MLA-C01 — Frequently Asked Questions

ECR stores container images, not ML model metadata. It lacks built-in versioning, approval workflows, and lineage tracking provided by the SageMaker Model Registry.

Model Groups (Model Package Groups) are a structured way to organize related models and automatically increment version numbers. Tags are simple key-value pairs and do not provide the same organizational or versioning structure.

No. It optimizes model execution speed and memory usage during training/inference, not the initial provisioning of instances.

No. Spot Instances are used to reduce costs. They do not affect the time required to provision and start infrastructure.

Model Monitor only evaluates data quality, model quality, bias and feature attribution on live traffic. It raises alerts about drift but holds no approval status, so it cannot block or authorize a deployment.

The pipeline registers the model version as PendingManualApproval, a reviewer confirms it, and an AWS SDK UpdateModelPackage call sets the status to Approved before the deployment step runs.

Athena queries tables that already exist in the Glue Data Catalog. It cannot automatically scan raw, unlabeled files to create the schema and catalog entries; that is the specific function of AWS Glue Crawlers.

No. SageMaker Batch Transform is primarily used for running inference on trained models against large datasets. For data cleaning and feature engineering of raw source data, AWS Glue DataBrew is the appropriate managed tool.

Event Notifications trigger actions (like starting a pipeline) when data changes. Lifecycle Rules manage the storage class and expiration of objects over time.

Model Registry stores versions for tracking and approval. An Endpoint provides the actual infrastructure to serve real-time inference requests from applications.

A Token is the raw unit of text (word/subword) processed by the model. An Embedding is the numerical vector representation of that token used to capture semantic meaning.

Temperature controls the randomness of the output. The descriptions provided refer to data units, vector conversions, and retrieval methods, none of which describe output randomness.

Linear Learner assumes linear relationships between features and the target. It cannot capture the complex, non-linear interdependencies mentioned in the scenario.

No, K-means is an unsupervised algorithm used for clustering unlabeled data. Fraud detection with transaction logs and profiles is a supervised classification task.

SGD updates repeatedly overshoot the minimum when the step size is too large, so loss falls and rebounds; both curves can oscillate until the rate is reduced.

Early stopping only halts training at a selected point; it does not correct the oversized weight updates that cause the oscillation.

Comprehend custom models are shared through the ImportModel API; manually copying artifacts in S3 does not carry the model's Comprehend metadata and adds development work.

Yes, importing a Comprehend custom model requires the source and destination to be in the same Region, matching this scenario in us-east-1.

More layers or neurons raise model capacity, which speeds up memorization of training noise and makes validation performance peak earlier and fall further — it worsens overfitting rather than mitigating it.

High bias shows up as poor performance on both training and validation sets (underfitting). Here the model fits well at first, so the problem is variance, and early stopping plus dropout target it directly.

SageMaker real-time endpoints require at least one instance at all times; autoscaling can add instances but cannot reduce the minimum below one. Serverless Inference scales to zero when idle.

Yes. The payload and runtime are within Serverless Inference limits, and MaxConcurrency 1 matches the single nightly invocation while allowing scale-to-zero between runs.

Parameter Store can hold the tokens as SecureString values but has no built-in rotation scheduler, so you would have to build and maintain the entire rotation workflow yourself.

KMS only manages encryption keys, not arbitrary secrets like API tokens, and its automatic key rotation cannot be scheduled to call an external embedding provider every 3 months.

Ground Truth is a human-in-the-loop labeling service that produces annotations, not a data frame. It offers no built-in imputation, deduplication, or outlier transforms.

Yes. Data Wrangler ships built-in transforms for imputing missing values, dropping duplicates, and scaling or clipping outliers, all inside one data flow.

SageMaker Training natively supports Amazon EFS and FSx for Lustre as file-system data sources; FSx for ONTAP is accessed by mounting it to a SageMaker instance in the same VPC instead.

Data Wrangler is a data preparation tool; a catalog or direct connection lets analysts explore FSx for ONTAP data but does not mount the volume for SageMaker training code.

Mechanical Turk draws on a public, anonymous crowd, so confidential images would leave the organization's controlled workforce and breach the employee-only access requirement.

Rekognition automates detection using pretrained or custom models but gives no employee-reviewed ground truth for your specific taxonomy, so it does not meet the highest-accuracy labeling goal.

Debugger profiles and debugs training jobs; it does not compare live endpoint predictions with ground truth or baseline feature attributions, so it cannot catch this drift.

Clarify's bias and feature-attribution drift monitors run through SageMaker Model Monitor; among the listed options, D is the only one that continuously monitors production model behavior.

Savings Plans require a spend commitment for predictable usage; because these jobs tolerate interruption, Spot with checkpoints captures deeper discounts and lower total cost.

Reserved capacity suits steady, uninterrupted workloads and still charges for unused or interrupted capacity; Spot is more cost-effective for fault-tolerant 48–72 hour jobs.

Batch transform scores a whole dataset on a schedule and writes results to S3; it has no per-request callback, so users would not be alerted when their individual image finishes.

SageMaker inference reads input objects from Amazon S3, and EFS is a shared file system for compute instances rather than the object store used for inference payloads.

Canary shifts a fixed percentage of traffic (e.g., 10%) to the new model immediately and keeps it there for monitoring. Linear gradually increases the percentage of traffic to the new model over a specified duration.

Use All at once for non-critical models, testing environments, or when you need to instantly switch all traffic to the new version without any gradual transition or rollback window.

A longer cooldown keeps the endpoint waiting longer before another scaling action, which reduces responsiveness when demand is volatile; 300 seconds supports faster adjustment.

Yes. SageMaker endpoint target tracking can use high-resolution metrics at 10-second intervals, which detect traffic spikes sooner than standard one-minute metrics.

Because the problem type (e.g., detection) determines the model architecture and output format. You cannot select a valid algorithm without knowing what task it needs to perform.

No, because different models require different input formats and may need specific augmentations. Preprocessing should be tailored to the selected model's requirements.

k-means only partitions records into k similar groups and needs k chosen upfront; it produces cluster membership, not the per-record anomaly score required to flag irregular patients.

It sets how many random cut trees form the forest. Larger values (well above 100) reduce variance in the anomaly scores at the cost of more compute and memory.

SMOTE interpolates between numeric feature vectors, so applying it to raw pixels often creates unrealistic blended images rather than meaningful variations of the minority class.

With one extremely small class, high accuracy can hide poor minority-class predictions; F1 balances precision and recall so improvement on the rare class is actually measured.

lambda is the L2 regularization term; raising it constrains leaf weights and reduces variance, so the model generalizes better to real-world validation data.

Adding more representative real-world data can help, but it requires data acquisition and retraining, while changing lambda is a single hyperparameter adjustment.

Canary shifts a small percentage initially to validate the new model before moving all traffic, whereas Linear shifts traffic gradually in increments over a specified time period.

Use All at once when you need to switch 100% of traffic to the new model immediately without any gradual transition or validation phase.

It can, but larger batch sizes often allow for faster convergence and better generalization if tuned correctly, improving speed without necessarily hurting final performance.

Spot Instances are primarily for cost reduction. They can interrupt training due to price spikes, potentially increasing total wall-clock time or causing job failures.