AI-300 ML Model Lifecycle And Operations — 30 Questions

Question 1

You are orchestrating a multi-step ML pipeline in Azure Machine Learning. You need to ensure that a downstream step only executes if the upstream model training step finishes successfully, while allowing the pipeline to continue even if a non-critical logging step fails. Which configuration should you use?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Configure pipeline run settings with 'continue_on_step_failure' set to True for the logging step.. In Azure Machine Learning pipelines, the 'continue_on_step_failure' setting on a step (or at the pipeline run level) allows the pipeline to proceed even when that step fails. By setting it to True specifically for the non-critical logging step, the pipeline continues past a logging failure, while the default behavior (False) for the training step ensures downstream steps only run if training succeeds. This gives the exact conditional execution behavior required.

Answer

Configure a 'WaitStep' to monitor the training job status.

Answer

Set the 'PipelineParameter' to execute only on success.

Answer

Use an 'Estimator' class with 'allow_reuse' set to False.

Question 2

You are deploying a model to an Azure Kubernetes Service (AKS) cluster. You need to ensure that the deployment handles traffic spikes by scaling based on GPU usage. Which setting must be enabled in the inference configuration?

Accepted Answer

Set 'autoscale_enabled' to True and 'target_utilization' for GPU metrics.. To scale based on GPU usage, you must define an 'autoscale' configuration within the 'InferenceConfig' that references 'target_utilization' for custom metrics like GPU.

Answer

Use 'enable_gpu' in the 'DeploymentConfig'.

Answer

Configure 'max_concurrent_requests'.

Answer

Enable 'cluster_purpose' as 'FastProd'.

Question 3

You need to ensure that your Azure ML models are deployed with high availability across multiple regions. Which deployment strategy should you configure?

Accepted Answer

Managed Online Endpoints with multiple deployments. Managed Online Endpoints support multi-region traffic distribution by using managed deployments behind a single endpoint.

Answer

Kubernetes Online Endpoints

Answer

Batch Endpoints

Answer

Azure Container Instances (ACI)

Question 4

You are managing model versioning in Azure Machine Learning Registry. You need to promote a model from 'Staging' to 'Production' without creating a new asset version. Which command or action should you perform?

Accepted Answer

Update the model asset by adding a 'Production' tag via the SDK or CLI.. In Azure Machine Learning Registry, model versions are immutable assets; you cannot change a version's stage by editing the version itself. Instead, you promote a model by updating its tags — specifically adding or updating a 'Production' tag on the existing model version via the SDK or CLI. This preserves the version number and lineage while signaling the new stage.

Answer

Change the 'run_id' in the model metadata.

Answer

Delete the old version and re-register as 'Production'.

Answer

Run 'az ml model create' with a new version number.

Question 5

A team uses Azure Machine Learning to track experiments. You need to ensure that every run is associated with a specific git commit hash to ensure reproducibility. Where should this be configured?

Accepted Answer

Within the 'experiment.start_logging()' call using the 'tags' parameter.. The 'tags' parameter in the experiment logging call (e.g., via MLflow's start_run or Azure ML's experiment.start_logging) allows arbitrary key-value metadata to be attached to a run, including the git commit hash. This makes the commit hash queryable and visible in the Azure ML studio run details, ensuring reproducibility. Configuring it here ties the metadata directly to the run object rather than to workspace or environment configuration.

Answer

In the 'workspace.json' file.

Answer

By modifying the compute cluster configuration.

Answer

Inside the 'conda_dependencies.yml' file.

Question 6

You are troubleshooting a model deployment failure where the container fails to start due to missing environment variables. Where do you find the logs to identify the cause?

Accepted Answer

The 'get-logs' command for the online deployment.. In Azure Machine Learning, when a model deployment fails due to container startup issues such as missing environment variables, the deployment logs are retrieved using the 'get-logs' command for the online deployment. This command provides container logs that include startup errors and environment variable issues.

Answer

The Key Vault access logs.

Answer

The pipeline run history.

Answer

The workspace diagnostic logs in Azure Monitor.

