9. Glossary and Cheatsheet

Glossary

Term	One line
Workspace	The regional ARM resource that owns everything. Check its kind before assuming what it is
kind	Default (Azure ML), Hub / Project (Foundry classic), FeatureStore. Same resource type, different products
Dependent resources	The storage account, Key Vault, Application Insights, and container registry beside every workspace. They survive workspace deletion
Datastore	A saved, named connection to a storage service with its auth attached
workspaceblobstore	The default datastore on the workspace's own storage account. Job outputs and snapshots land here
Data asset	A named, versioned pointer to data — uri_file, uri_folder, or mltable. Not a copy, not a snapshot
MLTable	A spec file describing how to load a folder of data — schema, delimiters, partitions
Environment	A versioned image + dependency spec, materialised and cached in ACR
Curated environment	Microsoft-maintained, prebuilt, in the shared azureml registry. Fast; you don't control what's in it
Job	One tracked execution. Types: command, sweep, pipeline, automl, spark
Experiment	A name you group jobs under. Nothing more than that
Code snapshot	The upload of the job's code directory at submission. A real copy, unlike a data asset. Governed by .amlignore
Component	A reusable, versioned pipeline step with a typed input/output interface
Pipeline job	A DAG of components, with step-level caching and per-step compute
Sweep	A hyperparameter search over a command job, with an early-termination policy
AutoML	Azure ML searching models and hyperparameters for you
Model	A versioned artifact: custom_model, mlflow_model (prefer this), or triton_model
Registry	A separate ARM resource sharing models, environments, and components across workspaces and regions. The promotion mechanism
Compute instance	A single-user managed dev VM. Billed per hour while running. Set idle shutdown
Compute cluster / AmlCompute	An autoscaling job pool. Set min_instances = 0
Serverless compute	Nodes Azure ML manages; you specify size and count on the job
Kubernetes compute	Your AKS or Arc cluster with the Azure ML extension, attached to the workspace
Spot / low-priority	Cheap, evictable nodes. Separate quota pool. Right for sweeps, wrong for one long run
Online endpoint	A stable HTTPS address with an auth mode. Managed or Kubernetes
Deployment	Model + environment + scoring script + instance type + count, behind an endpoint
Traffic split	Percentages across an endpoint's deployments. Your canary and your rollback
Batch endpoint	Submit a job over a dataset; runs on a cluster, writes outputs. Pays only while running
Scoring script	init() once per instance (load the model here), run() per request
init() / run()	Loading the model in run() is the classic latency bug
Managed VNet	Azure-run network isolation for compute. Disabled / AllowInternetOutbound / AllowOnlyApprovedOutbound
AllowOnlyApprovedOutbound	Egress only to declared FQDNs and private endpoints. Provisions a billed managed firewall
Model data collection	A per-deployment switch writing request/response payloads to storage, for drift analysis
Data drift	Production input distribution has moved away from the training baseline. The failure infra telemetry won't show you
AzureML Data Scientist	Data-plane role: run jobs, manage assets. Cannot create compute
AzureML Compute Operator	Control-plane role: start/stop/manage compute without full Contributor
Soft delete	Workspaces (and Key Vaults) go to a recycle bin; the name stays reserved
Purge protection	Key Vault: a one-way door. Once on, a deleted vault cannot be purged early
azapi	The Terraform provider that talks raw ARM, for resource types azurerm hasn't modelled — here, online deployments and registries

Commands you will actually type

Setup

az extension add -n ml                    # or: az extension update -n ml
az configure --defaults group=<rg> workspace=<ws>

Quota — check before anything else

az ml compute list-usage --location eastus -o table

Workspace

az ml workspace create -n <ws> -g <rg> -l <loc>
az ml workspace show   -n <ws> -g <rg>
az ml workspace list-deleted -o table
az ml workspace purge  -n <ws> -g <rg> -l <loc>     # frees a soft-deleted name

Compute

az ml compute create -n cpu-cluster --type AmlCompute \
  --size Standard_DS3_v2 --min-instances 0 --max-instances 4 \
  --idle-time-before-scale-down 300

az ml compute list -o table
az ml compute update -n cpu-cluster --max-instances 8
az ml compute stop  -n <instance-name>              # compute INSTANCES only
az ml compute start -n <instance-name>
az ml compute delete -n cpu-cluster --yes

Jobs

az ml job create -f job.yml --web
az ml job create -f job.yml --set inputs.n_estimators=500     # override without editing
az ml job list -o table
az ml job stream -n <job-name>
az ml job cancel -n <job-name>
az ml job download -n <job-name> --output-name model --download-path ./out

Assets

az ml data create        -f data.yml
az ml environment create -f environment.yml
az ml model create -n iris-rf --version 1 --type mlflow_model \
  --path azureml://jobs/<job>/outputs/artifacts/model
az ml model list -o table
az ml model archive -n iris-rf --version 1        # soft-hide, not delete

Registries — promotion

az ml model create --registry-name reg-mlops -f model.yml
az ml model share  -n fraud-rf --version 12 \
  --registry-name reg-mlops --share-with-name fraud-rf --share-with-version 12

⚠️ Registry command surface has changed across releases — verify against az ml model --help.

