AI-300 — frequently asked questions

URI_FOLDER indicates a directory containing multiple files, suitable for datasets like image classification folders. FILE indicates a single specific file location.

SDK v2 uses 'entities' for standardized definitions of inputs, outputs, and assets. 'data' is associated with the older SDK v1 Dataset abstraction.

The dashboard focuses on interpretability and fairness. Causal analysis is a separate, more complex statistical field not included in the standard dashboard components.

Explanation ranks feature importance based on current data, while Counterfactuals simulates hypothetical changes to inputs to predict different outcomes.

Models are outputs, not workflows. Components define the reusable process logic that produces those outputs.

Environments define software dependencies, while Components wrap code, environments, and I/O into a reusable, versioned task.

FAILED means the run ran into an error during execution. KILLED means the run was stopped manually or by the system before completion.

No. Cancelled or aborted runs have the status KILLED. Only runs that errored out have the status FAILED.

A Data Asset registers the raw local files within the Foundry workspace. The Vector Index operates on these registered assets to chunk and embed them; without the asset, there is no source data to index.

No. The Vector Index is a component that resides within an Azure AI Search resource (the Search Resource). You must provision the search capability first before defining how data is indexed within it.

Logs provide the actual retrieved context needed by evaluators to score relevance and groundedness. Without logs, there is no data to measure.

No. Vector embeddings are generated based on chunks. Changing the chunking strategy requires rebuilding the index so the new segments are embedded correctly.

User feedback is unstructured text or ratings that vary by individual interpretation. Measurable evaluation requires consistent, algorithmic scoring of attributes like Groundedness against a ground truth dataset.

Use automated evaluation metrics such as Groundedness (support from source), Relevance (alignment with query), and Faithfulness to quantitatively assess RAG performance changes.

Factory methods create new definitions from paths. Since the asset is already registered, you must load the existing asset reference instead.

Yes, mltable.load() works with registered assets regardless of the underlying file format, as long as the asset is properly defined in the workspace.

wasbs is used for Azure Blob Storage containers without hierarchical namespaces. abfss is used for Azure Data Lake Storage Gen 2, which supports hierarchical namespaces and file-system-like operations.

No, azureml:// is used for data assets stored within the Azure ML workspace's managed storage. External public or private containers require wasbs:// or https:// depending on authentication.

az ml model create registers a model artifact in the workspace for versioning. az ml online-endpoint create provisions the infrastructure required to serve predictions from that model.

Generally, no. You must register the model artifact in the workspace before you can reference it when creating a deployment or endpoint.

Compute Clusters are ephemeral and do not provide a persistent, pre-configured development environment like Jupyter or VS Code. They are designed for batch job execution, not interactive coding sessions.

Yes, but it is not optimal for scalable training. Instances are single-node resources and cannot auto-scale. For large jobs, a Compute Cluster is required to distribute the load across multiple nodes efficiently.

SSH is preferred in cloud environments like Azure ML because it allows for key-based authentication without storing passwords in scripts or environment variables, enhancing security.

No, unless you have previously configured and saved an SSH key pair that is already trusted by GitHub. For a fresh workspace, you must generate and register a new key.

No. Azure ML Environments are immutable. You cannot modify the base image of an existing environment record. You must create a new version of the environment with the updated image.

Tags are user-defined labels for organization and filtering (metadata). The Docker image reference determines the actual software stack run during training. Changing a tag does not change the image.

A Hub workspace provides centralized management for shared resources, policies, and identity across the organization. Project workspaces are isolated environments linked to the Hub, allowing teams to collaborate on specific models without affecting other teams.

Resource Groups provide Azure-level billing and access isolation, but they do not offer the ML-specific governance features like shared datasets, compute targets, and unified policy enforcement that Hub and Projects provide.

Yes, but the question specifically asks about automating evaluations for pull requests, which is a CI task typically handled by GitHub Actions.

A blocklist is a static filter for specific terms, while an evaluation uses models to assess safety and quality dynamically against defined thresholds.

An Environment packages the software dependencies (Python packages, Docker image) needed to run code. A Job Configuration defines how the job runs (compute target, script path, parameters) but does not contain the software itself.

This is often caused by inconsistent runtime dependencies. Using an Azure ML Environment ensures that every run uses the exact same versions of libraries and system tools, leading to reproducible results.

While possible, isolating dependencies in a new Conda environment prevents version conflicts with core Azure ML libraries and ensures reproducibility for your specific model workflows.

If you skip activation, conda install will default to the base environment. Consequently, your new kernel will either fail to launch due to missing dependencies or lack the specific packages defined in requirements.txt.

Managed identities provide automatic, rotation-free authentication via Azure AD without storing secrets in code or variables, significantly reducing the attack surface.

Yes, Azure ML workspaces can use managed identities to authenticate to Azure Repos and GitHub Enterprise Server, depending on the specific integration configuration.

In the URL format https://<account>.blob.core.windows.net/<container>, the container name is the segment directly following the domain suffix. For example, in https://storage1.blob.core.windows.net/data1, the container is 'data1'.

