Online recommendation systems generate recommendations for users based on real-time contextual information, such as user behaviors and item catalog data. These systems consist of several components, including candidate generation, feature retrieval, filtering, model inference, pointwise scoring and ranking, and listwise ranking. Candidate generation narrows down a vast number of possible candidates to a small set that can be ranked, while feature retrieval fetches data for the user and candidates being considered. Filtering removes candidates based on fetched data or model predictions, and model inference sends feature vectors to a model service to get predictions. Pointwise scoring and ranking scores items in isolation, while listwise ranking orders items in context with other items in the list. The goal of these components is to provide diverse and high-quality recommendations that meet user preferences and objectives.