Bare-metal and on-site hardware

Sometimes the cloud is simply off the table. The data isn't allowed to leave the building, or the site has no reliable link to anywhere, or the power fails often enough that trusting a provider stops being an answer. Then the servers come to you. The work is the same Kubernetes, just running on hardware you own.

What changes is that nobody else holds the floor under the system. The storage is yours, the network is yours, and so is the plan for when the power goes. The release has to install the same way on a rack in a room you can walk into as it does in a data center. And this is the part people really underestimate: a platform that only starts correctly when somebody remembers the order of the steps isn't finished.

In Kyiv the grid can drop with no warning at all, and that's just a condition you plan for here. So a write has to be safe before the power goes. And when the power comes back, the whole system has to come up by itself, with nobody standing next to it.

The nearest public example is this site. Its frontend and its content backend run on infrastructure I host and operate myself, with Cloudflare in front. That's the self-hosted end of the layer rather than an on-site build, though the discipline is the same: I keep control over how the important parts run.