GCP Professional ML Engineer Practice Question 2: Feature Store Online Serving
Google Cloud Professional Machine Learning Engineer · Test 1 · question 2 of 10
Two teams share a customer-churn model. The data science team builds training sets with a BigQuery query that computes aggregates such as avg_order_value_30d, while the application team reimplemented those same aggregates in the serving service's Java code. The aggregates have already drifted apart twice, and the serving path needs feature lookups in single-digit milliseconds.
You need one authoritative definition of each feature, used by both training and online serving, with the least duplicated logic. What should you implement?
Show answer and explanation
Answer:
Feature Store on Gemini Enterprise Agent Platform (formerly Vertex AI Feature Store) is designed around exactly this split. The BigQuery tables or views holding the feature data *are* the offline store, the same data the training query reads, and you layer feature groups, features, an online store, and a feature view on top so the latest values are served online at low latency from those same sources. One definition, two consumption paths, which is the structural fix for training-serving skew rather than a promise that two codebases will stay in sync.
Option A gets the latency but not the governance: a nightly cache is a second copy with its own freshness semantics, no feature registry, no notion of a feature_timestamp, and nothing stopping the two teams from disagreeing about what the aggregate means.
Option B keeps the two implementations conceptually unified but is unusable at serving time. A BigQuery query is an analytical scan measured in hundreds of milliseconds to seconds; it will not meet a single-digit-millisecond lookup budget for per-request feature retrieval.
Option D is the same defect as A with different plumbing. Replicating to Cloud SQL creates another operational database to size, patch, and keep in step, and it still leaves the feature definitions unregistered and unshared.
Reference: About Feature Store on Gemini Enterprise Agent Platform