Home Lab¶
A five-node k3s cluster on Raspberry Pi over WiFi, one server and four workers, ~34 GiB RAM and 20 cores between them. It runs my own workloads and is where I practice Kubernetes. local-path for storage, the control-plane board tainted so workloads stay on the four workers, the components k3s bundles swapped out for my own (Istio for ingress, frp reverse tunnels to reach it from outside). The whole design assumes the network can change under it, so the cluster is anchored to a DNS name that a WiFi DHCP reshuffle never breaks.
One page per decision, written generically so the pattern is reusable without my network's specifics:
- Kubernetes installation with k3s: the five boards, the minimal component set, the DNS-anchored API, and why the control-plane board is tainted.
- Remote access with frp reverse tunnels: SSH to every board and
kubectlagainst the API from anywhere, without opening the home router. - GitOps for the cluster with ArgoCD: helmfile, Kustomize, app-of-apps, and ArgoCD, deployed in a fixed order from one Git repo.
- Service mesh and ingress with Istio: a revisioned Istio install, GHCR-mirrored images, and a default-deny public gateway.
- Exposing web services with frp-operator: a dedicated traffic relay, PROXY protocol to keep the real client IP, and declarative DNS.
- Pointing CoreDNS at a reachable upstream resolver: the DNS fault that took down every public ingress at once, and the one-line k3s fix.
- A backup access path with Tailscale: a break-glass path into the cluster for when frp is the thing that is down.