Endpoints and deployments — the blue/green loop

az ml online-endpoint create -n <ep> --auth-mode AADToken

az ml online-deployment create -f deployment-green.yml       # created at 0% traffic
az ml online-endpoint invoke -n <ep> --deployment-name green --request-file sample.json
az ml online-endpoint update -n <ep> --traffic "blue=90 green=10"
az ml online-endpoint show   -n <ep> --query traffic

# ROLLBACK
az ml online-endpoint update -n <ep> --traffic "blue=100 green=0"

az ml online-deployment get-logs --endpoint-name <ep> --name green --lines 200
az ml online-deployment delete  --endpoint-name <ep> --name green --yes
az ml online-endpoint delete -n <ep> --yes

Batch

az ml batch-endpoint create -n <bep>
az ml batch-deployment create -f batch-deployment.yml --set-default
az ml batch-endpoint invoke -n <bep> --input azureml:my-data:3

Teardown

az group delete -n <rg> --yes --no-wait
az ml workspace list-deleted -o table     # confirm nothing lingers

Resource ID and URI shapes

# Workspace
/subscriptions/{sub}/resourceGroups/{rg}/providers/Microsoft.MachineLearningServices/workspaces/{ws}

# Compute
.../workspaces/{ws}/computes/{name}

# Online endpoint and deployment
.../workspaces/{ws}/onlineEndpoints/{ep}
.../workspaces/{ws}/onlineEndpoints/{ep}/deployments/{name}

# Registry (a separate top-level resource, not under a workspace)
/subscriptions/{sub}/resourceGroups/{rg}/providers/Microsoft.MachineLearningServices/registries/{reg}

Asset references inside YAML and the SDK use the azureml: scheme:

azureml:my-model:7                          # workspace asset, pinned version
azureml:my-model@latest                     # workspace asset, latest — never in prod
azureml://registries/{reg}/models/{name}/versions/{n}
azureml://registries/azureml/environments/sklearn-1.5/labels/latest   # curated
azureml://jobs/{job}/outputs/artifacts/model
azureml://datastores/{ds}/paths/{path}

Endpoints:

Workspace data plane : https://{region}.api.azureml.ms
Scoring              : https://{endpoint}.{region}.inference.ml.azure.com/score
Studio               : https://ml.azure.com

MLflow in three lines

import mlflow
mlflow.set_tracking_uri(ml_client.workspaces.get(ws).mlflow_tracking_uri)
mlflow.sklearn.autolog()

Inside a job submitted to Azure ML, the tracking URI is already set — mlflow.log_metric(...) just works with no Azure imports.

Limits worth knowing — with their scope

A number without a scope is useless in Azure. All of these vary by subscription type and region; read the live values rather than trusting any table, including this one.

Limit	Counted per	How to check
Dedicated vCPU, per VM family	subscription × region × family	az ml compute list-usage, portal Usage + quotas
Low-priority (Spot) vCPU	subscription × region (separate pool)	same
Managed online endpoint quota	subscription × region (separate pool again)	portal Usage + quotas
Endpoints per workspace	workspace	Azure docs / portal
Deployments per endpoint	endpoint	Azure docs
Nodes per compute cluster	cluster (and bounded by family quota)	cluster config
Workspaces per resource group	resource group	Azure docs
Workspace soft-delete retention	workspace	az ml workspace list-deleted
Key Vault soft-delete retention	vault	vault properties

⚠️ Every figure above is a shape, not a value. Verify against current Azure docs and your own subscription before designing to any of them.

Decision cheatsheet

Question	Answer
Training my own model, or calling a foundation model?	Own → Azure ML. Calling → Foundry
Authoring or running?	Authoring → compute instance (with idle shutdown). Running → cluster or serverless
Cluster min_instances?	0. If you're arguing otherwise, measure first
Spot or dedicated?	Sweeps and checkpointed work → Spot. One long run your release depends on → dedicated
Real-time or bulk?	Sub-second per request → managed online endpoint. Large batches, latency-tolerant → batch endpoint (much cheaper)
Spiky traffic, cost-sensitive?	Not a managed online endpoint — there's no scale-to-zero. Batch endpoint, or a Container App
Key auth or Entra ID on the endpoint?	Entra ID (AADToken). Key auth is for the first hour
One deployment or two?	Two. blue and green. The second one is your rollback
How do I roll back?	az ml online-endpoint update --traffic "blue=100 green=0"
Promoting a model between environments?	A registry. Not a file copy
@latest in a spec?	Fine in dev, never in prod
Terraform or az ml?	ARM resources → Terraform. Assets, models, deployments, traffic → az ml
Where's my money going?	A running compute instance, a cluster floor, or an idle online deployment. In that order

---

Next: back to the Azure Machine Learning overview, or on to the next topic in the article.

← Back to the Azure Machine Learning overview · ← Previous: Interview Questions