Netflix Unveils Slow-Fast Architecture for Notifications

Netflix engineers designed a new notification architecture inspired by Daniel Kahneman's fast/slow thinking framework to manage hundreds of millions of personalized messages. Their previous system used a single causal model to decide both whether to send a message and which one to send, based on short-term engagement predictions — a design that overlooked long-term member fatigue and tightly coupled frequency control with content selection.

The new hierarchical system separates these concerns: a 'Slow' policy sets a member's weekly messaging plan (channel frequency and pacing), while a 'Fast' policy makes real-time tactical decisions about which message to send at each opportunity. The Slow policy optimizes a utility function balancing positive engagement signals against fatigue/opt-out risk, with an added constant message cost to prevent over-sending. The two policies communicate asynchronously via a low-latency feature store.

This decoupling gives engineers a consistent member experience ('stickiness') and lets them independently tune send frequency and message relevance — a pattern echoing plan-vs-act designs seen in robotics and LLM agent systems.