Prepare a model for deployment
Free Microsoft Certified: Azure Data Scientist Associate practice — 6 questions on Prepare a model for deployment, with explanations. No sign-up. Full 12-question mixed test →
You deployed an MLflow model to a managed online endpoint using no-code deployment. Inference requests fail with a schema validation error because the JSON payload's column order differs from the signature recorded in the model's MLmodel file. What is the BEST fix?
No-code MLflow deployments auto-generate the scoring logic based on the signature logged with the model. The root cause is a client-side payload mismatch, so aligning the payload to the signature (column names, order, types) resolves the error without adding unnecessary custom code.
In the Responsible AI dashboard's fairness assessment, Group A has a selection rate of 0.42 and Group B has a selection rate of 0.30. Using the common rule of thumb that a demographic parity difference greater than 0.1 signals a fairness concern, what is the demographic parity difference and does it flag a concern?
Demographic parity difference is the absolute difference between selection rates: 0.42 − 0.30 = 0.12. Since 0.12 exceeds the 0.1 rule-of-thumb threshold, this flags a fairness concern that should be investigated before deployment.
You are preparing an MLflow-logged model for deployment, but the model requires specific CUDA libraries for GPU inference that are not available through a conda-only environment definition. Using the Python SDK v2, what is the correct way to package the environment?
When a model needs GPU libraries beyond what conda can provide, SDK v2 supports defining an Environment with a custom Docker image (containing the CUDA base) plus a conda_file layering the Python/MLflow dependencies on top — this is the supported pattern for custom GPU environments.
A media company must score 50 million video metadata records once per night. Each scoring run may take up to 4 hours, and the compute must scale to zero when idle to control cost. Which deployment target BEST fits these requirements?
Batch endpoints are designed for large-volume, long-running, asynchronous scoring jobs and run on AmlCompute clusters that can scale down to zero nodes when idle, minimizing cost between nightly runs.
You want to register a model using the Python SDK v2 so that Azure ML automatically generates the scoring script and environment for no-code deployment to a real-time endpoint. Which code correctly registers the model for this behavior?
Registering a model with type=AssetTypes.MLFLOW_MODEL preserves the MLflow flavor metadata (signature, environment, dependencies), which Azure ML uses to auto-generate the scoring script and environment for no-code deployment.
Which visualization in the Responsible AI dashboard is specifically designed to identify cohorts of data where the model exhibits disproportionately high error rates compared to the overall dataset?
The Error Analysis component's tree map and heat map visualizations partition the dataset into feature-based cohorts and highlight which slices have error rates significantly higher than the overall average, guiding targeted debugging before deployment.
Ready for the real thing?
The full course has two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed answer explanations.
Start my full course on Udemy →