AZ MLO study plan
AZ MLO · Give each week a clear focus.
8 weeks · About 7 hours per week
Design and prepare a machine learning solution & Explore data and train models
Map each of the five exam domains to the specific Azure Machine Learning workspace assets it governs, writing a one-page crosswalk
Provision a workspace with compute, datastores, and a key vault, then record every configuration choice and its rationale
Configure role-based access for a data scientist, a compute operator, and an endpoint consumer, and verify each permission boundary
Build a reusable environment by pinning dependencies in a conda or requirements file and registering it as a versioned asset
Your checkpoint
A reproducible workspace setup script plus a one-page domain crosswalk committed to a personal study repo
Explore data and train models
Profile a tabular dataset for missing values, cardinality, and skew, and document the cleaning decisions taken
Engineer at least five candidate features and justify each against a stated modeling objective
Split data into train, validation, and test sets with a fixed seed and explain how leakage was prevented
Train three baseline models and compare them on a metric chosen to match the business objective
Your checkpoint
A notebook or job script that trains three baselines and prints a comparison table with the chosen metric
Explore data and train models & Manage and monitor models and experiments
Convert exploratory notebook code into a parameterized training script that accepts hyperparameters as arguments
Log parameters, metrics, and artifacts with the tracking client and confirm they appear in the experiment run history
Run a sweep over at least two hyperparameter dimensions and identify the best trial by a stated criterion
Register the winning model with tags, a description, and lineage back to its run
Your checkpoint
A completed sweep whose best run is registered as model version 1 with full lineage visible
Deploy and operationalize machine learning models & Manage and monitor models and experiments
Package the registered model as a reusable component with typed inputs and outputs
Author a pipeline that chains data preparation, training, and evaluation as separate steps
Deploy the model to a managed online endpoint and send a scored request from a client
Configure a batch endpoint and run a scoring job over a file-based input
Your checkpoint
One online endpoint and one batch endpoint, each demonstrated with a captured request and response
Deploy and operationalize machine learning models & Implement responsible machine learning practices
Implement a blue-green deployment by adding a second deployment to an online endpoint and shifting traffic percentages
Define a rollback procedure and rehearse it so the previous deployment is restored within minutes
Document the data collected by endpoint logging and state which fields are needed for drift and fairness checks
Identify personally identifiable information in the scoring payload and apply masking before logging
Your checkpoint
A written runbook covering traffic split, rollback steps, and PII handling, validated by one live rehearsal
Manage and monitor models and experiments & Implement responsible machine learning practices
Enable data collection on the production endpoint and confirm scored payloads land in storage
Configure a scheduled job that computes feature drift against the training baseline
Set an alert rule that fires when drift exceeds a stated threshold and test it with synthetic data
Draft a model card capturing intended use, limitations, and evaluation results by subgroup
Your checkpoint
A working drift monitor with a tested alert plus a first-draft model card for the registered model
Implement responsible machine learning practices
Evaluate model performance across at least two sensitive subgroups and quantify the gap in the primary metric
Apply a mitigation technique such as reweighting or threshold adjustment and re-measure the gap
Generate an interpretability report explaining global feature importance and at least three individual predictions
Record the findings in the model card and state the residual risk and monitoring commitment
Your checkpoint
An updated model card with subgroup metrics, a mitigation comparison, and interpretation exhibits
Design and prepare a machine learning solution & Deploy and operationalize machine learning models & Manage and monitor models and experiments
Rebuild the full solution from scratch in a clean workspace using only the committed scripts and assets
Time a full retrain-and-redeploy cycle and record each step and its duration
Resolve every flagged review item in the study notes by confirming the fact against primary documentation
Complete a timed practice set and log every missed item by domain for final review
Your checkpoint
A clean-room rebuild that ends in a redeployed endpoint, plus a scored practice log with gaps annotated
Milestones
Week 2: Workspace is reproducible from script, and three baseline models are trained with a documented metric comparison
Week 4: A registered model is served by both an online and a batch endpoint, with lineage from data to deployment intact
Week 6: Drift monitoring, alerting, and a first-draft model card are all in place for the deployed model
Week 8: Clean-room rebuild succeeds end to end, and practice scores show no domain below a self-set pass threshold