Week 2: The mechanics of dispatch

Inside the three models that actually run a delivery platform — ETA prediction, order batching, and rating thresholds

A promise made twice#

The number the customer sees is a live optimisation problem, re-solved every few seconds as new information arrives — a kitchen running behind, a rider taking a wrong turn, another order dropped into the same route. Uber’s own engineering writeup on its arrival-time model, “DeepETA: How Uber Predicts Arrival Times Using Deep Learning,” describes exactly this: a prediction retrained continuously against live signal, explicitly not a fixed promise anyone signed off on. The number the rider is actually held to, by contrast, behaves much more like one: fall behind it too often and a rating or an acceptance-rate metric moves against you. The gap between those two numbers — one soft and constantly renegotiated, one hard and counted against you — is where most of this course’s tension lives, and it’s exactly what your dispatch-lab log is built to catch happening in real time.

Batching is a margin decision wearing a routing costume#

When a platform assigns you a second pickup mid-route, it’s presented as efficient routing — one rider, two nearby orders, less total driving. It’s also, inescapably, a decision that the first order’s promised time is now negotiable, made without asking the customer who was quoted that time or the rider now responsible for keeping it. A second, subtler version of the same move shows up in how a platform chooses which restaurant to route a rider through first when two pickups are on the same street but ready at different times: whichever order goes second effectively subsidises the other’s on-time delivery with its own lateness risk, and the rider absorbs whichever choice the batching algorithm made on their behalf. Whether batching is efficient logistics or a quiet cost transfer depends entirely on who absorbs the slippage when the second stop takes longer than the model assumed — and that’s a design choice, not a law of routing.

The threshold that never has to explain itself#

No manager reviews a late delivery and decides to say something. A rating drops, or a lateness counter increments, and the response is automatic and pre-specified — which is exactly what makes it discipline without a disciplinarian. Nobody has to defend the threshold in the moment it fires, and by the time a rider notices the pattern, there’s no single conversation left to have about it. That’s the shape algorithmic management keeps taking all semester: a decision that used to require someone to say it out loud, now happening without anyone having to.

Reading before the session#

A public writeup of a dispatch or ETA system (a platform engineering blog post is fine — Uber’s DeepETA post above is one valid choice) — bring one design decision from it that you think a rider would experience differently than an engineer intended.