mrhein/terraform-hcloud-kube-hetzner

Kube-Hetzner

A highly optimized, easy-to-use, auto-upgradable, HA-default & Load-Balanced, Kubernetes cluster powered by k3s-on-MicroOS and deployed for peanuts on Hetzner Cloud 🤑 🚀

About The Project

Hetzner Cloud is a good cloud provider that offers very affordable prices for cloud instances, with data center locations in both Europe and the US.

This project aims to create a highly optimized Kubernetes installation that is easy to maintain, secure and automatically upgrades both the nodes and Kubernetes. We aimed for functionality as close as possible to GKE's Auto-Pilot. Please note that we are not affiliates of Hetzner; but we do strive to be an optimal solution for deploying and maintaining Kubernetes clusters on Hetzner Cloud.

To achieve this, we built up on the shoulders of giants by choosing openSUSE MicroOS as the base operating system and k3s as the k8s engine.

Why OpenSUSE MicroOS (and not Ubuntu)?

Optimized container OS that is fully locked down, most of the filesystem is read-only!
Hardened by default with automatic ban for abusive IPs on SSH for instance.
Evergreen release, your node will stay valid forever, as it piggy-backs into OpenSUSE Thumbleweed's rolling-release!
Automatic updates by default, and if something breaks, it automatically rolls-back, thanks to it uses of BTRFS snapshots.
Supports Kured to properly drain and reboot nodes in an HA fashion.

Why k3s?

Certified Kubernetes Distribution, it is automatically synced to k8s source.
Fast deployment, it is deployed with a single command.
Comes batteries included, with its in-cluster helm-controller.
Easy automatic update, via the system-upgrade-controller.

Features

It uses Terraform to deploy as it's easy to use, and Hetzner provides a great Hetzner Terraform Provider.

Getting Started

Follow those simple steps, and your world's cheapest Kube cluster will be up and running.

✔️ Prerequisites

First and foremost, you need to have a Hetzner Cloud account. You can sign up for free here.

Then you'll need to have terraform, kubectl cli and hcloud the Hetzner cli. The easiest way is to use the homebrew package manager to install them (available on Linux, Mac, and Windows Linux Subsystem).

brew install terraform
brew install kubectl
brew install hcloud

💡 [Do not skip] Creating your kube.tf file

Create a project in your Hetzner Cloud Console, and go to Security > API Tokens of that project to grab the API key. Take note of the key! ✅
Generate a passphrase-less ed25519 SSH key pair for your cluster; take note of the respective paths of your private and public keys. Or, see our detailed SSH options. ✅
Prepare the module by copying kube.tf.example to kube.tf in a new folder which you cd into, then replace the values from steps 1 and 2. ✅
(Optional) Many variables in kube.tf can be customized to suit your needs, you can do so if you want. ✅
At this stage you should be in your new folder, with a fresh kube.tf file, if it is so, you can proceed forward! ✅

A complete reference of all inputs, outputs, modules etc. can be found in the terraform.md file.

It's important to realize that your kube.tf needs to reside in a NEW folder, not a clone of this git repo (the module by default will be fetched from the Terraform registry). All you need, is to re-use the kube.tf.example file to make sure you get the format right.

🎯 Installation

terraform init --upgrade
terraform validate
terraform apply -auto-approve

It will take around 5 minutes to complete, and then you should see a green output confirming a successful deployment.

Once you start with Terraform, it's best not to change the state of the project manually via the Hetzner UI; otherwise, you may get an error when you try to run terraform again for that cluster (when trying to change the number of nodes for instance).

Usage

When your brand-new cluster is up and running, the sky is your limit! 🎉

You can immediately kubectl into it (using the clustername_kubeconfig.yaml saved to the project's directory after the installation). By doing kubectl --kubeconfig clustername_kubeconfig.yaml, but for more convenience, either create a symlink from ~/.kube/config to clustername_kubeconfig.yaml or add an export statement to your ~/.bashrc or ~/.zshrc file, as follows (you can get the path of clustername_kubeconfig.yaml by running pwd):

export KUBECONFIG=/<path-to>/clustername_kubeconfig.yaml

If chose to turn create_kubeconfig to false in your kube.tf (good practice), you can still create this file by running terraform output --raw kubeconfig > clustername_kubeconfig.yaml and then use it as described above.

You can also use it in an automated flow, in which case create_kubeconfig should be set to false, and you can use the kubeconfig output variable to get the kubeconfig file in a structured data format.

You can view all kinds of details about the cluster by running terraform output kubeconfig or terraform output -json kubeconfig | jq.

