Netflix is known for their recommendation system so I knew they'd go deep on this.

The homepage has multiple rows, each representing a different recommendation strategy. I broke it down into: candidate generation (collaborative filtering + content-based), ranking (deep learning model), and row assembly.

Spent significant time on the ML architecture. Discussed using a two-tower model for candidate generation, with user features (watch history, ratings, demographics) and item features (genre, cast, metadata) encoded separately.

The interesting part was discussing real-time vs batch processing. New releases need to surface quickly (real-time), but historical preferences can be computed in batch. Proposed a lambda architecture with a batch layer (Spark) and a speed layer (Flink).

Follow up was about A/B testing recommendations. Discussed interleaving experiments where you mix results from control and treatment models in a single page view to reduce variance.

The interviewer asked a great question about filter bubbles and how to ensure diversity in recommendations. Discussed exploration-exploitation tradeoffs and using Thompson sampling for diversification.