AI-300 Practice Questions

The Azure Machine Learning model registry acts as a central repository within a workspace to manage the lifecycle of machine learning models. It allows you to store, version, and track your trained models along with their associated metadata, such as metrics and tags. This functionality is crucial for reproducibility, governance, and deploying the correct model version to production environments.

A. A Dataset is a reference to data in a datastore used for training; it is not the trained model itself.

C. A Pipeline is a workflow of ML steps; it can produce and register a model but is not the storage component.

D. A Compute cluster provides the processing resources for training models but does not store the resulting artifacts.

1. Microsoft. (2023). Manage models with Azure Machine Learning. Microsoft Learn. Retrieved from https://docs.microsoft.com/en-us/azure/machine-learning/concept-model. (Section: "What is a model?")

2. Microsoft. (2023). What is Azure Machine Learning?. Microsoft Learn. Retrieved from https://docs.microsoft.com/en-us/azure/machine-learning/overview-what-is-azure-machine-learning. (Section: "Assets")

A Canary deployment is a progressive exposure technique where a new model version is rolled out to a small, controlled subset of live traffic. The performance of this "canary" release is closely monitored against key metrics. If the new model performs as expected without introducing issues, traffic is gradually shifted from the old version to the new one until the rollout is complete. This strategy minimizes the risk and blast radius of deploying a potentially flawed model to the entire user base.

A. Blue/green: This strategy involves switching 100% of traffic from the old (blue) environment to the new (green) environment at once, not gradually.

C. Batch: This refers to a method of processing data (batch inference) rather than a strategy for deploying a real-time service.

D. Static: This is not a standard industry term for a progressive deployment strategy in the context of MLOps or DevOps.

1. Microsoft Learn. (2023). Safe rollout for online endpoints. "With a canary strategy, you can do a controlled rollout by first directing a small percentage of traffic to the new deployment. Then, you can increase the percentage of traffic once you're confident in the performance of the new deployment." Section: "Canary strategy".

2. Microsoft Azure. (2023). Cloud Design Patterns: Canary Release Pattern. "The Canary Release pattern is a deployment strategy that releases an application or service increment to a small subset of users. This pattern helps to verify the new version with a real-world workload before rolling it out to all users."

3. Fowler, M. (2010). CanaryRelease. martinfowler.com. "A canary release is a technique to reduce the risk of introducing a new software version in production by slowly rolling out the change to a small subset of users before rolling it out to the entire infrastructure."

A batch endpoint is the ideal deployment type for scheduled scoring jobs. It is designed to handle large volumes of data asynchronously. You invoke the endpoint with a pointer to the input data, and it initiates a job that processes the data in parallel on a compute cluster. This is efficient and cost-effective for tasks that are not time-sensitive and run on a recurring schedule, such as nightly or weekly batch predictions.

A. Real-time endpoint: This is designed for low-latency, on-demand inference on single or small batches of data, which is unsuitable for large, scheduled jobs.

C. Streaming endpoint: This is not a standard deployment type in Azure Machine Learning. Stream processing is typically handled by services like Azure Stream Analytics.

D. Manual deployment: This describes the action of deploying, not the type of endpoint or service created for scoring.

1. Microsoft. (2023). What are Azure Machine Learning endpoints? Microsoft Azure Documentation. Retrieved from https://learn.microsoft.com/en-us/azure/machine-learning/concept-endpoints. Section: "Batch endpoints".

2. Microsoft. (2023). Deploy models for batch inference with batch endpoints. Microsoft Azure Documentation. Retrieved from https://learn.microsoft.com/en-us/azure/machine-learning/how-to-use-batch-endpoint. Section: "When to use batch endpoints".

The primary purpose of Continuous Integration and Continuous Deployment (CI/CD) in MLOps is to automate the end-to-end machine learning lifecycle. This includes automatically building, testing, and deploying ML models and their associated application components. By automating these processes, MLOps aims to create a more reliable, repeatable, and rapid workflow for delivering machine learning solutions to production, reducing manual errors and accelerating the time to market.

