advanced 2 min answer

When is first-principles reasoning the right tool for an architect, and when is it a waste of time?

first-principlesphysicsassumptionsjudgementnvidiaconceptual
Show the full answer Hide the answer

When it is the right tool

When the received answer is expensive and the underlying constraint is physical or arithmetic.

  • Latency floors. Light in fibre between distant continents sets a round-trip floor that no protocol removes. Reasoning from that changes the design question from "make the round trip fast" to "make the user not wait for a round trip" — which is solvable.
  • Bandwidth and storage arithmetic. How many bytes, how often, at what cost per gigabyte. This routinely overturns intuitions about what is affordable.
  • Queueing behaviour. Delay grows as 1 / (1 − utilisation), so a utilisation target follows from a latency requirement rather than from cost preference.
  • Amplification arithmetic. Three retries at three layers is 27×, not 3×. Fan-out to twenty components means the p99 is set by each component's p99.95.
  • Capacity of a single node. What one machine can actually do, which is usually far more than assumed and frequently removes the case for distribution entirely.

In each case the calculation is simple, decisive, and rarely done.

When it is a waste of time

When the received answer encodes accumulated experience you do not have. Rederiving why databases use write-ahead logs, why TLS works as it does, or why a particular consensus protocol has the structure it has is not architecture — it is study, and the conclusions are already available.

When the constraint is organisational. No amount of first-principles reasoning changes how many teams there are, what expertise exists, or what the regulator requires.

When it is used to justify ignoring known failure modes. "In principle this should work" against a pattern the industry abandoned for good reasons is usually a rediscovery in progress.

The practical technique

Do the arithmetic before the design review. How many requests, how many bytes, how many round trips, what does one node do, what is the utilisation at target latency. It takes minutes, it is frequently decisive, and it converts an argument about preferences into an argument about numbers.

The most common finding is that the workload is far smaller than the proposed architecture assumes — which is the single most valuable thing first-principles reasoning does for an architect.