This is a CLI tool to quickly create and manage Kubernetes clusters in Hetzner Cloud using the lightweight Kubernetes distribution k3s from Rancher.
Hetzner Cloud is an awesome cloud provider which offers a truly great service with the best performance/cost ratio in the market and locations in both Europe and USA.
k3s is my favorite Kubernetes distribution because it uses much less memory and CPU, leaving more resources to workloads. It is also super quick to deploy and upgrade because it's a single binary.
Using hetzner-k3s
, creating a highly available k3s cluster with 3 masters for the control plane and 3 worker nodes takes 2-3 minutes only. This includes
- creating all the infrastructure resources (servers, private network, firewall, load balancer for the API server for HA clusters)
- deploying k3s to the nodes
- installing the Hetzner Cloud Controller Manager to provision load balancers right away
- installing the Hetzner CSI Driver to provision persistent volumes using Hetzner's block storage
- installing the Rancher System Upgrade Controller to make upgrades to a newer version of k3s easy and quick
- installing the Cluster Autoscaler to allow for autoscaling node pools
Also see this wiki page for a tutorial on how to set up a cluster with the most common setup to get you started.
I'm a Senior Backend Engineer and DevOps based in Finland and working for event management platform Brella.
I also write a technical blog on programming, DevOps and related technologies.
All that is needed to use this tool is
-
an Hetzner Cloud account
-
an Hetzner Cloud token: for this you need to create a project from the cloud console, and then an API token with both read and write permissions (sidebar > Security > API Tokens); you will see the token only once, so be sure to take note of it somewhere safe
-
kubectl installed
Before using the tool, be sure to have kubectl installed as it's required to install some components in the cluster and perform k3s upgrades.
brew install vitobotta/tap/hetzner_k3s
You need to install these dependencies first:
- libssh2
- libevent
- bdw-gc
- libyaml
- pcre
- gmp
wget https://github.com/vitobotta/hetzner-k3s/releases/download/v1.1.2/hetzner-k3s-mac-amd64
chmod +x hetzner-k3s-mac-amd64
sudo mv hetzner-k3s-mac-amd64 /usr/local/bin/hetzner-k3s
wget https://github.com/vitobotta/hetzner-k3s/releases/download/v1.1.2/hetzner-k3s-mac-arm64
chmod +x hetzner-k3s-mac-arm64
sudo mv hetzner-k3s-mac-arm64 /usr/local/bin/hetzner-k3s
wget https://github.com/vitobotta/hetzner-k3s/releases/download/v1.1.2/hetzner-k3s-linux-x86_64
chmod +x hetzner-k3s-linux-x86_64
sudo mv hetzner-k3s-linux-x86_64 /usr/local/bin/hetzner-k3s
I recommend using the Linux binary under WSL.
The tool requires a simple configuration file in order to create/upgrade/delete clusters, in the YAML format like in the example below (commented lines are for optional settings):
---
hetzner_token: <your token>
cluster_name: test
kubeconfig_path: "./kubeconfig"
k3s_version: v1.26.4+k3s1
public_ssh_key_path: "~/.ssh/id_rsa.pub"
private_ssh_key_path: "~/.ssh/id_rsa"
use_ssh_agent: false
ssh_allowed_networks:
- 0.0.0.0/0
api_allowed_networks:
- 0.0.0.0/0
private_network_subnet: 10.0.0.0/16
schedule_workloads_on_masters: false
# image: rocky-9 # optional: default is ubuntu-22.04
# autoscaling_image: 103908130 # optional, defaults to the `image` setting
# snapshot_os: microos # otional: specified the os type when using a custom snapshot
masters_pool:
instance_type: cpx21
instance_count: 3
location: nbg1
worker_node_pools:
- name: small-static
instance_type: cpx21
instance_count: 4
location: hel1
# image: debian-11
# labels:
# - key: purpose
# value: blah
# taints:
# - key: something
# value: value1:NoSchedule
- name: big-autoscaled
instance_type: cpx31
instance_count: 2
location: fsn1
autoscaling:
enabled: true
min_instances: 0
max_instances: 3
# additional_packages:
# - somepackage
# post_create_commands:
# - apt update
# - apt upgrade -y
# - apt autoremove -y
# enable_encryption: true
# existing_network: <specify if you want to use an existing network, otherwise one will be created for this cluster>
# kube_api_server_args:
# - arg1
# - ...
# kube_scheduler_args:
# - arg1
# - ...
# kube_controller_manager_args:
# - arg1
# - ...
# kube_cloud_controller_manager_args:
# - arg1
# - ...
# kubelet_args:
# - arg1
# - ...
# kube_proxy_args:
# - arg1
# - ...
Most settings should be self explanatory; you can run hetzner-k3s releases
to see a list of the available k3s releases.
If you don't want to specify the Hetzner token in the config file (for example if you want to use the tool with CI or want to safely commit the config file to a repository), then you can use the HCLOUD_TOKEN
environment variable instead, which has predecence.
If you set masters_pool.instance_count
to 1 then the tool will create a non highly available control plane; for production clusters you may want to set it to a number greater than 1. This number must be odd to avoid split brain issues with etcd and the recommended number is 3.
