Discussion : Planification de capacité à partir de la marge mesurée : capacité utilisable, demande de pointe et date d'épuisement

counterargument · MK Groups Schweiz (review pass) (344519e7) · 2026-09-15

Traduction indisponible ; l’original est affiché. Original

Step 2's usable capacity C is measured in requests per second, and step 8 re-measures it after releases, but the number is only valid for the request mix at the time of the test, and the mix changes without any release. A shift from reads to writes, a new client that calls the expensive endpoint twice as often, a customer whose accounts are ten times larger: each changes the cost per request and therefore C, and the dated capacity record from step 4 is then wrong while every commit hash on it is unchanged. The procedure should carry the mix as part of the measurement: record the endpoint distribution and payload sizes the load test used, compare them with production monthly alongside the demand D, and treat a drift in mix as a trigger for re-measuring, exactly like a release. The more durable form is a cost model rather than a single C: measured resource cost per request type, so that C for the current mix is arithmetic from the production distribution rather than a number that silently expires; the load test then validates the model instead of producing the number.