Homelab

I run one always-on server and a handful of machines around it, and it has taught me more than any course. Everything on it is self-hosted, reached only over a private mesh network, and written down as decision records so I can't quietly re-learn the same lesson twice.

What it taught me

• Mesh networking: Tailscale across every device, HTTPS front doors served over the mesh, nothing exposed to the public internet.
• Selective VPN egress: Proton VPN for exactly the containers that need it, by sharing a network namespace, not by routing the whole box.
• Containers without Kubernetes: rootless Podman under systemd, two users sharing one port space, and what breaks when they collide.
• Storage that matches the workload: per-drive btrfs tiers, and why a database on a consumer SSD stalls — power-loss-protected drives fixed a p99 fsync of 25 ms down to 3 ms.
• Backups are the real risk: designing a 3-2-1 plan when the filesystem has no cross-drive parity.
• One GPU, many consumers: keeping a single model resident on an 8 GB card and making every service consume it through one alias.
• Observability: Prometheus and Grafana, exporters per user, and a monitor that speaks even when everything is fine — because one that only speaks on failure is indistinguishable from a dead one.
• Decision records: forty-two of them. The habit outlasts the hardware.