QuestionQ165
Advanced Feature Store ConceptsA Machine Learning Engineer is building a real-time fraud-detection pipeline in Databricks. The existing pipeline uses historical transaction features stored in the Databricks Feature Store and serves a trained model through MLflow and Databricks Model Serving.
The business team now asks to add dynamic, recent-event-based features, such as the number of suspicious transactions in the last 5 minutes, to improve model accuracy. These features must be calculated and served efficiently at inference time. The engineer uses the Databricks Feature Engineering client to define transformation logic and logs the model with MLflow. They need to integrate these dynamic features into the inference pipeline by using the Databricks Feature Store. They must also ensure low latency and follow Databricks best practices.
Which approach satisfies these requirements?
- A Store the dynamic features in a separate real-time database and query them during inference, combining them with Feature Store features manually in the serving layer.
- B Define custom transformation functions using the Databricks Feature Engineering client, register them with the Feature Store, and log the model with MLflow.
- C Backfill the custom dynamic features into the Feature Store using batch jobs after each inference so they are available for future predictions.
- D Persist raw transaction events in a Delta table and use model serving’s memory or local feature cache to join them with historical features at inference time using custom logic.
Community Discussion