Feature serving

After you have created features, you can make those features available for model training and inference. This document describes the feature serving capabilities available in BigQuery ML.

Point-in-time correctness

The data used to train a model often has time dependencies built into it. Time sensitive feature tables include a timestamp column to represent the feature values as they existed at a given time for each row. You can use point-in-time lookup functions when querying data from time sensitive feature tables in order to ensure that there is no data leakage between training and serving. This enables point-in-time correctness.

Use the following functions to specify point-in-time cutoffs when retrieving time sensitive features:

• ML.FEATURES_AT_TIME
• ML.ENTITY_FEATURES_AT_TIME

You can use retrieved features to perform the following tasks:

• Train models and run inference in BigQuery ML.
• Directly serve online features in Vertex AI Feature Store.

Online serving with Vertex AI Feature Store

In addition to the built-in feature serving support in BigQuery ML, seamless integration with Vertex AI Feature Store lets you manage and serve features with low latency. More specifically, you can use the point-in-time lookup functions to create feature tables or views that you can serve directly, or you can manually create feature tables and register them with Vertex AI Feature Store for online serving. For more information, see Prepare data source.