Pass the AWS ML Engineer Associate — the First Time

1. A team trains SageMaker models on a 4 TB tabular dataset stored as gzip-compressed CSV in Amazon S3. Analysts also run frequent Amazon Athena queries that scan only 6 of the 80 columns. Which change to the storage format will MOST reduce both Athena scan cost and SageMaker data-loading time?

   • A. Convert the data to a single uncompressed JSON Lines file
   • B. Re-store the data as RecordIO-protobuf shards
   • C. Keep CSV but increase the gzip compression level
   • D. Convert the data to Apache Parquet with Snappy compression

   Answer: D. Parquet is a columnar format, so Athena reads only the 6 needed columns instead of scanning all 80, and Snappy-compressed Parquet is splittable for parallel reads. JSON Lines is row-oriented and verbose, increasing scan size. RecordIO-protobuf suits certain built-in algorithms but is not queryable by Athena and does not enable columnar pruning. Higher gzip levels do not enable column pruning, and a single gzip CSV is not splittable.

   Exam domain: Data Preparation for ML

2. Clickstream events arrive continuously and must be delivered to Amazon S3 as Parquet, partitioned by event date, with no application code to manage and built-in buffering and format conversion. The team wants the simplest fully managed pipeline. Which AWS service should they use?

   • A. Amazon Kinesis Data Streams with a custom consumer running on Amazon EC2
   • B. AWS Glue streaming job writing micro-batches to S3
   • C. Amazon Data Firehose with record format conversion to Parquet and dynamic partitioning
   • D. Amazon MSK with a self-managed Kafka Connect S3 sink

   Answer: C. Amazon Data Firehose (formerly Kinesis Data Firehose) is fully managed, buffers records, can convert JSON to Parquet, and supports dynamic partitioning to S3 with no servers to run. Kinesis Data Streams plus a custom EC2 consumer requires you to build and operate the delivery and conversion logic. MSK with Kafka Connect is self-managed and operationally heavier. A Glue streaming job is more code and infrastructure than needed for straightforward delivery to S3.

   Exam domain: Data Preparation for ML

3. A distributed SageMaker training job needs a shared POSIX file system that many GPU instances can mount simultaneously for very high-throughput, low-latency reads of training data already staged from S3. Which storage option BEST fits this requirement?

   • A. Amazon S3 accessed over the public endpoint
   • B. Amazon S3 Glacier Instant Retrieval
   • C. Amazon EFS Infrequent Access storage class
   • D. Amazon FSx for Lustre linked to the S3 bucket

   Answer: D. FSx for Lustre is a high-performance, POSIX-compliant parallel file system that integrates with S3 and is purpose-built for high-throughput ML and HPC training across many instances. EFS is POSIX but delivers lower throughput than Lustre for large-scale GPU training, and the IA class is for cost savings on rarely accessed files. S3 and Glacier are object stores, not shared POSIX file systems.

   Exam domain: Data Preparation for ML

4. A team wants a visual, low-code tool to profile a CSV dataset, identify missing values and outliers, and apply 250+ prebuilt cleaning and normalization transformations without writing Spark or pandas code. Which AWS service is designed for this?

   • A. AWS Lambda
   • B. Amazon EMR on EC2
   • C. Amazon Athena
   • D. AWS Glue DataBrew

   Answer: D. AWS Glue DataBrew is a visual data-preparation tool with hundreds of prebuilt transformations and built-in profiling for missing values and outliers, requiring no code. Athena queries data with SQL but does not provide a visual cleaning UI. EMR requires writing Spark code, and Lambda is general-purpose compute, not a data-prep tool.

   Exam domain: Data Preparation for ML

5. A high-cardinality categorical column 'product_id' has 50,000 distinct values and will feed a gradient-boosted tree model. One-hot encoding would create 50,000 sparse columns. Which encoding approach is the MOST appropriate to control dimensionality while preserving predictive signal?

   • A. Label encoding that assigns arbitrary sequential integers used directly as the feature
   • B. Dropping the column entirely because it is high cardinality
   • C. Target (mean) encoding computed only from the training split
   • D. One-hot encoding with all 50,000 categories retained

   Answer: C. Target encoding replaces each category with a statistic (such as the mean of the target) and keeps the feature one-dimensional, which works well for very high-cardinality columns; computing it only on the training split avoids leakage. One-hot at 50,000 levels explodes dimensionality. Arbitrary integer label encoding imposes a false ordinal relationship that misleads many models. Dropping the column discards potentially useful signal.

