AI-300 - Microsoft Certified Machine Learning Operations Engineer Associate Cheat Sheet

Updated 2026-08-07

Next Topic: AI-901 - Microsoft Azure AI Fundamentals Cheat Sheet

🎯Take a practice test on this topic5 practice tests · 199 questions→

🧠Study flashcards on this topic165 cards · spaced repetition→

AI-300 (Microsoft Certified: Machine Learning Operations Engineer Associate) validates your ability to design and run MLOps and GenAIOps infrastructure on Azure, together branded AIOps, covering everything from provisioning an Azure Machine Learning workspace with Bicep and GitHub Actions to deploying, evaluating, and optimizing generative AI agents built on Microsoft Foundry. The exam spans five domains weighted toward the model lifecycle (25-30%, training through production monitoring) and GenAIOps infrastructure (20-25%), with lighter but still-tested coverage of generative AI quality assurance and observability plus RAG and fine-tuning optimization (10-15% each). Infrastructure as code and CI/CD are the connective tissue across every domain - a candidate comfortable deploying both Machine Learning workspaces and Foundry resources with Bicep, Azure CLI, and GitHub Actions can carry that same automation mindset from workspace setup through model retraining triggers. Expect scenario questions that test which Azure or Foundry feature fits a given operational requirement, such as a real-time versus batch endpoint, a provisioned-throughput versus serverless deployment, or MLflow tracking versus a custom logger, rather than pure recall.

What This Cheat Sheet Covers

This topic spans 15 focused tables and 167 indexed concepts, 165 flashcards, 5 practice tests with 199 questions. Below is a complete table-by-table outline of this topic, spanning foundational concepts through advanced details.

Table 1: Provisioning and Securing Machine Learning WorkspacesTable 2: Managing Data Assets, Environments, and Components in Machine LearningTable 3: Implementing Infrastructure as Code for Machine LearningTable 4: Tracking Experiments and Exploring Models with MLflow and AutoMLTable 5: Running and Scaling Model Training Jobs and PipelinesTable 6: Registering and Governing Machine Learning ModelsTable 7: Deploying Machine Learning Models to Production EndpointsTable 8: Monitoring Machine Learning Models in ProductionTable 9: Configuring Foundry Environments, Identity, and Network SecurityTable 10: Deploying and Managing Foundation Models for ProductionTable 11: Versioning and Managing Prompts with Source ControlTable 12: Evaluating Generative AI Quality, Safety, and RiskTable 13: Observing Generative AI Applications and Agents in ProductionTable 14: Optimizing RAG Retrieval and Embedding PerformanceTable 15: Fine-Tuning and Customizing Generative AI Models

Quick IndexSubscribe to unlock

A jump-to index of every table row in this cheat sheet.

Mind MapSubscribe to unlock

An interactive map of every table and concept in this topic.

Table 1: Provisioning and Securing Machine Learning Workspaces

Domain 1, Task "Create and manage resources in a Machine Learning workspace": the top-level workspace resource, its datastores, its compute targets, and the identity/access control model that secures all of them.

