Building a Home Kubernetes Cluster with Raspberry Pi 5

The Raspberry Pi 5 finally has the I/O and memory headroom to run a real cluster. This guide walks through the full setup I use: three control-plane-capable nodes, k3s with the embedded etcd datastore, Traefik ingress, and NFS-backed persistent volumes.

Why k3s and Not k8s

k3s ships a certified Kubernetes distribution in a single binary, with sane defaults for ARM64. Everything you learn transfers directly to any managed Kubernetes you will touch professionally.

# On the first node (control plane with embedded etcd)
curl -sfL https://get.k3s.io | sh -s - server \
  --cluster-init \
  --node-taint "CriticalAddonsOnly=true:NoExecute"

# Grab the join token for the other nodes
sudo cat /var/lib/rancher/k3s/server/token

# On every additional node
curl -sfL https://get.k3s.io | sh -s - server \
  --server https://192.168.1.10:6443 \
  --token "$K3S_TOKEN"

Bootstrapping the Cluster

1. Flash Raspberry Pi OS Lite (64-bit) and set static DHCP reservations for every node.
2. Boot from NVMe if your budget allows — SD cards die under etcd write load.
3. Install the first server with --cluster-init, then join the rest.
4. Copy kubeconfig to your workstation and verify with kubectl get nodes.

[!WARNING] Do not run more than one etcd member on SD cards. etcd's write amplification will kill the cards within weeks — ask me how I know.

Persistent Storage

For a homelab, an NFS provisioner on a NAS or a dedicated node is the pragmatic choice. The local-path provisioner bundled with k3s is fine for experiments, but it pins pods to nodes.

apiVersion: storage.k8s.io/v1
kind: StorageClass
metadata:
  name: nfs-client
provisioner: cluster.local/nfs-subdir-external-provisioner
parameters:
  archiveOnDelete: "true"

What Runs on It

Cert-manager for TLS, Traefik for ingress, ArgoCD for GitOps, and a monitoring stack with kube-prometheus-stack. Total idle draw: about 22W for three nodes.