We detail YouTube's recommendation system, which leverages deep learning to personalize video suggestions for over a billion users. It addresses the challenges of scale, freshness, and noise inherent in the platform's vast and dynamic content library. The system employs a two-stage approach: candidate generation using deep neural networks to narrow down the massive video corpus, followed by a ranking model that refines the suggestions based on user and video features. Key innovations include using weighted logistic regression to optimize for expected watch time and incorporating the age of training examples to promote fresh content. The paper highlights the practical lessons learned while designing and maintaining such a large-scale recommendation system.