Wouldn't people also want to rent the machines during holiday season as much as they do during the rest of the year?
jkorn
@williamx presumably you will get a heavier load during peak seasons, so you would need to scale out. You would need as many machines as you use under normal load, plus more to accommodate the extra load you get during peak seasons.
mattlkf
@williamx I think the key point here is that the cost of over-provisioning is proportional to the ratio of peak demand to lowest demand, not the difference.
For example, if the peak demand at the end of the year requires 2 machines but for most of the year the demand could be served by just 1 machine, then for most of the year you are paying for double the amount of machines you need. But if the peak demand is 11 machines and for the rest of the year you can get by with 10 machines, then for most of the year you are paying for only 1.1x the amount of machines you need.
In other words, when Amazon rents machines out year-round to a customer with uniform demand, the entire blue graph shifts up by a constant amount and the ratio of peak demand to normal demand lessens, so the cost of over-provisioning is less.
axiao
@williamx Another idea that could work is to just increase the price of renting machines during periods of high demand. This would allow them to better control how many machines they rent out at any given time, while also improving their ability to capture $$$ from the increased demand during periods like the holiday season.
Amazon does have something like this called Amazon EC2 Spot Instances, which provisions out spare compute power to the highest bidder, essentially creating a market for compute power.
Wouldn't people also want to rent the machines during holiday season as much as they do during the rest of the year?
@williamx presumably you will get a heavier load during peak seasons, so you would need to scale out. You would need as many machines as you use under normal load, plus more to accommodate the extra load you get during peak seasons.
@williamx I think the key point here is that the cost of over-provisioning is proportional to the ratio of peak demand to lowest demand, not the difference.
For example, if the peak demand at the end of the year requires 2 machines but for most of the year the demand could be served by just 1 machine, then for most of the year you are paying for double the amount of machines you need. But if the peak demand is 11 machines and for the rest of the year you can get by with 10 machines, then for most of the year you are paying for only 1.1x the amount of machines you need.
In other words, when Amazon rents machines out year-round to a customer with uniform demand, the entire blue graph shifts up by a constant amount and the ratio of peak demand to normal demand lessens, so the cost of over-provisioning is less.
@williamx Another idea that could work is to just increase the price of renting machines during periods of high demand. This would allow them to better control how many machines they rent out at any given time, while also improving their ability to capture $$$ from the increased demand during periods like the holiday season.
Amazon does have something like this called Amazon EC2 Spot Instances, which provisions out spare compute power to the highest bidder, essentially creating a market for compute power.