SYSTEM Cited by 1 source
Netflix notification platform¶
Overview¶
Netflix's notification platform sends hundreds of millions of personalized notifications daily across push, email, and in-app channels to help members discover content. As of 2026, it uses a hierarchical Slow/Fast architecture inspired by Kahneman's dual-process theory.
Architecture¶
Slow Policy (Strategic Layer)¶
- Cadence: Runs at weekly granularity per member.
- Input: Long-term engagement patterns, channel preferences, fatigue signals.
- Output: Personalized "Pacing Plan Action" — one of ~O(100) discrete cross-channel frequency strategies.
- Optimization: Maximizes a multi-objective utility function: engagement reward minus message cost + fatigue penalty.
- Storage: Writes plan to feature store.
Fast Policy (Tactical Layer)¶
- Cadence: Executes on every notification send opportunity (real-time).
- Input: Pacing plan (from feature store) + candidate messages + member context.
- Output: Which specific message to send (or skip this opportunity per pacing plan).
- Optimization: Maximizes immediate relevance and engagement within Slow Policy constraints.
Previous System¶
Single causal model predicting single-message incrementality (short-horizon). Frequency controlled implicitly via a global relevance threshold calibrated to target aggregate send rates. This coupled ranking with pacing and limited personalization.
Key Results¶
- Largest production metric lift to date after migrating to hierarchical architecture.
- Strongest gains among infrequent viewers ("casual viewers").
- Enables independent A/B testing of pacing strategies vs. content ranking.