This is an awful oversimplification, but in my experience, there are two different kinds of applications:
- User-centeric (think: an e-commerce site): your costs are going to grow somewhat linearly with your userbase/traffic. Do a load test to figure out how many concurrent users each dyno supports and then model your costs with that in mind.
- Data-centric (think: an analytics platform): your costs are to going to grow linearly with the amount of data points each customer is going to be using. Figure out how many data points a single worker dyno (running 24x7) can process and then model your costs with that in mind.
- User-centeric (think: an e-commerce site): your costs are going to grow somewhat linearly with your userbase/traffic. Do a load test to figure out how many concurrent users each dyno supports and then model your costs with that in mind.
- Data-centric (think: an analytics platform): your costs are to going to grow linearly with the amount of data points each customer is going to be using. Figure out how many data points a single worker dyno (running 24x7) can process and then model your costs with that in mind.