8 of the 28 questions in this set, with the correct answer marked and every option explained.
1. An ML engineer is preparing a large tabular training dataset that will be read repeatedly by a distributed SageMaker AI training job. Read speed and compression both matter, and only a handful of the dataset's 200 columns are used in any single training run. Which data format should the engineer choose?
✓Apache Parquet
Parquet is columnar and compressed, so a job that only touches a handful of columns can skip reading the rest of the file entirely — exactly the access pattern described, and it reads much faster than a row-oriented format at this scale.
✗CSV
A plain-text, row-oriented format with no compression or column pruning. Every read has to parse all 200 columns even when only a few are needed, which is the opposite of what this workload wants.
✗JSON
Row-oriented and even more verbose than CSV per record, with no native compression. It doesn't let a job skip unused columns either.
✗RecordIO
RecordIO packages records for efficient sequential streaming into SageMaker AI's built-in algorithms, but it doesn't provide column pruning for wide tables with mostly-unused columns the way a columnar format does.
2. A retailer streams clickstream events through Amazon Kinesis Data Streams and wants the computed features (rolling session counts, recent-purchase flags) available with low latency to both a real-time fraud model and a batch retraining job, without recomputing the features twice. What should the ML engineer use?
✓SageMaker Feature Store, writing to it from a stream-processing consumer
Feature Store keeps an online store for low-latency single-record lookups (what the real-time model needs) and an offline store in S3 for batch access (what retraining needs), fed from one write path — so the same computed feature never has to be built twice.
✗Recomputing the features independently in each downstream job
Guarantees the two jobs can drift apart if the logic is ever updated in only one place, and duplicates the compute cost of feature engineering for no benefit.
✗Storing the features only in Amazon Redshift
Redshift is a strong fit for the offline, batch side, but it isn't built for the single-digit-millisecond lookups a real-time fraud model needs at inference time.
✗Storing the features only in Amazon S3 as CSV files
Works for the offline/batch side but has no low-latency lookup path at all, so the real-time model would have nowhere fast to read a feature from.
3. A team is training a large computer vision model on millions of small image files. The training job repeatedly re-reads the full dataset each epoch and is bottlenecked on storage I/O throughput. Which storage choice would most directly relieve that bottleneck for the training instances?
✓Amazon FSx for Lustre, linked to the S3 bucket holding the images
FSx for Lustre is a high-throughput, low-latency parallel file system built for exactly this pattern — repeated, high-throughput reads of the same large dataset across many training epochs — and it can lazily load objects from S3 as they're first accessed.
✗Amazon S3 Glacier Deep Archive
The cheapest S3 storage class, but built for data retrieved rarely with retrieval taking hours — the opposite of what a storage-I/O-bound, repeatedly-read training job needs.
✗Amazon DynamoDB
A low-latency key-value store for discrete items accessed by key, not a file system built for high-throughput sequential or parallel reads of a large image dataset.
✗Standard Amazon EBS General Purpose (gp3) volume attached to a single instance
A single EBS volume's throughput is a per-instance ceiling and doesn't parallelize across a distributed training job's many instances the way a shared parallel file system does.
4. An ML engineer needs to explore, visualize and clean a raw tabular dataset in Amazon S3 — handling missing values, detecting outliers, and applying transformations — using a visual, low-code interface directly inside SageMaker AI Studio. Which tool fits this requirement?
✓SageMaker Data Wrangler
Data Wrangler is purpose-built for this exact workflow: import data, explore it visually, apply built-in and custom transforms, and export a reusable pipeline, all inside SageMaker AI Studio without hand-writing a Spark job.
✗Amazon Athena
A SQL query engine for data in S3. It can filter and aggregate, but it has no visual data-cleaning interface and doesn't live inside the SageMaker AI Studio workflow.
✗AWS Glue crawlers
Crawlers infer schema and populate the Glue Data Catalog with metadata — they don't clean, transform or visualize the data itself.
✗Amazon QuickSight
A business-intelligence and dashboarding tool for visualizing already-clean data, not a data-cleaning or feature-engineering tool for preparing an ML training set.
