Unless you’re legally required to handle clients’ data in the country in which they reside (more and more common), an on-prem have almost always been cheaper to run (and often faster).
It does require people with different skills to run it, which has to be taken into account.
Perhaps one thing that on-prem is bad at is efficiency: historically, companies often have too much capacity, mostly to cover for peaks of activity, but this leads to lower average server utilisation.
If you're starting a new Rails app and live in the EU, the optimal strategy is pretty clear:
Get a 64-thread box on Hetzner for 250 euro a month, allocate 20% of it to the DB, 40% to web, 40% to bg, then just forget about scaling until you're making money hand over fist.
At a certain point, self-hosting is so much cheaper than PaaS that you can way overprovision your server and not worry about it until you're making serious money.
Just sucks if you are a service provider and need some kind of SLA to be taken seriously (see the flak Fly.io / Resend are getting) since it's so difficult to find good resources on solid setups. Think for every guide for bare metal there are a few hundred for cloud and both assume a handful of users
Unless you’re legally required to handle clients’ data in the country in which they reside (more and more common), an on-prem have almost always been cheaper to run (and often faster).
It does require people with different skills to run it, which has to be taken into account.
Perhaps one thing that on-prem is bad at is efficiency: historically, companies often have too much capacity, mostly to cover for peaks of activity, but this leads to lower average server utilisation.
This made me think of this recent tweet from Nate Berkopec, a developer well-known in the Rails community as a specialist in performance and scaling. https://x.com/nateberkopec/status/1846429195670552635
At a certain point, self-hosting is so much cheaper than PaaS that you can way overprovision your server and not worry about it until you're making serious money.Just sucks if you are a service provider and need some kind of SLA to be taken seriously (see the flak Fly.io / Resend are getting) since it's so difficult to find good resources on solid setups. Think for every guide for bare metal there are a few hundred for cloud and both assume a handful of users