Debugging ML Models: Build a Feature Platform That Actually Works

This talk explores how to build an effective feature platform to solve machine learning model debugging challenges. Ivan Burmistrov shares practical insights from developing a feature platform for Moj, a short-video app with over 100 million monthly active users. Discover the core concepts of feature platforms, their high-level architecture, and strategies for overcoming critical challenges including real-time stream processing upgrades, performance degradation, database contention, and data model optimization for scale. The presentation provides actionable knowledge for ML engineers and data scientists looking to improve their debugging workflows and build more reliable machine learning systems.