Azure DP-100: Data Scientist Associate Exam Guide

Domain 1 — Design a machine learning solution (20–25%)

This domain tests your ability to define the right Azure ML architecture before writing a single line of training code. Key topics include:

• Selecting the appropriate compute type: compute instances for interactive development, compute clusters for training jobs, serverless compute for cost optimisation, and Kubernetes-attached compute for hybrid scenarios.
• Configuring Azure ML workspace components: datastores, data assets, environments, and the role of the workspace registry for sharing assets across teams.
• Designing for cost efficiency: choosing between on-demand vs. dedicated clusters, low-priority VMs, and autoscaling policies.
• Connectivity and security: private endpoints, virtual network integration, managed identity for data access, and role assignments in the workspace.

Domain 2 — Explore data and train models (35–40%)

The heaviest domain, and the one where most of the SDK v2 content lives. You need to be comfortable with the full training workflow:

• Data access patterns — Creating and consuming mltable and uri_file/uri_folder data assets; reading from Azure Blob Storage, Azure Data Lake Gen2, and Azure SQL via registered datastores.
• AutoML — Configuring AutoML jobs for classification, regression, and time-series forecasting; understanding featurization, validation splits, exit criteria, and primary metric selection. AutoML appears more heavily on the exam than many candidates expect.
• Custom training jobs — Writing command and sweep jobs with the SDK v2; structuring training scripts to accept arguments and log metrics via mlflow.log_metric(); building reusable environments from curated images vs. custom Docker images.
• Hyperparameter tuning — Configuring sweep jobs with random, grid, and Bayesian sampling; early termination policies (Bandit, Median, Truncation Selection); understanding the trade-off between exploration and exploitation.
• MLflow experiment tracking — Logging parameters, metrics, artefacts, and models with MLflow autologging vs. explicit logging; comparing runs in the Azure ML Studio UI.
• Pipelines — Defining multi-step pipelines with the @pipeline decorator or YAML; passing data between steps; scheduling and triggering pipeline runs.

Domain 3 — Prepare a model for deployment (20–25%)

Before a model ships, it needs to be evaluated, registered, and packaged. This domain covers:

• Model registration — Registering models from training runs, from local files, or from external sources; versioning and tagging; the difference between registering as mlflow_model vs. custom_model.
• Responsible AI evaluation — Using the Azure ML Responsible AI dashboard to assess fairness (disaggregated metrics across cohorts), model interpretability (SHAP-based feature importances), error analysis (tree-based error explorer), and counterfactual analysis. The exam tests your ability to read the dashboard and draw the right conclusion — not the underlying mathematics.
• Model packaging — Writing scoring scripts (init() and run() functions) for custom deployment; understanding when MLflow model packaging removes the need for a scoring script entirely.

Domain 4 — Deploy and retrain a model (10–15%)

The smallest domain by weight but high-stakes in production scenarios:

• Managed online endpoints — Deploying to managed online endpoints with blue/green traffic splitting; configuring instance count, SKU, and autoscaling; testing with the Azure ML Studio test UI and REST client.
• Batch endpoints — When to use batch vs. online endpoints; configuring parallel scoring for large datasets; monitoring batch endpoint run status.
• Model monitoring — Setting up data drift and prediction drift monitors; configuring baseline datasets; understanding alert thresholds and the signal types (data drift, prediction drift, feature attribution drift).
• Retraining triggers — Automating retraining with pipeline schedules or monitoring alerts; the pattern of re-registering a retrained model and updating endpoint traffic allocation.

Where DP-100 sits in the Azure AI certification stack

• AZ-900 — Azure Fundamentals (prerequisite knowledge, not required)
• DP-100 — Azure Data Scientist Associate ← you are here
• AI-102 — Azure AI Engineer Associate (Azure AI Services + Azure OpenAI)
• DP-203 — Azure Data Engineer Associate (Synapse, Data Factory, Databricks)
• AZ-305 — Azure Solutions Architect Expert (enterprise architecture, includes AI/ML patterns)

High-value study areas based on the 2025–2026 objective weights

• SDK v2 job syntax (YAML and Python) — You need to recognise correct vs. incorrect command job configurations, input/output declarations, and environment references. Practice writing jobs from scratch, not just reading them.
• AutoML configuration — AutoML questions appear disproportionately often relative to the time most candidates spend on it. Know featurization options, primary metrics per task type, and how to interpret the AutoML leaderboard.
• MLflow tracking vs. Azure ML logging — The exam tests whether you know when to use mlflow.autolog(), when to call mlflow.log_metric() explicitly, and how these appear in the Azure ML Studio UI.
• Responsible AI dashboard interpretation — Practice reading cohort analysis results, identifying model fairness issues from disaggregated metrics, and recommending mitigation strategies. You do not need to code the dashboard — you need to interpret it.
• Endpoint types and traffic management — Know when to use online vs. batch endpoints, how to configure blue/green splits, and what happens when an endpoint autoscales.