Thoughts on Colocation

After a week of slogging servers around northern California, I thought a brain dump on colocation might be useful to readers, and to future me.

I wrote about the difference between colocation, leased servers and other kinds of hosting in an earlier post. This one is strictly about colocation, the 'Condo' approach where you own a bunch of hardware and need a place to put it.

What you are after is not complicated:

• A physical enclosure
• Some kind of Internet connection
• Power
• Security guards

However, buying it is a pain in the neck.

Money

It's typical to sign a one- or three-year contract. Right off the bat this introduces an element of pressure, since you're making a fairly binding, long-term decision without knowing what you're doing. Unless you live in the Bay Area, colocation is a cost on par with your rent or mortgage.

Worse, I've found that the cost for identical configurations can vary by a factor of three or more. You really do need to shop around. And you need to be especially careful of punitive terms for things like overstepping bandwidth or power requirements. These are things you have to dig out of the fine print of contracts at a moment when you just want the whole process to be over.

Salespeople

Renting colo space feels like buying a car. You typically decide on a specific configuration you want, and then ask for quotes. For the salespeople involved, this is just the start of a long conversation they want to have with you. They'll be very curious about your budget, and want to talk about the hosting equivalent of an underbody clearcoat (various "hybrid cloud solutions") and extended warranty.

Although colo space is a commodity, salespeople become tetchy if you treat it as such. They will insist on talking to you over the phone and bristle at the suggestion that their job could be replaced by a web form. It is a good idea not to think about how much their salary or commission adds to your costs.

The reason I say colo space is a commodity is not because all facilities are the same, but because small-time clients will have no practical way of assessing their quality. There are certainly some facilities that are obviously bad, but most data centers have sane policies, look nice if you visit, and talk eloquently about uptime. The only way to evaluate a data center is to go through a series of small and big outages together, but by then you're already in a multi-year contract. So in practice, there is not enough information to pick a "better" data center, just a bunch of anecdotes on message boards. I believe you're better off getting space in two cheap places than trying to pick one high-end one.

Resellers

A lot of colo space is resold through intermediaries. This is often the only way to get a smaller amount of space than a full cabinet. Someone rents a bunch of cabinets at a data center and then parcels them out by the slice to clients, in pieces as small as 1u.

There are two things to watch out for in this arrangement:

1. A bunch of people will have access to your physical equipment. Good resellers will take pains to introduce the various people in a cabinet to one another (or at least provide contact info).
2. If your provider gets in a dispute with the data center, you may not be able to physically take out your hardware. It may even be seized without advance warning.

Power

Power is the great bugbear in hosting. You need to know how much your equipment uses, and how much you're likely to need. It is also often the limiting factor.

A useful rule of thumb in my case has been 1 rack unit = 1 amp. However, it is quite difficult to estimate the power consumption of a server before you buy it. You end up having to plug it in to a Kill-A-Watt meter under normal load.

A full cabinet is somewhere around 42 units, but a typical full cabinet power allotment is 20 amps. So you can't just fill a cabinet with servers.

The power situation is even worse than it sounds. That 20 amp figure is strictly theoretical, since you aren't allowed to use the full amount. You are limited to 80% of this figure, so a "20A" cabinet has 16A usable power, enough for eight pinboard-style servers.

Finding Datacenters

Finding information is another pain in the neck. The place of choice is an awful forum called WebHosting Talk. There is an open business opportunity to anyone who can stick a front-end interface on this site that lets you enter number of servers, bandwidth, physical location, and spit out a list of offers.

Another business opportunity is to make an authoritative directory of Bay Area data centers, since there is a bewildering assortment of sellers, re-sellers, and re-re-sellers offering the same physical space. Conversely, some large providers maintain multiple facilities.

Earthquakes

Nobody likes to talk about earthquakes. But anything in the Bay Area or Seattle is going to come crashing down at some point. Another thing that has proven impossible is finding out what facilities are at highest risk. It's one thing to go offline when the Big One comes (half the Internet will be down with you). But losing a rack full of hardware into the maw of the earth is worse.

So here's my plea to hackers: figure out where stuff is physically hosted, correlate it with seismic hazard maps, and make a nice web form that lets people shop for specific power/bandwidth/space configurations without talking to salespeople. Charge money for it! I will pay you! Others will pay you!

—maciej on August 03, 2013