Question

A machine learning team trains models in Vertex AI using a custom training pipeline. Their features are computed daily in BigQuery and include rolling 30-day aggregations of customer purchase behavior. The team needs to serve these same features to an online prediction endpoint with single-digit millisecond latency, and they want to avoid recomputing the aggregations at request time. Which approach best meets these requirements?

Accepted Answer

Ingest the precomputed features from BigQuery into a Vertex AI Feature Store online store, then read them by entity ID during online prediction.

Answer

Materialize the daily aggregations into a BigQuery table and query it directly from the online prediction service at request time.

Answer

Store the features in a BigQuery materialized view and enable BI Engine to accelerate the online prediction lookups.

Answer

Recompute the 30-day rolling aggregations inside a BigQuery ML TRANSFORM clause so serving stays consistent with training.