Concept	Example	Description
Machine Learning Workspace	`ws = Workspace(name="mlw-basic-prod",` `location="eastus",` `display_name="Basic workspace")` `ml_client.workspaces.begin_create(ws)`	• Top-level resource that groups jobs, pipelines, data assets, models, and endpoints for a team. • Create one workspace per project for clean cost reporting, not one shared workspace for everything. • Not to be confused with a resource group, which just holds Azure resources without any ML-specific organization.
Workspace-Associated Resources	Auto-created on workspace creation: `Storage Account • Container Registry` `• Key Vault • App Insights`	• Four Azure resources Azure Machine Learning provisions by default: a Storage account (artifacts, uploads), Container Registry (custom Docker images), Key Vault (secrets for compute targets), Application Insights (endpoint monitoring). • You can bring your own existing resources instead.
Hub Workspace	`Project(name="myexampleproject",` `hub_id=created_hub.id)`	• Groups multiple project workspaces under shared security settings, connections, and compute for centralized governance. • Same resource type as a Microsoft Foundry hub. • Not a billing consolidation tool; cost reporting still tracks at the project workspace.
Workspace Deletion	`ml_client.workspaces.begin_delete(` `name=ws.name,` `delete_dependent_resources=True)`	• Default delete is a soft delete of the workspace object only; it does not remove the linked Storage account, Key Vault, ACR, or App Insights. • Pass `delete_dependent_resources=True` to also remove those.
Datastore	`azureml://datastores/<name>/paths/` `<folder>/<file>.parquet`	• A secure reference to an existing Azure storage account (Blob, Files, ADLS Gen1/Gen2), not a copy of the data. • Creating a datastore never creates or moves the underlying storage.
Credential-Based vs Identity-Based Access	Credential-based • SAS, account key, service principal Identity-based • Microsoft Entra ID or managed identity	• Two ways a datastore authenticates to storage. Identity-based is Microsoft's preferred pattern: no secret is stored, so a workspace Reader can't read out a key. • Credential-based secrets ARE visible to anyone with Reader access on the workspace.
Default Datastores	`workspaceblobstore` • `workspaceartifactstore` `workspaceworkingdirectory` • `workspacefilestore`	Every workspace auto-creates four datastores over its default storage account: `workspaceblobstore` (uploads, job snapshots, pipeline cache), `workspaceartifactstore` (metrics, models, components), `workspaceworkingdirectory` (notebooks, compute instance, prompt flow files), `workspacefilestore` (alternate upload container).
Compute Instance	`ComputeInstance(name="ci-basic",` `size="STANDARD_DS3_v2")`	• A fully managed, single-owner cloud workstation for interactive development (notebooks, JupyterLab, VS Code). • Single node only, billed while running, so enable idle shutdown to avoid paying for an inactive instance. • Not to be confused with a compute cluster.
Compute Cluster	`AmlCompute(name="cpu-cluster",` `min_instances=0, max_instances=2,` `idle_time_before_scale_down=120)`	• A managed, autoscaling single- or multi-node cluster for training and batch inference jobs, shared across a workspace's users. • Set `min_instances=0` to scale down to zero nodes between jobs, but idle nodes still count against subscription quota.
Serverless Compute	`job = command(command="echo hi",` `environment="azureml://...sklearn-1.5")` `# omit 'compute' to use serverless`	• A training compute target that needs no cluster to create, size, or manage; Azure Machine Learning handles the full lifecycle. • Omitting the `compute` parameter on a job submission uses serverless compute automatically.
Attached Compute	Attach an existing • Azure Databricks cluster • Azure HDInsight • remote VM	• An unmanaged compute resource created and maintained outside Azure Machine Learning, then connected to the workspace. • The workspace does not create, scale, or patch it; you're responsible for its lifecycle.
Managed Identity (System-Assigned vs User-Assigned)	`az ml workspace show --query identity` `# "type": "SystemAssigned,UserAssigned"`	• The workspace identity it uses to talk to its Storage account, Key Vault, and ACR. Microsoft's recommendation for that scenario is the default system-assigned identity (SAI); user-assigned (UAI) is the alternative for an independent lifecycle across multiple resources. • A workspace can be upgraded from SAI to SAI+UAI, but not reverted back to SAI-only.
Azure RBAC Default Roles	`AzureML Data Scientist` `+ AzureML Compute Operator`	• Built-in workspace roles: AzureML Data Scientist (all actions except creating/deleting compute or changing the workspace), AzureML Compute Operator (create/manage/delete compute only), plus the general Reader / Contributor / Owner roles. • Combine Data Scientist + Compute Operator for self-service compute creation without full Contributor access.
Storage Blob Data Contributor Role	Grant `Storage Blob Data Contributor` to the workspace managed identity on the storage account	• The Azure RBAC data-plane role a workspace's managed identity needs to read and write blob data when the default storage account is configured for identity-based access. • Missing this role assignment is a common cause of "access denied" errors on an otherwise correctly identity-configured workspace.

Table 2: Managing Data Assets, Environments, and Components in Machine Learning

AI-300 Domain 1, Task "Create and manage assets in a Machine Learning workspace": versioned data assets and the storage references they sit on, curated versus custom compute environments and how Azure Machine Learning caches their Docker images, reusable pipeline components, and registries for sharing assets across workspaces.

