In theory, that's the idea, but don't host a webserver or anything on it that can't die. This is traditionally used for distributed jobs like data churning, rather than services, since if the spot price goes over your max bid, your server turns off.
Also, just making sure you have your numbers right, it's $0.085 per hour for on-demand, not $0.85. Might make a difference in your projections :)
Let's go over your questions.
So basically I want my original instance to be running at all times. Then when it starts going over capacity I want the Auto Scaling Group to start launching instances and the Load Balancer to distribute the load across them. Is my thinking here sound?
I'd say yes, but I do have a couple reservations. If I understand correctly, you've placed your "main" instance outside of the auto scaling group. Theoretically, that would ensure that auto scaling doesn't kill off your original instance. There are a couple of things I'd like to mention:
- You're not making full use of the possibilities of Auto Scaling. Auto Scaling not only enables your setup to scale, but it can also ensure limits. If, for whatever reason, your "main" instance dies, your auto scaling policy won't come into action. If you keep your instance in an auto scaling group with a
min-size of 1, Auto Scaling automatically replaces the failed instance.
- When auto scaling, it's often best practise to treat your instances as being "disposable", because that's how you build resilient systems. Don't depend on one instance to always be available.
- You can set the termination policy for your auto scaling group so that it always terminates the newest instances first. That would ensure your "main" instance will be kept (as long as it's healthy). My previous comment still applies though.
When I make code and data changes to my original instance, do I have to remake the image my Launch Configuration uses?
I'd say no, but that's more of a design issue. Your image should describe the state of your server, but it shouldn't be responsible for code distribution. Consider a situation where you'd have to update your application because of an urgent bug, but your servers are under high load. Does updating your main server, creating an AMI, updating your launch config and killing off your auto scaled servers so they can be respawned with the latest AMI sound like an attractive solution? My answer to that would be no (again). Look into source code version control and deployment strategies (I'm a Rails developer 60% of the time and use git and capistrano, for instance).
There are situations where your approach would work as well and there is a lot of middle ground here (I would recommend also looking into Chef and userdata scripts). I myself actually rarely use custom AMIs, thanks to Chef.
What needs to be down with DNS names and IPs? I'm currently using Route 53, do I make that point to my Load Balancer and that's it?
Basically, yes. You can select the loadbalancer(s) that should be attached to new instances when creating your auto scaling group. I haven't used the GUI for Auto Scaling yet, but I'm quite sure it's in there somewhere. If not, the CLI still supports it. Point your Route53 record to your ELB alias and that's basically it.
Response to additional questions (2014/02/23):
If you're developing using Rails, I can highly recommend Capistrano for deployments, which can take a specific version of your app in your preferred version control system (like git) and deploy it to a number of servers in a specific environment. There are a bunch of tutorials out there, but Ryan Bates' revised (and pro) Railscasts on the subject are very concise, although you need a subscription to his website to watch both of them. Most of the other tutorials will get you going as well though. Fire up a new server with your AMI and a launch script that pulls a temporary clone of your git repo and runs a local Capistrano command to get your app going. This ensures that, later on, you can also deploy new versions of your application using Capistrano with just one command to all running servers.
Capistrano also provides a couple of other benefits, including enabling you to execute specific tasks on all or just one of your servers and roll back a deployment, which is very hard to accomplish using just git.
Best Answer
Lambda works somewhat like this, though instead it makes the user wait while it spins up an app. In Lambda though you don't have an instance to manage -- its all done under the hood.
I'd also check if you could do something involving Lambda and a Route53 failover check, such that users hit Lambda if the instance is down and the instance if it is up. This might not work of Lambda requires a specific Host header though.
Or use R53 failover again to run a reserved-instance t2.nano to do the 'waiting room', then proxy or redirect to another instance once it is up.