   Exam domain: Data Preparation for ML

6. A data scientist has a tabular dataset in Amazon S3 with a labeled target column and wants SageMaker to automatically explore feature preprocessing, candidate algorithms, and hyperparameters to produce a leaderboard of trained models with minimal code. Which SageMaker capability is the BEST fit?

   • A. Amazon SageMaker Debugger
   • B. Amazon SageMaker JumpStart foundation models
   • C. Amazon SageMaker Model Monitor
   • D. Amazon SageMaker Autopilot (AutoML)

   Answer: D. Autopilot (AutoML) automatically handles preprocessing, algorithm selection, and hyperparameter tuning on tabular data, producing a ranked leaderboard of candidate models. JumpStart provides pretrained foundation/proprietary models rather than AutoML over your tabular data. Debugger inspects training jobs for issues, and Model Monitor watches deployed endpoints for drift, neither of which builds candidate models.

   Exam domain: ML Model Development

7. A retail company must forecast daily demand for thousands of SKUs across many stores, where some items have only a few months of sales history. The team wants a managed approach that captures cross-series patterns rather than fitting one model per SKU. Which approach best frames and solves this problem?

   • A. Treat it as binary classification using the SageMaker XGBoost built-in algorithm
   • B. Treat it as a time-series forecasting problem and use the SageMaker DeepAR built-in algorithm
   • C. Treat it as clustering using the SageMaker K-Means built-in algorithm
   • D. Treat it as anomaly detection using the SageMaker Random Cut Forest built-in algorithm

   Answer: B. This is a time-series forecasting problem, and DeepAR is designed to train a single global model across many related time series, which helps cold-start SKUs with little history by learning shared patterns. XGBoost classification, K-Means clustering, and Random Cut Forest anomaly detection address fundamentally different problem types and would not produce per-SKU demand forecasts.

   Exam domain: ML Model Development

8. A team wants to build a customer-support summarization feature. They want to start from a high-quality pretrained large language model deployable on SageMaker, then optionally fine-tune it on their own ticket data. Which SageMaker capability most directly provides access to such pretrained foundation models?

   • A. Amazon SageMaker JumpStart
   • B. Amazon SageMaker Ground Truth
   • C. Amazon SageMaker Clarify
   • D. Amazon SageMaker Feature Store

   Answer: A. SageMaker JumpStart provides a catalog of pretrained foundation models (including LLMs) that can be deployed and fine-tuned on SageMaker, making it the direct fit for an NLP summarization use case. Feature Store stores features, Ground Truth labels data, and Clarify analyzes bias and explainability, none of which provide pretrained foundation models.

   Exam domain: ML Model Development

9. During a SageMaker training job for a neural network, the model fits the training data well but the validation loss begins increasing while training loss keeps decreasing. The team wants the training job to automatically stop and keep the best checkpoint when validation stops improving. Which technique should they configure?

   • A. Increase the learning rate used in training
   • B. Remove the validation channel from the job
   • C. Switch across to a larger training instance type
   • D. Enable early stopping on the validation metric

   Answer: D. Rising validation loss alongside falling training loss is overfitting, and early stopping halts training when the monitored validation metric stops improving, retaining the best model. Increasing the learning rate or scaling up instances does not address overfitting, and removing the validation channel would eliminate the very signal needed to detect the problem.

   Exam domain: ML Model Development

10. A team uses SageMaker Automatic Model Tuning (AMT) to find good hyperparameters for an XGBoost model. They want to minimize the number of training jobs while letting the tuner learn from completed jobs to choose promising next configurations. Which tuning strategy best meets this goal?

    • A. Grid search over the space
    • B. Random search of the space
    • C. Bayesian optimization search
    • D. Manual search by the team

    Answer: C. Bayesian optimization builds a probabilistic model of the objective and uses results from completed jobs to select promising next hyperparameter configurations, typically reaching good results with fewer training jobs. Grid and random search do not learn from prior results, and manual search is not an AMT strategy and is inefficient.

    Exam domain: ML Model Development

Want more? Finish the quiz above to unlock 10 more free questions, or open the full MLA-C01 practice app.