Concept	Example	Description
Data Asset	`az ml data create --path ./data --name titanic --version 1 --type uri_folder`	A versioned, friendly-named reference to specific data; the file itself stays in its original storage location. • Types: uri_file, uri_folder, mltable • Not required for ad hoc access, a Datastore URI works without one

The rest of this section is part of the full cheat sheet

Table 3: Implementing Infrastructure as Code for Machine Learning

Domain 1, Task "Implement IaC for Machine Learning": authenticating GitHub Actions to Azure, deploying workspaces with Bicep and the Azure CLI, automating provisioning through workflows, restricting workspace network access, and managing source control with Git.

Concept	Example	Description
Workload Identity Federation (OIDC)	`permissions: {id-token: write}` on the job, then `azure/login@v2` using `AZURE_CLIENT_ID`, `AZURE_TENANT_ID`, `AZURE_SUBSCRIPTION_ID` secrets, no client secret stored	• Microsoft's recommended way for a GitHub Actions workflow to authenticate to Azure without a stored secret. • Not to be confused with the older service principal `--json-auth` secret pattern, which still works but is discouraged.

The rest of this section is part of the full cheat sheet

Table 4: Tracking Experiments and Exploring Models with MLflow and AutoML

AI-300 Domain 2, Orchestrate model training: configuring MLflow experiment tracking, using notebooks and compute instances for interactive exploration, letting Automated ML search across algorithms for the strongest model, and comparing model performance across jobs in the studio and through the MLflow SDK.

Concept	Example	Description
MLflow Tracking URI (Auto-Configuration)	`import mlflow` `mlflow.set_tracking_uri(mlflow_tracking_uri)` only needed outside Azure compute	• On an Azure ML notebook, compute instance, or compute cluster, MLflow already knows the workspace, so `set_tracking_uri()` is skipped. • Working locally, on Azure Databricks, or on Azure Synapse Analytics still needs it. • Not to be confused with hosting an MLflow server: Azure ML uses the MLflow APIs directly, there's no separate server to stand up.

