Problem
Lyft's pricing engine and driver heatmaps read a precomputed table of travel times between geohash-6 cells, which are grid cells a few city blocks wide. For years large parts of it ran on ETAs from 2018–2019. The old data under-estimated travel times, so drivers looked closer than they were. It also lacked matchable pairs, and the system then acted as if no driver could be there. Each consuming team kept its own denylists and allowlists, and small version updates never justified switching.
Example from the article: the real-time map showed demand over open water in the SFO region because the pipeline never checked whether a cell was drivable.
Solution in brief
Lyft rebuilt the cell set from the road network, prunes it gently, and estimates pair ETAs from trip history with a simulator as fallback. Publish extra fields so consumers can filter the data themselves. Prove the change to the main consumer with a side-by-side test, then refresh on a fixed cycle and split the files by time of day.
Questions this case answers
- How to keep a precomputed pairwise distance table accurate and complete when downstream systems depend on its exact structure and every consumer patches its gaps separately
- How to estimate pairwise values for sparse entries with a simulator when observed history is too thin to trust
- How to refresh a stale travel-time lookup table that a pricing engine optimizes over without every consumer patching it?
- How to decide which map cells belong in a precomputed origin-destination ETA dataset?
Tools and methods
- Neighborhood Reachability Matrix — pairwise ETA in minutes between geohash-6 cell centers.
- Neighborhood Centers — list of valid cells per region.
- Drivable-geohash set — cells derived from road segments; positive signal and final denylist.
- Routing Simulation System (RSS) — traffic-aware simulated ETA for sparse pairs.
- Apache Airflow — scheduled DAG that builds both files per region.
Google or a one-time email code. No password.