Amazon Web Services announced on September 8, 2026, that Amazon SageMaker Feature Store now supports feature-level writes through a new UpdateRecord API. Amazon SageMaker Feature Store is a managed repository for machine learning features, the processed data used to train models and generate predictions. UpdateRecord lets customers update one or more feature values in a single call without reading or rewriting the full record, and the capability is available for both the Standard online store tier backed by Amazon DynamoDB and the In-Memory online store tier backed by Amazon ElastiCache.

Before UpdateRecord, changing one feature in a Feature Group required a read-modify-write cycle with GetRecord and PutRecord. A fraud-scoring pipeline refreshing a customer’s risk_score had to read the full record, merge the new value in application code, and write the entire record back. AWS said that pattern added latency, consumed Read Capacity Units (RCUs), and created race conditions when multiple pipelines updated different features in the same record, including a lost-update problem where one write could silently overwrite another.

UpdateRecord accepts the target FeatureGroupName, RecordIdentifierValueAsString, and a Features list containing one or more feature values, with up to 100 features per call. The target record must already exist because UpdateRecord is not an upsert. Customers can optionally pass TtlDuration to set or override a per-record time-to-live, but EventTime must also be present when TtlDuration is used.

The service validates AWS Identity and Access Management (IAM) permissions, checks EventTime ordering, and applies an atomic merge so omitted features remain unchanged. If a request includes EventTime and that timestamp is newer than or equal to the current EventTime, the update is applied and EventTime advances. If the timestamp is earlier, the full call is rejected with an HTTP 409 ConflictException. If EventTime is omitted, feature changes are applied while the existing EventTime stays unchanged.

AWS also introduced a Standard_V2 storage format for the Standard tier. In-Memory feature groups need no new storage type, and feature-level writes work on all existing In-Memory feature groups out of the box. Standard tier users must opt into Standard_V2 when creating a feature group or migrate existing Standard feature groups to Standard_V2.

AWS described 2 migration strategies for existing Standard feature groups. Strategy A uses the Feature Processor SDK to read records from the old feature group and re-ingest them into a new Standard_V2 feature group; it preserves rollback options but requires a managed migration job, full upfront read and write costs, and client updates to the new feature group name. Strategy B uses UpdateFeatureGroup with OnlineStoreConfig.StorageType = “Standard_V2” for an in-place controlled switchover; AWS said this has zero downtime, charges migration cost only for touched records, keeps the same feature group name and API endpoints, but is irreversible and leaves untouched cold records in the legacy format until written.

AWS listed use cases including streaming feature hydration, backfilling new features, error correction at scale, high-velocity feature updates, and multi-producer feature groups. In one example, a data-quality job corrects customer_segment for 50,000 records by updating one field across thousands of records. In another, an enterprise keeps millions of records and hundreds of features in a single feature group while different teams, batch ETL jobs, streaming analytics, and real-time scoring engines update only the features they own.

UpdateRecord adds fine-grained IAM controls through 2 condition keys: sagemaker:IsUpdateRecord, which distinguishes partial updates from full PutRecord writes, and sagemaker:UpdatableFeatures, which restricts the feature names a principal can update. AWS said existing IAM policies that deny sagemaker:PutRecord automatically block UpdateRecord as well. Updates made through UpdateRecord also flow to the offline store through the same replication pipeline as PutRecord, emitting a complete record snapshot for training datasets and historical analytics.

UpdateRecord follows the same pricing model as PutRecord. For the Standard tier, DynamoDB write capacity unit (WCU) charges are based on item size after the update, while customers can avoid the read capacity previously required by read-modify-write. AWS說d,

Amazon Web Services announced on September 8, 2026, that Amazon SageMaker Feature Store now supports feature-level writes through a new UpdateRecord API. Amazon SageMaker Feature Store is a managed repository for machine learning features, the processed data used to train models and generate predictions. UpdateRecord lets customers update one or more feature values in a single call without reading or rewriting the full record, and the capability is available for both the Standard online store tier backed by Amazon DynamoDB and the In-Memory online store tier backed by Amazon ElastiCache.

Before UpdateRecord, changing one feature in a Feature Group required a read-modify-write cycle with GetRecord and PutRecord. A fraud-scoring pipeline refreshing a customer’s risk_score had to read the full record, merge the new value in application code, and write the entire record back. AWS said that pattern added latency, consumed Read Capacity Units (RCUs), and created race conditions when multiple pipelines updated different features in the same record, including a lost-update problem where one write could silently overwrite another.

UpdateRecord accepts the target FeatureGroupName, RecordIdentifierValueAsString, and a Features list containing one or more feature values, with up to 100 features per call. The target record must already exist because UpdateRecord is not an upsert. Customers can optionally pass TtlDuration to set or override a per-record time-to-live, but EventTime must also be present when TtlDuration is used.

The service validates AWS Identity and Access Management (IAM) permissions, checks EventTime ordering, and applies an atomic merge so omitted features remain unchanged. If a request includes EventTime and that timestamp is newer than or equal to the current EventTime, the update is applied and EventTime advances. If the timestamp is earlier, the full call is rejected with an HTTP 409 ConflictException. If EventTime is omitted, feature changes are applied while the existing EventTime stays unchanged.

AWS also introduced a Standard_V2 storage format for the Standard tier. In-Memory feature groups need no new storage type, and feature-level writes work on all existing In-Memory feature groups out of the box. Standard tier users must opt into Standard_V2 when creating a feature group or migrate existing Standard feature groups to Standard_V2.

AWS described 2 migration strategies for existing Standard feature groups. Strategy A uses the Feature Processor SDK to read records from the old feature group and re-ingest them into a new Standard_V2 feature group; it preserves rollback options but requires a managed migration job, full upfront read and write costs, and client updates to the new feature group name. Strategy B uses UpdateFeatureGroup with OnlineStoreConfig.StorageType = “Standard_V2” for an in-place controlled switchover; AWS said this has zero downtime, charges migration cost only for touched records, keeps the same feature group name and API endpoints, but is irreversible and leaves untouched cold records in the legacy format until written.

AWS listed use cases including streaming feature hydration, backfilling new features, error correction at scale, high-velocity feature updates, and multi-producer feature groups. In one example, a data-quality job corrects customer_segment for 50,000 records by updating one field across thousands of records. In another, an enterprise keeps millions of records and hundreds of features in a single feature group while different teams, batch ETL jobs, streaming analytics, and real-time scoring engines update only the features they own.

UpdateRecord adds fine-grained IAM controls through 2 condition keys: sagemaker:IsUpdateRecord, which distinguishes partial updates from full PutRecord writes, and sagemaker:UpdatableFeatures, which restricts the feature names a principal can update. AWS said existing IAM policies that deny sagemaker:PutRecord automatically block UpdateRecord as well. Updates made through UpdateRecord also flow to the offline store through the same replication pipeline as PutRecord, emitting a complete record snapshot for training datasets and historical analytics.

UpdateRecord follows the same pricing model as PutRecord. For the Standard tier, DynamoDB write capacity unit (WCU) charges are based on item size after the update, while customers can avoid the read capacity previously required by read-modify-write. AWS說d,