Why Your Uber Driver Is Across the Street
The pin looks right. The pickup goes wrong. How GPS uncertainty, stale updates, and road geometry turn a simple ride into a system design problem.
· 3 min read · Tanmay Singh
Your phone says “Driver arriving.” You can see the car. The driver can see you.
There is just one problem: a road divider separates you. Now your “one minute away” ride needs another lap around the block.
How does an app that coordinates entire cities get the last few metres wrong?
Because knowing where someone is and knowing where to meet them are different problems.
This is a proposed design inspired by a real product problem. Uber's documented techniques are linked below; the architecture and example numbers are illustrative.
1. The blue dot is an estimate
Between tall buildings, satellite signals can be blocked or reflected. A phone may place you on the opposite pavement even when you have not moved. Uber's engineering explanation of urban GPS describes errors of 50 metres or more and an approach using 3D building maps to improve estimates.
So a location update should carry more than coordinates:
latitude + longitude + accuracy + captured_at + sequence
Accuracy describes uncertainty. Capture time tells us how old the observation is. A sequence number helps reject updates that arrive out of order within a device session.
Product decision: when the estimate spans both sides of a street, ask the rider to confirm an entrance. More decimal places will not resolve the ambiguity.
2. Across the street can mean five minutes away
A distance calculation sees two nearby points. A routing engine sees one-way roads, turn restrictions, dividers, and legal stopping places.
Map matching estimates which road segment a driver occupies using recent positions, heading, and plausible movement. Simply snapping to the nearest road can choose the wrong carriageway. Riders need pedestrian paths and entrances, too: snapping a person onto a driving lane solves the wrong problem.
Consider two hypothetical drivers:
Use a spatial index to shortlist nearby drivers, then compare road travel time to a valid pickup point. Straight-line distance is a useful filter, not the final decision.
3. A perfect coordinate can still be wrong now
Imagine a car travelling at 10 metres per second. A location captured eight seconds ago may be 80 metres behind its current position, even with perfect GPS.
Mobile networks delay, duplicate, and reorder messages. In our design, each driver session gets an ordered stream; older sequence numbers cannot overwrite newer state. Capture time and server receipt time help assess freshness, with clock skew accounted for. Stale drivers expire from the matching index.
Animating a car smoothly does not make its position current. During a connection gap, briefly predict movement if confidence allows, then show “Location updating” rather than endlessly driving an imaginary car along the map.
4. Design a meeting point, not just a pin
Here is a compact architecture for the problem:
Partition location processing by driver ID to preserve order; use geographic cells to narrow searches, including neighbouring cells at boundaries.
For each candidate meeting point, compare driver arrival time, rider walking time, and uncertainty. If both travel simultaneously, the earliest meeting is approximately the larger of their two travel times. First exclude prohibited stopping spots and inaccessible walking routes; then rank feasible options.
Uber already documents designated pickup areas and rider-selected meeting locations. In our design, confirmation creates a shared, versioned pickup record. If the rider changes entrances, both apps receive the new version and acknowledge it. A drifting GPS estimate should not silently move an agreed meeting point.
The metric that matters: did you actually meet?
A fast location API can still deliver a frustrating pickup. Track time from driver arrival to trip start, pickup-related cancellations, and how often riders must change the meeting point. Compare these by venue and accessibility needs so a better average does not hide worse experiences.
The next time your driver is across the street, remember the missing pieces: a trustworthy position, a fresh update, a legal route, and an agreed place to stop.
What would you choose: a two-minute walk to an easier pickup, or a five-minute wait at your door? What should the app ask before deciding? Share your take below.
Want the bigger picture? Read Designing Uber, end to end.