SYSTEM Cited by 1 source

Amazon SageMaker Feature Store

Amazon SageMaker Feature Store is a managed repository inside SageMaker AI for storing, updating, and retrieving machine-learning feature vectors. It offers two storage modes with different SLA shapes:

• Online store — low-latency, low-throughput lookup of only the latest valid feature vector per record; used for interactive inference and rapid per-request feature retrieval.
• Offline store — backed by Amazon S3 in append mode; optimised for cost-efficient high-throughput data IO; latency "in the order of minutes"; used for batch pipelines, archiving, debugging, long-term retention.

Stub page — expand as deeper SageMaker Feature Store internals are ingested from dedicated AWS posts.

Seen in

• sources/2025-06-29-zalando-building-a-dynamic-inventory-optimisation-system-a-deep-dive — canonical first wiki disclosure of SageMaker Feature Store used with both online and offline stores in one production ML system. Zalando ZEOS's replenishment recommender stores per-SKU feature vectors in both modes; the online store guarantees 10–20 ms latency per SKU for both read and write, enabling the partner-portal-driven interactive re-scoring path. The offline store is S3-backed, append-mode, storing daily datapoints plus user-triggered feature-vector updates — the combination ensures that the online and offline optimisation subsystems always agree. First wiki instance of the explicit dual-mode feature-store architecture — see concepts/online-vs-offline-feature-store and patterns/online-plus-offline-feature-store-parity.

Related

• systems/aws-sagemaker-ai — parent SageMaker product suite.
• systems/aws-s3 — storage substrate for the offline store.
• concepts/online-vs-offline-feature-store — the concept page codifying the dual-mode distinction.
• patterns/online-plus-offline-feature-store-parity — pattern for keeping both modes in sync.