Basically, you don't. It is the wrong way to think about scalability and reliability.
A single EC2 instance is a single point of failure. Changing instance types requires a reboot, so your site will be unavailable while that happens. Even if you do take this approach, you would be limited to instances in a single AZ, which again becomes a single point of failure.
Finally, since you want to minimise down time, you would be under or over provisioned for much of the time - so you'd have poor performance for you users, or you'd be paying too much.
I would recommend starting by reading The Well Architected Framework to get a better understanding of why you want to do this.