Amazon SageMaker Feature Store now supports feature-level writes through a new UpdateRecord API. Applications can change as many as 100 selected values while preserving every omitted feature, replacing the previous read-modify-write cycle required for even a one-field update.
The service applies the partial change atomically, reducing extra reads, latency and the risk that two pipelines updating different fields overwrite each other. It checks event-time ordering and rejects an update when a supplied timestamp is older than the stored value. A full resulting snapshot is still copied to the offline store so training datasets remain consistent with online records.
The feature works immediately with existing In-Memory groups backed by ElastiCache. Standard groups backed by DynamoDB need the new Standard_V2 storage format. Customers can create a new group and bulk-copy records or switch an existing group in place without downtime. The in-place declaration is irreversible, and older untouched records migrate only when a later write updates them, so teams need to weigh rollback needs against migration cost.