Implement machine learning model lifecycle and operations
Free Microsoft MLOps Engineer Associate (AI-300) practice — 6 questions on Implement machine learning model lifecycle and operations, with explanations. No sign-up. Full 12-question mixed test →
A data science team uses Azure Machine Learning to train multiple models with MLflow tracking. They need to package a model for deployment that automatically retrieves features from an Azure Machine Learning feature store at inference time, ensuring the exact feature engineering logic used during training is applied. The feature retrieval specification must be versioned with the model artifact. Which approach should they implement?
Azure Machine Learning supports packaging feature retrieval specifications directly with MLflow models using the FeaturesMetadata class and set_model_features() method. This approach ensures the feature store references, transformations, and versions are immutably bound to the model artifact, guaranteeing consistent feature engineering between training and inference without manual coordination.
An MLOps team runs distributed training for a large transformer model across 8 GPU nodes in an Azure Machine Learning compute cluster. Training frequently stalls when one node fails mid-training. They need a fault-tolerant distributed training strategy that automatically handles node failures by redistributing work to remaining nodes without restarting the entire job from scratch. Which solution should they implement?
PyTorch Elastic Training (torch.distributed.elastic) with Azure ML's elastic training support enables true fault tolerance by dynamically adjusting the training group size when nodes fail. The etcd-v2 rendezvous backend maintains coordination state, and configuring min_nodes/max_nodes in ScaleSettings allows the job to continue with fewer nodes without full restart. This is specifically designed for the scenario described.
A financial services company deploys a fraud detection model to an Azure Machine Learning managed online endpoint with two deployments: v1 (90% traffic) and v2 (10% traffic). After 24 hours, v2 shows 15% higher latency than v1 but 8% better accuracy. The business requires <200ms p95 latency. Monitoring shows v2's latency is 245ms p95 due to additional feature engineering. They need to safely roll back v2 while preserving the ability to redeploy an optimized version later. What should they do?
Setting traffic to 0% provides immediate rollback while preserving the deployment configuration and model registration for future optimization. Monitoring v1 for 48 hours ensures stability, and disabling (not deleting) v2 keeps the deployment definition intact for easy reactivation. This follows safe rollback best practices: preserve rollback capability, validate stability, and maintain deployment artifacts.
An MLOps engineer monitors a deployed loan approval model and detects that the feature 'credit_utilization_ratio' has shifted from a training mean of 0.42 (σ=0.18) to a current mean of 0.31 (σ=0.15) over 30 days. Model accuracy dropped from 89% to 84%. Azure Machine Learning model monitoring shows data drift detected with a drift coefficient of 0.23. The business requires maintaining >87% accuracy. Which action should the engineer take FIRST?
Before retraining, the engineer must determine if drift represents real distribution shift or data quality issues (pipeline bug, data collection change, incorrect feature computation). The correlation between feature drift and accuracy drop needs investigation—retraining on bad data perpetuates errors. This is the scientifically sound first step in drift response: diagnose before treating.
A machine learning team uses Azure Machine Learning pipelines to orchestrate training workflows. They need to compare the performance of 12 different model architectures trained across 4 hyperparameter configurations each (48 total training runs), then automatically select the best model based on F1-score for registration. The comparison must track all metrics, parameters, and artifacts for reproducibility. Which approach should they implement?
Using MLflow's parent-child run hierarchy provides structured experiment tracking with automatic parameter/metric/artifact logging. The MLflow search API enables programmatic querying and comparison across all runs with filtering and sorting by F1-score. This approach maintains full reproducibility, enables automated selection, and follows Azure ML's recommended pattern for comparing multiple training runs within orchestrated pipelines.
A retail company evaluates a new product recommendation model for potential bias before deployment. Using Azure Machine Learning's Responsible AI dashboard, they discover that the model shows a 12% accuracy disparity between two customer demographic groups for the sensitive feature 'age_group', with the 65+ segment having 76% accuracy versus 88% overall. The model uses 23 features including purchase_history, browsing_behavior, and age_group. Which remediation approach should they implement to address this fairness issue while maintaining model utility?
Fairlearn's GridSearch with EqualizedOdds constraint is designed specifically for this scenario—it trains multiple mitigated models that reduce disparity between demographic groups while optimizing the fairness-accuracy tradeoff. The approach directly addresses the 12% disparity using established fairness mitigation techniques, and validation through Responsible AI dashboard confirms both disparity reduction and acceptable accuracy impact.
Ready for the real thing?
The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.
$129.99 $34.99 with code FREETEST33 — valid through September 2.
Get my $34.99 deal →