Reverse engineering Netflix’s sub-100ms thumbnail engine : Computer vision + Contextual Bandits

How a system extracts thousands of frames, learns your preferences, and serves the perfect thumbnail in milliseconds.

1. Beyond Global A/B Testing: Netflix abandons traditional static A/B testing—which identifies a single best thumbnail for all users—in favor of Contextual Multi-Armed Bandits.

2. Decoupled Two-Tier Architecture (AVA vs. Real-Time Inference): Heavy computer vision tasks are shifted entirely offline to the AVA (Automated Visual Analysis) pipeline, which extracts, filters, and rates thousands of frames to curate a candidate pool of ~10 pristine images per title.

3. Risk-Free Validation via Counterfactual Offline Replay: To evaluate new bandit algorithms without risking live user engagement or app performance, Netflix uses Offline Replay. .

4. Edge Caching & Low-Latency Feature Stores: Because dynamic personalization breaks conventional edge caching (since every user receives a unique combination of thumbnails), image binary assets are cached globally by static IDs.