Thinking Fast & Slow for a Personalized Notification System

This post details the evolution of Netflix's personalized notification system by introducing a hierarchical 'Slow-Fast' architecture. The 'Slow' policy defines a personalized weekly messaging plan (frequency and pacing) by maximizing a utility function that balances engagement rewards with explicit message costs to manage long-term fatigue and opt-out risk. The 'Fast' policy then uses this plan from a feature store to make tactical, real-time send decisions. This decoupling allows for independent optimization of pacing strategies and message selection, leading to significant metric improvements.