A recommendation model is trained from nightly batch aggregates but scored using live event streams. You need the same feature definitions and values at training and serving time, with correct point-in-time history for training. What design is right?
Show answer and explanation
Answer: C
A feature store lets you define a feature once, materialize its history into an offline store for training with point-in-time lookups, and sync the latest values to an online store for low-latency serving. Because both paths share one definition, training and serving see consistent values. Duplicated feature groups, notebook recomputation and hand-copied CSV files each reintroduce skew.
Why the other options are wrong
- A. Separate groups for the batch and online paths duplicate the logic. The two copies will drift, which is the train-serve skew the design should prevent.
- B. Recomputing from raw logs in a notebook is manual and inconsistent with training. It is not a reliable way to serve features in production.
- D. Hand-loaded CSV files are stale and error-prone. They provide no online serving path and no point-in-time history for training.