You already know: graphs and shortest paths (D2), regression and quantile losses (D3), streaming aggregation (D16) and the surge case's lesson that the marketplace depends on fast, trustworthy signals. This case adds ETA prediction at marketplace scale: an ETA is used both by dispatch (to pick the best driver, hundreds of thousands of times a second) and by riders (as a promise they plan around). The design questions: which ETA, how to keep road speeds fresh, how to combine a routing engine with ML, how to serve ETAs at dispatch scale and how to make the promise honest.
How to read this lesson. No code, on purpose. Numbers marked assumed are assumptions you state out loud.
1. The prompt, and the questions that change the design
A mid-level answer: "train a model on past trips to predict duration". A staff candidate first asks which ETA (driver to pickup, pickup to destination, or the whole delivery including food preparation), who consumes it (dispatch ranking or a rider-facing promise), what accuracy means (is late worse than early?) and at what scale. Those answers produce a decomposed ETA, a routing engine with an ML correction, separate fast and accurate tiers, and a quantile promise.
Assumptions you state: