3 ms·

It's easy to say this when you're hacking on a side project and just need the smallest instance possible, or just a medium instance. But when you're scaling for

It's easy to say this when you're hacking on a side project and just need the smallest instance possible, or just a medium instance. But when you're scaling for millions of users, then those costs add up.. really really quick. AWS, especially nickel and dime you for the smallest of things, such as total # of I/O requests. Plus I find, the whole myth about able to scale faster.. somewhat misleading. I use EngineYard, and you need to ALLOCATE a certain amt of space. Sure you can increase it in the future, but you'd need to shut down the instance, and create a new one with the extra space... not exactly fast nor convenient. And it's not like you can't add an extra hard drive if you hosted your own servers...

druiid 14y ago
I think there is always going to be some amount of overhead with physical or cloud services for scaling. I can say though that one thing you can do with physical servers which is pretty neat is shut down entire physical nodes when you don't need them to save on resources. Your scaling software has to know how to do this, but IPMI is fun stuff like that.
Corrado 14y ago
I agree that the cloud can make the best case for hacking side projects. It's quick to deploy and quick to destroy. However, scaling your system depends on how you architect the system. Building in the cloud is not the same as building in your colo cabinet. Your example of moving your instance to a bigger one is a classic example of doing it wrong (so to speak). In the cloud you don't scale up, you scale out; many machines doing the work. And bringing those machines on-line takes minutes, that is certainly faster than shutting down and rebuilding. I've played with EngineYard & Heroku and while I really like their services, they tend to re-enforce the "old world" way of doing something. Break out and play with "raw" AWS and I think you'll find a lot to like.

druiid 14y ago

I think there is always going to be some amount of overhead with physical or cloud services for scaling. I can say though that one thing you can do with physical servers which is pretty neat is shut down entire physical nodes when you don't need them to save on resources. Your scaling software has to know how to do this, but IPMI is fun stuff like that.

Corrado 14y ago

I agree that the cloud can make the best case for hacking side projects. It's quick to deploy and quick to destroy. However, scaling your system depends on how you architect the system. Building in the cloud is not the same as building in your colo cabinet. Your example of moving your instance to a bigger one is a classic example of doing it wrong (so to speak). In the cloud you don't scale up, you scale out; many machines doing the work. And bringing those machines on-line takes minutes, that is certainly faster than shutting down and rebuilding. I've played with EngineYard & Heroku and while I really like their services, they tend to re-enforce the "old world" way of doing something. Break out and play with "raw" AWS and I think you'll find a lot to like.