Sign in
Home/AWS/Associate (MLA-C01)/Free questions

AWS Certified Machine Learning Engineer - Associate (MLA-C01) — Free Practice Questions

10 free sample questions from a bank of 106, with the correct answers and explanations. No signup required — start practising right now.

1Case Study - A company is building a web-based AI application by using Amazon SageMaker. The application will provide the following capabilities and features: ML experimentation, training, a central model registry, model deployment, and model monitoring. The application must ensure secure and isolated use of training data during the ML lifecycle. The training data is stored in Amazon S3. The company needs to use the central model registry to manage different versions of models in the application. Which action will meet this requirement with the LEAST operational overhead?
  • Create a separate Amazon Elastic Container Registry (Amazon ECR) repository for each model.
  • Use Amazon Elastic Container Registry (Amazon ECR) and unique tags for each model version.
  • Use the SageMaker Model Registry and model groups to catalog the models.
  • Use the SageMaker Model Registry and unique tags for each model version.
Answer: C

The short version

C — SageMaker Model Registry with model groups is the managed model catalog. It versions, approves, and tracks models centrally with least operational overhead, while ECR only stores container images.

Key concepts in this question

  • SageMaker Model Registry: Fully managed central catalog for ML models, supporting registration, versioning, approval workflows, and deployment lineage.
  • Model Groups and Model Packages: A Model Group collects related model versions, each Model Package gets an auto-incremented version starting at 1, enabling clear lineage across retraining iterations.
  • ECR versus Model Registry: ECR stores Docker container images used for training and inference code, it does not track model metrics, approvals, or lineage as first-class objects.
  • Least operational overhead: Means preferring a purpose-built managed registry with built-in versioning over manually organizing repositories or tags.

Why C is correct

SageMaker Model Registry with model groups directly satisfies the central registry requirement for different model versions. Each trained model becomes a Model Package inside a Model Group, versioned numerically and automatically, with metadata and approval status retained. Docs describe this grouping and incrementing behavior with minimal effort.

Why the others are wrong

  • A. Separate ECR repository per model stores container images, not model lineage, and multiplies repository management without version governance.
  • B. ECR plus unique tags still manages images, not registered models, and relies on manual tag conventions with no approval or lineage tracking.
  • D. Registry plus unique tags misstates the mechanism, because versioning is handled by Model Groups with automatic numeric versions, not by manual tagging alone.

MLA-C02 exam tip

When the stem says central model registry plus versions, choose Model Registry plus model groups, when it says container images, choose ECR.

2Case study - An ML engineer is developing a fraud detection model on AWS. The training dataset includes transaction logs, customer profiles, and tables from an on-premises MySQL database. The transaction logs and customer profiles are stored in Amazon S3. The dataset has a class imbalance that affects the learning of the model's algorithm. Additionally, many of the features have interdependencies. The algorithm is not capturing all the desired underlying patterns in the data. Which AWS service or feature can aggregate the data from the various data sources?
  • Amazon EMR Spark jobs
  • Amazon Kinesis Data Streams
  • Amazon DynamoDB
  • AWS Lake Formation
Answer: D

The short version

D — AWS Lake Formation aggregates sources into a governed data lake. It centralizes S3 logs, S3 profiles, and on-premises MySQL tables for unified discovery.

Key concepts in this question

  • Data aggregation across sources: Combining S3 objects plus relational tables into one queryable foundation, including schema registration, crawling, and permission enforcement.
  • AWS Lake Formation: Managed data lake service that ingests, catalogs, cleans, and secures data from S3 and relational sources, building a central Glue Data Catalog with fine-grained access.
  • EMR Spark versus lake governance: Spark is a powerful distributed processing engine for ETL, but it does not by itself create a governed catalog or cross-source permissions layer.
  • Streaming and NoSQL mismatch: Kinesis Data Streams handles real-time ingestion and DynamoDB is a key-value store, neither aggregates batch S3 plus MySQL into a lake.

