AWS MLS-C01 Practice Test

MLS-C01 practice test: what's covered

• ML Specialty
• SageMaker
• Data Eng
• Modeling
• MLOps

Free MLS-C01 sample questions

1. A data engineering team is building a real-time analytics pipeline to capture clickstream events from a high-traffic e-commerce web application, process the stream with low latency, and store continuous micro-batches in Amazon S3 for subsequent ML feature store ingestion. Which AWS solution provides the MOST scalable, low-maintenance approach?

   • A. Amazon Kinesis Data Streams to ingest raw clickstream records, connected directly to Amazon Kinesis Data Firehose to aggregate micro-batches and sink data into Amazon S3 in Apache Parquet format.
   • B. Amazon Managed Streaming for Apache Kafka (MSK) configured with self-managed EC2 consumer instances running custom Python scripts to periodically dump CSV files into Amazon S3.
   • C. AWS Glue ETL batch jobs scheduled to execute every 60 seconds to query transactional database logs and output compressed JSON files into Amazon S3.
   • D. Amazon Simple Queue Service (SQS) standard queues polled by AWS Lambda functions that write raw uncompressed text files sequentially to an EBS volume attached to a single EC2 instance.
   Show answer

   Correct: A. Amazon Kinesis Data Streams provides high-throughput real-time streaming, and pairing it with Kinesis Data Firehose offers a fully managed, serverless mechanism to buffer, transform, format (Parquet/ORC), and store micro-batch data directly into S3 with minimal operational overhead. Option B is incorrect because managing custom EC2 consumers on MSK adds significant infrastructure management overhead compared to serverless Firehose delivery. Option C is incorrect because Glue batch jobs running every minute create high job latency, overhead, and cost compared to continuous streaming delivery. Option D is incorrect because writing SQS messages to an EC2 EBS volume creates a single point of fai

2. A machine learning engineer is training a deep learning model in Amazon SageMaker using a multi-terabyte dataset of high-resolution images stored in Amazon S3. Initial training runs suffer from severe GPU starvation because downloading the full dataset to EBS volumes at job startup takes over two hours. Which TWO data input modes or storage configurations will eliminate GPU idle time and optimize throughput? (Select TWO)

   • A. Configure the SageMaker Training Job to use Fast File Mode to stream data directly from Amazon S3 on demand without downloading files upfront.
   • B. Mount an Amazon FSx for Lustre file system pre-linked to the S3 bucket and configure the training job input channel to use File Mode pointing to FSx for Lustre.
   • C. Configure the training job to use Pipe Mode and store image files as uncompressed, individual PNG files inside nested S3 subdirectories.
   • D. Copy the multi-terabyte dataset to Amazon Elastic Block Store (EBS) cold HDD (sc1) volumes attached to each training worker node.
   Show answer

   Correct: A and B. Fast File Mode streams S3 objects directly into the training container using POSIX file system interfaces without waiting for upfront data download, eliminating GPU startup delays. Amazon FSx for Lustre delivers sub-millisecond latency and high throughput for high-performance computing, pre-loading data from S3 to maximize GPU utilization. Option C is incorrect because Pipe Mode requires streaming single sequential record formats (e.g. TFRecords or RecordIO); raw individual image files in nested directories are inefficient for Pipe Mode streaming. Option D is incorrect because EBS cold HDD (sc1) volumes deliver low throughput and cannot provide the bandwidth needed for multi-GPU deep learnin

3. A machine learning engineer is analyzing a dataset for a credit card fraud detection model. The dataset contains 100,000 transactions, where only 200 records (0.2%) represent fraudulent activity. If the team trains a standard logistic regression model without adjusting for class imbalance, what is the primary risk during model training and evaluation?

