DP-100 Designing and Implementing a Data Science Solution on Azure - Practice Test 1

A compute instance is a managed cloud workstation optimized for machine learning development. It comes pre-installed with Jupyter, JupyterLab, VS Code, and RStudio. A compute cluster is for scalable training jobs. An inference cluster (AKS) is for deploying models. Attached compute is for bringing your own resources. See more: Design and Prepare a ML Solution

When you create an Azure ML workspace, four associated resources are automatically provisioned: Azure Storage Account, Azure Key Vault, Azure Container Registry, and Application Insights. Azure Key Vault stores secrets and credentials used by the workspace. See more: Design and Prepare a ML Solution

A data asset of type uri_file is a versioned reference to a single file (like a CSV). A datastore is just a connection to a storage service, not a versioned reference to specific data. An mltable defines a tabular schema across files. An environment defines software dependencies. See more: Design and Prepare a ML Solution

A datastore is a connection reference to an Azure storage service (Blob Storage, Data Lake Storage Gen2, Azure SQL Database, etc.). It stores the connection information so you can access data without exposing credentials in your code. See more: Design and Prepare a ML Solution

The NC-series and ND-series VMs include NVIDIA GPUs and are designed for deep learning, computer vision, and other GPU-intensive workloads. D-series is general purpose, F-series is compute optimized (CPU), and E-series is memory optimized. See more: Design and Prepare a ML Solution

Azure ML Designer is a drag-and-drop interface for building ML pipelines visually. It supports data preparation, model training, and evaluation components. Automated ML automates algorithm selection but is not a pipeline builder. CLI v2 uses YAML files. Notebooks require code. See more: Explore Data and Train Models

The Split Data component divides a dataset into two parts (e.g., 80/20 for training and testing). It supports random split, percentage split, and stratified sampling. Select Columns filters columns, Clean Missing Data handles null values, and Normalize Data scales features. See more: Explore Data and Train Models

AUC_weighted is the recommended primary metric for imbalanced classification datasets because it accounts for the area under the ROC curve weighted by class frequency. Accuracy can be misleading with imbalanced data since predicting the majority class yields high accuracy. r2_score is for regression tasks. See more: Explore Data and Train Models

Azure ML uses MLflow as the default tracking and logging framework. MLflow provides APIs for logging parameters, metrics, and model artifacts. Azure ML Studio natively displays MLflow-logged data. TensorBoard can be used additionally but is not the default. See more: Explore Data and Train Models

Automated ML automatically iterates through multiple algorithms and hyperparameters to find the best model. A sweep job only searches hyperparameters for a single specified algorithm. Designer requires manual configuration. Pipeline jobs chain steps but don't auto-select algorithms. See more: Explore Data and Train Models

In Azure ML SDK v2, script parameters are passed via the command string using the [null] syntax. The inputs dictionary maps these placeholders to actual values. Environment variables and Key Vault are for secrets, not training parameters. See more: Prepare a Model for Deployment

A sweep job performs hyperparameter tuning by running multiple trials with different parameter combinations. It supports random, grid, and Bayesian sampling. A command job runs a single script execution. A pipeline job chains steps. An AutoML job selects algorithms automatically. See more: Prepare a Model for Deployment

In Azure ML pipelines, data flows between steps by connecting the output of one step to the input of the next. Azure ML handles serialization, storage, and deserialization automatically. Steps do not communicate via environment variables, databases, or REST calls. See more: Prepare a Model for Deployment

Publishing a pipeline as a pipeline endpoint makes it callable via a REST API. This enables automation, scheduling, and triggering from external systems. It does not deploy models for inference or share across workspaces. See more: Prepare a Model for Deployment

Managed online endpoints provide Azure-managed infrastructure for real-time inference with built-in auto-scaling, blue-green deployments, and monitoring. Batch endpoints are for bulk scoring. Compute instances are for development. Pipeline endpoints trigger training pipelines. See more: Deploy and Retrain a Model

Azure ML scoring scripts must define init() and run(). The init() function is called once when the service starts to load the model. The run() function is called for each inference request to process input and return predictions. See more: Deploy and Retrain a Model

