The document discusses recommendations for a practical recommender system that can scale to large datasets. It proposes a two-tier architecture with a real-time serving layer and batch computation layer. It recommends using matrix factorization with alternating least squares (ALS) for the batch layer, as ALS can be parallelized across large datasets. It describes how to implement ALS in a way that allows for real-time updates and "fold-in" of new user data. The system described has been implemented in Myrrix, an open-source recommender engine.