irfan in irfanadziri.dev in ~/homelab on main via astro v7.4.12
Homelab
One Proxmox node that runs my apps, my dev machine and my agents, reachable only over a Tailscale mesh, with nothing exposed to the internet.
The homelab is where I run what I build for myself, and where I try things before I’d do them for anyone else. It has run since October 2024 on one machine that I administer alone.
Topology
Text version
My laptop and phone connect to the Tailscale mesh, which is the only way into the lab. The mesh reaches a single Proxmox node. On that node run self-hosted apps in LXC containers, the dev VM I write code on, agent and automation workloads, the monitoring dashboard and a Proxmox Backup Server. The apps, the dev VM and the agents are backed up to that backup server. A dashed edge from the internet to the node is marked blocked: there are no open ports.
One Proxmox hypervisor runs everything. Small services and apps live in LXC containers, because a container costs a fraction of a VM’s memory. The machine I develop on is a full VM, so it behaves like any other Linux server. Agent and automation workloads, and the dashboard that watches the node, run beside them.
Access model
Tailscale-only access. No WAN exposure. That’s the design rule, and everything else follows from it.
- A device of mine joins the tailnet with its own identity.
- The mesh routes it to the node, which nothing outside the house can otherwise reach.
- The router forwards no ports. There is no public IP to scan and no login page facing the internet.
The reasoning is about who can knock. A lab run by one person, in spare hours, will lag behind on patches; anything reachable from the internet will eventually meet that gap. With no open ports, an unpatched service is still a problem, but only for devices I already trust.
It’s also why analytics for this site is a hosted service rather than something on the lab: self-hosting it would have meant opening a port.
Backups
Proxmox Backup Server takes backups of the guests: the app containers, the dev VM and the agent workloads. I have restored a guest from it, so the backups are known to restore, not just known to run.
Monitoring
The node is watched by a monitoring service I wrote with nothing but the Python standard library, reachable only over the mesh. How it works, and why it isn’t Grafana, is written up as a system doc: Monitoring with zero dependencies.
What this isn’t, and what I’d change
It isn’t highly available. There is one node, so a hardware fault stops everything until I fix it, and nothing fails over. It’s a lab, and that trade is fine for a lab.
The backups are the weak point, and I know it. Proxmox Backup Server runs on the same node it protects, so a dead disk or a dead machine takes the backups with it. They cover mistakes, a broken upgrade or a deleted container, not hardware loss. The first change I’d make is a second copy off the node: a sync to another box or to object storage, so one failure can’t remove both the guests and their backups.