You need to develop a custom TensorFlow model for online predictions with training data stored in BigQuery. You want to apply instance-level data transformations to the data consistently during both model training and serving. How should you configure the preprocessing routine?