AWS ML Engineer Associate Course Notes

SageMaker Feature Store is chosen when features must be defined once, reused across models, and served consistently for both training and inference. The Online Store supports low-latency reads for real-time serving, while the Offline Store keeps historical feature records for training, backtesting, and analytics. If the scenario mentions training-serving consistency, shared feature definitions, or feature reuse across projects, Feature Store is usually the AWS service being tested.

• Using the Offline Store for live feature lookups when the scenario needs millisecond serving from the Online Store.

• Online Store means low-latency serving; Offline Store means training history and analytics.

SageMaker deployment questions usually hinge on how predictions arrive. Real-time endpoints are chosen for low-latency request and response traffic. Serverless inference fits intermittent traffic when the model can tolerate cold-start behavior, while provisioned concurrency is added when serverless endpoints need predictable startup latency. Batch transform is selected for offline scoring of stored datasets.

• Choosing a real-time endpoint for stored offline scoring when Batch Transform is the intended SageMaker inference mode.

• Real-time is immediate, async is delayed, batch transform is offline, and serverless is intermittent.

Automatic Model Tuning is chosen when the model algorithm and training data are known but the best hyperparameter values are not. A tuning job runs multiple training jobs across a search space and compares them using an objective metric such as validation accuracy, F1, RMSE, or another metric emitted by the training job. The exam clue is usually improving model performance through hyperparameter search rather than changing the dataset or endpoint mode.

• Changing endpoint configuration when the question asks for hyperparameter optimization through Automatic Model Tuning.

• Automatic Model Tuning needs a search space and objective metric.

SageMaker Pipelines is chosen when ML steps need to run as a repeatable workflow instead of one-off notebook commands. Processing steps prepare or validate data, training steps build models, transform steps run batch inference, and model registration steps move approved artifacts toward deployment. The exam usually describes reproducibility, auditability, or repeatable execution as the clue.

• Running notebook steps manually when the requirement is repeatable SageMaker Pipelines execution.

• Use Pipelines for repeatable ML workflows with auditable steps.

Deployment guardrails reduce the risk of replacing a production model endpoint. Blue/green deployments create a new fleet, shift traffic, monitor alarms, and roll back when configured conditions fail. Choose blue/green when the scenario asks for safer endpoint updates with automatic rollback based on CloudWatch alarms or bake-time checks.

• Calling shadow deployment an A/B test even though shadow predictions are not returned to users.

• Blue/green is safe replacement with rollback controls.

Model Monitor is chosen when a deployed model needs continuous checks against a baseline. Data capture records production requests and responses so monitoring jobs can evaluate data quality, model quality, bias drift, or feature attribution drift. Without data capture, the monitoring workflow lacks the production evidence needed for reports and violations.

• Expecting Model Monitor to work without endpoint data capture or a usable baseline.

• Data capture comes before production monitoring reports.

Responsible AI questions test whether model release decisions are documented, explainable, and reviewed. Model Cards document model purpose, intended use, risk considerations, metrics, and evaluation results for a specific model. AI Service Cards describe AWS AI service behavior, use cases, limitations, and responsible design considerations at the service level.

• Using Model Cards when the question asks about AWS service transparency, which points to AI Service Cards.

• Model Cards document a model; AI Service Cards document an AWS AI service.

The Well-Architected Machine Learning Lens is chosen when the question asks how to evaluate an ML workload across architecture, operations, security, reliability, performance, and cost. It pushes ML systems beyond model accuracy into production concerns such as observability, repeatability, access control, resilience, and continuous improvement.

• Using Savings Plans to solve poor endpoint sizing when right-sizing or Inference Recommender is the better first step.

• Well-Architected ML Lens means architecture review, not a single SageMaker job.

Production ownership means choosing the AWS path that matches who builds, deploys, and operates the ML solution. SageMaker Studio is the integrated development environment for building, training, evaluating, and managing ML work. Canvas is selected when business users need no-code model building or prediction workflows without writing notebooks or training code.

• Choosing Canvas for full-code ML engineering workflows that belong in SageMaker Studio.

• Studio is full-code ML development; Canvas is no-code ML.