CNI

The default is Flannel, but you can also choose Calico or Cilium, by setting the cni_plugin variable in kube.tf to "calico" or "cilium".

As Cilium has a lot of interesting and powerful configurations' possibility. We give you the possibility to configure your Cilium with the helm cilium_values variable (see the cilium specific helm values) before you deploy your cluster.

Scaling Nodes

Two things can be scaled: the number of nodepools or the number of nodes in these nodepools. You have two lists of nodepools you can add to your kube.tf, the control plane nodepool and the agent nodepool list. Combined, they cannot exceed 255 nodepools (you are extremely unlikely to reach this limit). As for the count of nodes per nodepools, if you raise your limits in Hetzner, you can have up to 64,670 nodes per nodepool (also very unlikely to need that much).

There are some limitations (to scaling down mainly) that you need to be aware of:

Once the cluster is up; you can change any nodepool count and even set it to 0 (in the case of the first control-plane nodepool, the minimum is 1); you can also rename a nodepool (if the count is to 0), but should not remove a nodepool from the list after once the cluster is up. That is due to how subnets and IPs get allocated. The only nodepools you can remove are those at the end of each list of nodepools.

However, you can freely add other nodepools at the end of each list. And for each nodepools, you can freely increase or decrease the node count (if you want to decrease a nodepool node count make sure you drain the nodes in question before, you can use terraform show to identify the node names at the end of the nodepool list, otherwise, if you do not drain the nodes before removing them, it could leave your cluster in a bad state). The only nodepool that needs to have always at least a count of 1 is the first control-plane nodepool.

Autoscaling Node Pools

We are supporting autoscaling node pools by deploying the k8s cluster autoscaler (CA). By default, this feature is disabled. You can control the feature via adding a pool description to the following variable in kube.tf (by default this array is empty):

autoscaler_nodepools = [
    {
      name        = "autoscaler"
      server_type = "cpx21" # must be same or better than the control_plane server type (regarding disk size)!
      location    = "fsn1"
      min_nodes   = 0
      max_nodes   = 5
    }
  ]

By adding at least one map to the array of autoscaler_nodepools the feature will be enabled. The nodes are booted based on a snapshot that is created from the initial control_plane. So please ensure that the disk of your chosen server type is at least the same size as the one of the first control_plane.

See the CA documentation for more configuration options.

High Availability

By default, we have three control planes and three agents configured, with automatic upgrades and reboots of the nodes.

If you want to remain HA (no downtime), it's essential to keep a count of control planes nodes of at least three (two minimum to maintain quorum when one goes down for automated upgrades and reboot), see Rancher's doc on HA.

Otherwise, it is essential to turn off automatic OS upgrades (k3s can continue to update without issue) for the control-plane nodes (when two or fewer control-plane nodes) and do the maintenance yourself.

Automatic Upgrade

The Default Setting

By default, MicroOS gets upgraded automatically on each node and reboot safely via Kured installed in the cluster.

As for k3s, it also automatically upgrades thanks to Rancher's system upgrade controller. By default it will be set to the initial_k3s_channel, but you can also set it to stable, latest, or one more specific like v1.23 if needed or specify a target version to upgrade to via the upgrade plan (this also allows for downgrades).

You can copy and modify the one in the templates for that! More on the subject in k3s upgrades.

Turning Off Automatic Upgrade

If you wish to turn off automatic MicroOS upgrades (Important if you are not launching an HA setup which requires at least 3 control-plane nodes), you need to set:

automatically_upgrade_os = false

Alternatively ssh into each node and issue the following command:

systemctl --now disable transactional-update.timer

If you wish to turn off automatic k3s upgrades, you need to set:

automatically_upgrade_k3s = false

Alternatively, you can either remove the k3s_upgrade=true label or set it to false. This needs to happen for all the nodes too! To remove it, apply:

kubectl -n system-upgrade label node <node-name> k3s_upgrade-

Alternatively, you can disable the k3s automatic upgrade without individually editing the labels on the nodes. Instead, you can just delete the two system controller upgrade plans with:

kubectl delete plan k3s-agent -n system-upgrade
kubectl delete plan k3s-server -n system-upgrade

Individual Components Upgrade

Rarely needed, but can be handy in the long run. During the installation, we automatically download a backup of the kustomization to a kustomization_backup.yaml file. You will find it next to your clustername_kubeconfig.yaml at the root of your project.

First create a duplicate of that file and name it kustomization.yaml, keeping the original file intact, in case you need to restore the old config.
Edit the kustomization.yaml file; you want to go to the very bottom where you have the links to the different source files; grab the latest versions for each on GitHub, and replace. If present, remove any local reference to traefik_config.yaml, as Traefik is updated automatically by the system upgrade controller.
Apply the updated kustomization.yaml with kubectl apply -k ./.

