AWS adds feature-level writes to SageMaker Feature Store

AWS announced feature-level writes for Amazon SageMaker Feature Store through a new UpdateRecord API. The service is a fully managed repository for storing, sharing and managing machine-learning features, which AWS describes as processed data used to train models and generate predictions.

With UpdateRecord, an application can update one or more feature values in a single call without reading or rewriting the entire record. AWS says Feature Store applies the requested changes atomically to the existing record, while features omitted from the request remain unchanged.

Replacing full-record updates

Before this API, changing even one feature required a read-modify-write process using GetRecord and PutRecord. An application first read the complete record, merged the new value in application code and then wrote the entire record back.

AWS gives a fraud-scoring pipeline as an example. To refresh a customer’s risk_score, the pipeline had to retrieve all features in the record before writing the updated record.

According to AWS, that process added latency to each update and consumed additional read capacity. The company also says concurrent pipelines updating different features in the same record could create race conditions. In the worst case, one pipeline’s write could silently overwrite another’s, a situation AWS identifies as a lost-update problem.

The extra GetRecord requests also generate Read Capacity Units charges, AWS says. For customers operating wide feature groups with frequent updates, the company says those costs can accumulate.

How the new API works

A client application sends an UpdateRecord request containing only the changed features. AWS says the Feature Store service validates AWS Identity and Access Management permissions, checks EventTime ordering to reject stale writes and performs an atomic merge.

The partial write is applied to the online store. AWS says a full record snapshot is automatically replicated to the offline store so that training datasets remain accurate. The announcement describes multiple pipelines, including clickstream, purchases and scoring, independently writing their own features to the same record without coordination.

UpdateRecord is available for both Amazon SageMaker Feature Store online store tiers: the Standard tier backed by Amazon DynamoDB and the In-Memory tier backed by Amazon ElastiCache.

Source: AWS, “Amazon SageMaker Feature Store introduces UpdateRecord for feature-level writes,” published September 8, 2026.