Question 7

You want to enforce a policy that all models must be registered before being deployed. Which feature should you use to implement this constraint?

Accepted Answer

Azure Policy definitions.. Azure Policy definitions are the native governance mechanism in Azure for enforcing organizational rules such as 'all models must be registered before deployment.' A policy definition with a deny effect can block deployments that reference unregistered models, and it can be assigned at management group, subscription, or resource group scope. This provides the centralized, auditable enforcement that the scenario requires.

Answer

Virtual Network peering.

Answer

Role-Based Access Control (RBAC).

Answer

Compute Instance quotas.

Question 8

A model is exhibiting data drift. You have created a drift monitor. What is the next step to automate the retraining?

Accepted Answer

Configure an Event Grid subscription to trigger a pipeline.. After creating a monitor, you set up an 'Event Grid' subscription to notify an Azure Function or Logic App to trigger the pipeline.

Answer

Update the dataset version.

Answer

Manually restart the compute cluster.

Answer

Create a 'RetrainingTrigger' in the model registry.

Question 9

You are running a distributed training job using the 'PyTorch' framework on Azure Machine Learning. You need to configure the 'DistributionConfiguration'. Which setting is mandatory for multi-node training?

Accepted Answer

'process_count' or 'node_count' in the configuration.. For multi-node distributed training in Azure ML with PyTorch, the DistributionConfiguration must specify either process_count (processes per node) or node_count (number of nodes) so the job knows how to shard across the cluster. Without these, the job cannot coordinate multi-node execution. The framework field is set to 'pytorch', not TensorFlow, and shm_size/GPU flags are optional tuning parameters.

Answer

'framework' set to 'TensorFlow'.

Answer

'shm_size' set to 1GB.

Answer

'enable_gpu' set to False.

Question 10

You are configuring an 'Azure Machine Learning Compute Cluster' for a heavy training job. You want to ensure it shuts down automatically when no jobs are running. What setting should you configure?

Accepted Answer

'min_nodes' set to 0.. The 'idle_seconds_before_scaledown' property in the compute configuration determines how long the cluster waits before removing idle nodes.

Answer

'max_nodes' set to 0.

Answer

'cluster_priority' to 'LowPriority'.

Question 11

You have an Azure Machine Learning pipeline that uses 'PipelineData' to pass information between steps. You want to share data between a training step and a scoring step. What is the recommended way to persist this data?

Accepted Answer

Use an 'OutputDataBinding' to a registered Datastore.. 'PipelineData' allows intermediate data passing, but 'OutputDataBindings' are preferred for persisting artifacts that need to be accessed later, such as model files.

Answer

Embed the data in the run metadata.

Answer

Write the file to the local temp directory.

Answer

Use an environment variable to pass the file path.

Question 12

You are reviewing the 'Run History' in Azure Machine Learning. You want to compare the training time of two different experiments. Which UI feature should you use?

Accepted Answer

The 'Experiments' tab, then 'Compare' button.. The 'Charts' or 'Metrics' comparison view in the Azure ML Studio allows you to visualize and compare metrics across multiple runs.

Answer

The 'Models' registry list.

Answer

The 'Endpoints' dashboard.

Answer

The 'Notebooks' file browser.

Question 13

You are debugging an Azure ML pipeline. You want to see the stdout of a specific step that failed. How do you access this?

Accepted Answer

Run 'az ml job get-logs' with the job ID.. You can access the logs for each step by clicking on the 'Outputs + logs' tab of that specific step run in the Azure ML Studio interface.

Answer

Check the 'Experiment' dashboard settings.

Answer

Download the 'run.json' file.

Answer

Look in the 'Workspace' logs folder.

Question 14

You have an Azure ML environment that needs to access a private Azure SQL database. How should you configure the connection?

Accepted Answer

Use a Managed Identity for the compute cluster.. You should use a 'Datastore' that stores the credentials in Azure Key Vault to securely access the SQL database from the compute environment.

Answer

Enable public access for the SQL database.

Answer

Hardcode the connection string in the script.