Customizing the Cluster Components

Most cluster components of Kube-Hetzner are deployed with the Rancher Helm Chart yaml definition and managed by the Helm Controller inside k3s.

By default, we strive to give you optimal defaults, but if wish, you can customize them.

Before deploying

In the case of Traefik, Rancher, and Longhorn, we provide you with variables to configure everything you need.

On top of the above, for Nginx, Rancher, Cilium, Traefik and Longhorn, for maximum flexibility, we give you the ability to configure them even better via helm values variables (e.g. cilium_values, see the advanced section in the kube.tf.example for more).

After deploying

Once the Cluster is up and running, you can easily customize many components like Traefik, Nginx, Rancher, Cilium, Cert-Manager and Longhorn by using HelmChartConfig definitions. See the examples section, for more information.

For other components like Calico and Kured (which uses manifests), we automatically save a kustomization_backup.yaml file in the root of your module during the deployment, so you can use that as a starting point. This is also useful when creating the HelmChartConfig definitions, as both HelmChart and HelmChartConfig definitions are very similar.

Adding Extras

If you need to install additional Helm charts or Kubernetes manifests that are not provided by default, you can easily do so by using Kustomize. This is done by creating the extra-manifests/kustomization.yaml.tpl directory/file besides your kube.tf.

This file needs to be a valid Kustomization manifest, but it supports terraform templating! (The templating parameters can be passed via the extra_kustomize_parameters variable (via a map) to the module).

All files in the extra-manifests directory including the rendered version of kustomization.yaml.tpl will be applied to k3s with kubectl apply -k (which will be executed after and independently of the basic cluster configuration).

You can use the above to pass all kinds of kubernetes YAML configs, including HelmChart and/or HelmChartConfig definitions (see the previous section if you do not know what those are in the context of k3s).

Examples

Useful Cilium commands

With Kube-Hetzner, you have the possibility to use Cilium as a CNI. It's very powerful and has great observability features. Below you will find a few useful commands.

Check the status of cilium with the following commands (get the cilium pod name first and replace it in the command):

kubectl -n kube-system exec --stdin --tty cilium-xxxx -- cilium status
kubectl -n kube-system exec --stdin --tty cilium-xxxx -- cilium status --verbose

Monitor cluster traffic with:

kubectl -n kube-system exec --stdin --tty cilium-xxxx -- cilium monitor

See the list of kube services with:

kubectl -n kube-system exec --stdin --tty cilium-xxxx -- cilium service list

For more cilium commands, please refer to their corresponding Documentation.

Ingress with TLS

You have two solutions, the first is to use Cert-Manager to take care of the certificates, and the second is to let Traefik bear this responsibility.

We advise you to use the first one, as it supports HA setups without requiring you to use the enterprise version of Traefik. The reason for that is that according to Traefik themselves, Traefik CE (community edition) is stateless, and it's not possible to run multiple instance of Traefik CE with LetsEncrypt enabled. Meaning, you cannot have your ingress be HA with Traefik if you use the community edition and have activated the LetsEncrypt resolver. You could however use Traefik EE (enterprise edition) to achieve that. Long story short, if you are going to use Traefik CE (like most of us), you should use cert-manager to generate the certificates. Source here.

Via Cert-Manager (recommended)

In your module variables, set enable_cert_manager to true, and just create your issuers as described here https://cert-manager.io/docs/configuration/acme/.

Then in your Ingress definition, just mentioning the issuer as an annotation and giving a secret name will take care of instructing cert-manager to generate a certificate for it! It simpler than the alternative, you just have to configure your issuer(s) first with the method of your choice.

Ingress example:

apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  name: my-ingress
  annotations:
    cert-manager.io/cluster-issuer: letsencrypt
spec:
  tls:
  - hosts:
    - '*.example.com'
    secretName: example-com-letsencrypt-tls
  rules:
  - host: '*.example.com'
    http:
      paths:
      - path: /
        pathType: Prefix
        backend:
          service:
            name: my-service
            port:
              number: 80

⚠️ In case of using Ingress-Nginx as ingress controller, if you choose to use the HTTP challenge method you need to do an additional step of adding this annotation load-balancer.hetzner.cloud/hostname to the Nginx service definition. And you set it equal to a FQDN that points to your LB address.

This is to circumvent this known issue cert-manager/cert-manager#466, also see kube-hetzner#354. Otherwise, you can just use the DNS challenge, which does not require any additional tweaks to work.

