the machine room doesn't fit in a pocket — open alflora research on a real computer.
two servers, one electricity bill, every mistake documented
Server 1 — daily driver: Intel Xeon E5-2686 v4, 64GB DDR4, Intel Arc A770 16GB, X99 board. Works the day job and the night shift.
Server 2: dual Xeon CPUs, 128GB RAM, lots of storage. The heavy lifter.
Finding: last decade's enterprise hardware is this decade's homelab bargain. The cloud rents you less for more.
The entire operation — both servers, the GPU, git, CI, storage, status page, and an AI chat — costs about 1000 rubles (~$10) a month in electricity, at full power, 24/7. Hardware owned, so that's the whole bill.
Finding: price/performance is the point. Less than a streaming subscription, more uptime responsibility than most jobs.
Two stacks, one docker network, and a postgresql alias that resolved to the wrong database: Gitea connected to authentik's postgres. Everything "worked" until it very much didn't.
Finding: a network alias is a global variable. One alias per consumer, named like it hates ambiguity — because in a shared network, it will be shared.
The server cannot reach its own public IP — hairpin NAT isn't happening on this network. Anything that "phoned home" through the public hostname simply didn't.
Finding: everything internal now talks over container networks or 127.0.0.1. Public hostnames are for the public; localhost is for family.
Externally hosted assets started returning 401s, and a deploy UI froze on them. The assets weren't ours, the outage wasn't ours, and the fix being "wait for someone else" was unacceptable.
Finding: everything is self-hosted now. An external dependency is an outage you haven't scheduled yet — see the design notes on why near-zero images is policy.
If you can't fix it at 3am, you don't own it — it owns you. Everything on these servers has been broken at the worst possible hour at least once, and everything got fixed by the same pair of hands.
Finding: that's not a bug of self-hosting; that's the entire feature. Uptime is a personality trait now.
The 3am rule: everything on these servers has broken at the worst possible hour at least once — and got fixed by the same pair of hands. That's not a bug of self-hosting; that's the entire feature.
This server cannot reach its own public IP — hairpin NAT simply isn't happening on this network. Everything that phoned home through the public hostname silently didn't, until we rewired it all.
Both servers, the GPU, git, CI, storage, the status page and an AI chat — at full power, 24/7, for about 1000 rubles of electricity. Hardware owned, so the bill is the whole bill.
Browser-based Android emulation with a free compute hour a month. Built, deployed, verified, and killed in one day. The stack worked perfectly — that was the problem.