A team uses Vertex AI Feature Store to serve features for online predictions. They notice that the online serving latency is high for certain features. The features are stored in a BigQuery source with high cardinality. What is the best practice to reduce latency?