Via Traefik CE (not recommended)

Here is an example of an ingress to run an application with TLS, change the host to fit your need in examples/tls/ingress.yaml and then deploy the example:

kubectl apply -f examples/tls/.

apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  name: my-ingress
  annotations:
    traefik.ingress.kubernetes.io/router.tls: "true"
    traefik.ingress.kubernetes.io/router.tls.certresolver: le
spec:
  tls:
    - hosts:
        - example.com
  rules:
    - host: example.com
      http:
        paths:
          - path: /
            pathType: Prefix
            backend:
              service:
                name: my-service
                port:
                  number: 80

Single-node cluster

Running a development cluster on a single node without any high availability is also possible.

When doing so, automatically_upgrade_os should be set to false, especially with attached volumes the automatic reboots won't work properly. In this case, we don't deploy an external load-balancer but use the default k3s service load balancer on the host itself and open up port 80 & 443 in the firewall (done automatically).

Use in Terraform cloud

To use Kube-Hetzner on Terraform cloud, use as a Terraform module as mentioned above, but also change the execution mode from remote to local.

Configure add-ons with HelmChartConfig

For instance, to customize the Rancher install, if you choose to enable it, you can create and apply the following HelmChartConfig:

apiVersion: helm.cattle.io/v1
kind: HelmChartConfig
metadata:
  name: rancher
  namespace: kube-system
spec:
  valuesContent: |-
    **values.yaml content you want to customize**

The helm options for Rancher can be seen here https://github.com/rancher/rancher/blob/release/v2.6/chart/values.yaml.

Same goes for all add-ons, like Longhorn, Cert-manager, and Traefik.

Debugging

First and foremost, it depends, but it's always good to have a quick look into Hetzner quickly without logging in to the UI. That is where the hcloud cli comes in.

Activate it with hcloud context create Kube-hetzner; it will prompt for your Hetzner API token, paste that, and hit enter.
To check the nodes, if they are running, use hcloud server list.
To check the network, use hcloud network describe k3s.
To look at the LB, use hcloud loadbalancer describe traefik.

Then for the rest, you'll often need to log in to your cluster via ssh, to do that, use:

ssh root@xxx.xxx.xxx.xxx -i ~/.ssh/id_ed25519 -o StrictHostKeyChecking=no

Then, for control-plane nodes, use journalctl -u k3s to see the k3s logs, and for agents, use journalctl -u k3s-agent instead.

Last but not least, to see when the previous reboot took place, you can use both last reboot and uptime.

Takedown

If you want to take down the cluster, you can proceed as follows:

terraform destroy -auto-approve

And if the network is slow to delete, just issue hcloud load-balancer delete clustername to speed things up! As the load-balancer is usually the resource that is the slowest to get deleted on its own.

Also, if you had a full-blown cluster in use, it would be best to delete the whole project in your Hetzner account directly as operators or deployments may create other resources during regular operation.

History

This project has tried two other OS flavors before settling on MicroOS. Fedora Server, and k3OS. The latter, k3OS, is now defunct! However, our code base for it lives on in the k3os branch. Do not hesitate to check it out, it should still work.

There is also a branch where openSUSE MicroOS came preinstalled with the k3s RPM from devel:kubic/k3s, but we moved away from that solution as the k3s version was rarely getting updates. See the microOS-k3s-rpm branch for more.

Contributing

🌱 This project currently installs openSUSE MicroOS via the Hetzner rescue mode, making things a few minutes slower. To help with that, you could take a few minutes to send a support request to Hetzner, asking them to please add openSUSE MicroOS as a default image, not just an ISO. The more requests they receive, the likelier they are to add support for it, and if they do, that will cut the deployment time by half. The official link to openSUSE MicroOS is https://get.opensuse.org/microos, and their OpenStack Cloud image has full support for Cloud-init, which would probably very much suit the Hetzner Ops team!

Code contributions are very much welcome.

Fork the Project
Create your Branch (git checkout -b AmazingFeature)
Commit your Changes (`git commit -m 'Add some AmazingFeature')
Push to the Branch (git push origin AmazingFeature)
Open a Pull Request targeting the staging branch.

Acknowledgements

k-andy was the starting point for this project. It wouldn't have been possible without it.
Best-README-Template made writing this readme a lot easier.
Hetzner Cloud for providing a solid infrastructure and terraform package.
Hashicorp for the amazing terraform framework that makes all the magic happen.
Rancher for k3s, an amazing Kube distribution that is the core engine of this project.
openSUSE for MicroOS, which is just next level Container OS technology.

mrhein / terraform-hcloud-kube-hetzner