This is a written record of how I set up a k3s cluster on some bare metal machines.

References

1. https://docs.k3s.io/quick-start
2. https://docs.k3s.io/datastore/ha

Components

Three physical servers were used in this setup.

These are the VMs that I run on those hosts. (IPs are arbitrary)

The VMs all run Debian 12, and the following packages are installed

1. nfs-common (on all VMs)
2. nfs-kernel-server (on kshare01, kshare02)

Also, before doing any of the following, I created a PostgreSQL database on kshare02 to use as a external data source for my master nodes.
In this example, assume it’s running on port 5432 of kshare02, and the following database is set up and ready to accept connections.

Setting up the first master node

Install k3s, telling it to use external data source

1

curl -sfL https://get.k3s.io | sh -s - server --datastore-endpoint="postgres://k3s:pass@192.168.1.10:5432/master"

Then edit k3s service

1
2
3
4
5

ExecStart=/usr/local/bin/k3s \
    server \
    --node-ip 192.168.1.2 \
    --datastore-endpoint="postgres://k3s:pass@192.168.1.10:5432/master" \
    --disable servicelb \

and reload+restart it

Setting up the second master node

Set up another master node, and join it to the cluster

From kmaster01, copy the server token

1
2

sudo cat /var/lib/rancher/k3s/server/token
# copy for later use

on kmaster02

1
2
3

  curl -sfL https://get.k3s.io | sh -s - server \
    --datastore-endpoint="postgres://k3s:pass@192.168.1.10:5432/master" \
    --token "<token>"

To confirm both nodes are there, log in to kmaster01 and run the command:

1

sudo kubectl get node

Confirm that both kmaster01 and kmaster02 are shown.

Join worker nodes

Log in to each worker node and run the following command:

1

curl -sfL https://get.k3s.io | K3S_URL=https://192.168.1.2:6443 K3S_TOKEN="<token>" sh -

Again, to confirm all nodes are added, log in to kmaster01 (or kmaster02) and run:

1

sudo kubectl get node

Conclusion

It was fun, and far easier than I thought to set up a kubernetes cluster with baremetal machines. Currently I use my cluster to run this blog, a personal portfolio website, and a few other pods for my own use.

There is one issue, though, which is that although there are two master nodes in this setup, the DB is still the single point of failure. Once I figure out a redundant solution, I will post it here.