Azure ML managed online endpoints support blue-green deployments. You deploy a new version (green) alongside the current (blue), gradually shift traffic from blue to green, and then remove the old deployment once validated. See more: Deploy and Retrain a Model

MLTable defines a tabular schema on top of one or more files. It supports column selection, type casting, and data transformations. uri_file points to a single file. uri_folder points to a directory. custom_model is a model asset type. See more: Design and Prepare a ML Solution

The Execute Python Script component in Azure ML Designer allows you to add custom Python code to your pipeline. It accepts DataFrames as input and outputs DataFrames, enabling custom feature engineering, transformations, or model logic. See more: Explore Data and Train Models

Recall (also called sensitivity or true positive rate) measures the proportion of actual positive cases correctly identified: TP / (TP + FN). Precision measures TP / (TP + FP). F1 is the harmonic mean of precision and recall. Accuracy is overall correct predictions / total predictions. See more: Explore Data and Train Models

With min_instances=0, the cluster scales down to zero nodes after idle_time_before_scale_down expires (default 120 seconds). At zero nodes, you incur no compute charges. The cluster is not deleted; it scales back up when a job is submitted. See more: Design and Prepare a ML Solution

An Azure ML environment encapsulates the software dependencies (conda/pip packages) and Docker base image for running training and scoring scripts. Hardware specs are defined by V size on compute targets. Credentials are managed by datastores. Network configuration is at the workspace level. See more: Prepare a Model for Deployment

The Azure ML Python SDK v2 uses the azure-ai-ml package. The primary entry point is MLClient from azure.ai.ml. The azureml.core import was for SDK v1 which is now legacy. See more: Design and Prepare a ML Solution

Batch endpoints are designed for asynchronous, high-throughput scoring of large datasets. They process data in parallel and are ideal for overnight or scheduled batch processing. Online endpoints are for real-time, low-latency requests. Pipeline endpoints trigger training pipelines. See more: Deploy and Retrain a Model

The blocked algorithms setting in Automated ML lets you exclude specific algorithms from consideration. Exit criteria control timeout and iteration limits. Featurization settings configure preprocessing. Cross-validation folds control model validation. See more: Explore Data and Train Models

Application Insights is an associated workspace resource that monitors deployed web services, collecting telemetry data like request rates, response times, and failure rates. Azure Monitor is a broader monitoring service. Key Vault stores secrets. Container Registry stores Docker images. See more: Design and Prepare a ML Solution

When a model is logged using MLflow (e.g., mlflow.sklearn.log_model()), you should register it with type MLFLOW_MODEL. This enables Azure ML to automatically generate a scoring script and environment for deployment. CUSTOM_MODEL is for models not logged with MLflow. See more: Deploy and Retrain a Model

Bayesian sampling builds a probabilistic model of the objective function and uses it to select the most promising hyperparameters for the next trial. Random sampling picks values randomly. Grid sampling exhaustively searches all combinations. Sobol is a quasi-random sequence. See more: Prepare a Model for Deployment

The init() function is called once when the service starts. It is used to load the model and any required resources into memory. The run() function is called for each inference request. See more: Deploy and Retrain a Model

The ml_client.online_endpoints.invoke() method sends a test request to a managed online endpoint. It accepts the endpoint name and a request file or data payload. The other methods listed do not exist in the Azure ML SDK v2. See more: Deploy and Retrain a Model

R-squared (R²) measures the proportion of variance in the target variable explained by the model. An R² of 1.0 means perfect prediction. MAE measures average absolute errors. RMSE penalizes larger errors more heavily. Relative Squared Error is normalized by the total variance. See more: Explore Data and Train Models

The Score Model component takes a trained model and a test dataset and generates predictions. Train Model creates the trained model. Evaluate Model compares predictions to actual labels to compute metrics. Select Columns filters columns. See more: Explore Data and Train Models

The Clean Missing Data component handles missing values. You can configure it to replace missing values with mean, median, mode, a custom value, or remove rows entirely. The "Replace using median" option substitutes nulls with the column median. See more: Explore Data and Train Models

