Building Real-time ETA Prediction at Scale

This post covers how we built a real-time ETA prediction system that handles 50,000 requests per day with p99 latency under 100ms.

The Problem

Our ride-hailing platform needed accurate ETA predictions for:

• Driver-passenger matching
• Estimated pickup times
• Dynamic pricing calculations

Initial Approach:

eta = distance / average_speed

But this gave us ~5 minute errors on average.

The Solution

We built a machine learning pipeline using:

1. XGBoost for primary predictions
2. CatBoost for edge cases
3. Feature engineering including:
   • Historical travel times
   • Time of day patterns
   • Weather data
   • Traffic incidents
   • Driver availability zones

Results

Lessons Learned

1. Feature engineering > model complexity
2. Cache aggressively with expiration
3. Fallback to simple heuristics for edge cases
4. Monitor distribution shifts in production

Future Work

• Add real-time traffic incorporation
• Experiment with deep learning for pattern recognition
• A/B test different model versions