AccountKeyConfiguration uses the primary or secondary storage account key, which grants full, permanent access to the storage account. It does not support expiring tokens, making it unsuitable for scenarios requiring limited-duration access.

load_component parses a local YAML file into a Python object, while create_or_update sends that object to the Azure ML service to register it.

Yes, you can define components programmatically using Python classes, but this question specifically requires handling a local YAML definition.

SSH keys allow for passwordless, automated authentication which is faster and more secure than entering credentials repeatedly via HTTPS, especially important for CI/CD pipelines.

No, you need a compute resource like a Compute Instance or Compute Cluster to host the environment where the Git repository will be cloned and managed.

Yes, the 'count' row in the output of .describe() represents the number of non-null (non-empty) values for each column.

You can select only numerical columns using df.select_dtypes(include=['number']) before calling .describe() to exclude categorical data.

By default, .describe() provides the 25th, 50th (median), and 75th percentiles along with count, mean, std, min, and max.

Data drift monitors changes in the input features compared to the training baseline, while prediction drift monitors changes in the output scores. Both can impact model performance.

Yes, by configuring alerts on drift metrics and linking them to an Azure Machine Learning pipeline that triggers automatic model retraining when thresholds are exceeded.

Pre-processing modifies input data before training to remove bias correlations, while post-processing adjusts model outputs (like thresholds) after training to satisfy fairness constraints without changing the model itself.

ExponentiatedGradient is a reduction method that wraps the base estimator and modifies the training loop, effectively requiring retraining or iterative updates to the model based on sample weights.

Intermittent 5xx errors are often caused by resource exhaustion when traffic spikes exceed the current instance count. Adjusting autoscaling settings helps the service handle load more effectively.

Unexpected responses usually result from scoring script errors or data format mismatches. Reviewing deployment logs provides detailed stack traces and error messages to pinpoint the exact failure point in your code.

Feature count is a static attribute of the model's input schema. It does not change during inference and does not reflect the endpoint's processing speed or resource utilization.

Connections active measures concurrent sessions at any given moment, indicating immediate load. Requests per minute measures the rate of incoming requests over time, indicating overall throughput trends.

No, autoscaling only adjusts compute resources. Retraining is triggered separately via alert rules configured for specific data or model drift thresholds.

Monitoring metrics will be collected but no automated notifications or downstream integrations will occur when thresholds are exceeded.

The MLflow Tracking URI directs experiment runs to be logged to the Azure ML workspace, enabling centralized storage and comparison of run histories.

Use MLflow metrics logging functions like mlflow.log_metric to explicitly record performance values, which then appear in the run history for analysis.

Global error rates mask disparities within specific subgroups. A model can have high overall accuracy but perform poorly for protected groups, violating fairness principles.

Feature importance reveals which input variables most significantly influence predictions, aiding in transparency and helping detect if protected attributes or their proxies are driving decisions.

Yes. You can update the traffic allocation weights to direct 100% of traffic to the previous stable deployment, effectively rolling back without deleting the endpoint.

The failed deployment remains in the endpoint but receives no traffic. You can investigate it later or delete it once resolved.

No. Calculating a value like accuracy only stores it in a local variable. You must explicitly call mlflow.log_metric() to send that data to the experiment tracking service.

mlflow.set_experiment defines the container (the experiment name) where data will go. mlflow.sklearn.log_model saves the actual model binary/artifact into the current run's history.

No. While MLflow tracks runs and artifacts, you must explicitly register the model in the Azure ML workspace to create a versioned entry in the registry.

Tagging an experiment adds metadata to the run for easier searching and filtering. Registering a model creates a persistent, versioned object in the workspace that can be deployed and reused.

No. The question requires specific configuration to ensure 'significant drops' are caught. Default values may not align with your business requirements for accuracy.

You need to identify which metrics are available and relevant from production traffic before you can set meaningful thresholds for them.

Archiving keeps the model registered in the workspace but marks it as non-deployable, retaining all artifacts for audit. Unregistering completely removes the model from the workspace registry.

No, once a model version is archived, it cannot be deployed. To deploy it again, you would need to restore it, if supported, or register a new version from the same artifacts.

Traffic splitting allows you to deploy multiple model versions to the same endpoint and assign weight percentages to each. This enables canary testing by directing a small percentage of live traffic to the new version.

A staging endpoint provides isolation but does not test the model against the actual production load or user behavior in a mixed environment. Traffic splitting allows for controlled validation within the production infrastructure.

Regression is defined as a decrease in performance on any evaluated metric compared to the baseline. Even if the primary goal metric improves, a drop in another metric constitutes regression.

If some metrics improve and others degrade, the model has not universally improved. You must answer 'No' for statements claiming universal improvement or lack of regression if any metric got worse.

Online endpoints are for real-time, low-latency inference, typically used for APIs and chat applications. Batch endpoints are for asynchronous, high-volume processing where results are returned later.

Traffic splitting allows you to direct a portion of requests to a new model version while keeping the old one active. If the new version fails, you can shift 100% of traffic back to the old version instantly.