Hardware
The machines this instance runs on, and the choices made for them. The generic requirements (operating systems, access, the install options) are in the ohmlab documentation.
Equipment
- 9 x Raspberry Pi 4 model B, 8GB: the K3s nodes
- 1 x Raspberry Pi 4 model B, 4GB: the gateway
- 3 x TP-Link TL-SG105E switches
- 4 x 2TB HDD (HDTP320EK3AA), USB-attached to four of the workers
Layout
The hosts and their roles are in ansible/inventory/hosts.yml:
| Group | Hosts | Notes |
|---|---|---|
| gateway | pi-gw | HAProxy (80, 443, 6443) in front of the cluster |
| masters | pi0, pi3, pi4 | K3s control plane and embedded etcd |
| workers | pi1, pi2, pi5 to pi8 | pi1, pi2, pi5 and pi6 carry a disk (additional_disk: true) and run Longhorn |
Operating system
The nodes run Debian 13 (trixie) on the Debian generic arm64 kernel (the opt-in kernel-upgrade Ansible tag), upgraded in place from Raspberry Pi OS (64-bit).
Networking
The cluster runs Cilium with K3s's ServiceLB disabled. Traefik is exposed through a static externalIPs list, and the gateway's HAProxy forwards public 80/443 onto the node addresses.
It started on flannel and moved to Cilium one node at a time with ohmlab's scripts/cilium-migrate-node.sh.
Storage
Longhorn uses the four USB disks. They are spinning disks without UAS or TRIM, so I/O peaks (snapshots, backups) are the cluster's main bottleneck.