Free MLA-C01 practice test

SageMaker Feature Store is a repository to store, share, and serve curated features consistently for training (offline) and inference (online). S3 Glacier (A) is cold archival storage, Lambda (C) runs code, and SNS (D) is messaging — none is a feature repository.

SageMaker Ground Truth provides managed data-labeling workflows (with human labelers and optional automated labeling). Model Monitor (C) watches production models, Glue DataBrew (D) is visual data prep, and Athena (A) queries S3 with SQL — none is a labeling service.

The training set fits the model, the validation set guides tuning/model selection, and the held-out test set gives an unbiased estimate of real-world performance. Option D is false (splitting doesn't add data), and options A and C invent nonexistent reasons.

Compressed columnar Parquet plus sensible S3 partitioning reduces bytes scanned and speeds reads, lowering cost and training time. A single giant uncompressed CSV (B) and millions of tiny objects (C) are both inefficient, and SNS (D) is a messaging service, not storage.

Sentinel error codes like -9999 aren't real measurements; treat them as invalid/missing and remove or impute them so they don't skew the model. Keeping them (D) injects extreme false values, and options B and C corrupt the feature.

SageMaker Clarify detects bias in datasets (and models) and provides explainability. Neo (B) optimizes models for target hardware, Edge (A) manages edge deployments, and Debugger (D) inspects training runs for issues — not the bias-detection tool for datasets.

RMSE is in the target's units and squares errors, penalizing large deviations more — a standard regression metric. Precision, recall, and F1 (B, C, D) are classification metrics and don't apply to predicting a continuous price.

XGBoost (a gradient-boosted tree algorithm, available as a SageMaker built-in) is a go-to for tabular classification/regression with strong results and modest tuning. A CNN (B) targets images, speech-to-text (C) targets audio, and a DNS resolver (D) is not an ML algorithm.

SageMaker Automatic Model Tuning runs training jobs across hyperparameter ranges (using strategies like Bayesian search) to find the best configuration. Model Monitor (C) watches production, Ground Truth (A) labels data, and Feature Store (B) manages features — none tunes hyperparameters.

With severe imbalance, always predicting the majority class yields high accuracy while completely failing on the important minority class — so precision, recall, and F1 are more informative. Option B is false, option C is false (accuracy is computable), and option D is false (accuracy is a classification metric).

Bayesian optimization models the relationship between hyperparameters and results, using prior trials to pick promising next values more efficiently than random or grid search. Manual guessing (D), leaving values fixed (A), and deleting layers (C) are not tuning strategies.

Match the inference option to each need: a real-time endpoint for low-latency live requests, and batch transform for cost-efficient offline bulk scoring. Forcing one endpoint for both (D) is costly/awkward, batch transform can't meet live latency (C), and per-record training (A) isn't inference.

Containerization packages code, libraries, and runtime for reproducible, portable execution when built-in images don't fit your needs. It doesn't shrink model files (D), and it neither bypasses IAM (B) nor monitoring (A) — those remain in force.

Amazon ECR is the managed container registry where you push custom Docker images that SageMaker pulls for training or inference (bring-your-own-container). Glacier (B) is archival storage, Route 53 (A) is DNS, and Secrets Manager (D) stores secrets — none is a container registry.

A canary deployment sends a small slice of traffic to the new version first, monitoring before increasing rollout — limiting blast radius. A big-bang replacement (B) is riskier, batch transform (C) is offline scoring, and 'blue with no green' (D) isn't a valid strategy.

Alarms plus auto scaling add capacity to protect service quality while you investigate the underlying cause. Ignoring the problem (C) harms users, deleting the model (A) is destructive and premature, and turning off monitoring (D) hides the issue rather than fixing it.

End-to-end ML security is layered: least-privilege IAM, encryption at rest (KMS) and in transit (TLS), network isolation (VPC), and monitoring/audit (CloudWatch, CloudTrail). Option B lists anti-practices, option D is insufficient, and option A wrongly assumes the model secures the surrounding system.

SageMaker Clarify supports bias detection and explainability, and integrates with Model Monitor to watch for bias drift and feature-attribution drift in production. Neo (A) optimizes models for hardware, Trusted Advisor (B) gives account best-practice checks, and QuickSight (D) is BI — none does production bias/explainability monitoring.

SageMaker Model Monitor runs scheduled checks against a baseline to detect data-quality and model-quality drift (plus bias and feature attribution drift) on endpoints. Ground Truth (A) labels data, Glue DataBrew (D) preps data, and Comprehend (C) is NLP — none monitors production model quality.

Data and artifacts should be encrypted at rest (KMS) and in transit, with IAM restricting access. Public buckets (B), disabled encryption (C), and anonymous access (D) all expose sensitive data and violate security best practices.