You already know: the system design method (lesson 01): clarify, frame, estimate, design, evaluate, defend. Every case in this bank includes a back-of-the-envelope section. This lesson teaches the method behind those sections: which numbers to estimate, how to get them in minutes, which reference values to carry in your head, how to size ML training and serving, and how to check that an estimate is sane. The goal is not precision; it is numbers that change decisions.
How to read this lesson. No code, on purpose. All values are approximate and marked as assumptions or orders of magnitude; real systems must be measured.
1. What estimation decides
Estimation exists to decide things. Four kinds of numbers drive most decisions:
The staff habit: before computing a number, say what it will change. "If this is above about 10,000 requests per second, we need a cache in front of the model." If no answer would change the design, skip the number.