AWS Certified Machine Learning Engineer – Associate 1-10

The correct answer is B.
The root cause of the problem is that individual Parquet files are too large to be loaded into the memory of the training instances.
By using Apache Spark on Amazon EMR to repartition the files into smaller units, SageMaker can stream and process them efficiently, resolving the memory issue at its source.
Option A adds an EBS volume that increases storage capacity rather than memory, so it does not relieve the memory constraint.
Option C uses a memory-optimized instance, which increases cost and will hit the limit again as the data volume grows.
Option D uses SMDDP, a mechanism for distributing gradients and computation, which does not solve the problem of the file size itself.
Therefore, B, which optimizes the data layout, is the best choice.
Apache Spark – Amazon EMR

The correct answer is A.
AWS Glue is an AWS-native ETL service that provides schema discovery, ETL job execution, and metadata management in a single managed service.
A Glue crawler automatically crawls the data source in S3 to infer the schema and automatically registers the result in the AWS Glue Data Catalog, so there is no need to build your own schema-inference logic or metadata-management mechanism.
Option B’s SageMaker Data Wrangler is primarily intended for visual data preparation and is not suited for catalog management.
Option C’s combination of Athena and Step Functions requires building your own orchestration, which adds effort.
Option D’s approach of using scikit-learn on EC2 has the highest operational overhead.
Therefore, A, which achieves schema discovery and metadata management with the least manual effort, is correct.
What is AWS Glue – AWS Glue

The correct answer is C.
Amazon Kendra is an AI-powered search service specialized in semantic search, and by using the Amazon Kendra S3 connector you can ingest documents in the S3 bucket directly.
Because Kendra itself provides semantic search over the ingested documents through built-in functionality, there is no need to generate embeddings or build and operate your own vector database, so the requirements can be met with minimal operational overhead.
Option A’s combination of AWS Batch and Glue and Option B’s approach using a SageMaker notebook and Feature Store both require building your own embedding-generation and vector-search mechanisms, resulting in high operational overhead.
Option D’s Textract is a text-recognition service and does not provide a search function.
Therefore, C, which has the least operational overhead, is correct.
What is Amazon Kendra – Amazon Kendra

The correct answer is A.
An Amazon SageMaker real-time inference endpoint is a managed hosting feature optimized for low-latency inference.
By configuring auto scaling, the number of instances is automatically increased or decreased even in response to unpredictable request spikes, achieving both high availability and low latency.
Option B’s schedule-based scaling cannot keep up with sudden, unpredictable changes in demand.
Option C’s approach of scaling on EKS based on a memory metric is operationally complex and does not directly reflect changes in request volume.
Option D’s Spot Instances offer a cost advantage but carry a risk of interruption and are not suited to the high-availability requirement.
Therefore, A, which is managed and supports auto scaling, is the best choice.
Real-time inference – Amazon SageMaker AI

The correct answer is C.
By setting an event pattern in Amazon EventBridge that matches uploads to S3 and invoking the SageMaker pipeline directly as the target of the rule, you can automatically start the pipeline in an event-driven manner without implementing additional polling processes or custom code.
Option A’s S3 Lifecycle rule is a feature intended for transitioning storage classes or deleting objects, and cannot be used as a trigger to start a pipeline.
Option B’s approach of scanning the bucket with a Lambda function requires implementing scheduled execution or polling, which increases operational overhead.
Option D’s Amazon MWAA requires building and maintaining an Airflow environment, which is excessive for these requirements.
Therefore, C, which has the least operational overhead, is correct.
Using EventBridge – Amazon S3

The correct answer is C.
An Amazon EMR cluster is composed of primary nodes, core nodes, and task nodes, each with a different role.
The primary node manages the entire cluster, core nodes store data in HDFS while running tasks, and task nodes handle only computation without storing data.
Because task nodes do not store data, there is no risk of data loss even if a Spot Instance interruption occurs, so you can safely and inexpensively add task nodes to scale processing capacity.
Option A’s addition of primary nodes is not supported by the design of EMR.
Option B’s core nodes also store data, so there is a risk of data loss due to Spot Instance interruptions.
Option D’s approach of adding core nodes with On-Demand Instances is more expensive.
Therefore, C, which reduces processing time cost-effectively, is the best choice.
Understanding Amazon EMR node types (primary, core, and task) – Amazon EMR

The correct answer is B.
The performance degradation in production is caused by variations in image quality across cameras.
By adding impulse noise with the Data Wrangler Corrupt Image transform to create training data that mimics the noise and variability that occur with real cameras, you can improve the model’s robustness in a short time without collecting additional data.
Option A’s collection of additional images is effective but takes time to collect and prepare, so it is not the fastest response.
Option C’s contrast enhancement transform is effective for contrast bias but cannot fully address overall image-quality variability such as noise.
Option D’s resize transform is intended to standardize image size and does not resolve image-quality variability.
Therefore, B, which is expected to improve accuracy in the shortest time, is correct.
Prepare ML data with Amazon SageMaker Data Wrangler – Amazon SageMaker AI

The correct answer is B.
In an imbalanced dataset where the minority class is extremely underrepresented, accuracy becomes high simply by predicting the majority class, so it is not an appropriate evaluation metric.
The F1 score is a metric that balances precision and recall, making it suitable for evaluating minority-class performance in imbalanced data.
For image data, techniques that increase minority-class samples through image augmentation such as rotation, flipping, and cropping are effective.
The accuracy used in Options A and C is inappropriate as an evaluation metric for imbalanced data.
Options C and D’s SMOTE is a synthetic-sample-generation technique for tabular data and is not suited to image pixel data.
Therefore, B, which combines optimizing for the F1 score with image augmentation, is the best solution.
Metrics and validation – Amazon SageMaker AI

The correct answer is A.
In a regression model that predicts house prices, extreme outliers have a large impact on the loss function and reduce prediction accuracy for the majority of typical data.
The very large mansion and extremely small apartment included in this dataset are clear outliers in Square Meters.
By removing these outliers and then applying a logarithmic transformation to the Square Meters variable, you can suppress the distortion caused by the wide range of magnitudes and improve prediction accuracy for the majority of typical houses and apartments.
Option B’s normalization alone cannot fundamentally eliminate the impact of the outliers themselves.
Options C and D’s One-Hot Encoding is a technique for categorical variables and is not suited to a continuous numeric variable such as Square Meters.
Therefore, A, which combines outlier removal with a logarithmic transformation, is the most appropriate preprocessing.
Prepare ML data with Amazon SageMaker Data Wrangler – Amazon SageMaker AI