How to Ace the DP-100: Azure Data Scientist Associate

Question 1 · Design and prepare a machine learning solution

Your team is designing an Azure Machine Learning workspace for a large-scale financial services project. Which authentication method provides the most secure and scalable access to datastores?

• Use a service principal with role-based access control (RBAC)
• Use shared access signature (SAS) tokens
• Use account keys with global access
• Configure basic username/password authentication

Correct answer: Use a service principal with role-based access control (RBAC)

A service principal provides fine-grained, programmatic access with the ability to assign specific permissions and rotate credentials without disrupting overall system access. It supports enterprise security requirements and follows the principle of least privilege.

Question 2 · Explore data and train models

You are developing an AutoML experiment for a binary classification problem with an imbalanced dataset. Which configuration would best address potential performance issues?

• Enable cross-validation and configure primary metric as AUC_weighted
• Use default settings with accuracy as primary metric
• Disable feature engineering and use basic sampling
• Apply random oversampling without additional preprocessing

Correct answer: Enable cross-validation and configure primary metric as AUC_weighted

Using AUC_weighted as the primary metric helps handle class imbalance by considering performance across all classes. Cross-validation ensures robust model evaluation and reduces overfitting in imbalanced datasets.

Question 3 · Prepare a model for deployment

You are training a large machine learning model using Azure Machine Learning and want to optimize training costs. Which approach would most efficiently reduce compute expenses?

• Configure spot instances with automatic fallback to standard instances
• Use only premium GPU instances
• Run training jobs continuously without interruption
• Manually terminate and restart training jobs

Correct answer: Configure spot instances with automatic fallback to standard instances

Spot instances provide significant cost savings by utilizing unused Azure compute capacity at a lower price. Configuring automatic fallback ensures job completion if spot instances are preempted, balancing cost optimization and training reliability.

Question 4 · Monitor models

You have deployed a machine learning model in Azure ML that predicts customer churn. After two months, you notice the model's accuracy has dropped from 92% to 78%. What is the most likely cause?

• The Azure ML compute cluster needs to be scaled up
• The model's hyperparameters were incorrectly set during deployment
• The prediction latency has increased
• The model requires retraining due to data drift

Correct answer: The model requires retraining due to data drift

A significant drop in model accuracy over time is a classic indicator of data drift, where the statistical properties of input data change. This is one of the primary reasons models degrade in production. Scaling compute or hyperparameter issues would not cause accuracy degradation in this manner.

Question 5 · Design and prepare a machine learning solution

You are configuring a compute cluster for distributed machine learning training. Which configuration ensures cost-effective and flexible compute resources?

• Enable auto-scaling with a minimum of 0 and maximum of 10 nodes
• Set a fixed number of 5 nodes permanently
• Use a single large GPU instance
• Disable auto-scaling completely

Correct answer: Enable auto-scaling with a minimum of 0 and maximum of 10 nodes

Auto-scaling with a range of 0-10 nodes allows the cluster to dynamically adjust compute resources based on workload, minimizing idle time and reducing overall compute costs while maintaining training flexibility.