Amazon SageMaker Feature Store released the UpdateRecord API, allowing users to update individual feature values without reading or rewriting entire records. This new capability is available for both Standard and In-Memory tiers of the online store. The update eliminates the need for a read-modify-write cycle, reducing latency and improving efficiency.
Previously, updating even a single feature required a full read-modify-write process, which introduced latency, consumed unnecessary read capacity, and created race conditions. The new API allows users to specify only the features they want to change, with the service applying the updates atomically. This approach prevents data overwrites and ensures consistency across multiple pipelines.
The UpdateRecord API also supports temporal ordering by validating EventTime, ensuring that updates are only applied if the new timestamp is later than the existing one. This helps maintain data integrity and prevents lost-update issues. The API data flow includes writing only the changed features to the online store while replicating a full record snapshot to the offline store.
"You provide only the features that you want to change, and Amazon SageMaker Feature Store applies the updates atomically to the existing record," said the source. This atomic merge ensures that updates are consistent and reliable, even when multiple pipelines modify the same record.
The introduction of UpdateRecord follows Amazon's ongoing efforts to enhance the SageMaker Feature Store for better scalability and performance. The company did not specify the exact cost savings for customers, and the feature-level writes require the new Standard_V2 storage format for the Standard tier. Migration strategies are available for existing feature groups.
Source: awsml
