A music streaming company needs a feature pipeline that stores features for offline training and for online inference, preserves feature history, and gives data scientists searchable access. Which option meets these needs with the least operational overhead?

Amazon SageMaker Feature Store is designed specifically for managing, storing, and serving machine learning features for both training and inference, directly addressing all requirements with minimal operational overhead. It automatically handles the creation of online and offline stores, preserves feature history, and provides searchable access to feature groups. The IAM role ensures secure and controlled access for data scientists. Using a single online store for both training and inference (second option) is inefficient for large-scale training, as online stores are optimized for low-latency, small-batch reads, not bulk data access. Manually managing S3 buckets (third option) requires significant custom development for versioning, search, and integrating with SageMaker training, increasing operational overhead. Using DynamoDB tables (fourth option) also involves substantial custom development for historical data management, data transfer to S3 for training, and search capabilities, leading to higher operational complexity compared to SageMaker Feature Store.