Question 12
Domain 1: Databricks Machine LearningA fraud model must look up customer features within milliseconds during each API request. Which feature-serving approach best fits?
Correct answer: B
Explanation
An online feature store is built for low-latency retrieval, so it can return customer features "within milliseconds" during each API request. Offline stores are for training and batch access, while online feature serving supports real-time inference by exposing features through fast key-value lookups.
Why each option is right or wrong
A. Offline feature table lookups during nightly batch scoring
Nightly batch scoring supports offline processing, not per-request millisecond feature retrieval.
B. Online feature serving backed by an online feature store
The requirement is sub-second, per-request retrieval during inference, which aligns with an online feature store exposed through low-latency key-value access rather than batch-oriented storage. In practice, online stores are designed to serve features in milliseconds for API calls, whereas offline feature stores are used for training and historical backfills and cannot meet real-time latency constraints.
C. Downloading the feature table to a local laptop before each request
Local laptop downloads are manual and impractical for scalable, production API inference.
D. Using only notebook display outputs
Notebook outputs are for exploration and visualization, not operational feature serving.