Spot fleet is good, but it may not fit every use case. E.g you can not easily shutdown the non-prod environment with spot fleet, due to its scalable target only allows 1 – 3000. In some scenarios, running autoscaling group on spot instance is a better way to go.
And essentially, it is all about the launch configurations and spot instance request. I developed a bot (named Huggies) to manipulate it. Here is a quick demo.
This is how a typical autoscaling group looks like. Lets say I name my autoscaling group as newsapi-jackie for convenience.
Now I am telling Huggies to use spot instance for my environment by giving the bidding request (bid m1.small for 0.001 per hour).
Under the hood, Huggies makes a copy of the existing launch configurations then add the bidding parameters into it.
To spin up new instances, I need to tell Huggies to start my environment.
Start/Stop environment is just a matter of updating the minimum and maximum size of the autoscaling group.
However my bid price $0.001 is too low to get a m1.small instance. But fortunately Huggies is smart enough to know it, and make a new bid price for me.
How does Huggies know it?
Huggies logs it as:
What if I want to change back to on-demand instance or the spot instance is out of capacity?
Huggies will switch back to the original launch configuration if spot instance is out of capacity or user asks so.
Jackie Chen's IT Workshop 