The rest of this section is part of the full cheat sheet

Table 1: Provisioning and Securing Machine Learning Workspaces

Concept	Example	Description
Machine Learning Workspace	`ws = Workspace(name="mlw-basic-prod",` `location="eastus",` `display_name="Basic workspace")` `ml_client.workspaces.begin_create(ws)`	• Top-level resource that groups jobs, pipelines, data assets, models, and endpoints for a team. • Create one workspace per project for clean cost reporting, not one shared workspace for everything. • Not to be confused with a resource group, which just holds Azure resources without any ML-specific organization.
Workspace-Associated Resources	Auto-created on workspace creation: `Storage Account • Container Registry` `• Key Vault • App Insights`	• Four Azure resources Azure Machine Learning provisions by default: a Storage account (artifacts, uploads), Container Registry (custom Docker images), Key Vault (secrets for compute targets), Application Insights (endpoint monitoring). • You can bring your own existing resources instead.
Hub Workspace	`Project(name="myexampleproject",` `hub_id=created_hub.id)`	• Groups multiple project workspaces under shared security settings, connections, and compute for centralized governance. • Same resource type as a Microsoft Foundry hub. • Not a billing consolidation tool; cost reporting still tracks at the project workspace.
Workspace Deletion	`ml_client.workspaces.begin_delete(` `name=ws.name,` `delete_dependent_resources=True)`	• Default delete is a soft delete of the workspace object only; it does not remove the linked Storage account, Key Vault, ACR, or App Insights. • Pass `delete_dependent_resources=True` to also remove those.
Datastore	`azureml://datastores/<name>/paths/` `<folder>/<file>.parquet`	• A secure reference to an existing Azure storage account (Blob, Files, ADLS Gen1/Gen2), not a copy of the data. • Creating a datastore never creates or moves the underlying storage.
Credential-Based vs Identity-Based Access	Credential-based • SAS, account key, service principal Identity-based • Microsoft Entra ID or managed identity	• Two ways a datastore authenticates to storage. Identity-based is Microsoft's preferred pattern: no secret is stored, so a workspace Reader can't read out a key. • Credential-based secrets ARE visible to anyone with Reader access on the workspace.
Default Datastores	`workspaceblobstore` • `workspaceartifactstore` `workspaceworkingdirectory` • `workspacefilestore`	Every workspace auto-creates four datastores over its default storage account: `workspaceblobstore` (uploads, job snapshots, pipeline cache), `workspaceartifactstore` (metrics, models, components), `workspaceworkingdirectory` (notebooks, compute instance, prompt flow files), `workspacefilestore` (alternate upload container).
Compute Instance	`ComputeInstance(name="ci-basic",` `size="STANDARD_DS3_v2")`	• A fully managed, single-owner cloud workstation for interactive development (notebooks, JupyterLab, VS Code). • Single node only, billed while running, so enable idle shutdown to avoid paying for an inactive instance. • Not to be confused with a compute cluster.
Compute Cluster	`AmlCompute(name="cpu-cluster",` `min_instances=0, max_instances=2,` `idle_time_before_scale_down=120)`	• A managed, autoscaling single- or multi-node cluster for training and batch inference jobs, shared across a workspace's users. • Set `min_instances=0` to scale down to zero nodes between jobs, but idle nodes still count against subscription quota.
Serverless Compute	`job = command(command="echo hi",` `environment="azureml://...sklearn-1.5")` `# omit 'compute' to use serverless`	• A training compute target that needs no cluster to create, size, or manage; Azure Machine Learning handles the full lifecycle. • Omitting the `compute` parameter on a job submission uses serverless compute automatically.
Attached Compute	Attach an existing • Azure Databricks cluster • Azure HDInsight • remote VM	• An unmanaged compute resource created and maintained outside Azure Machine Learning, then connected to the workspace. • The workspace does not create, scale, or patch it; you're responsible for its lifecycle.
Managed Identity (System-Assigned vs User-Assigned)	`az ml workspace show --query identity` `# "type": "SystemAssigned,UserAssigned"`	• The workspace identity it uses to talk to its Storage account, Key Vault, and ACR. Microsoft's recommendation for that scenario is the default system-assigned identity (SAI); user-assigned (UAI) is the alternative for an independent lifecycle across multiple resources. • A workspace can be upgraded from SAI to SAI+UAI, but not reverted back to SAI-only.
Azure RBAC Default Roles	`AzureML Data Scientist` `+ AzureML Compute Operator`	• Built-in workspace roles: AzureML Data Scientist (all actions except creating/deleting compute or changing the workspace), AzureML Compute Operator (create/manage/delete compute only), plus the general Reader / Contributor / Owner roles. • Combine Data Scientist + Compute Operator for self-service compute creation without full Contributor access.
Storage Blob Data Contributor Role	Grant `Storage Blob Data Contributor` to the workspace managed identity on the storage account	• The Azure RBAC data-plane role a workspace's managed identity needs to read and write blob data when the default storage account is configured for identity-based access. • Missing this role assignment is a common cause of "access denied" errors on an otherwise correctly identity-configured workspace.

Table 2: Managing Data Assets, Environments, and Components in Machine Learning

Concept	Example	Description
Data Asset	`az ml data create --path ./data --name titanic --version 1 --type uri_folder`	A versioned, friendly-named reference to specific data; the file itself stays in its original storage location. • Types: uri_file, uri_folder, mltable • Not required for ad hoc access, a Datastore URI works without one

Table 3: Implementing Infrastructure as Code for Machine Learning

Concept	Example	Description
Workload Identity Federation (OIDC)	`permissions: {id-token: write}` on the job, then `azure/login@v2` using `AZURE_CLIENT_ID`, `AZURE_TENANT_ID`, `AZURE_SUBSCRIPTION_ID` secrets, no client secret stored	• Microsoft's recommended way for a GitHub Actions workflow to authenticate to Azure without a stored secret. • Not to be confused with the older service principal `--json-auth` secret pattern, which still works but is discouraged.

Table 4: Tracking Experiments and Exploring Models with MLflow and AutoML

Concept	Example	Description
MLflow Tracking URI (Auto-Configuration)	`import mlflow` `mlflow.set_tracking_uri(mlflow_tracking_uri)` only needed outside Azure compute	• On an Azure ML notebook, compute instance, or compute cluster, MLflow already knows the workspace, so `set_tracking_uri()` is skipped. • Working locally, on Azure Databricks, or on Azure Synapse Analytics still needs it. • Not to be confused with hosting an MLflow server: Azure ML uses the MLflow APIs directly, there's no separate server to stand up.