AWS adds feature-level UpdateRecord writes to SageMaker Feature Store

AWS introduced the UpdateRecord API for Amazon SageMaker Feature Store on Tuesday, giving applications a way to atomically change selected values in an existing online-store record without reading and rewriting the entire record. AWS said the operation is available in every region where Feature Store is offered.

The API removes the GetRecord-plus-PutRecord sequence previously required to change part of a shared record. AWS says eliminating that read can reduce read-capacity costs and the risk of concurrent producers overwriting one another’s changes. AWS’s launch post does not include independent benchmarks or customer measurements for those claims. DataPhoenix recently covered another SageMaker operations release, AWS’s HyperPod InstantStart control plane.

According to the UpdateRecord API reference, one request can change between one and 100 features; omitted features remain unchanged. The operation is not an upsert. The record must already exist in the online store, and a missing or soft-deleted record returns a ResourceNotFound error.

UpdateRecord supports only Standard_V2 and InMemory online stores. Standard-tier feature groups must use Standard_V2, while existing InMemory groups need no change in storage type. AWS’s online-store documentation says Standard groups can migrate in place through UpdateFeatureGroup, without being recreated or having their data reingested. The migration is one-way, and UpdateRecord can take several minutes to become available after it finishes.

EventTime ordering blocks stale writes. A timestamp older than the record’s stored EventTime produces an HTTP 409 conflict; an equal or newer value is accepted. If EventTime is omitted, the existing value remains in place. Adding a time-to-live duration requires EventTime in the same request.

AWS also added IAM conditions for more granular write permissions. Its Feature Store security documentation says UpdateRecord is authorized through sagemaker:PutRecord rather than a separate action, so existing PutRecord allow and deny policies also apply. The sagemaker:IsUpdateRecord condition distinguishes partial updates from full writes, and sagemaker:UpdatableFeatures can limit which feature names a principal may change.

If a feature group has an offline store, omitting TargetStores applies the update to all associated stores. AWS says a complete record snapshot is replicated to the offline store. The company also says UpdateRecord follows PutRecord pricing, with Standard-tier write-capacity charges based on the item’s size after the update.