Databricks Machine Learning Professional Question 204
Single answerA company is deploying a machine learning model to predict customer churn in real-time. The model requires specific feature values such as the customer's most recent activity, current subscription status, and their last transaction amount, which are not available in the pre-computed feature store. What is the best approach to ensure the model receives these feature values for real-time deployment?
- A
Pre-compute all feature values and store them in a feature store for batch inference.
- B
Use Just-In-Time (JIT) feature computation to retrieve and compute features at the time of prediction.
- C
Rely only on historical, pre-computed feature values, even if some features are outdated.
- D
Disable real-time prediction and switch to batch processing to avoid missing feature values.
Show answer and explanation
Correct answer: B
Explanation
In real-time machine learning deployments, JIT feature computation is critical for ensuring the model receives the most up-to-date feature values at the time of prediction. This is especially important when the required feature values cannot be pre-computed or stored in a feature store, as is the case with features like recent activity or last transaction amount. JIT computation ensures that these values are dynamically retrieved and computed during the prediction process, enabling accurate and timely predictions.
- A. Incorrect.
Pre-computing all feature values in a feature store is suitable for batch inference but does not address the need for real-time predictions requiring up-to-date feature values.
- B. Correct.
Just-In-Time (JIT) feature computation is ideal for real-time deployment, as it ensures that the most up-to-date feature values are retrieved and computed when predictions are made.
- C. Incorrect.
Using outdated, pre-computed feature values can negatively impact model accuracy in real-time scenarios, where the latest data is essential for correct predictions.
- D. Incorrect.
Disabling real-time prediction and switching to batch processing defeats the purpose of real-time deployment and does not solve the issue of obtaining up-to-date feature values.