Built with Rust.
Measured in milliseconds.

Every driver location ping writes ten H3 cell IDs at once, resolutions 0 through 9. Geography is never computed at request time — the reverse index is already sitting in Redis at whatever scale the search turns out to need.

Reach is derived from elapsed time, not a retry counter, so a re-entrant search can never un-widen. Opening at the 10 km ceiling would offer the very first round to the whole city while the passenger is still reading “Finding drivers near you”.

The engine takes the finest resolution whose ring of cells still fits a 61-read budget. The hexagons overhang the search badly, which is the trade: a fixed read cost, then every candidate re-measured against the true radius.

Online but not busy. No trip in progress. Pickup inside the driver's own radius setting. Not already holding an offer, not previously declined, and on the same test lane as the passenger.

Four weighted factors the operator sets: proximity on an exponential curve, rating, GPS freshness, minus a cancellation penalty. Here the nearest driver — 310 m out, 3.9 stars, one ride in three cancelled — loses to one twice as far away.

The top-ranked driver gets a 30-second window, clamped so an offer can never outlive the booking it belongs to. The attempt is written to the ledger before the push notification leaves — the ledger is what the next round reads.

When an offer lapses the search resumes from the ledger, never re-offering the same driver, and reaches further. At 30 seconds in there is no room for another full window, so the last round drops the distance limit — answering with continent-sized cells for the same 61 reads.