How Lyft Uses ML to Make 100 Million Predictions A Day

Lyft processes 100 million ML predictions daily through their LyftLearn Serving platform, which addresses both data plane performance and control plane complexity. The system uses isolated microservices where each team owns their repository, deployment pipeline, and runtime environment. Key components include an HTTP serving layer with Flask/Gunicorn, a core serving library handling model lifecycle, custom ML code injection points, and integration with Kubernetes/Envoy infrastructure. The platform features automated config generation, built-in model self-testing, and supports any Python-compatible ML framework while maintaining strict isolation between teams.