You can specify any number of worker node pools, static or autoscaled, and have mixed nodes with different specs for different workloads.
At the moment Hetzner Cloud has five locations: two in Germany (nbg1
, Nuremberg and fsn1
, Falkenstein), one in Finland (hel1
, Helsinki) and two in the USA (ash
, Ashburn, Virginia, and hil
, Hillsboro, Oregon). Please keep in mind that US locations only offer instances with AMD CPUs at the moment, while the newly introduced ARM instances are only available in Falkenstein-fsn1 for now.
For the available instance types and their specs, either check from inside a project when adding a server manually or run the following with your Hetzner token:
curl -H "Authorization: Bearer $API_TOKEN" 'https://api.hetzner.cloud/v1/server_types'
To create the cluster run:
hetzner-k3s create --config cluster_config.yaml
This will take a few minutes depending on the number of masters and worker nodes.
By default, the image in use is ubuntu-22.04
for all the nodes, but you can specify a different default image with the root level image
config option or even different images for different static node pools by setting the image
config option in each node pool. This way you can, for example, have some node pools with ARM instances use the correct OS image for ARM. To do this and use say Ubuntu 22.04 on ARM instances, set image
to 103908130
with a specific image ID. With regard to autoscaling, due to a limitation in the Cluster Autoscaler for Hetzner it is not possible yet to specify a different image for each autoscaled pool, so for now you can specify the image for all autoscaled pools by setting the autoscaling_image
setting if you want to use an image different from the one specified in image
.
To see the list of available images, run the following:
export API_TOKEN=...
curl -H "Authorization: Bearer $API_TOKEN" 'https://api.hetzner.cloud/v1/images?per_page=100'
Besides the default OS images, It's also possible to use a snapshot that you have already created from an existing server. Also with custom snapshots you'll need to specify the ID of the snapshot/image, not the description you gave when you created the template server.
I've tested snapshots for openSUSE MicroOS but others might work too. You can easily create a snapshot for MicroOS using this tool. Creating the snapshot takes just a couple of minutes and then you can use it with hetzner-k3s by setting the config option image
to the ID of the snapshot, and snapshot_os
to microos
.
- if possible, please use modern SSH keys since some operating systems have deprecated old crypto based on SHA1; therefore I recommend you use ECDSA keys insted of the old RSA type
- if you use a snapshot instead of one of the default images, the creation of the servers will take longer than when using a regular image
- the setting
api_allowed_networks
allows specifying which networks can access the Kubernetes API, but this only works with single master clusters currently. Multi-master HA clusters require a load balancer for the API, but load balancers are not yet covered by Hetzner's firewalls - if you enable autoscaling for one or more nodepools, do not change that setting afterwards as it can cause problems to the autoscaler
- autoscaling is only supported when using Ubuntu or one of the other default images, not snapshots
- worker nodes created by the autoscaler must be deleted manually from the Hetzner Console when deleting the cluster (this will be addressed in a future update)
- SSH keys with passphrases can only be used if you set
use_ssh_agent
totrue
and use an SSH agent to access your key. To start and agent e.g. on macOS:
eval "$(ssh-agent -s)"
ssh-add --apple-use-keychain ~/.ssh/<private key>
The create
command can be run any number of times with the same configuration without causing any issue, since the process is idempotent. This means that if for some reason the create process gets stuck or throws errors (for example if the Hetzner API is unavailable or there are timeouts etc), you can just stop the current command, and re-run it with the same configuration to continue from where it left.
To add one or more nodes to a node pool, just change the instance count in the configuration file for that node pool and re-run the create command.
Important: if you are increasing the size of a node pool created prior to v0.5.7, please see this thread.
To make a node pool smaller:
- decrease the instance count for the node pool in the configuration file so that those extra nodes are not recreated in the future
- delete the nodes from Kubernetes (
kubectl delete node <name>
) - delete the instances from the cloud console if the Cloud Controller Manager doesn't delete it automatically (make sure you delete the correct ones ๐คญ)
In a future release I will add some automation for the cleanup.
- delete the node from Kubernetes (
kubectl delete node <name>
) - delete the correct instance from the cloud console
- re-run the
create
command. This will re-create the missing node and have it join to the cluster
It's easy to convert a non-HA with a single master cluster to HA with multiple masters. Just change the masters instance count and re-run the create
command. This will create a load balancer for the API server and update the kubeconfig so that all the API requests go through the load balancer.
If it's the first time you upgrade the cluster, all you need to do to upgrade it to a newer version of k3s is run the following command:
hetzner-k3s upgrade --config cluster_config.yaml --new-k3s-version v1.27.1-rc2+k3s1
So you just need to specify the new k3s version as an additional parameter and the configuration file will be updated with the new version automatically during the upgrade. To see the list of available k3s releases run the command hetzner-k3s releases
.
Note: (single master clusters only) the API server will briefly be unavailable during the upgrade of the controlplane.
To check the upgrade progress, run watch kubectl get nodes -owide
. You will see the masters being upgraded one per time, followed by the worker nodes.
