generate load from the terminal

Open a new instance of your favorite terminal (BTW: I prefer iTerm) and execute the following command (each line committed by RETURN)

while true
wget -q -O-]/api/httpfunction/?name=Thorsten

You can cancel the requests by simply hitting CTRL+C

Option 2: Generate load using a pod

Open a new instance of your favorite terminal. As described in Kubernetes docs, you can use the busybox image with interactive tty to generate a Pod which will put some load on the Azure Function. Spinning up this pod using kubectl is straightforward.

kubectl run -i
  --tty azure-fn-loadgenerator
  --image=busybox /bin/sh

# Hit RETURN for command prompt

while true; do wget -q -O-]/api/httpfunction/?name=Thorsten; done

If you want to stop the pod, remember that kubectl run will always create a deployment. That said, you’ve to exit the containers prompt and delete the deployment using kubectl delete deploy azure-fn-loadgenerator.

Inspect the cluster

Keep the load generator running in the background and move to the second terminal instance or tab. If you’re using an Azure Container Services Kubernetes cluster, you can use kubectl get horizontalpodautoscalers or the shortcut kubectl get hpa to see the utilization and the ongoing scaling. Unfortunately, this doesn’t work – at least with my – minikube installation.

But there is another way how to see the autoscaling action. You can have a look at the Kubernetes dashboard. Just execute kubectl proxy and open the reported URL in the browser. Depending on your Kubernetes version you may have to manually add /ui to the URL in order to access the dashboard.

Inside the Kubernetes dashboard you find all information about your k8s cluster. As you can see in the picture below, Kubernetes already scaled my deployment, I’ve right now 10 instances of my Pod running and dealing with the incoming load. If you pay attention to the Age column, you see that those pods have different ages.

Kubernetes cluster with autoscaled deployment

Stop the Load and verify downscaling

Stop the load generation as described above. Depending on your custom downscale delay it’ll take some time until Kubernetes will kill some of your pods. Navigate to the k8s Dashboard and see the pods disappearing πŸ˜€

Kubernetes cluster with downscaled deployment


Kubernetes autoscaling is really powerful and from a developers perspective it makes fun using it. It’s just amazing to see Kubernetes spinning up a new instance of your pods to enforce the desired state. I hope you enjoyed the read and got some new ideas.