I'm not sure how many servers Dropbox is up to now, but we've got over 200 physical machines (probably nearing 300, growing quickly), and much more when you break them down to how many virtual services there are running. We do that with 3 operations people.
The biggest advantage for us has been the ability to scale up, rather than worrying about how to scale out right now. For a startup, I'd say this is even more important. Horizontal scaling is one of the most difficult parts of my job. That's not saying that it's hard, but the fact is its straight up not as easy as just buying better hardware.
> 1000s of servers
When was the last time someone needed to add multiple 1000s of servers on-the-fly (i.e. 5 minutes ago)? Am I just naive to think that this would be a huge spike? If you're talking about getting 1000s of servers over a longer time frame, I'm pretty sure that AWS would be able to keep up.Before that, for anything that fits on less than, say, 20 servers, you're way better off renting servers.
Besides, this article is about scaling. If your needs are static, who cares what you use. It's about where you go from there, and I'd rather be on the cloud before I have to.
Management tools have different needs from 2 servers to 10 to 50.
- Sometimes you just don't know whether your traffic is going to spike 10% or 100%. And I'm not talking about one or two computers, I'm talking about adding hundreds. Are you sure your supplier has everything in stock for a rushed overnight order? Your exact hard drive model? Your aging CPU? You're seriously going to experiment with a new batch of shit and tell your team to spend all night wiring all of those racks? Do you even have the space left in your cage? Enough space on your routers? Enough power? Even if you're in managed hosting like SoftLayer, these are now their issues and Not Their Problem if they can't turn around for you in the time you need.
- In the time we were on the cloud we were better able to understand our hardware needs so that we could actually spec out machines optimally. Even better, technology improved considerably to bring us low-cost SSDs which wouldn't have been possible at the start.
- There was no way we could manage these servers and a datacenter without a dedicated network engineer and an SRE. And even that was pushing it. If you've ever tried to hire these positions, it's even harder to find good ones than software engineering. We got really lucky. Also, you've spent a lot of time on your engineering interview process and you have it down -- now do the same for two more positions that you know much less about.
- There is a huge engineering cost to moving off and building your own tools. Two servers? OK. Two thousand? Different ball game.
- I would argue that even a company like Google uses essentially a cloud solution that they've built internally and made available to their teams. AWS helps make a piece of that accessible for the rest of us.
TLDR: I thought I was hot shit too when I ran a Newegg server off my parents' internet connection, but come back when you're pulling down a megawatt and tell me the cloud sucks.