AWS ML Engineer Associate Practice Test

Enable Amazon SageMaker managed warm pools in the training job configuration. is correct because SageMaker managed warm pools let you keep training infrastructure warm after a training job completes. The cited source, Reduce training start time with SageMaker Managed Warm Pools, supports this answer for the Model Training and Tuning scenario rather than the adjacent distractors.

The offline store persists all historical feature values over time along with metadata columns, including write_time (when the record was written to the offline store) and api_invocation_time. To perform time-travel queries and prevent data leakage during model training or backtesting, you can query the offline store using Amazon Athena and filter records based on write_time or event time to reconstruct the features as they existed at a specific point in time.

Deploying, automating, monitoring, and maintaining the model reliably on AWS is the production-minded concern that comes next. The certification scope is aimed at engineers who move ML work from promising experiments into operational AWS systems.

SageMaker Savings Plans provide a flexible pricing model where you commit to a consistent amount of compute usage (measured in hourly spend) for a 1-year or 3-year term. This offers significant savings (up to 64%) on SageMaker hosting (including real-time, serverless, and async endpoints), processing, and training jobs compared to on-demand pricing.

In SageMaker Pipelines, a data cleansing script runs as a ProcessingStep. Model training is executed using a TrainingStep. Model evaluation on a test set runs as another ProcessingStep (or a dedicated evaluation script). A ConditionStep is then used to evaluate the model's performance metric against a threshold, conditionally executing the nested RegisterModel step.

Multi-model endpoints match that profile because they are designed to host many models behind shared resources and can reduce cost when traffic is uneven across models. AWS also notes that occasional cold-start-related latency is a tradeoff teams should tolerate for this pattern.

A Model Monitor baseline with statistics and constraints is the correct starting point because monitoring needs a reference definition of normal behavior. AWS uses those baselines to compare production traffic and detect drift or violations later.

SageMaker Shadow Deployments allow you to test a new model (shadow model) by routing a portion of production traffic to it in parallel with the current model (production model). The shadow model's responses are discarded and only the production model's response is returned to the user, allowing safe validation of performance, latency, and correctness.

AWS AI Service Cards are a transparency resource designed to help customers understand the intended use, limitations, and performance characteristics of AWS AI services (like Amazon Rekognition, Amazon Transcribe, etc.) as part of AWS's commitment to responsible AI.

SageMaker Clarify bias analysis is correct because SageMaker Clarify bias analysis is the strongest fit because it is designed to evaluate bias in data and models before or after deployment. The cited source, Run an Amazon SageMaker Clarify processing job, supports this answer for the Model Training and Tuning scenario rather than the adjacent distractors.