Question 15

You are configuring a 'Managed Online Endpoint' with SSL termination. Where do you manage the SSL certificates?

Accepted Answer

Via the Azure front-end load balancer/gateway service.. The SSL certificates are managed at the Azure 'Front Door' or 'Application Gateway' level, or by using the built-in certificate management if using Azure-provided domains for endpoints.

Answer

In the 'Python' score script.

Answer

Inside the Endpoint configuration YAML.

Answer

In the 'workspace.json' file.

Question 16

You want to perform hyperparameter tuning using the 'HyperDrive' service. You have a requirement to stop poor-performing runs early to save compute costs. Which policy should you use?

Accepted Answer

'MedianStoppingPolicy'.. The 'BanditPolicy' is the standard early-termination policy that stops runs based on a slack factor/amount compared to the best-performing run.

Answer

'TruncationSelectionPolicy'.

Answer

'RandomParameterSampling'.

Question 17

You are configuring a 'Managed Online Endpoint' for a very large model (10GB+). The deployment is failing during the 'pulling image' phase. What is the most likely cause?

Accepted Answer

The 'readiness_probe' timeout is too short.. The 'pulling image' phase is when the managed online endpoint pulls the model's Docker image from Azure Container Registry to the compute instance. For a 10GB+ model, the pull can take many minutes. The default readiness_probe timeout is often too short, causing the deployment to fail before the image finishes pulling and the container becomes ready. Increasing the readiness_probe timeout (and possibly the startup timeout) resolves this.

Answer

The instance count is too high.

Answer

The ACR is private.

Answer

The compute is too small.

Question 18

You are monitoring model drift. Which TWO features are required to configure a Data Drift Monitor?

Accepted Answer

A target dataset to compare against.. In Azure Machine Learning, a Data Drift Monitor requires two datasets: a baseline dataset (option C), which represents the reference distribution the model was trained on, and a target dataset (option B), which is the newer production data compared against the baseline to compute drift metrics. The monitor calculates drift by comparing the feature distributions of the target dataset against the baseline dataset over time, so both are mandatory inputs. Option A is incorrect because an Azure SQL database is not required — datasets can reside in various sources such as Azure Blob Storage or Azure Data Lake, and the monitor does not depend on SQL. Option D is incorrect because, while a model may be associated with the monitor for organization, a model version is not a required configuration input for the drift monitor itself. Option E is incorrect because Data Drift Monitor uses built-in drift metrics (e.g., population stability index, Jensen-Shannon distance) and does not require a custom Python script.

Answer

An Azure SQL database.

Answer

A model version.

Answer

A custom Python script.

Question 19

You have an automated deployment pipeline. You want to run an integration test on the model after deployment. Which tool is best suited for this?

Accepted Answer

Azure DevOps pipelines.. 'Azure DevOps' (or GitHub Actions) is the industry-standard tool for orchestrating post-deployment testing as part of a CI/CD pipeline.

Answer

A manual 'Endpoint' health check.

Answer

The training script.

Answer

Azure Machine Learning Studio 'Tests' tab.

Question 20

Which THREE types of information are found in a Run Object in Azure ML?

Accepted Answer

Tags.. In Azure Machine Learning, a Run object is the record of a single execution of an experiment and it captures the metadata produced during that execution. Option A (Tags) is correct because runs support arbitrary key-value tags that you can add for labeling, filtering, and organizing runs in the studio and via the SDK. Option C (Metrics) is correct because logged metrics (for example, via run.log or mlflow.log_metric) are stored on the run and used for charting and comparing model performance. Option E (Parameters) is correct because run parameters (for example, via run.log or the ScriptRunConfig arguments) are recorded on the run to document the inputs and hyperparameters used. Option B (Billing history) is not part of a Run object; cost and billing data are surfaced through Azure Cost Management and the subscription/billing APIs, not the run record. Option D (Network topology map) is not stored in a Run object; network configuration belongs to the workspace, compute, and virtual network resources rather than to an individual run.

Answer

Billing history.

Answer

Network topology map.