Uber·Data Scientist·Onsite - System Design / Architecture·Senior

Senior

Jul 2026

Summary

Uber DS interview focused entirely on a deep system design problem around ETA prediction. The scope was massive and I felt like I was constantly playing catch-up trying to cover all the angles they wanted.

Questions Asked (1)

Design a full end-to-end system to predict pickup and drop-off ETAs at the time a ride is requested. Walk through your data sources, modeling choices, uncertainty handling, training and serving infrastructure, evaluation strategy, and how you'd roll it out.

System DesignA/B Testing & ExperimentationTechnical Trade-offs

Author's notes

This was basically seven questions duct-taped into one.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Structure your answer around the ML system design lifecycle: start with problem framing and data sources, then move through feature engineering, modeling with uncertainty quantification, and finally training/serving infrastructure, evaluation, and rollout. Emphasize trade-offs between accuracy, latency, and scalability, and how you'd handle real-world constraints like sparse data and dynamic conditions.

Pro tip: Show that you understand Uber's specific challenges: real-time traffic, driver supply, and the need for calibrated uncertainty (e.g., prediction intervals) to inform rider communication and pricing. Mention how you'd leverage Uber's existing data infrastructure (e.g., H3, Michelangelo) to avoid reinventing the wheel.

1. Problem Framing & Data Sources

Define the prediction task: estimate pickup ETA (time until driver arrives) and drop-off ETA (time until rider reaches destination) at request time. Identify data sources: historical trip data, real-time traffic, driver locations, road network, weather, events, and rider/driver attributes.

2. Feature Engineering & Modeling

Engineer features like distance, route complexity, time of day, driver experience, and real-time traffic. Choose models: gradient boosting for tabular data, deep learning for spatial-temporal patterns, and quantile regression or Bayesian methods for uncertainty.

3. Training & Serving Infrastructure

Design a pipeline for batch training and online serving with low latency. Use feature stores for consistency, model versioning, and scalable inference (e.g., via Michelangelo). Ensure real-time feature updates and fallback strategies.

4. Evaluation & Uncertainty Handling

Evaluate with metrics like MAE, RMSE, and calibration of prediction intervals. Use offline evaluation with temporal splits and online A/B tests. Quantify uncertainty via quantile regression or ensembles, and communicate it to users.

5. Rollout & Monitoring

Roll out gradually: shadow mode, then A/B test on a small percentage, monitoring business metrics (e.g., cancellation rate, rider satisfaction). Set up alerts for drift and retrain regularly.

Key Points to Mention

Use of real-time and historical data sources, including traffic, weather, and driver supply.
Model choice: gradient boosting vs. deep learning, and handling of spatial-temporal data.
Uncertainty quantification: quantile regression, prediction intervals, and calibration.
Infrastructure: feature store, low-latency serving, and integration with existing ML platforms.
Evaluation: offline metrics, online A/B testing, and business impact metrics.
Rollout strategy: shadow mode, gradual rollout, monitoring, and retraining.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.