Pipelines provide modularity, reproducibility, and reusability. Each step can be independently executed, tracked, and reused across experiments. GPU optimization is a compute feature. Hyperparameter tuning uses sweep jobs. Model deployment uses endpoints. See more: Prepare a Model for Deployment

When auth_mode is "key", you authenticate by including the endpoint's primary or secondary key in the Authorization header as a Bearer token: "Authorization: Bearer <key>". The key can be retrieved using ml_client.online_endpoints.get_keys(). See more: Deploy and Retrain a Model

df.describe() returns descriptive statistics: count, mean, std, min, 25%, 50%, 75%, and max for numeric columns. df.info() shows column types and non-null counts. df.head() shows first rows. df.shape returns row and column counts. See more: Explore Data and Train Models

For NLP text classification tasks, specify task type as "text_classification". The standard "classification" type is for tabular data. AutoML for NLP supports text_classification, text_classification_multilabel, and text_ner (named entity recognition). See more: Explore Data and Train Models

AZUREML_MODEL_DIR is the environment variable that contains the path to the directory where the model is deployed. In the init() function, you use os.getenv("AZUREML_MODEL_DIR") to locate and load the model file. See more: Deploy and Retrain a Model

Azure Container Instances (ACI) provides a quick, low-cost way to deploy and test models during development. AKS is for production-scale deployments. Compute clusters are for training. Data Lake Storage is for data storage. See more: Deploy and Retrain a Model

Managed online endpoints have built-in traffic routing. Set the traffic dictionary to allocate percentages: {"blue": 90, "green": 10}. This sends 90% of requests to blue and 10% to green. No external load balancer or separate endpoints are needed. See more: Deploy and Retrain a Model

mlflow.log_param() logs a single key-value parameter (e.g., learning rate, number of epochs). mlflow.log_metric() logs numeric metrics (e.g., accuracy, loss). mlflow.log_artifact() logs files. mlflow.set_tag() adds metadata tags to the run. See more: Explore Data and Train Models

Azure ML Studio at ml.azure.com is the purpose-built web interface for managing all workspace resources including data assets, experiments, models, endpoints, and compute. Azure Portal is for general Azure resource management. See more: Design and Prepare a ML Solution

Z-Score normalization (standardization) transforms data to have zero mean and unit standard deviation: (x - mean) / std. Min-Max scales values to a range [0,1]. Both are available in the Normalize Data component. See more: Explore Data and Train Models

The default_compute parameter in the @pipeline decorator specifies the compute target used by pipeline steps that don't explicitly define their own compute. Individual steps can override this with their own compute setting. See more: Prepare a Model for Deployment

Predicting house prices is a regression task because the target variable (price) is a continuous numeric value. Classification predicts categories. Forecasting is for time-series data. Clustering groups similar items. See more: Explore Data and Train Models

train_test_split() from sklearn.model_selection splits arrays or matrices into random train and test subsets. KFold is for k-fold cross-validation iterators. cross_val_score evaluates a model with cross-validation. split_data does not exist in sklearn. See more: Explore Data and Train Models

ml_client.compute.begin_delete() permanently deletes a compute resource. begin_stop() stops a compute instance but doesn't delete it. The remove() and shutdown() methods don't exist in the SDK v2. See more: Prepare a Model for Deployment

ml_client.jobs.stream() streams the console output of a running job in real-time. ml_client.jobs.get() retrieves job metadata and status. The logs() and watch() methods don't exist in SDK v2. See more: Prepare a Model for Deployment

df.corr() computes pairwise correlation coefficients between numeric columns. Values range from -1 (perfect negative correlation) to +1 (perfect positive correlation). df.describe() gives summary statistics. value_counts() counts unique values. groupby() groups data for aggregation. See more: Explore Data and Train Models

The recommended approach is blue-green deployment: create a new deployment with the updated model, test it with a small percentage of traffic, then gradually shift all traffic to the new deployment. This ensures zero downtime and allows rollback if issues are detected. See more: Deploy and Retrain a Model