Design ETA prediction for Uber rides

Quick Overview

This question evaluates competency in end-to-end machine learning system design for real-time spatiotemporal ETA prediction, encompassing feature engineering from heterogeneous data sources, uncertainty quantification, model selection, and production serving and monitoring.

Design ETA prediction for Uber rides

Company: Uber

Role: Data Scientist

Category: Machine Learning

Difficulty: hard

Interview Round: Technical Screen

Design an end‑to‑end system to predict pickup and drop‑off ETAs at request time. Specify: (1) data sources/features (road graph and segment metadata, GPS traces, historical segment speeds, live traffic incidents/speeds, weather, events, supply‑demand density, driver behavioral features, device/network latency, road closures); (2) modeling approaches comparing a map‑matching + dynamic shortest‑path baseline to gradient‑boosted trees and a spatiotemporal deep model (sequence/graph), and how you’ll represent routes and capture heteroskedasticity; (3) uncertainty outputs and calibration (e.g., quantile regression, conformal prediction) with 50/90% intervals; (4) training/serving architecture (feature store, streaming updates, drift detection, online refresh cadence, canary models, latency/availability SLAs); (5) leakage/censoring controls (e.g., excluding post‑dispatch signals, handling reroutes, cancellations) and strategies for cold‑start regions and rare events; (6) offline evaluation metrics (MAE/MAPE, P50/P90 error, coverage of 90% intervals, CRPS) with stratified slices by city, time‑of‑day, weather, and surge; and (7) online evaluation via A/B test with guardrails (cancellation rate, pickup wait, driver idle time, reliability of intervals) and rollout criteria. Conclude with trade‑offs and select an MVP model for initial deployment.

Overview: This question evaluates competency in end-to-end machine learning system design for real-time spatiotemporal ETA prediction, encompassing feature engineering from heterogeneous data sources, uncertainty quantification, model selection, and production serving and monitoring.

Uber logo
Uber
Oct 13, 2025
hardData ScientistTechnical ScreenMachine Learning
16
0

System Design: Real‑Time Pickup and Drop‑off ETA Prediction

Context: You’re designing an end‑to‑end system that predicts pickup and drop‑off ETAs at trip request time for a large ride‑hailing marketplace. At request time, the driver is not yet assigned; predictions must serve within tight latency and remain reliable during traffic spikes, incidents, and rare events.

Specify the following:

  1. Data Sources and Features
    • Road graph and segment metadata
    • GPS traces and map‑matched trips
    • Historical segment speeds and travel times (time‑of‑day/weekday seasonality)
    • Live traffic incidents and real‑time speeds
    • Weather and major events
    • Supply‑demand density and marketplace state
    • Driver behavioral features (e.g., acceptance, cruising patterns)
    • Device/network latency and dispatch latency
    • Road closures, construction, and routing constraints
  2. Modeling Approaches
    • Baseline: map‑matching + dynamic shortest‑path
    • Gradient‑boosted trees
    • Spatiotemporal deep model (sequence/graph)
    • How you represent routes for each model
    • How you capture heteroskedasticity (variance changing with context)
  3. Uncertainty and Calibration
    • Produce 50% and 90% prediction intervals
    • Methods: quantile regression, conformal prediction
    • Calibration strategy
  4. Training and Serving Architecture
    • Feature store (offline/online parity)
    • Streaming updates for traffic and incidents
    • Drift detection and monitoring
    • Online refresh cadence and deployment (canary, blue/green)
    • Latency and availability SLAs
  5. Leakage and Censoring Controls
    • Exclude post‑dispatch signals
    • Handling reroutes and cancellations
    • Cold‑start regions and rare events strategies
  6. Offline Evaluation
    • Metrics: MAE/MAPE, P50/P90 error, coverage of 90% intervals, CRPS
    • Stratified slices by city, time‑of‑day, weather, and surge
    • Proper temporal/geographical holdout to avoid leakage
  7. Online Evaluation and Rollout
    • A/B test with guardrails: cancellation rate, pickup wait, driver idle time, reliability of intervals
    • Rollout criteria and monitoring plan

Conclude with trade‑offs and select an MVP model for initial deployment.