Lyft's Feature Store is a key component of its Data Platform, focusing on managing and deploying Machine Learning features at scale. It centralizes feature engineering, ensuring consistency across various models and workflows that drive data-driven decisions in ridesharing. The system streamlines the entire lifecycle of features, from creation and storage to real-time access, which is crucial for effective model training and inference. Over the past few years, Lyft has made significant architectural improvements to enhance efficiency, scalability, and user experience, particularly as AI and Large Language Model applications become more prevalent. The Feature Store supports batch feature ingestion, which relies on Hive data tables. Engineers define features using Spark SQL and JSON configurations, leading to automatically generated Airflow DAGs that are production-ready. These DAGs handle data quality checks and store results in both offline and online paths. The offline path is used for historical analysis and model training, while the online path ensures low-latency access for real-time inference. For real-time feature serving, Lyft utilizes a system called `dsfeatures`, built on top of AWS data stores like DynamoDB. This setup includes a performance cache to speed up data retrieval and an OpenSearch integration for specific embedding features. The configuration files contain crucial metadata, allowing for effective monitoring and versioning of features. Versioning tracks changes over time, ensuring that developers use the correct feature versions, while lineage tracking provides insights into the origin and transformation of features. Discoverability of features is facilitated by integrating with Amundsen, Lyft's data discovery platform. This integration helps users search for existing features, reducing duplicated efforts and optimizing engineering time. By enhancing the way features are managed and accessed, the Feature Store plays a vital role in the machine learning development lifecycle at Lyft, strengthening collaboration with the Machine Learning Platform team.