The costs skyrocket if you’re not doing it right - otherwise it tends to be linear with request count. The nice thing, though, is that you don’t have to panic or do anything at any scale - it’ll just keep running along.

Once you hit high levels of traffic, instead of panicking and trying to add more servers and fixing DB structure, you do a cost benefit analysis and see what to do next. Either work on reducing your cost slope, or make it logarithmic by moving to regular servers.

The problem is the per request cost is going to be astronomical compared to what it needs to be once you e.g. get above the initial 1 GB/month tier. Bandwidth at AWS costs about 50 times as much as with Hetzner for example.

> Once you hit high levels of traffic, instead of panicking and trying to add more servers and fixing DB structure, you do a cost benefit analysis and see what to do next. Either work on reducing your cost slope, or make it logarithmic by moving to regular servers.

...and that's a good reason to try to avoid the lock-in of AWS specific services. Relatively entry level setups on alternative platforms will let you scale well past the system where AWS costs starts to get painful and people start panicking about how to cut costs. When you hit that point you really don't want to be facing a major re-engineering effort to be able to even use other options to bring the cost down just as the costs are soaring.

> Bandwidth at AWS costs about 50 times as much as with Hetzner for example.

It's actually closer to 70x

Hetzner: 1.25 USD per TB (after 30 TB free per server) [1]

AWS: 90 USD per TB [2]

So, 100 TB at Hetzner will cost you 125 USD (not counting the 30 TB free). You would pay 9000 USD at AWS.

[1] https://wiki.hetzner.de/index.php/Traffic/en

[2] https://aws.amazon.com/ec2/pricing/on-demand/

Hetzner traffic isn't even that cheap compared to some unmetered offers, for example online.net.

I had to migrate away from Hetzner because I was actually paying 300 dollars on traffic alone.

The problem with some of the unmetered offers is that it's hard to know how much you'll actually be able to use, and for how long (before they oversell), so there's some appeal in paying for actual capacity. But certainly, you will pay a premium for that.

Thanks for pulling the actual numbers - I was too lazy to verify, so figured it best to err on the conservative side (and also depending on how high you go, the AWS prices per TB drop quite a bit).

Exactly this. You could host cheaper with a really low tier VPS, but as soon as you get posted to hacker news, game over. If traffic increases to such a degree that this architecture is no longer the best choice, we can move off without constantly fighting to keep our tiny server(s) alive.

The big issue is reserved instances - in order to get that level of savings you need to lock in for a year. At that point you're basically running on-prem servers since you're committed.