50 Microsoft Certified: Azure Data Scientist Associate Practice Questions: Question Bank 2025

You are planning an Azure Machine Learning project and must ensure that data scientists can reproduce results across runs. Which action BEST supports reproducibility?

You have a tabular dataset in Azure Machine Learning where one class represents only 3% of the observations. You want a quick baseline model and evaluation that accounts for the imbalance. Which metric should you prioritize?

You trained a model using an Azure Machine Learning pipeline and want to ensure that the trained model artifact is easy to locate and reuse for deployment. What should you do at the end of training?

A managed online endpoint is receiving requests, but you need to review request rate, latency, and failures over time. Where should you look first in Azure Machine Learning?

You must train a model on data stored in Azure Data Lake Storage Gen2. Company policy forbids storing account keys in code or notebooks. What is the recommended approach for the training job to access the data?

You are evaluating a regression model in Azure Machine Learning. The residuals show a clear non-linear pattern when plotted against predicted values. What is the most likely implication?

You are using Azure Machine Learning pipelines. One step performs heavy feature engineering and produces a transformed dataset that rarely changes. Training runs are taking too long because the feature engineering step executes every time. What should you do to improve performance?

You want to deploy a model, but the scoring script requires custom Python libraries not included in curated environments. Which approach is recommended to ensure consistent deployments across dev/test/prod?

You deploy a model to a managed online endpoint. Calls to the endpoint return HTTP 503 errors intermittently during traffic spikes. The model loads slowly on startup due to a large artifact. What is the BEST mitigation?

Your team must meet a requirement that only approved models can be deployed, and every deployment must be traceable to the exact training code, data version, and environment. Which design BEST satisfies this requirement in Azure Machine Learning?

You are preparing data for a classification model in Azure Machine Learning. Several columns are stored as strings but represent categories (for example, "Low", "Medium", "High"). You want to ensure a consistent, reproducible transformation is applied during both training and inference. What should you use?

You have a registered model in Azure Machine Learning and need to deploy it for real-time scoring to a managed compute target with minimal infrastructure management. Which compute target should you choose?

You want to compare several training runs for a model and quickly identify which run produced the best validation AUC. You are using Azure Machine Learning SDK v2. Which feature should you use?

A data science team needs to ensure that only approved, curated datasets are used for model training. They want dataset versioning and a central place to discover and reuse datasets across projects in the same Azure Machine Learning workspace. What should they implement?

You are using Automated ML in Azure Machine Learning for a time-series forecasting problem. The dataset contains multiple stores, each with its own demand history. You want the model to learn from all stores while producing forecasts per store. Which configuration is required?

You need to deploy a model that requires custom Python packages not included in the default Azure Machine Learning base images. You want a repeatable deployment process for both testing and production. What is the recommended approach?

Your team is standardizing on MLflow for experiment tracking. You want to log parameters, metrics, and the trained model artifact from training runs executed in Azure Machine Learning. What should you do?

A real-time endpoint deployed in Azure Machine Learning shows intermittent 504/timeout errors during peak load. CPU utilization is high and request latency spikes. What is the best next step to improve reliability while maintaining real-time scoring?

You must deploy a model in Azure Machine Learning to a production environment where inbound traffic must remain on a private network and public internet access is not allowed. Which design best meets this requirement?

A regulated organization requires that every production inference request can be traced back to the exact model version, code, and data preprocessing steps used. They also want to reproduce the training run that produced the deployed model. What should you implement in Azure Machine Learning?

You are building a training pipeline in Azure Machine Learning. You need to ensure the same input data is used across repeated training runs and can be traced for audit purposes. What should you use?

You trained a classification model in Azure Machine Learning and want to compare multiple runs and select the best model based on AUC. Which Azure ML feature is designed for this?

You deployed a real-time endpoint and want to ensure the scoring script can access secrets (for example, a database password) without hardcoding them in code or environment variables stored in source control. What is the recommended approach?

Your team needs to train a model using a large dataset stored in Azure Data Lake Storage Gen2. You want to avoid copying data during training and allow parallel reads. Which dataset access approach best fits?

