Databricks Certified Machine Learning Associate free question 1 / 10 · Domain 1: Databricks Machine Learning
In the MLOps workflow that Databricks recommends, an organization cannot give data scientists read-only access to the prod catalog while they work in the development stage. Which alternative does the documentation describe for developing and evaluating ML code in that situation?
- Place a copy of the production data in the staging catalog, and have the development workspace read the staging catalog instead
- Write a snapshot of the production data to the dev catalog, and build the project code against that snapshot
- Give the data scientists write access to the prod catalog, so that they can create their temporary and feature tables there
- Run the training job written in development in the production workspace, and register the resulting model to the dev catalog for evaluation
Show the answer and explanation
Answer: B
Why this is correct
In the recommended workflow the development environment is the dev catalog, where data scientists have read-write access and create temporary data and feature tables. Ideally they can also read production data, inference tables, and metric tables in the prod catalog. The page then states the fallback directly: if read-only access to the prod catalog cannot be granted, a snapshot of production data can be written to the dev catalog so that data scientists can develop and evaluate project code.
Why the other options are wrong
"Run the training job written in development in the production workspace, and register the resulting model to the dev catalog for evaluation" … Data scientists typically do not have write or compute access in the production environment, so running their development job in the production workspace is not what the workflow describes. The workflow also registers a model to the catalog of the environment the pipeline ran in, so a model trained there would land in the prod catalog, not in the dev catalog.
"Place a copy of the production data in the staging catalog, and have the development workspace read the staging catalog instead" … The staging catalog exists for testing ML pipelines and registering models during the staging stage, and assets written to it are generally temporary and kept only until testing is complete. The development environment may need access to it for debugging, but it is not described as the place to develop from.
"Give the data scientists write access to the prod catalog, so that they can create their temporary and feature tables there" … The workflow gives data scientists read-write access to the dev catalog, not the prod catalog, and notes that they typically do not have write or compute access in production. Production access for them is limited to read-only visibility of test results, logs, model artifacts, pipeline status, and monitoring tables.
Key point
The boundary between development and production is drawn in both workspaces and catalogs. Read-only access to the prod catalog is the ideal; a snapshot written to the dev catalog is the documented substitute when that access cannot be given.