A. CI/CD is the practice of automating workflows to eliminate manual deployment and integration tasks.

C. Data labeling is a data preparation activity that may be part of an MLOps pipeline but is not the purpose of CI/CD.

D. Model visualization is part of the analysis and monitoring phase, not the core function of the CI/CD pipeline.

1. Microsoft. (2023). What is MLOps?. Microsoft Learn. Retrieved from https://docs.microsoft.com/en-us/azure/machine-learning/concept-mlops. (Section: "MLOps and CI/CD")

2. Microsoft. (2023). Set up MLOps with Azure DevOps. Microsoft Learn. Retrieved from https://docs.microsoft.com/en-us/azure/machine-learning/how-to-setup-mlops-azureml. (Section: "CI/CD architecture")

Ensemble learning is a machine learning technique that combines the predictions from two or more models to produce a more accurate or robust prediction than any individual model. The core idea is that by aggregating the "votes" or predictions of multiple models, the weaknesses of individual models can be mitigated, leading to improved overall performance and generalization. Common ensemble methods include bagging (e.g., Random Forest), boosting (e.g., Gradient Boosting), and stacking.

B. Data partitioning: This is the process of splitting a dataset into subsets (e.g., train, test, validate) for model development and evaluation, not combining models.

C. Feature scaling: This is a data preprocessing technique used to normalize the range of features, which aids in model training but does not involve combining models.

D. Model deployment: This is the operational process of making a trained model available for use, which is distinct from the training technique of combining models.

1. Dietterich, T. G. (2000). Ensemble Methods in Machine Learning. In Multiple Classifier Systems, First International Workshop, MCS 2000 Cagliari, Italy, June 21–23, 2000 Proceedings (pp. 1-15). Springer Berlin Heidelberg. https://doi.org/10.1007/3-540-45014-91 (Section 1, "Introduction").

2. Microsoft. (2023). Automated machine learning (AutoML). Azure Machine Learning Documentation. Retrieved from https://learn.microsoft.com/en-us/azure/machine-learning/concept-automated-ml. (The documentation notes that AutoML uses "voting and stacking ensemble models" as a final iteration to improve results).

3. Ng, A. (n.d.). Machine Learning. Stanford University Courseware (CS229). Retrieved from http://cs229.stanford.edu/notes2019fall/cs229-notes-ensemble.pdf (Section 1, "Ensemble learning").

An Azure Machine Learning compute cluster is a managed-compute infrastructure designed for scalable training workloads. It consists of a cluster of CPU or GPU virtual machine nodes that can automatically scale up when a job is submitted and scale down when idle, making it cost-effective and powerful. This auto-scaling capability and its design for parallel processing make it the best option for training models on large datasets or for computationally intensive tasks that require distributed training.

A. A compute instance is a single-node, managed workstation primarily used for development and debugging, not for scalable training jobs.

C. Notebook VM is an older term for what is now the compute instance; it serves the same purpose and is not designed for scalability.

D. A data store is a reference to a storage service (like Azure Blob Storage) and is not a compute resource used for training models.

1. Microsoft Documentation, "What are compute targets in Azure Machine Learning?". Under the "Training compute targets" section, it describes a "Compute cluster" as a "managed-compute infrastructure that allows you to easily create a cluster of CPU or GPU compute nodes in the cloud." It explicitly mentions its auto-scaling capabilities.

2. Microsoft Documentation, "azureml.core.compute.AmlCompute class". The documentation for the class used to create compute clusters states it is "A managed compute target that you can use to train models and run scripts at scale."

3. Microsoft Documentation, "What is an Azure Machine Learning compute instance?". This document clarifies that a compute instance is a "managed cloud-based workstation for data scientists" and is "optimized for machine learning development workflows."

In Azure Machine Learning, pipelines can be automated to run on a schedule or in response to data changes. This is accomplished using pipeline triggers. There are two types: a schedule trigger, which runs the pipeline based on a specified time schedule (e.g., daily, weekly), and a data change trigger, which initiates a pipeline run when new data is created or modified in a specified datastore location. This functionality is essential for automating retraining and batch inference workflows.

B. Data store: A data store is a reference to a storage service where data is kept; it does not schedule pipeline execution, though it can be part of a data-based trigger.

C. Endpoint scaling: This feature adjusts the compute resources for a deployed model to handle request loads, not for scheduling training or batch pipelines.

D. Model registry: The model registry is a repository for storing and versioning trained models, not a scheduling mechanism.

1. Microsoft. (2023). Trigger a machine learning pipeline from a Logic App. Azure Machine Learning Documentation. Retrieved from https://learn.microsoft.com/en-us/azure/machine-learning/how-to-trigger-pipeline-logic-app. (This document explains how to trigger pipelines, implicitly covering the concept of triggers).