If the upgrade gets stuck for some reason, or it doesn't upgrade all the nodes:
- Clean up the existing upgrade plans and jobs, and restart the upgrade controller
kubectl -n system-upgrade delete job --all
kubectl -n system-upgrade delete plan --all
kubectl label node --all plan.upgrade.cattle.io/k3s-server- plan.upgrade.cattle.io/k3s-agent-
kubectl -n system-upgrade rollout restart deployment system-upgrade-controller
kubectl -n system-upgrade rollout status deployment system-upgrade-controller
You can also check the logs of the system upgrade controller's pod:
kubectl -n system-upgrade \
logs -f $(kubectl -n system-upgrade get pod -l pod-template-hash -o jsonpath="{.items[0].metadata.name}")
A final note about upgrades is that if for some reason the upgrade gets stuck after upgrading the masters and before upgrading the worker nodes, just cleaning up the resources as described above might not be enough. In that case also try running the following to tell the upgrade job for the workers that the masters have already been upgraded, so the upgrade can continue for the workers:
kubectl label node <master1> <master2> <master2> plan.upgrade.cattle.io/k3s-server=upgraded
- consider adding a temporary node during the process if you don't have enough spare capacity in the cluster
- drain one node
- update etc
- reboot
- uncordon
- proceed with the next node
If you want to automate this process I recommend you install the Kubernetes Reboot Daemon ("Kured"). For this to work properly, make sure the OS you choose for the nodes has unattended upgrades enabled at least for security updates. For example if the image is Ubuntu, you can add this to the configuration file before running the create
command:
additional_packages:
- unattended-upgrades
- update-notifier-common
post_create_commands:
- sudo systemctl enable unattended-upgrades
- sudo systemctl start unattended-upgrades
Check the Kured documentation for configuration options like maintenance window etc.
To delete a cluster, running
hetzner-k3s delete --config cluster_config.yaml
This will delete all the resources in the Hetzner Cloud project created by hetzner-k3s
directly. At the moment instances created by the cluster autoscaler must be deleted manually. This will be addressed in a future update.
Once the cluster is ready, you can already provision services of type LoadBalancer for your workloads (such as the Nginx ingress controller for example) thanks to the Hetzner Cloud Controller Manager that is installed automatically.
There are some annotations that you can add to your services to configure the load balancers. I personally use the following:
service:
annotations:
load-balancer.hetzner.cloud/hostname: <a valid fqdn>
load-balancer.hetzner.cloud/http-redirect-https: 'false'
load-balancer.hetzner.cloud/location: nbg1
load-balancer.hetzner.cloud/name: <lb name>
load-balancer.hetzner.cloud/uses-proxyprotocol: 'true'
load-balancer.hetzner.cloud/use-private-ip: "true"
I set load-balancer.hetzner.cloud/hostname
to a valid hostname that I configure (after creating the load balancer) with the IP of the load balancer; I use this together with the annotation load-balancer.hetzner.cloud/uses-proxyprotocol: 'true'
to enable the proxy protocol. Reason: I enable the proxy protocol on the load balancers so that my ingress controller and applications can "see" the real IP address of the client. However when this is enabled, there is a problem where cert-manager fails http01 challenges; you can find an explanation of why here but the easy fix provided by some providers - including Hetzner - is to configure the load balancer so that it uses a hostname instead of an IP. Again, read the explanation for the reason but if you care about seeing the actual IP of the client then I recommend you use these two annotations.
The annotation load-balancer.hetzner.cloud/use-private-ip: "true"
ensures that the communication between the load balancer and the nodes happens through the private network, so we don't have to open any ports in the filrewall for the nodes (other than the port 6443 for the Kubernetes API server).
The other annotations should be self explanatory. You can find a list of the available annotations here.
Once the cluster is ready you can create persistent volumes out of the box with the default storage class hcloud-volumes
, since the Hetzner CSI driver is installed automatically. This will use Hetzner's block storage (based on Ceph so it's replicated and highly available) for your persistent volumes. Note that the minimum size of a volume is 10Gi. If you specify a smaller size for a volume, the volume will be created with a capacity of 10Gi anyway.
I recommend that you create a separate Hetzner project for each cluster, because otherwise multiple clusters will attempt to create overlapping routes. I will make the pod cidr configurable in the future to avoid this, but I still recommend keeping clusters separated from each other. This way, if you want to delete a cluster with all the resources created for it, you can just delete the project.
If the tool hangs forever after creating servers and you see timeouts, this may be caused by problems with your SSH key, for example if you use a key with a passphrase or an older key (due to the deprecation of some crypto stuff in newwer operating systems). In this case you may want to try setting use_ssh_agent
to true
to use the SSH agent. If you are not familiar with what an SSH agent is, take a look at this page for an explanation.
Please create a PR if you want to propose any changes, or open an issue if you are having trouble with the tool - I will do my best to help if I can.
Contributors:
This tool is available as open source under the terms of the MIT License.
Everyone interacting in the hetzner-k3s project's codebases, issue trackers, chat rooms and mailing lists is expected to follow the code of conduct.