A single binary against a runtime, costed

A recurring proposal to move a service to a compiled language, costed properly rather than argued about.

the service: a webhook receiver. 400 requests a
minute, 12ms of work, no database.

  PHP-FPM today       220 MB across 20 workers,
                      p95 18ms, deployed by the same
                      pipeline as everything else
  a compiled binary   ~15 MB, p95 estimated 3ms, a
                      second build toolchain, a second
                      deployment shape, and one person
                      who can change it

the saving: 15ms at the p95 on an endpoint whose
caller has a 30-second timeout.

The performance difference is real and is irrelevant to the caller, which is the measurement that ends the discussion — a fifteen-millisecond improvement on a path with a thirty-second budget buys nothing. The proposal recurs annually and the decision record now contains the numbers, which is the only thing that makes the fourth conversation shorter than the first.