2. Microsoft. (2023). Schedule a pipeline job. Azure Machine Learning Documentation. Retrieved from https://learn.microsoft.com/en-us/azure/machine-learning/how-to-schedule-pipeline-job. (This document explicitly details the Schedule object and its use for time-based triggers).

3. Microsoft. (2023). Create event-driven machine learning workflows. Azure Machine Learning Documentation. Retrieved from https://learn.microsoft.com/en-us/azure/machine-learning/how-to-use-event-grid. (This document covers data-change based triggers using Event Grid).

An Azure Machine Learning compute cluster is a managed-compute infrastructure of virtual machines that automatically scales up or down based on workload. When a training job is submitted, the cluster provisions nodes up to the configured maximum. Once the job completes, it scales down to the minimum node count, which can be zero to minimize costs. This autoscaling capability is ideal for handling variable training demands without manual intervention, making it the correct choice for scaling compute resources automatically during training.

A. Compute instance: This is a single-node development workstation and does not automatically scale for training jobs.

C. Dataset: This is a reference to data used for training or inference; it is not a compute resource.

D. Model registry: This is a repository for storing and versioning trained models, not for executing training jobs.

- Microsoft. (2023). What are compute targets in Azure Machine Learning? Azure Machine Learning Documentation. Retrieved from https://learn.microsoft.com/en-us/azure/machine-learning/concept-compute-target. In the "Azure Machine Learning compute cluster" section, it states, "This compute infrastructure is created and managed by Azure Machine Learning. It scales automatically when a job is submitted."

- Microsoft. (2023). Create an Azure Machine Learning compute cluster. Azure Machine Learning Documentation. Retrieved from https://learn.microsoft.com/en-us/azure/machine-learning/how-to-create-attach-compute-cluster. The document details the minnodes and maxnodes parameters that control the autoscaling behavior.

Blue/green deployment is a release strategy that minimizes downtime and risk by running two identical production environments, referred to as "Blue" and "Green." Only one environment is live at any time. To deploy a new version, you deploy it to the inactive environment (e.g., Green) while the current version runs on the live one (Blue). After testing the Green environment, you switch the router to direct all incoming traffic to it. If any issues arise with the new version, you can perform an instant rollback by simply switching the router back to the Blue environment, which is still running the previous stable version.

B. Canary: This involves a gradual rollout to a subset of users. Rollback is possible but is not an instant, single-action switch like in blue/green.

C. Shadow: This strategy runs a new version alongside the old one without serving its responses to users, primarily for testing and not for live deployment with instant rollback.

D. Batch: This is a method for processing data for inference and is not a deployment strategy for online endpoints that facilitates rollback.

1. Microsoft. (2023). Safe rollout for online endpoints. Microsoft Learn. Retrieved from https://learn.microsoft.com/en-us/azure/machine-learning/how-to-safely-rollout-online-endpoints. (Under the "Blue-green deployment" section, it states: "If there's an issue, you can revert to the previous version... by switching traffic back to the blue deployment.")

2. Fowler, M. (2010). BlueGreenDeployment. MartinFowler.com. Retrieved from https://martinfowler.com/bliki/BlueGreenDeployment.html. (This is a foundational article from a reputable source in software engineering that defines the pattern and highlights the ability to "rapidly roll back if anything goes wrong.")

Azure Machine Learning integrates with MLflow, an open-source platform for managing the ML lifecycle. MLflow Tracking is a component of MLflow that is used to log and track training run metrics, parameters, artifacts, and models. When you submit a training job in Azure Machine Learning, it can automatically log this information, which can then be viewed and compared in the Azure Machine Learning studio, providing a comprehensive history and audit trail for all experiments.

Pipelines are used to define and automate machine learning workflows. While a pipeline run is an experiment that is tracked, the pipeline itself is the workflow definition.

Compute instances are cloud-based workstations for development. Experiments are run on them, but they are not the tracking feature itself.

Data labeling is a feature for annotating data, which is a step in data preparation, not experiment tracking.

1. Microsoft. (2023). Track ML experiments and models with MLflow. Microsoft Learn. Retrieved from https://learn.microsoft.com/en-us/azure/machine-learning/how-to-use-mlflow-cli-runs. (See the "Introduction" section, Paragraph 1, which states "Azure Machine Learning uses MLflow Tracking for logging and tracking experiments.").

2. Microsoft. (2023). Log & view metrics and log files. Microsoft Learn. Retrieved from https://learn.microsoft.com/en-us/azure/machine-learning/how-to-log-view-metrics. (This page details how MLflow is the backend for logging).