You are using automated ML for a binary classification problem with highly imbalanced classes. Your model performs well on accuracy but misses most positive cases. What should you change first to better reflect business goals during training?

A training script runs successfully on your local machine but fails in an Azure ML job with a missing library error. You want a repeatable fix that ensures the same dependencies are used for every run. What should you do?

You deployed a managed online endpoint. Requests sometimes fail due to intermittent timeouts when the input payload is large. Metrics show CPU saturation on the deployment. What is the best first action to improve reliability?

You must ensure that a deployed model cannot be called from the public internet and is only reachable from resources inside your virtual network. Which deployment configuration best meets this requirement?

You need to deploy a model that requires custom pre-processing code and non-Python system dependencies (for example, OS packages). You also need reproducibility across dev/test/prod. What is the best approach?

Your regulated organization requires end-to-end lineage: which code, data, environment, and parameters produced the model currently in production. Which combination of Azure ML practices best satisfies this requirement?

You need to ensure the same Python dependencies are used during training and batch inference in Azure Machine Learning. You want the most maintainable approach for reusing the environment across jobs. What should you do?

You are training a model with a large dataset stored in Azure Data Lake Storage Gen2. You want to avoid copying the data to the compute target and instead stream it efficiently during training using Azure ML. What should you use?

You deployed an online endpoint in Azure Machine Learning. You want to view request latency, request count, and failed requests over time. Which approach should you use?

You want to compare two classification models in Azure Machine Learning using a single number that reflects performance across all thresholds. Which metric should you use?

A team is using Azure ML pipelines and wants each step to rerun only when its inputs change to reduce compute usage. What should they implement?

You trained a model and want to deploy it to a managed online endpoint with minimal operational overhead, including managed scaling and TLS termination. What compute option should you choose for the deployment?

Your dataset has missing values in multiple numeric features. You want to handle missingness in a way that prevents data leakage during cross-validation. What is the recommended approach?

You need to deploy a model that requires a custom inference server and nonstandard network configuration. You still want to manage it through Azure ML endpoints. What deployment approach should you choose?

A regulated organization requires that training jobs cannot exfiltrate data to the public internet. You must run training in Azure ML while meeting this requirement. What should you implement?

After deploying a new model version to an online endpoint, you notice intermittent 500 errors under load. You suspect the issue occurs only when multiple requests are processed concurrently. What is the best next step to confirm and isolate the cause in Azure ML?

You are exploring a dataset in Azure Machine Learning. Several columns contain long-tailed numeric values, and your model is overly influenced by a few extreme outliers. You want a simple transformation that reduces the impact of outliers without removing rows. What should you do?

You want to run a one-time training job in Azure Machine Learning using a curated environment and a managed compute cluster. Which asset best represents the execution of your training script and produces outputs like metrics and model artifacts?

You deployed a real-time endpoint in Azure Machine Learning and need to collect request/response payloads for later debugging. You must avoid storing any sensitive raw payloads but still want to detect schema drift. What should you configure?

A classification dataset contains 2% positive labels. Your initial model shows high accuracy but poor recall for the positive class. Which evaluation approach best reflects model performance for the minority class?

You are training a model with the Azure Machine Learning Python SDK and want to ensure the same random train/validation split is used across runs for reproducibility. What should you do?

Your team needs to standardize feature engineering so that training and batch scoring apply identical transformations. They want versioning and reuse across multiple models. What should you implement in Azure Machine Learning?

You must train a model using data stored in a private Azure Storage account that blocks public access. Training will run on a managed compute cluster. Which approach is recommended to allow secure access from Azure Machine Learning?

You deployed a model to a managed online endpoint. After deployment, requests intermittently return HTTP 500 errors. Application Insights shows the container is restarting due to out-of-memory conditions. What is the most appropriate first action?

You need to publish a batch scoring pipeline that can be triggered on a schedule and writes predictions back to a datastore. The pipeline must be reusable across environments (dev/test/prod) with minimal code changes. Which design best meets these requirements?

A regulated organization requires that model artifacts, training data references, hyperparameters, code version, and environment details are all auditable for every production model. They also require the ability to recreate a past model exactly. Which approach best satisfies these requirements in Azure Machine Learning?