The rungs
A handful of physical numbers set the floor under every design, and no amount of clever code beats them. Reading from the processor's own fastest cache takes about a nanosecond. Main memory is around 100 nanoseconds. A random read from a solid state disk costs roughly 100 microseconds, a thousand times more than memory. A round trip inside one data centre is around 500 microseconds. A round trip across an ocean is about 100 milliseconds, and nothing you deploy will make light move faster through glass.
The jumps matter more than the figures, because each jump is where an architecture wins or loses. Memory to disk is a factor of a thousand, and that gap is the entire case for caching. Same data centre to another continent is a further factor of 200, and that gap is the entire case for content delivery networks and regional copies. Move the data near the user, because you cannot move the user near the data.
Killing designs before anyone writes code
The ladder kills designs before anyone writes code. A page making 40 database queries one after another at a millisecond each inside the data centre costs 40 milliseconds, which is fine. Run that same chatty loop across regions at 100 milliseconds a trip and the page takes 4 seconds by construction, no matter how good the code is.
That is a structural fix, not a tuning one. Batch the 40 calls into one, or move the computation to where the data already lives.
Counting the round trips on the critical path before you draw anything, and multiply. It takes thirty seconds, it kills bad designs early, and interviewers notice when you do it unprompted.
Worked example
Fatima is designing a profile page for a social app with users in Sydney and servers only in Oregon, 140 ms away. The naive flow makes 3 sequential API calls, each one crossing the Pacific: 420 ms of pure round trips before rendering. She walks the ladder for options. Batching the 3 calls into one endpoint gets it to 140 ms. Adding a read replica in Sydney puts data 2 ms from those users, and the same three calls now cost 6 ms; even unbatched, the page feels instant. The replica costs about 400 dollars a month, which the team happily pays, since the alternative was physics. Her rule of thumb afterward: count ocean crossings first, database indexes second.