Posted On: Jun 30, 2023
Amazon SageMaker Feature Store now supports the ability to set a time to live (TTL) for records in the online store. The online store contains the most recent features and is used for low latency model serving. With this launch, you can now set a TTL duration after which a record will be hard deleted, so you have more flexibility when it comes to managing online store data.
TTL can be set at the feature group level which will be applied to all records in that group, or at the individual record level. When a record is permanently deleted from the online store upon TLL expiry, Feature Store will also write a tombstone record to the offline store. The offline store contains historical features and is used for training and batch inference. Tombstone records can be identified and removed from the offline store in order to provide more accurate or more recent data for model training, or to reduce storage costs.