| > Don't start a conversation with an opening generalization like that if you want something constructive. Especially when the rest of your post is clearly based on the single anecdote of your experience. You are doing as much, if not more generalization by way of the assumptions you're making. > >Running your own hardware is AWFUL.
> Maybe for you. Not for any sysadmin with even just a couple of years of experience. Sysadmins aren't real estate attorneys or facilities managers or security personnel. A lot of them aren't even tech ops who physically manage the DC hardware and are oncall 24/7 to fix problems on-site if needed. You need all of those things to run your own DC, and potentially a lot more if you're physically building the center itself (architects, contractors, civil engineers, etc.). And if you're serious about latency, availability, and durability, you're going to need those things in multiples for however many datacenters are needed to meet your targets. How many organizations have the millions of dollars of capex needed to get that off the ground and keep it all running? > Netflix doesn't run 1/3 of the Internet traffic off of AWS, only a tiny subset because of the aforementioned shitty economics. To how many organizations do the economics of serving 1/3 of Internet traffic apply? 2? How is that a counterexample to his point about datacenters only making sense for the very largest? Even if you sidestep all those costs by renting instead of building and even if we take for granted that your "shitty economics" are still shitty down numerous orders of magnitude from Netflix-size, you're still burning money making your devs design and operate your system twice - once for the DC, once for AWS, and however much work it is to glue the 2 together. The end result may be cheaper infrastructure-wise but it will also be unavoidably less reliable and more complex purely by virtue of having more than twice as many moving parts. Let's say implementing things this way takes a single dev time-and-a-half compared to just doing it on one or the other. Let's say (very conservatively) you pay your dev $100k a year + $50k benefits. You're now $75k in the hole from the get-go. That's enough to pay for roughly 83 m5.large EC2 instances on-demand (no RI) for a year. Your company has to be very large for the marginal savings of using a DC to outweigh that kind of deficit. |
I think you are exaggerating when talking about "millions". Companies running in old school DC will not build the actual DC. They will rent a 42U cabinet or part of it.
You can typically rent 1/4 of a cabinet for a few hundred bucks a month.
I you operate at a small scale, you typically don't have that many servers, maybe 4 or 5. A decent server is in the range of 4000 to 6000$ and will last generally around 5 years.
And these servers rarely breaks (we have a fleet of more 300 servers, and maybe 1 or 2 "wake the on call" crashes a year).
Throw in 1 or 2 decent switches at about 1000 to 2000$ each and you are mostly good to go.
You end up with a capex of ~50 000$ to get you started, with a depreciation over 5 years.