My thoughts on colocation? When choosing a data center, look at your neighbors. Mine were Dropbox, Netflix, Splunk, and Etsy. Best tools for finding a good data center: www.peeringdb.com, www.datacentermap.com.
Yup. I've moved on but my cage is there, close to Dropbox. Nice digs and nice choice on your part. We were one of the first 10 or so in the building. Pretty full now. How's the new wing looking?
For "finding datacenters", traceroute is useful. You can just pick some websites and see where they host. For example, is seems that pinboard.in is hosted at http://he.net/colocation.html, Fremont 2, or at least that is where I am being routed.
I don't think he was making an argument for ease of maintenance, but rather, reliability per dollar. Even if you've got two crappy datacenters, you're at least redundantly crappy.
You need a capable sysadmin anyways. It makes sense to pay her/him to design and scale your "platform" properly, especially if you're making any money from it.
this is only useful for clients who aren't running their own switches & bgp. if they are running their own network, they can peer with ISPs that don't have physical locations in their dns names. this is a thing you can request, for obvious reasons.
I started out dedicated, and switched to colo about 8 years ago.
Not that I really need colo, could easily go cloud, back to dedicated, or even dirt cheap with some shared hosting provider for hosting needs, but I've always liked having my own customized setup in colo.
Here's why:
1) $50 per 1U (power included)
2) /27 IP range
3) 250GB monthly bandwidth
4) 24/7 DC tech support
Pretty basic setup, couple of Dell R610s with gigabyte switch and Cisco ASA. VMware ESXi runs on top of the bare metal, so basically have my own VPS environment.
It's a nice break from coding to learn a bit about the systems admin side of the fence, I like it ;-)
Of course, the times they are a changing, at some point I'm going to need to go Cloud. For now pretty content though...
Sure, although if you have a 99.99% uptime SLA, I'd look elsewhere.
Provider is SagoNetworks, based out of Tampa, and have a brand new-ish facility in Atlanta.
Started with them around 2003 and have not looked elsewhere, DC guys bailed me out early days while learning the ropes -- tech support is generally very good.
Couple of times a year the shit hits the fan, core router goes down or other mini-disaster, and you get nailed with a few hours downtime. Give them a ring, they say it's blah blah, we're working on it.
Otherwise, everything hums along, no news is good news ;-)
That depends heavily on your company's skills and where you live. Colo is Much cheaper than good (read: useful) managed hosting, but you have to either have the money for a staff of 3+ local sysadmins so that someone can always be on call, or have the skills to admin the boxes and swap hardware yourself and a backup person if you ever want to go on vacation.
I'm assuming that you're in a position where you can't tolerate a few hours of downtime in the middle of the night if something breaks.
I feel like for most companies with a tech staff smaller than about 8-10, you're probably going to have all programmers/DBAs who may have some server skills, but not the competencies to go chase down complex hardware issues at 2am. Past a certain size real sysadmins will have plenty of work do the 99.99% of the time when you're not down, so it's natural to switch over to a colo situation.
There are also tax and accounting pro/cons to buying the hardware.
If you have more than one or two servers, you don't necessarily need to have someone turn up at the colo at 2am. Like with The Cloud, the physical servers just become a resource to be managed. If the discount for colo is enough, for the price of N cloud servers you can have, say, N+2 in a colo (maybe even N x 2). Then you don't have to worry about fixing anything that breaks right away, since you've got some spares that you can fix at your leisure.
That's definitely true, although I'd still be worried about system problems further up the chain. (Load balancers, firewalls, etc). Hopefully even with managed hosting you're still not at the mercy of any one hardware/software failure.
Part of it just boils down to risk tolerance I think, and how comfy you are with sysadmin skills.
I'm responsible a medium sized (< 10 servers) web site that use managed hosting for. Their advice/service is invaluable to me, partially because I trust them enough that I can just hire web programmers without worrying about their advanced linux/networking skills being super high. The host has also helped us debug some fairly deep problems that we don't have a chance at in-house like hypervisor config issues, and even a processor 'errata', AKA processor hardware bug.
It's also nice to have an ops team to ping questions off of, since we're not nearly large enough to hire a dedicated sysadmin otherwise (and they'd be without sysadmin work 50% of the time anyway).
NB: I'm treating 'sysadmin' like it's just one skill set to administer routers and server hardware, configure firewalls and networking, optimize database boxes, etc. This is probably not totally true, but fits for purposes of discussion. YMMV.
I'd love to see a similar study to the Google study[1] done on power supplies. Intuitively, one would think there might be some events/environmental causes for PSU failure (like poor cooling, power spikes) -- so that the likelihood that two servers drop dead at the same time are higher than it would be if the outages were random. Now, "same time" could of course be a few hours, or even days -- giving you time to get a replacement up, before all your database servers are down.
Needing multiple servers, or the ability to easily grow is a good excuse to rent out 1/2 or 1/4 cabinet. One managed server at a place like Rackspace can cost $800/month. If you lease 1/2 cabinet you can put whatever you want in it, but you also have to fix it, or pay someone to.
If you've had a box there for a few years, there's no guarantee that there's empty space near it for more servers or a load balancer.
For a lot of small companies that have 2-3 servers, 'fanatical support' simply isn't worth it. We still end up having to call them to reboot a box or tell them what to do. Or call them and ask why it's down, when they should be calling and telling us.
It pretty quickly starts making sense. I've had clients paying 50k per month to AWS when they have no earthly business paying more than 5k for someone versed in all this.
I don't believe most software folks realize just how much can be saved with colo strategies, building custom servers, buying directly from distributors, component sharing strategies, etc.
It's a great competitive advantage, 10x less burn rate or 10x the hardware of others. Or fall somewhere in between.
