Amazon SageMaker Feature Store has introduced two new APIs to address operational challenges in feature pipeline management. The BatchWriteRecord API allows up to 25 records to be written across multiple feature groups in a single API call, significantly improving throughput for high-volume data ingestion. This is particularly beneficial for pipelines that must sustain high API call rates, such as a fraud-detection system processing 10,000 records per second across five feature groups. The ListRecords API enables teams to enumerate record identifiers within a feature group, offering support for both Standard and In-Memory storage tiers. These updates aim to streamline data management and reduce overhead for machine learning workflows.
The BatchWriteRecord API tackles the limitations of single-record ingestion by allowing multiple records to be processed in one call. This approach reduces connection overhead and tail latency, which previously limited throughput when using the PutRecord API. The API preserves EventTime-based ordering guarantees and provides partial-success semantics, meaning individual record failures do not halt the entire request. It also supports time-to-live (TTL) control for records, ensuring they are automatically removed after a specified period. The ListRecords API complements these improvements by enabling teams to query and manage records in the online store, even when using the In-Memory storage tier.
Amazon SageMaker Feature Store is a fully managed repository for storing, sharing, and managing features for machine learning models. It offers low-latency online serving for real-time inference and an offline store for historical data retention. The new APIs aim to enhance the platform's efficiency by addressing common operational gaps, such as the need for high-throughput data ingestion and the inability to browse records in the In-Memory tier. These updates are part of Amazon's ongoing efforts to improve the scalability and usability of its machine learning tools.
Source: awsml
