I modeled this as a fleet scheduling problem using Queueing Theory and Dynamic Programming over a finite horizon. The standard practice of throttling once the number of jobs in server exceeds certain threshold is in fact suboptimal. The optimal policy consists in segmenting the part of the traffic that is retry sensitive, from those that are not. For example, an user doing a basic data parsing might still do well under a weaker model, but a power user will certainly feel the degradation and ask more.
Demo: Just a toy instance to illustrate the issue. The user can create their own policies and see how they perform against the industry standard and the optimal one. It is roughly 100 lines of Flask + JS frontend.
Paper with proofs: https://arxiv.org/abs/2608.23986
For those of you who have worked in inference infra, does this match anything you have seen?