However, it also points out much lower-end hardware solutions that cost under $10k but perform much better than the largest EC2 instance, for I/O.
ETA: This is why I tend to roll my eyes at the notion of "commodity" hardware. The article's low-end array is 400MB/s, but rolling ones own can yield over twice that for the same or lower price tag. All this well before reaching the unscalable cliff of enterprise pricing.
The main difference is that, even in startup companies, you have a supported solution and can call someone at 2 am when your storage dies and expect to find replacement parts and support. Good luck trying to drive to Fry's and buy replacement hard drives for a server that was built 2 years ago by someone that no longer works there.
This may not be the case at a startup outside of a technology hub and is almost certainly not the case where nobody in the company is competent with hardware[1].
Otherwise, a failure isn't just easy to correct but cheap, too. It's so cheap that keeping cold spares around is a no-brainer, unlike with the "gold-plated"[2] products.
The other danger is that the "engineer" who comes out to hand-deliver and install the replacement part gets it wrong. If he pulls the working, rather than failed, part[3], there's not much consequence for him personally or the vendor, unlike with startup founders or even staff.
In fact, even engineering it with enough hot spares to ride out its useful life. 4% AFR for your 88 disks? Just add 8 hot spares. Worried about cables or the controller card? Double up. Both together raises a $9k 1 gigaBYTE/second (4GB/s peak) array to a whopping $10k.
Even just doubling everything is likely to cost far less than a year of tarnished-bronze support from a big vendor.
[1] By which I mean assembling discrete consumer parts, not soldering or anything lower level.
[2] Personally, I prefer to refer to "enterprise" targeted pricing as "hookers and blow," but it is, admittedly, without a catchy adjective.
[3] Yes, I've had this happen. I've also watched a colleague pull the wrong drive out of an array, against his better judgment, at the insistence of the vendor's phone support.
What's that? You can't buy that exact drive so now your homemade RAID 5 is running in degraded mode and you hope it will stay up long enough to copy your data off onto another system? Sucks to be you, you tried to save a few bucks and got burned.
In the enterprise, we pay big bucks because we want to KNOW that we can call an 800 number and get an exact replacement hard drive, even if they stopped selling them 3 years ago.
With arrays I build, I don't have onerous constraints like requiring identical size[1]. Moreover, if I'm not already already retiring disks at the 3 year mark, I'm very much remiss in my duties.
* Sucks to be you, you tried to save a few bucks and got burned.*
We're not talking about a few bucks. We're talking hundreds of thousands of dollars. That's enough to pay a salary for those 3 years as well as having replaced with something less than a couple generations old.
[1] I assumethat's what you mean, since true geometry is all but impossible to detect on modern drives.