It's kind of like a fluid model, in many ways.

Thing is, I don't think it's desirable that persistent delays exist anywhere.

Now, autotuning codel is probably a good idea... but the implementation should be local. If you're a bottleneck, you may as well behave as if you're the only bottleneck, even if that isn't the case, because if you're in a distributed bottleneck situation your actions will help anyway. A global distributed solver like is proposed here could possibly work in a private network, but I can't see that solution flying in precisely the place you really need it: peering.

One way to do this locally would be to identify the impulse response of the stochastic DE describing the aggregate running through the link using something like https://arxiv.org/abs/1309.7857 and then tune to optimal behaviour for the resulting DE, along the lines of https://arxiv.org/abs/1510.08439.