What does he means, an hour to spin down? You pay for a minimum of an hour per node on EC2? I wasn't aware of that... why don't they charge per minute (or second, for that matter)?
They charge by the partial hour. E.g. when you shut down a small instance after 61 minutes you'll pay for two hours.
Frankly I fail to see how this matters today or what kind of app he envisions where it could matter in the future.
1000 large instances set you back a measly $800 per hour. If your app needs elasticity to a degree anywhere near that then the spindown overhead will probably be the least of your worries. After all I cannot imagine a traffic pattern that would "flicker" at a second or even subsecond rate.We're rather talking about huge surges (slashdot effect) or the classic bell curve pretty much everywhere.
When you compare that $800/hr to the operating costs of running your own 25 fully stuffed racks then you'll still come out ahead by a large margin - with every opportunity for a spindown only adding to your advantage.
> Frankly I fail to see how this matters today or what kind of app he envisions where it could matter in the future.
> After all I cannot imagine a traffic pattern that would "flicker" at a second or even subsecond rate.
The app he's describing in the article: testing of his product. If done on 400 machines in parallel, it would take 2 min 36 sec, or 1.040 machine minutes. But he don't want to pay for 24.000 machine minutes, so he's opting for the the closer to one hour option, saving a lot of money, but failing at his objective: make the deployment (of which testing is a sub-process) as short as possible.