5. A dataset has a categorical column with 50,000 unique product IDs. An ML engineer wants to encode this column for a neural network without exploding the number of input features. Which encoding approach is most appropriate?
✓Learn a low-dimensional embedding for the product IDs
An embedding layer maps each high-cardinality category to a dense, low-dimensional vector learned during training — keeping the input dimensionality small while still letting the model capture similarity between product IDs.
✗One-hot encoding
One-hot encoding would create 50,000 new binary columns, one per product ID — exactly the feature explosion the requirement is trying to avoid.
✗Ordinal encoding by assigning each product ID a random integer
Assigns an arbitrary numeric order to a category that has none, which teaches the model a false sense of magnitude or ranking between unrelated product IDs.
✗Dropping the column entirely
Discards a potentially predictive feature rather than solving the dimensionality problem — an information loss the question doesn't call for.
6. A company needs a human-in-the-loop workflow to create high-quality labeled training data for an image classification model, including the ability to route only low-confidence predictions to human labelers as the model improves. Which AWS service should the ML engineer use?
✓Amazon SageMaker Ground Truth
Ground Truth manages labeling workflows with human labelers (including Amazon Mechanical Turk or a private workforce) and supports active learning, where a model in training automatically labels high-confidence examples and only routes uncertain ones to humans.
✗Amazon Rekognition
A pre-trained computer vision AI service for tasks like object and label detection — it doesn't manage a human labeling workforce or workflow for building a custom training dataset.
✗AWS Glue DataBrew
A visual data-preparation tool for cleaning and transforming tabular data — it has no image-labeling or human-workforce capability.
✗Amazon Comprehend
A natural-language-processing AI service for text, not an image-labeling workflow tool.
7. An ML engineer runs SageMaker Clarify on a training dataset for a loan-approval model and finds a large difference in proportions of labels (DPL) between two demographic groups before any model is trained. What does this pre-training bias metric indicate, and what is an appropriate mitigation?
✓The training labels themselves are imbalanced across groups; mitigate with techniques like resampling or synthetic data generation before training
DPL measures whether the positive outcome (loan approved) appears at different rates across groups in the raw training data, independent of any model. Since the imbalance exists in the data itself, the fix happens before training — through resampling, reweighting or synthetic data generation — not by adjusting a trained model afterward.
✗The trained model is making biased predictions; mitigate by retraining with a different algorithm
DPL is a pre-training metric computed on the raw data, before any model exists — there's no trained model's predictions to have found bias in yet.
✗The dataset has missing values that need to be imputed
DPL measures label proportions across groups, not missingness. Imputation addresses a different data-quality problem entirely.
✗The evaluation metric chosen for the model is inappropriate for the business problem
Metric selection is a model-development decision made later; it has no bearing on a bias metric computed directly from the raw training data before training starts.
8. A healthcare company needs to prepare a dataset containing protected health information (PHI) for training an ML model, while meeting compliance requirements for data at rest and limiting exposure of identifying fields to the training pipeline. Which combination of actions best satisfies this?
✓Mask or tokenize direct identifiers in the dataset, and encrypt the data at rest with AWS KMS-managed keys
Masking or tokenizing fields like names and identifiers reduces how much identifying information the training pipeline is ever exposed to, and SSE-KMS protects the data at rest with auditable, centrally managed encryption keys — together addressing both the exposure and at-rest requirements.
✗Store the raw PHI dataset in a public S3 bucket for easy team access
A public bucket is the opposite of protecting PHI — it removes access control entirely rather than limiting exposure.
✗Leave the data unencrypted but restrict the S3 bucket to a single IAM user
IAM access control alone doesn't satisfy an at-rest encryption requirement, and it does nothing to reduce how much identifying information appears in the fields the training pipeline reads.
✗Convert the dataset to Parquet format for better compression
Parquet is a legitimate format choice for training efficiency, but format choice has no bearing on data anonymization or encryption — the two things this requirement is actually asking for.
20 more questions in the app
Practise the full 28-question set with a timer, scoring and progress tracking.