We use ELB (Elastic Load Balancer) with an auto-scaling group behind it. The instances get added or removed based on the latency reported by CloudWatch/ELB. During the course of a day, the number of EC2 instances running can vary from 8 to 25.
Usually 8 is enough, but when spikes happen or when the capacity of instances to process stuff drops (which does happen in cloud-computing, based on what neighbours you have and what they are doing), then new instances are started in a matter of minutes.
And this billing by hour does save us a lot of money. Because 8 instances is enough, until it isn't and the traffic is so huge that it can choke and freeze 8 instances.
Incidental, this ability is one reason we moved off Heroku.
Its computing and hardware, absolutely everything fails eventually.
"EBS would fail often too"
How often is often, and in what way did it fail ?I'm currently thinking about using RDS in a production environment and it would be interesting to know the limitations.
But, that's why AWS gives you RDS backups, AZ failover, etc. -- the backups won't fail (they're stored on S3), and with AZ failover, you can make downtime vanishingly small. But you need to actually do that.
A lot of people seem to confuse S3's 99.9999...% reliability with the EBS-backed stuff, which you must plan for eventual failure with.
plan for EBS to fail
plan for RDS to fail
'snapshot early snapshot often', as snapshots will help you in case of a failure
there is no automatic snapshot functionality ( yet? )
plan how to snapshot an EBS RAID (Filesystem, uptime)
I'm not heavily experienced in cloud offerings, but it's far easier to manage the AWS stuff than the rackspace - the DNS management is quite flexible yet couldn't be simpler with AWS's 'Route 53', but with the part of Rackspace we were using it wasn't 'all in one place', which made it hard to peruse or alter.
I found support at both places to be upbeat and knowledgable, though I don't know about timeliness since I've only really lodged low-priority tickets. AWS does need more domain knowledge in order to understand its flexibility, and I've gotten a good workout from my $50/mo support add-on.
I appreciated their honestly but they were lacking a few required services for us at the time.
If I need on-demand instances these days, I'll do it at Linode. They bill to the day, so you don't need to commit to a month for temporary instances. It's not as scriptable, but it works and doesn't cost a fortune over renting hardware either. More often, I just get servers with more CPU cores and more RAM than I need so there's plenty of room to absorb spikes.