When you request a ride on Lyft, dozens of machine learning models activate instantly. Price calculations, driver incentives, fraud detection, and ETA predictions all happen in milliseconds, repeated millions of times daily. Handling this scale required a dedicated solution.

Lyft built LyftLearn Serving, a platform designed to manage complex machine learning workloads across two key areas: the data plane (runtime performance) and the control plane (model updates and deployments). Early on, Lyft faced challenges with a monolithic system that caused team conflicts and deployment delays. The fix was a shift to microservices.

Each team, like Pricing or Fraud Detection, gets its own GitHub repository and independent service. Teams control their TensorFlow or PyTorch versions without affecting others. This isolated model improves ownership, speeds up releases, and reduces system-wide outages.

At runtime, Lyft uses Flask and Gunicorn behind Envoy load balancers to manage traffic. These components connect to a shared core library containing Lyft’s custom logic. By breaking up its architecture, Lyft streamlined machine learning operations while scaling for growth.