My Profile Photo

Stefan Prodan

DevOps Consultant. DX at Weaveworks. Passionate about Cloud Native tech. Loves programming in Go, JS and .NET

Installing Kubernetes on bare-metal with Terraform Scaleway provider and kubeadm

This is a step by step guide on setting up Kubernetes on Scaleway bare-metal ARM and x86-64. The main reason I’ve been working on this project is that I wanted to automate the creation of test environments for OpenFaaS and Weave Net on ARM. I was looking for a cheap solution to run integration tests and after trying out several cloud providers I’ve settled on Scaleway. Scaleway is a french cloud provider that offers bare-metal ARM and x86-64 servers at affordable prices. Using Terraform Scaleway provider along with kubeadm you can have a fully functional Kubernetes cluster in ten minutes.

Initial setup

Clone the repository and install the dependencies:

$ git clone
$ cd k8s-scw-baremetal
$ terraform init

Note that you’ll need Terraform v0.10 or newer to run this project.

Before running the project you’ll have to create an access token for Terraform to connect to the Scaleway API. Using the token and your access key, create two environment variables:



Create an ARMv7 bare-metal Kubernetes cluster with one master and two nodes:

$ terraform workspace new arm

$ terraform apply \
 -var region=par1 \
 -var arch=arm \
 -var server_type=C1 \
 -var nodes=2 \
 -var weave_passwd=ChangeMe \
 -var k8s_version=stable-1.9 \
 -var docker_version=17.03.0~ce-0~ubuntu-xenial

This will do the following:

  • reserves public IPs for each server
  • provisions three bare-metal servers with Ubuntu 16.04.1 LTS
  • connects to the master server via SSH and installs Docker CE and kubeadm armhf apt packages
  • runs kubeadm init on the master server and configures kubectl
  • downloads the kubectl admin config file on your local machine and replaces the private IP with the public one
  • creates a Kubernetes secret with the Weave Net password
  • installs Weave Net with encrypted overlay
  • installs cluster add-ons (Kubernetes dashboard, metrics server and Heapster)
  • starts the worker nodes in parallel and installs Docker CE and kubeadm
  • joins the worker nodes in the cluster using the kubeadm token obtained from the master

Scale up by increasing the number of nodes:

$ terraform apply -var nodes=3 

Tear down the whole infrastructure with:

terraform destroy -force

Create an AMD64 bare-metal Kubernetes cluster with one master and a node:

$ terraform workspace new amd64

$ terraform apply \
 -var region=par1 \
 -var arch=x86_64 \
 -var server_type=C2S \
 -var nodes=1 \
 -var weave_passwd=ChangeMe \
 -var k8s_version=stable-1.9 \
 -var docker_version=17.03.0~ce-0~ubuntu-xenial

Remote control

After applying the Terraform plan you’ll see several output variables like the master public IP, the kubeadmn join command and the current workspace admin config.

In order to run kubectl commands against the Scaleway cluster you can use the kubectl_config output variable:

Check if Heapster works:

$ kubectl --kubeconfig ./$(terraform output kubectl_config) top nodes

NAME           CPU(cores)   CPU%      MEMORY(bytes)   MEMORY%   
arm-master-1   655m         16%       873Mi           45%       
arm-node-1     147m         3%        618Mi           32%       
arm-node-2     101m         2%        584Mi           30% 

The kubectl config file format is <WORKSPACE>.conf as in arm.conf or amd64.conf.

In order to access the dashboard you’ll need to find its cluster IP:

$ kubectl --kubeconfig ./$(terraform output kubectl_config) \
  -n kube-system get svc --selector=k8s-app=kubernetes-dashboard

NAME                   TYPE        CLUSTER-IP      EXTERNAL-IP   PORT(S)   AGE
kubernetes-dashboard   ClusterIP   <none>        80/TCP    6m

Open a SSH tunnel:

ssh -L 8888:<CLUSTER_IP>:80 [email protected]<MASTER_PUBLIC_IP>

Now you can access the dashboard on your computer at http://localhost:8888.



Expose services outside the cluster

Since we’re running on bare-metal and Scaleway doesn’t offer a load balancer, the easiest way to expose applications outside of Kubernetes is using a NodePort service.

Let’s deploy the podinfo app in the default namespace. Podinfo has a multi-arch Docker image and it will work on arm, arm64 or amd64.

Create the podinfo nodeport service:

$ kubectl --kubeconfig ./$(terraform output kubectl_config) \
  apply -f

service "podinfo-nodeport" created

Create the podinfo deployment:

$ kubectl --kubeconfig ./$(terraform output kubectl_config) \
  apply -f

deployment "podinfo" created

Inspect the podinfo service to obtain the port number:

$ kubectl --kubeconfig ./$(terraform output kubectl_config) \
  get svc --selector=app=podinfo

NAME               TYPE       CLUSTER-IP      EXTERNAL-IP   PORT(S)          AGE
podinfo-nodeport   NodePort   <none>        9898:31190/TCP   3m

You can access podinfo at http://<MASTER_PUBLIC_IP>:31190 or using curl:

$ curl http://$(terraform output k8s_master_public_ip):31190

  arch: arm
  max_procs: "4"
  num_cpu: "4"
  num_goroutine: "12"
  os: linux
  version: go1.9.2
  app: podinfo
  pod-template-hash: "1847780700"
annotations: 2018-01-08T00:39:45.580597397Z api
  HOME: /root
  HOSTNAME: podinfo-5d8ccd4c44-zrczc
  PATH: /usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin


You can deploy OpenFaaS on Kubernetes with Helm or by using the YAML files form the faas-netes repository.

