CertKeen

Google Cloud Professional Cloud Architect · Free practice question 6 of 12

Feature Store against training-serving skew

A fraud model at Wyndham Card performs well offline but poorly in production because online feature values are computed differently from the training pipeline. The team wants a single managed source of features for both training and low-latency online serving. What should the architect recommend?

  1. A.Feature Store on Gemini Enterprise Agent Platform (formerly Vertex AI Feature Store), with features defined over BigQuery and served online from the same definitions
  2. B.A Memorystore cache populated by the serving application only
  3. C.Separate feature logic in the training notebook and in the serving application, reviewed together each quarter
  4. D.A larger model that is less sensitive to feature differences
Show answer and explanation

Correct answer: A. Feature Store on Gemini Enterprise Agent Platform (formerly Vertex AI Feature Store), with features defined over BigQuery and served online from the same definitions

Why: Feature Store manages feature definitions over BigQuery data and serves the same features online at low latency, so training and serving use consistent values. Separate implementations are the cause of the skew. A serving-only cache does not align with training data, and a bigger model does not fix inconsistent inputs.

More free Google Cloud Professional Cloud Architect questions