Select Two (Multi-Select) Questions

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.

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.

Apply feature selection to reduce the number of features.

Why wrong: Feature selection is optional and not a mandatory step for all models.

Remove outliers from the dataset.

Why wrong: Outlier removal is not always required; it depends on the data and model.

Apply one-hot encoding to each word

Why wrong: One-hot encoding is for categorical variables, not text preprocessing.

Compute TF-IDF vectors

Why wrong: TF-IDF is a feature extraction step, not preprocessing.

SMOTE to balance classes

Why wrong: Addresses class imbalance, not general overfitting.

One-hot encoding of all categorical features

Why wrong: Increases dimensionality, potentially worsening overfitting.

Log transformation of skewed features

Why wrong: Addresses skewness, not overfitting.

Enable SageMaker Clarify for explainability during training

Why wrong: Clarify is for bias detection and explainability, not for training setup.

Build a custom Docker container with PyTorch and Transformers

Why wrong: SageMaker provides built-in HuggingFace containers, so custom container is not needed.

Use SageMaker Processing to preprocess the data in parallel

Why wrong: Processing is for data preprocessing, not for fine-tuning setup.

Use an ml.c5.2xlarge instance with CPU only

Why wrong: CPU only would not provide GPU acceleration required.

Use SageMaker Batch Transform for inference

Why wrong: Batch Transform is for offline inference, not real-time.

Store all data in a single large file and use append operations

Why wrong: Single files are not scalable and appending is inefficient for S3.

Manually copy new files to the same S3 bucket

Why wrong: Manual processes are error-prone and not scalable.

Overwrite the entire existing dataset with the new data

Why wrong: Reprocessing all data is costly and unnecessary.

Target encoding of high-cardinality categorical features

Why wrong: Target encoding can cause data leakage and reduces interpretability.

PCA dimensionality reduction

Why wrong: PCA transforms features into uninterpretable components.

One-hot encoding of categorical features

Why wrong: One-hot encoding adds many columns, complicating interpretability.

Domain-specific feature engineering based on business rules.

Why wrong: This is often useful but not a general technique applicable to all datasets.

One-hot encoding all categorical features.

Why wrong: One-hot encoding high cardinality features creates a large sparse matrix.

Remove outliers from the minority class

Why wrong: This would further reduce the minority class, worsening imbalance.

Apply PCA to reduce dimensionality

Why wrong: PCA does not address class imbalance.

Use stratified splitting for train/test sets

Why wrong: Stratified splitting maintains class proportions but does not balance the data itself.

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.

Remove any rows with outlier values.

Why wrong: Outliers may be valid; removal should be justified, not automatic.

Apply target encoding to all categorical features regardless of cardinality.

Why wrong: Target encoding for low-cardinality may lead to overfitting; one-hot is often better.

Use uncompressed CSV files for simplicity

Why wrong: Uncompressed CSV leads to slower performance and higher costs.

Partition the data by every column to maximize filtering

Why wrong: Too many partitions cause many small files and overhead.

An AWS CodePipeline pipeline with approval stage

Why wrong: CodePipeline is a separate service; Step Functions can include approval without CodePipeline.

An Amazon SNS topic for notification of approval status

Why wrong: SNS is a notification service; while it may be used in the approval task, it is not a required component of the state machine itself.

Use a serverless endpoint configuration to automatically scale.

Why wrong: Serverless endpoints do not support GPU instances.

Use Spot instances for the endpoint to reduce cost.

Why wrong: Spot instances can be terminated at any time, making them unsuitable for low-latency real-time inference.

Provision multiple on-demand GPU instances behind a load balancer.

Why wrong: This is costly and may over-provision for normal traffic.

Store model artifacts in Amazon EBS volumes attached to the endpoint instances

Why wrong: Model artifacts should be stored in S3, not EBS; EBS is ephemeral.

Manually track model versions using tags because SageMaker Model Registry is not available for deployment

Why wrong: SageMaker Model Registry is available and should be used for versioning.

Disable CloudWatch Logs to reduce costs during inference

Why wrong: CloudWatch Logs are important for monitoring and troubleshooting; disabling is not best practice.

Disable the existing endpoint to prevent stale predictions during retraining

Why wrong: Disabling the endpoint would cause downtime; use a blue/green deployment instead.

Manually compare the model's predictions against a holdout validation set each week

Why wrong: Manual comparison is not automated and doesn't scale.

Convert the data to a CSV string for model input

Why wrong: SageMaker endpoints accept JSON; converting to CSV adds unnecessary complexity.

Re-train the model periodically using new data

Why wrong: Re-training is a separate offline process, not part of inference preprocessing.

Impute missing values using the mean of the entire dataset

Why wrong: Using the full dataset mean leaks information from test data.

Shuffle the data randomly before splitting

Why wrong: Shuffling ignores time order, causing future data to influence training.