Backup tunnel: surviving a VPS outage without losing clients

The hosting setup in question looks like this: physical machines with GPUs sit at home in Russia and are rented out through the vast.ai marketplace. That does not work directly — you need an exit abroad. Each rig holds a tunnel to a rented VPS overseas, and the whole world sees the machine at that VPS address.

This arrangement has one weak spot, and you remember it exactly once — when the provider emails you about scheduled maintenance. Say a forty-minute window, and all your VPS instances sit in the same data centre, which means they reboot together.

What exactly is lost

The VPS goes down, the rig goes with it. Not "runs slower" — it disappears from the catalogue entirely: the marketplace stops seeing it, active rentals break, clients move to other hosts.

Forty minutes of that is every machine you own dropping out of search results at once, and a reliability rating damaged on each of them.

The cost is asymmetric. A rating dip recovers by itself within hours. A client who leaves never comes back: they find another host, set their environment up there, and stop looking for you.

The idea

Rent a spare VPS from a different provider in a different country, and teach the rigs to hop over to it automatically. Sounds simple. In practice you immediately hit questions the marketplace documentation does not answer:

• Does the marketplace penalise an IP change — or only genuine unavailability?
• Does unlisting the machine before the switch help?
• How deep is the dip and how long does it take to recover?
• What happens to client containers during the move?

What follows is what came out of it. First the measurements on live machines with paid rentals, then a ready-made solution with scripts.

Every figure in this guide comes from six rigs: 2× RTX 5080, RTX 5090, 4× RTX 4090, RTX 3080, RTX 3080 Ti. Tunnels on Xray 26.3.27, VLESS + Reality + Vision, the kaalia host agent.