Databricks Machine Learning Professional Question 55
Select 3You are working on a machine learning project in Databricks, and you have deployed a model to production using MLflow. A new version of the model has been trained and registered in the MLflow Model Registry. Which of the following steps are necessary to safely update the production model to the new version while minimizing downtime?
- A
Set the new version of the model to the 'Production' stage in the MLflow Model Registry.
- B
Perform A/B testing with a subset of traffic routed to the new model version before promoting it to production.
- C
Archive the current production model version immediately after registering the new model.
- D
Validate the new model version using a staging environment before making it the production model.
- E
Directly delete the old production model version to avoid confusion.
Show answer and explanation
Correct answers: A, B, D
Explanation
To safely update a production model with a new version in Databricks using MLflow, it is crucial to follow a structured process that includes staging, validation, and testing (e.g., A/B testing). This ensures that the new model performs as expected and minimizes the risk of disruption to the production environment. Directly archiving or deleting the old model or skipping validation/testing can lead to unnecessary risks and downtime.
- A. Correct.
Setting the new version of the model to the 'Production' stage is necessary to update the production environment. However, this should only be done after validation and testing.
- B. Correct.
A/B testing allows you to evaluate the performance of the new model version in a controlled environment by routing partial traffic to it. This is a safe approach to mitigate risks.
- C. Incorrect.
Archiving the current production model immediately after registering the new model is not advisable. The current production model should remain active until the new model is validated and ready.
- D. Correct.
Validating the new model version in a staging environment ensures that it performs as expected before promoting it to production.
- E. Incorrect.
Deleting the old production model version immediately is risky and not recommended. It should remain available for rollback purposes in case issues arise with the new model.