1500 Questions | AWS Machine Learning Specialty 2026

Data processing and transformation

Data wrangling and feature engineering

Model training and hyperparameter tuning

Model evaluation and selection

Model monitoring and maintenance

Model versioning and rollback

Data storytelling and communication

Data management and governance

Model optimization and scalability

Model security and compliance

B) K-Means: Incorrect. K-Means is an unsupervised clustering algorithm used to group similar data points. It is not designed to predict a user's preference for an unpurchased item.

C) Factorization Machines: Correct. Factorization Machines are specifically designed to deal with sparse data and are the ideal built-in SageMaker algorithm for building recommendation systems when dealing with massive, sparse user-item interaction matrices.

D) BlazingText: Incorrect. BlazingText is used for natural language processing (NLP) tasks such as Word2Vec (word embeddings) and text classification. It is completely irrelevant to tabular purchase history recommendation tasks.

E) Random Cut Forest (RCF): Incorrect. RCF is an unsupervised algorithm used primarily for anomaly detection, such as identifying spikes in time-series data, not for recommendations.

F) Linear Learner: Incorrect. While Linear Learner can handle classification and regression, it struggles to capture the complex, non-linear interactions between users and items in a sparse matrix compared to Factorization Machines.

Overall Explanation: The core of this scenario focuses on Domain 2 (Machine Learning Implementation). Recommendation engines relying on user-purchase history inherently produce sparse matrices (most users only buy a tiny fraction of total products). SageMaker's Factorization Machines algorithm is explicitly built to capture interactions between variables in highly sparse datasets, making it the standard choice for this architecture.

B) Amazon SQS triggering an EC2 instance: Incorrect. This requires provisioning, patching, and managing EC2 instances manually, making it highly inefficient operationally compared to fully managed serverless options.

C) AWS Glue Crawler to Amazon Athena: Incorrect. Glue Crawlers and Athena are designed for querying data at rest in Amazon S3 (Data Lake architecture), not for continuously ingesting and loading real-time streams into Redshift.

D) Amazon Kinesis Data Streams directly to Redshift: Incorrect. Kinesis Data Streams does not have a native, fully managed delivery mechanism directly to Redshift without writing custom consumer applications. Firehose is the delivery service.

E) AWS Step Functions coordinating AWS Batch jobs: Incorrect. Step Functions and AWS Batch are suited for orchestrating complex, long-running batch computing workloads, not for continuous, near real-time stream processing.

F) Amazon EMR running Apache Spark Streaming directly to Redshift: Incorrect. While technically possible, deploying and maintaining an entire EMR cluster for a "lightweight" transformation is massive administrative overhead and overkill when serverless options exist.

Overall Explanation: This question tests Domain 1 (Data Ingestion, Storage, and Processing). The key requirement is operational efficiency (minimal management) for near real-time ingestion and transformation. Kinesis Data Firehose is specifically built for this exact use case, allowing you to attach a Lambda function to transform data on the fly before it hits the Redshift destination, requiring zero server management.

B) SageMaker Real-time Endpoints with Auto Scaling: Incorrect. While Auto Scaling handles spikes, Real-time endpoints require at least one instance to be running 24/7. During the "long periods of zero requests," you will still be charged for the idle compute.

C) SageMaker Serverless Endpoints: Correct. Serverless Endpoints automatically scale compute capacity based on traffic and scale down to zero when there are no requests. You only pay for the exact compute duration and data processed, perfectly matching the erratic traffic and cost constraints.

D) SageMaker Asynchronous Endpoints: Incorrect. Asynchronous endpoints are ideal for requests with large payload sizes or long processing times, but they still require configuring underlying instances and auto-scaling rules. They do not scale to zero natively in the same fully managed way Serverless does.

E) SageMaker Batch Transform: Incorrect. Batch Transform is used for offline inference on large datasets stored in S3. The scenario explicitly requires deployment for "production inference" dealing with application traffic spikes, implying a real-time/on-demand need.

F) Deploy the model on a fleet of Spot EC2 instances: Incorrect. Spot instances can be interrupted at any time with a 2-minute warning. They are highly discouraged for synchronous production inference workloads due to the risk of dropped requests.

Overall Explanation: This scenario evaluates Domain 3 (Model Operations and Deployment). The defining constraints are unpredictable traffic, periods of zero requests, and strict cost optimization. SageMaker Serverless Inference is the only option that scales down to exactly zero compute, ensuring you do not pay a single cent during the idle periods while still instantly handling traffic spikes.

You can retake the exams as many times as you want.

This is a huge original question bank.

You get support from instructors if you have questions.

Each question has a detailed explanation.

Mobile-compatible with the Udemy app.