Amazon SageMaker Feature Store Adds Feature-Level Write Support with UpdateRecord API

Amazon SageMaker Feature Store has introduced an UpdateRecord API that enables feature-level writes, simplifying the process of updating individual features within a record. Previously, updating a single feature required reading the full record, merging the change, and rewriting the entire record, which added latency, increased costs, and risked data inconsistencies.

The new API allows users to send only the features they wish to update, and the service atomically merges these changes into the existing record. This process eliminates the need for full record reads and writes, reducing latency and operational costs, especially at scale.

The UpdateRecord API supports both the Standard (DynamoDB-backed) and In-Memory (ElastiCache-backed) storage tiers. For the Standard tier, a new storage format called Standard_V2 is available, which supports feature-level writes. Existing feature groups can be migrated to this format with minimal disruption, either through a bulk migration or an in-place switch.

This capability is particularly useful for real-time data pipelines, backfilling features, correcting data discrepancies, and managing high-velocity feature updates across multiple data producers. It also integrates with IAM for fine-grained access control, ensuring secure and controlled feature updates.

The feature supports EventTime-based ordering, preventing stale updates from overwriting fresher data, and ensures updates are reflected in both online and offline stores, maintaining data consistency for training and analytics.

Available today in all AWS regions where SageMaker Feature Store operates, the UpdateRecord API enhances the efficiency and reliability of feature engineering workflows, enabling more scalable and accurate machine learning applications.