   • A. The model will achieve high loss variance during training and fail to converge within standard gradient descent epoch limits.
   • B. The model will suffer from extreme multi-collinearity and produce zero-value weights for all non-fraudulent features.
   • C. The model will learn a trivial decision boundary that predicts the majority class (non-fraud) for all instances, achieving 99.8% accuracy while failing to catch fraudulent transactions.
   • D. The model will automatically overfit to the minority class and produce 100% false positive alerts on legitimate customer transactions.
   Show answer

   Correct: C. In severely imbalanced classification tasks (e.g., 0.2% positive class), standard loss functions like cross-entropy are dominated by majority class examples. Minimizing loss often leads the model to predict the majority class for all inputs, yielding misleadingly high accuracy (99.8%) while recall on the fraud class drops to 0%. Option A is incorrect because logistic regression converges easily on imbalanced data; the issue is metric deception, not convergence failure. Option B is incorrect because class imbalance does not induce multi-collinearity (which stems from correlated independent variables). Option D is incorrect because without rebalancing or cost-sensitive learning, models predict

4. During exploratory data analysis for a retail customer churn project, a data scientist observes that several continuous numeric features (e.g., annual spend, account tenure) contain extreme positive outliers and right-skewed distributions. The data scientist plans to train a neural network model. Which data transformation strategy should be applied to normalize these features effectively?

   • A. Apply One-Hot Encoding across continuous numerical columns to split real values into discrete binary indicator variables.
   • B. Replace all continuous numerical values with the global mean of each feature without applying non-linear scaling functions.
   • C. Apply a Log transformation (or Box-Cox / Yeo-Johnson transform) to reduce right skewness, followed by StandardScaler (z-score normalization).
   • D. Apply Min-Max scaling directly to raw un-transformed right-skewed data to force all outlier values to exactly 0.0.
   Show answer

   Correct: C. Log transformation compresses high-magnitude values and converts right-skewed distributions toward a Gaussian distribution, making numerical features suitable for gradient-based models like neural networks. Applying StandardScaler afterwards centers data around zero mean and unit variance. Option A is incorrect because One-Hot Encoding is designed for categorical variables, not continuous numerical variables. Option B is incorrect because replacing continuous values with the mean destroys feature variance and predictive information. Option D is incorrect because applying Min-Max scaling directly to heavily skewed data squeezes non-outlier data into a tiny interval near 0 while outliers dicta

5. A data scientist calculates a feature correlation matrix for a tabular dataset and discovers two independent continuous features (monthly_bandwidth_usage and data_transfer_gb) with a Pearson correlation coefficient of r = 0.98. What model stability issue does this introduce, and how should it be resolved?

   • A. Multicollinearity, which causes unstable and high-variance coefficient estimates in linear models; resolve by dropping one feature or applying Principal Component Analysis (PCA).
   • B. Class imbalance, which causes tree-based algorithms to fail to split nodes; resolve by applying Synthetic Minority Over-sampling Technique (SMOTE).
   • C. Concept drift, which causes offline validation scores to decline over time; resolve by retraining the model on historical batch partitions.
   • D. Underfitting, which prevents non-linear decision trees from learning feature boundaries; resolve by increasing neural network layer depth.
   Show answer

   Correct: A. High correlation (r = 0.98) indicates severe multicollinearity, making it impossible for linear models to isolate the individual effect of each predictor and producing unstable coefficients with large standard errors. Dropping one of the redundant variables or using PCA compresses collinear dimensions into orthogonal components. Option B is incorrect because correlation between continuous predictors is unrelated to target class imbalance or SMOTE. Option C is incorrect because multicollinearity is a static feature property, whereas concept drift describes changes in the target relationship over time. Option D is incorrect because tree models handle correlated features well, but linear models

Last updated: 2026-10-19

MLS-C01 practice test FAQ

Related practice tests

• AWS MLA-C01 practice test
• AWS AIF-C01 practice test
• AWS DEA-C01 practice test
• AWS SAA-C03 practice test
• Databricks Machine Learning Associate practice test
• Microsoft DP-100 practice test