Select Two (Multi-Select) Questions

Upload the training data to an S3 bucket.

Why wrong: Training data upload is part of training, not deployment.

Register the model in the SageMaker Model Registry.

Why wrong: Model Registry is optional for versioning, not required for deployment.

Set up a SageMaker Notebook instance.

Why wrong: Notebook instances are for development and experimentation.

Create an AWS Glue DataBrew recipe to transform the data

Why wrong: DataBrew is not necessary for Athena queries.

Add each partition manually using ALTER TABLE ADD PARTITION

Why wrong: Manual addition is not scalable for many partitions.

Use network ACLs to restrict API calls to the SageMaker API.

Why wrong: Network ACLs are for network traffic, not API authorization.

Enable Multi-AZ deployment for the notebook instance.

Why wrong: Multi-AZ is not supported for notebook instances.

Associate the notebook instance with a public subnet that has an internet gateway.

Why wrong: A public subnet with internet gateway increases exposure.

Reduce the number of epochs to match the number of GPUs

Why wrong: Epochs are not related to distributed training; reducing epochs underfits.

Manually split the training data into shards and upload to S3

Why wrong: Manual sharding is unnecessary; SageMaker handles data distribution.

Set the instance count to 1 with a multi-GPU instance

Why wrong: Single instance, even with multiple GPUs, is not distributed training across instances.

Use Amazon Athena to create views with random sampling

Why wrong: Athena random sampling does not guarantee stratified splits.

Use AWS Glue's built-in random split transform

Why wrong: Random split does not preserve target distribution.

Use Amazon SageMaker's built-in SplitType parameter in a Processing Job

Why wrong: SplitType only supports random division.

Use a fixed random seed when sampling data to ensure repeatability.

Why wrong: Using a random seed is a good practice but is more about consistent sampling than overall data quality.

Shuffle the dataset before splitting into train and test sets.

Why wrong: Shuffling is a good practice for many ML tasks, but it is not specifically about data quality or reproducibility.

Manually inspect and clean data to remove outliers.

Why wrong: Manual processes are not scalable and not reproducible.

Use a multi-model endpoint

Why wrong: Multi-model endpoint is not required for auto scaling.

Enable data capture

Why wrong: Data capture is for logging, not scaling.

Configure a lifecycle configuration

Why wrong: Lifecycle configuration is for notebook instances, not endpoints.

Data validation can only be performed on training data, not inference data

Why wrong: Validation can be applied to any dataset.

Data validation steps cannot pass results to subsequent steps

Why wrong: Property values can be passed downstream.

Data validation requires a trained model to evaluate predictions

Why wrong: Validation is on raw data, not predictions.

Increase the model's complexity to improve accuracy on edge devices.

Why wrong: Increasing model complexity adds layers and parameters, which directly increases inference latency and memory consumption, contradicting the stem’s requirement for low latency and minimal memory footprint on edge devices. This option is tempting because deeper models often yield higher accuracy on powerful hardware, and would be correct for a cloud deployment where compute resources are abundant and latency constraints are relaxed.

Deploy the model directly as a SageMaker endpoint and have the edge devices call it over the internet.

Why wrong: Edge deployment means running locally, not relying on cloud inference.

Use Athena's UNNEST function on the raw data

Why wrong: Athena is not part of AWS Glue ETL.

Use Amazon SageMaker Processing with scikit-learn

Why wrong: SageMaker Processing is separate from Glue ETL.

Amazon Redshift

Why wrong: Redshift is a data warehouse, not primarily for data lake cataloging.

Amazon Athena

Why wrong: Athena is a query service, not for ingestion or cataloging.

AWS Glue ETL

Why wrong: Glue ETL is for transformation, but the question asks for ingestion and cataloging; S3 and Glue Data Catalog are the key components.

Include the raw PII in the training dataset and rely on the model to not memorize it.

Why wrong: Including PII is a privacy risk and may violate regulations.

Use AWS Glue to copy the data to a separate bucket without any transformations.

Why wrong: Copying without transformation does not address PII concerns.

Deploy a shadow endpoint for comparison

Why wrong: Shadow endpoint is for traffic shifting, not monitoring drift.

Use SageMaker Debugger for monitoring

Why wrong: Debugger is for training debugging, not production monitoring.

SageMaker Python SDK

Why wrong: The SDK is used to interact with SageMaker APIs, not for IaC management of Studio domains.

Boto3

Why wrong: Boto3 is the AWS SDK for Python, used for programmatic access, not IaC declarative management.

Use the largest instance type available

Why wrong: Largest instance types increase cost and may not yield proportional performance gains; optimization should consider cost-efficiency.

Increase the number of layers in the model to improve accuracy

Why wrong: Deeper models may improve accuracy but also increase training time and risk overfitting; this is not a general performance optimization practice.

Normalize text data to lowercase

Why wrong: Although sometimes useful, it is a specific step and not a general necessary part of data preparation for all datasets.

Remove all outliers blindly

Why wrong: Blind removal can discard valuable data and bias the model.

Use a large neural network to handle all transformations

Why wrong: Neural networks are models, not data preparation techniques.

The conditional step should be implemented as a separate Lambda function called from the pipeline.

Why wrong: SageMaker Pipelines has a native ConditionStep, no need for Lambda.

The pipeline will automatically retry the training step if the condition fails.

Why wrong: SageMaker Pipelines does not automatically retry steps based on conditions; it will fail the pipeline.

The model registration step should be placed before the condition step to ensure the model is always registered.

Why wrong: Registration should only happen if condition is met; placing before would register even if accuracy fails.

ModelQuality

Why wrong: ModelQuality is not a built-in CloudWatch metric; it requires custom metrics from Model Monitor.

CpuUtilization

Why wrong: CPU utilization reflects instance health, not directly model performance degradation.

InvocationCount

Why wrong: Invocation count shows traffic volume, not model quality or health.

Use Amazon SageMaker Data Wrangler to continuously ingest streaming data into the feature group

Why wrong: Data Wrangler is for interactive data preparation, not continuous streaming ingestion.

Use Amazon SageMaker Ground Truth to label incoming streaming data

Why wrong: Ground Truth is for manual labeling, not for streaming ingestion of features.

One-hot encoding

Why wrong: Typically used for categorical variables, not text.

Principal component analysis (PCA)

Why wrong: PCA reduces dimensionality after vectorization, not a text feature extraction method.