Clone the faas-netes repo:

git clone
cd faas-netes

Deploy OpenFaaS for ARM:

$ kubectl --kubeconfig ./$(terraform output kubectl_config) \
  apply -f ./namespaces.yml,./yaml_armhf

Deploy OpenFaaS for AMD64:

$ kubectl --kubeconfig ./$(terraform output kubectl_config) \
  apply -f ./namespaces.yml,./yaml

You can access the OpenFaaS gateway at http://<MASTER_PUBLIC_IP>:31112.

Horizontal Pod Autoscaling

Starting from Kubernetes 1.9 kube-controller-manager is configured by default with horizontal-pod-autoscaler-use-rest-clients. In order to use HPA we need to install the metrics server to enable the new metrics API used by HPA v2. Both Heapster and the metrics server have been deployed from Terraform when the master node was provisioned.

The metric server collects resource usage data from each node using Kubelet Summary API. Check if the metrics server is running:

$ kubectl --kubeconfig ./$(terraform output kubectl_config) \
 get --raw "/apis/" | jq
    "kind": "NodeMetricsList",
    "apiVersion": "",
    "metadata": {
      "selfLink": "/apis/"
    "items": [
        "metadata": {
          "name": "arm-master-1",
          "selfLink": "/apis/",
          "creationTimestamp": "2018-01-08T15:17:09Z"
        "timestamp": "2018-01-08T15:17:00Z",
        "window": "1m0s",
        "usage": {
          "cpu": "384m",
          "memory": "935792Ki"
        "metadata": {
          "name": "arm-node-1",
          "selfLink": "/apis/",
          "creationTimestamp": "2018-01-08T15:17:09Z"
        "timestamp": "2018-01-08T15:17:00Z",
        "window": "1m0s",
        "usage": {
          "cpu": "130m",
          "memory": "649020Ki"
        "metadata": {
          "name": "arm-node-2",
          "selfLink": "/apis/",
          "creationTimestamp": "2018-01-08T15:17:09Z"
        "timestamp": "2018-01-08T15:17:00Z",
        "window": "1m0s",
        "usage": {
          "cpu": "120m",
          "memory": "614180Ki"

Let’s define a HPA that will maintain a minimum of two replicas and will scale up to ten if the CPU average is over 80% or if the memory goes over 200Mi.

apiVersion: autoscaling/v2beta1
kind: HorizontalPodAutoscaler
  name: podinfo
    apiVersion: apps/v1beta1
    kind: Deployment
    name: podinfo
  minReplicas: 2
  maxReplicas: 10
  - type: Resource
      name: cpu
      targetAverageUtilization: 80
  - type: Resource
      name: memory
      targetAverageValue: 200Mi

Apply the podinfo HPA:

$ kubectl --kubeconfig ./$(terraform output kubectl_config) \
  apply -f

horizontalpodautoscaler "podinfo" created

After a couple of seconds the HPA controller will contact the metrics server and will fetch the CPU and memory usage:

$ kubectl --kubeconfig ./$(terraform output kubectl_config) get hpa

NAME      REFERENCE            TARGETS                      MINPODS   MAXPODS   REPLICAS   AGE
podinfo   Deployment/podinfo   2826240 / 200Mi, 15% / 80%   2         10        2          5m

In order to increase the CPU usage we could run a load test with hey:

#install hey
go get -u

#do 10K requests rate limited at 20 QPS
hey -n 10000 -q 10 -c 5 http://$(terraform output k8s_master_public_ip):31190

You can monitor the autoscaler events with:

$ kubectl --kubeconfig ./$(terraform output kubectl_config) describe hpa

  Type    Reason             Age   From                       Message
  ----    ------             ----  ----                       -------
  Normal  SuccessfulRescale  7m    horizontal-pod-autoscaler  New size: 4; reason: cpu resource utilization (percentage of request) above target
  Normal  SuccessfulRescale  3m    horizontal-pod-autoscaler  New size: 8; reason: cpu resource utilization (percentage of request) above target

After the load tests finishes the autoscaler will remove replicas until the deployment reaches the initial replica count:

  Type    Reason             Age   From                       Message
  ----    ------             ----  ----                       -------
  Normal  SuccessfulRescale  20m   horizontal-pod-autoscaler  New size: 4; reason: cpu resource utilization (percentage of request) above target
  Normal  SuccessfulRescale  16m   horizontal-pod-autoscaler  New size: 8; reason: cpu resource utilization (percentage of request) above target
  Normal  SuccessfulRescale  12m   horizontal-pod-autoscaler  New size: 10; reason: cpu resource utilization (percentage of request) above target
  Normal  SuccessfulRescale  6m    horizontal-pod-autoscaler  New size: 2; reason: All metrics below target


Thanks to kubeadm and Terraform, bootstrapping a Kubernetes cluster on bare-metal can be done with a single command and it takes just ten minutes to have a fully functional setup. If you have any suggestion on improving this guide please submit an issue or PR on GitHub at stefanprodan/k8s-scw-baremetal. Contributions are more than welcome!

comments powered by Disqus