Dropbox Dash leverages AI to streamline access to files, work chats, and company content, relying on a real-time machine learning ranking system. At the heart of this system is a feature store that manages data signals essential for predicting file relevance. With a vast number of potential documents, Dash must analyze user behavior across various formats to surface the most pertinent information quickly. The feature store is designed for speed and efficiency, addressing challenges posed by a hybrid infrastructure that includes both on-premises and cloud elements. Building the feature store required a tailored approach, as off-the-shelf solutions fell short due to the need for low-latency communication and extensive parallel processing for user queries. The ranking system processes numerous behavioral and contextual features for each search, demanding sub-100ms latency. The architecture relies on Feast for orchestration and serving, while integrating a custom Go service to optimize performance. This change improved feature retrieval speed, enabling the system to handle thousands of requests per second with latencies consistently in the 25-35ms range. To keep features relevant and fresh, Dropbox implemented a three-part ingestion strategy that balances real-time signals with batch processing. Batch ingestion utilizes intelligent change detection to minimize unnecessary data writes, cutting update times significantly. Streaming ingestion captures immediate user actions, ensuring the system remains responsive to real-time collaboration activities. This approach mitigates the risk of stale data and enhances the overall user experience, keeping Dash efficient and effective in meeting user needs.