Uber doesn't work with whole roads. It breaks the global road network into around 100 million small segments, and it measures the traffic itself. Every driver's app sends location data every four seconds, so every time a driver crosses a segment, Uber learns exactly how long that segment took right now.
But live measurements only tell you about right now. So Uber built DeepETT, a deep learning system that predicts the crossing time of every segment for the next three hours. If an accident just slowed a road down, it trusts the live signal over the old pattern, and it fills in quiet side streets by reading the busy roads around them. It answers around two million forecasts every second, which makes it the busiest model inside Uber.
The routing engine adds up the predicted segment times along the fastest path, and a second set of models trained on millions of real trips corrects the remaining error before the number hits your screen.
When Uber shipped this, long-trip accuracy improved by 6%, worth an estimated $100 million a year in gross bookings.
#SystemDesign #MachineLearning #DeepLearning #SoftwareArchitecture #BackendDevelopment