DP-100 Sample Questions & Answers

Performance disparities often arise from imbalanced data where the model has insufficient examples for minority groups. The most effective initial step is to address this root cause by collecting more representative data or using sampling techniques (like SMOTE or random oversampling) to create a more balanced training dataset. This allows the model to learn the patterns for the minority group more effectively.

The MLmodel file is the primary metadata file for an MLflow model. It contains essential information, including the model's flavor, creation time, and most importantly, the signature. The signature explicitly defines the input and output schemas (data types, names, and tensor shapes), which is used by deployment tools to validate requests and format data correctly for inference.

This is true. Azure ML compute clusters can be configured with a minimum number of nodes (typically zero). The cluster automatically scales up when a job is submitted and scales down to the minimum count when idle. You are only charged for the nodes when they are allocated and running, not for the cluster resource itself when it's scaled to zero nodes.

The az ml online-endpoint update command is used to modify the properties of an existing endpoint, including the traffic allocation. To route all traffic to the green deployment, you would use this command with the --traffic "green=100" parameter. This updates the endpoint's routing rules in place, ensuring a smooth transition with no downtime.

The azureml:// URI scheme is the standard way to reference data assets and paths within datastores in Azure ML SDK v2. By creating a data asset with a path pointing to this URI and then using pd.read_csv(my_path), the SDK handles the authentication and data access seamlessly, loading the CSV into a Pandas DataFrame.

This approach balances cost and performance. Starting with the smaller, cheaper model is a cost-effective experiment. If Model-B can be fine-tuned to meet the required performance benchmarks, there is no need to incur the higher costs associated with the larger Model-A. This iterative approach, starting with the simplest/cheapest viable option, is a common best practice in machine learning.

In the Azure ML SDK v2, you can directly reference a local YAML file that defines an environment. When you construct the CommandJob, you set the environment parameter to the path of this file (e.g., environment='./my-env.yml'). Azure ML will then build or retrieve the specified environment and use it to run the job.

An Azure Machine Learning Compute Cluster is the ideal choice for scalable, distributed training. It allows you to create a multi-node cluster that can automatically scale up and down. For distributed frameworks like Horovod, a multi-node GPU cluster provides the necessary parallel processing power and interconnectivity to efficiently train models on massive datasets.

For forecasting tasks, AutoML has powerful built-in time-series featurization capabilities. By correctly identifying the time column in the forecasting settings and leaving featurization on 'auto' (the default), AutoML will automatically generate a rich set of features from the date/time column, such as year, month, day, day of week, week of year, and more.

The semantic ranker (also known as semantic search) is a premium feature in Azure AI Search that uses deep learning models to understand the contextual meaning and intent behind a query. It re-ranks the initial set of results from keyword or vector search based on semantic relevance, significantly improving the quality of results for RAG applications.