When should I use AutoML vs custom notebooks?

AutoML accelerates baselines and standard tasks; custom notebooks are better for complex feature engineering, advanced methods or strict constraints.

How do I ship a model to production?

Register the model, create an Environment, define an Endpoint (online or batch) and use CI/CD for controlled releases with canary/blue‑green.

How do I monitor drift and quality?

Enable data/target drift detection, prediction logging, alerts on metrics (latency/accuracy), and scheduled retraining.

Azure Machine Learning

Studio, AutoML, registry, endpoints, pipelines and MLOps to bring AI to production.

Technology Cluster · Back to Azure AI Technologies · AI for Business

Key capabilities

Azure ML Studio & Designer

No‑code/low‑code environment and visual authoring for reproducible experiments.

AutoML

Automatic model/hyperparameter search; fast baselines and algorithm selection.

Registry & Feature store

Versioning of models, datasets and features for reuse and traceability.

Pipelines & orchestration

Modular steps (preprocess, train, validate, deploy) orchestrated on‑prem/cloud.

Deployment & Endpoints

Online (real‑time) or batch services on CPU/GPU, autoscaling and revisions.

MLOps & governance

CI/CD, quality gates, approvals, lineage, security and compliance.

Responsible AI & monitoring

Explainability, fairness, drift detection and metric/cost alerting.

Reference integration

Typical workflow

# 1) Register dataset
# 2) Experiment/AutoML
# 3) Register model
# 4) Create Environment (conda/docker)
# 5) Endpoint (online/batch) & test
# 6) CI/CD with approvals

Architecture

Data Lake + Feature Store → Training on Compute/AKS → Registry → Endpoint (AKS/ACI/Batch) with monitoring.

Quick comparison

Component	When to use	Output
Studio/Designer	No‑code, PoCs, business teams	Repeatable pipelines/experiments
AutoML	Fast baseline & standard tasks	Best model + metrics
Registry	Reuse and traceability	Model/dataset versions
Endpoints	Serve predictions	REST/Batch + scalability
MLOps	Production at scale	CI/CD, policies, audit

Best practices

Data & features

Version datasets/features, define schemas, validate quality and manage drift.

Scalability

Use profiling, autoscaling and request batching; separate training/serving resources.

Security

Managed identities, private networks, key‑vault, least‑privilege access policies.

Next steps

Prepare for DP‑100 Continue with AI‑102 Explore business use cases

FAQ

Online vs batch endpoints?

Online for low latency; batch for high volumes and overnight processing. They often coexist.

How to control training costs?

Spot/low‑priority, early stopping, reduced hyperparameter grids and profiling; archive non‑essential runs.

How to meet compliance?

Data locality, encryption at rest/in transit, access logging, approvals and explainability for critical models.