Microsoft DP-100 Practice Test

DP-100 practice test: what's covered

• Azure Data Scientist Associate

Free DP-100 sample questions

1. A data science team needs an interactive authoring surface in Azure Machine Learning studio for exploring data and prototyping training scripts in Jupyter and VS Code. Which compute target should you provision for this individual development work?

   • A. A compute instance, which is a managed single-node VM for interactive notebook authoring
   • B. A compute cluster configured with a minimum node count set to zero nodes
   • C. An attached Azure Databricks workspace running a shared Spark cluster
   • D. A Kubernetes inference cluster created from an existing AKS resource
   Show answer

   Correct: A. A compute instance is a managed, single-user VM designed for interactive development with Jupyter, JupyterLab, and VS Code integration, making it the right development workstation. A compute cluster is for scaled, non-interactive training jobs and is not meant for interactive authoring. Attached Databricks targets distributed Spark workloads, not lightweight notebook prototyping, and an inference cluster (AKS) is used only for deploying models to real-time endpoints.

2. You are connecting an Azure Machine Learning workspace to existing storage so training jobs can reference data securely without embedding credentials. Which statements about datastores and data assets are accurate? (Choose TWO)

   • A. A datastore stores the connection information and credentials to an Azure storage service
   • B. A data asset is a versioned reference to data that adds lineage and reproducibility
   • C. A datastore physically copies all source files into the workspace default container
   • D. A data asset can only point to a single CSV file and never to a folder of files
   • E. Datastores must use account keys and cannot authenticate with managed identity
   Show answer

   Correct: A and B. A datastore securely abstracts the connection and credentials to backing storage such as Blob or ADLS Gen2, and a data asset (uri_file, uri_folder, or mltable) is a versioned pointer that adds lineage and reproducibility. The copy claim is wrong because datastores reference storage rather than duplicating files into the workspace. A data asset is not limited to a single CSV; uri_folder and mltable cover folders and tables. The managed identity claim is incorrect because datastores support identity-based access in addition to keys and SAS tokens.

3. Your organization requires that all Azure Machine Learning training jobs run against an existing big-data Spark environment already governed by the data platform team. Which compute approach satisfies this requirement?

   • A. Attach the existing Azure Databricks cluster as an attached compute target
   • B. Recreate the workload on a serverless online endpoint instead
   • C. Provision a new GPU compute instance for each team member
   • D. Export the data and run everything on a local laptop kernel
   Show answer

   Correct: A. Attached compute lets the workspace use external resources such as Azure Databricks or HDInsight that are already managed elsewhere, which directly reuses the governed Spark environment. An online endpoint is for low-latency model serving, not Spark training, so it does not apply. Provisioning new compute instances ignores the requirement to use the existing platform, and running on a local laptop abandons governance, scalability, and reproducibility entirely.

4. A churn dataset has 4% positive cases, and stakeholders complain that a model reporting 96% accuracy still misses almost every churner. Which evaluation metric should you prioritize to reflect performance on the minority class?

   • A. The area under the precision-recall curve, which focuses on the minority positives
   • B. Plain classification accuracy averaged across both observed classes
   • C. Mean squared error computed on the predicted class indices
   • D. The R-squared coefficient of determination for the predictions
   Show answer

   Correct: A. Precision-recall AUC concentrates on the positive (minority) class and exposes weak recall that accuracy hides, which is exactly the stakeholders' concern. Plain accuracy is misleading on heavy class imbalance because predicting the majority class alone scores high. Mean squared error and R-squared are regression metrics and are not appropriate for measuring classification quality on imbalanced labels.

5. While preparing a train-test split for a fraud classifier with rare positive events, you must ensure both partitions retain the same class balance as the full dataset. Which sampling technique accomplishes this?

   • A. Stratified sampling that preserves the original class proportions in each split
   • B. Simple random sampling without regard to the target label distribution
   • C. Dropping all minority rows before partitioning the dataset
   • D. Sorting rows by the label column and slicing the first portion
   Show answer

   Correct: A. Stratified sampling keeps the proportion of each class identical across train and test sets, which is critical for rare-event problems and reliable evaluation. Simple random sampling can leave a split with too few positives and unstable metrics. Dropping minority rows destroys the signal you are trying to detect, and sorting then slicing by label creates partitions that are entirely one class, making evaluation meaningless.

Last updated: 2026-10-19

DP-100 practice test FAQ

Related practice tests

• Microsoft DP-600 practice test
• AWS MLA-C01 practice test
• Microsoft DP-900 practice test
• Databricks Machine Learning Associate practice test
• Microsoft DP-700 practice test
• Microsoft AI-102 practice test