Homelab Kubernetes Cluster
4-node k3s cluster on Raspberry Pi, driven by OpenTofu, with observability (VictoriaMetrics, Loki, OpenTelemetry, Grafana), Let's Encrypt and self-hosted services.
Context
I wanted to push my homelab beyond a simple NAS or single server. The goal: a resilient, reproducible infrastructure close to a real production environment, but on low-power ARM hardware.
Architecture
The cluster runs on 4 Raspberry Pi 4 (8 GB RAM each) connected over Gigabit Ethernet:
- 1 dedicated control plane
- 3 worker nodes with Longhorn distributed storage
- Pi-hole for DNS/DHCP and network filtering
- Traefik as ingress controller with automatic TLS certificates
Self-Hosted Services
- Immich: photo management, Google Photos style
- Gitea: private Git repository
- Docker Registry: private image registry
- Joplin: note synchronization
- Radicale: CalDAV/CardDAV
Infrastructure as Code and observability
In 2026 the cluster moved under OpenTofu: existing resources were imported without redeployment, and every new layer is enabled by a flag with a known RAM budget. An observability stack sized for Raspberry Pi (VictoriaMetrics, Loki, OpenTelemetry Collector, Grafana) finally gives a view of what is running, and cert-manager provides Let’s Encrypt certificates. Details, pitfalls and choices in the dedicated chapter.
Challenges
Longhorn recovery after reinstallation was the most complex challenge: I had to manually recreate volume mappings to preserve existing data without proprietary formats.
What I Learned
- Fine-grained resource management on constrained ARM nodes
- Deep network debugging (CNI, kube-proxy, CoreDNS)
- Backup strategies without vendor lock-in
- Bringing an existing cluster under infrastructure as code with no service interruption