Feature serving

This document describes your options for making features available for BigQuery ML model training and inference. For all options, you must save the features in BigQuery tables as a prerequisite first step.

Point-in-time correctness

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

Serve features in BigQuery ML

To train models and perform batch inference in BigQuery ML, you can retrieve features using one of the point-in-time lookup functions described in the Point-in-time correctness section. You can include these functions in the query_statement clause of the CREATE MODEL statement for training, or in the query_statement clause of the appropriate table-valued function, such as ML.PREDICT, for serving.

Serve features with Vertex AI Feature Store

To serve features to BigQuery ML models that are registered in Vertex AI, you can use Vertex AI Feature Store. Vertex AI Feature Store works on top of feature tables in BigQuery to manage and serve features with low latency. You can use online serving to retrieve features in real time for online prediction, and you can use offline serving to retrieve features for model training.

For more information about preparing BigQuery feature data to be used in Vertex AI Feature Store, see Prepare data source.