I am already blown away at this level already. I see companies pooring money into Heroku/AWS/etc and can't fathom spending a mortgage level type cost for some RAM (e.g. 92+ GB). Don't get me wrong, there IS a use case for these but once you are on solid footing, have large revenues, etc... it is so darn simple to save yourself a few hundred thousand a year. Even with employees.
My experience is that the 1-amp-per-rack-unit rule of thumb applies only if every server in the cabinet is under full load (CPU and all spinning disks). This is almost never the case, however: our average utilization is around 8 amps per phase on a 3-phase circuit. (Load limits are per phase, not total.)
You can also get 30-amp circuits in most data centers if you're concerned about it.
Keep in mind, too, that your load is usually balanced across 2 PDUs (assuming you're buying systems with redundant PSUs, which you should) and on 208-230V power, which is more efficient than 120V due to lower resistance. If you configure your systems correctly, the load will be shared across both PDUs under normal conditions. That said, you'll still need to ensure you don't overload the remaining circuit if redundancy is lost.
In summary, don't worry too much about overloading a 30A circuit; there are plenty of full cabinets in a DC for a reason.
As for the California earthquake risk, one option is to colocate your servers in Dallas or somewhere with no earthquake risk. The Maunder article talks about why colocation doesn't tie you geographically to that place.
I recently did the math with the help of a friend who uses Amazon AWS. I have three primary servers in a data center and by running them myself versus relying upon Amazon, I have saved my company about $5,000 per year, each year, over six years.
If it's just three servers, I'd be astonished if any meaningful amount of time was spent maintaining the hardware after the initial installation. You have to maintain the operating system in either case, so this is a clear win for dedicated machines.
Everybody forgets the E part of EC2. If your workload is not elastic, or is very small, it's incredibly likely that EC2 is not the most cost effective solution.
Indeed, and you can almost certainly buy a server for the cost of the reservation.
Don't get me wrong, EC2 is an amazingly useful platform and it has been a game changer for our industry in a lot of ways. Improved hosting economics isn't generally one of them.
Something I find annoying is that Data Centers often figure that if they get you in, then you will be either unable or unwilling to move so the 'renewal' can be a lot larger than than the original contract. Equinex tried to take us from $108/kva to $450/kva on the renewal. I was like "Really? You don't think we'll move out with a > 4x gain in cost of staying?"
this is true of all commercial leases; If they think it's hard for you to move, well, "market rate has gone up"
That is the thing, though, generally speaking? it costs a /lot/ to fill a space. 10% of the total value of the lease is not unusual. So they try to get you in with a low rate, then jack it up later.
Personally, I think a setup fee, then some sort of guaranteed low rate ongoing (maybe tied to the local price of electricity?) would make more sense. It's less costly for all involved if you sit at the same datacenter forever; it's just that the way that these things are sold, the owner of the data center wants to 'capture value' by raising your rates if it's hard to move.
The owner of the datacenter also wants to start you out with a low rate, for the same reason; if they can get you hooked so that it's hard to move, well hey, they might break even the first year, but after year two or three, they are looking pretty good.
Personally? I find this really frustrating. All parties, when working to maximize their own value, actually waste a whole lot of value for all involved. It's an example of capitalism destroying some pretty significant value.
You see this in almost all cases where switching service providers is expensive (and, I think, it's one of the reasons why "the cloud" is so appealing.)
Personally, I really like the idea of a 'condo' model for the data-center, and for renting in general. Yeah, you'd end up paying $100K+ for your rack, but it wouldn't take very long to make that up, and you'd end up owning part of the organization that controls the whole datacenter.
I think the big problem with the condo model is that in regular condos, usually you don't have individuals that own 100s of units, while that's not at all unusual in a data center.
Hrm, why don't people just rent dedicated servers? For less than $100 a month these days you can get a 32GB server with SSDs. If your server dies, the provider needs to replace it... Seems like all win to me.
Control over network configuration and hardware is one of the big reasons. I hate getting bitten because my three new servers are halfway across the facility and stuck with somebody else who is maxing out the uplink between our shared top of rack switch and the core...
Edited: Also, your "$100/month" quote is pretty fanciful. Who is handing out 32GB with SSDs for that price? I'd really love to know, because the best I've been able to get for a database quality machine (RAID-10 with BBU, SSDs, and lots of RAM) is closer to 10x that.
Despite that being a xeon, the SP series do NOT use ECC ram. I specifically asked them about that that not long ago, after noticing that the details listed ECC ram for the EG server class and above, but didn't mention ECC for the SP series.
I think the $100 price point might have been slightly hyperbolic, but you can get 32GB of memory with 320GB of SSD space from DigitalOcean for $320 per month.
While Digital Ocean is great, it isn't really a fit if you're considering running on your own bare metal server (especially for things like an actually loaded database server).
Bay Area colocation has been expensive since they started billing based on power, rather than physical space. I had a good thing going in the 90s to ~2005ish with a long term contract I signed giving me a full cabinet on a 100mbit unmetered line for ~$600/mo. When that contract expired it promptly jumped to ~$1500/mo.
I know that's not much space compared to the big boys, but it fit our needs at the time pretty well.
If you're looking for colocation stuff these days, and you're not huge and need to be in San Jose, I'd look at some of the spaces in Sacramento.
Other great spots around the country: Dallas, Chicago, Atlanta, and in Virginia around MAE-East.
Agreed, but you do have to sift through a lot of noise; i.e. no shortage of X (maybe shady) provider giving deal of the month offers.
Then again, outside of WHT, where else can you find the colo provider gems? Certainly the best rates will be found there in my experience; if you're lucky, provider tech support will be excellent as well.