Why D is correct

The fraud detection dataset spans S3 transaction logs, S3 customer profiles, and on-premises MySQL tables, so the first need is a unified governed lake before modeling. Lake Formation crawls and registers these sources, centralizes metadata, and enforces security as data is consolidated in S3 for analytics and ML. That aggregation role matches the question, with Lake Formation providing the central foundation.

Why the others are wrong

  • A. EMR Spark jobs process and transform at scale but do not by themselves create a governed central catalog across S3 and MySQL.
  • B. Kinesis Data Streams ingests streaming events, it does not batch-aggregate S3 objects plus on-premises relational tables.
  • C. DynamoDB is a NoSQL operational store, unsuitable as an aggregation lake for logs, profiles, and relational tables.

MLA-C02 exam tip

When the stem stresses many sources plus central governance and S3 plus JDBC, think Lake Formation, when it stresses distributed transform code, think EMR or Glue.

3A company is using Amazon SageMaker to develop ML models. The company stores sensitive training data in an Amazon S3 bucket. The model training must have network isolation from the internet. Which solution will meet this requirement?
  • Run the SageMaker training jobs in private subnets. Create a NAT gateway. Route traffic for training through the NAT gateway.
  • Run the SageMaker training jobs in private subnets. Create an S3 gateway VPC endpoint. Route traffic for training through the S3 gateway VPC endpoint.
  • Run the SageMaker training jobs in public subnets that have an attached security group. In the security group, use inbound rules to limit traffic from the internet. Encrypt SageMaker instance storage by using server-side encryption with AWS KMS keys (SSE-KMS).
  • Encrypt traffic to Amazon S3 by using a bucket policy that includes a value of True for the aws:SecureTransport condition key. Use default at-rest encryption for Amazon S3. Encrypt SageMaker instance storage by using server-side encryption with AWS KMS keys (SSE-KMS).
Answer: B

The short version

B — Private subnets plus S3 gateway endpoint keep training off the internet. Training jobs run without a public route and reach S3 through the AWS private network, so sensitive data never traverses the internet while training still has the access it needs.

Key concepts in this question

  • SageMaker network isolation: training containers can run inside a VPC so administrators control exactly which network paths exist for data access and egress.
  • Private subnets: subnets without an internet gateway route have no direct internet ingress or egress, which is the foundation for isolated training workloads.
  • S3 gateway VPC endpoint: a gateway endpoint adds a private prefix-list route to S3, keeping S3 traffic inside the AWS network without a NAT device or internet gateway.
  • NAT gateway effect: a NAT gateway provides outbound internet access from private subnets, which directly violates a requirement for isolation from the internet.

Why B is correct

B places training in private subnets and adds an S3 gateway VPC endpoint with routes for S3 traffic through that endpoint. That combination preserves isolation because there is no internet gateway or NAT path, while still allowing the job to read sensitive training data from S3 over the private AWS network. It is the standard AWS pattern for isolated SageMaker training that needs S3 access.

Why the others are wrong

  • A. Routes training through a NAT gateway, which creates internet egress and breaks isolation from the internet.
  • C. Uses public subnets with only security-group filtering, so internet routability remains and true isolation is not achieved.
  • D. Adds encryption controls for storage and transport but creates no network isolation and leaves connectivity unchanged.

MLA-C02 exam tip

If the stem says isolated from the internet yet needs S3, choose private subnets plus S3 gateway endpoint and reject any NAT option.

4A company needs to use Retrieval Augmented Generation (RAG) to supplement an open source large language model (LLM) that runs on Amazon Bedrock. The company's data for RAG is a set of documents in an Amazon S3 bucket. The documents consist of .csv files and .docx files. Which solution will meet these requirements with the LEAST operational overhead?
  • Create a pipeline in Amazon SageMaker Pipelines to generate a new model. Call the new model from Amazon Bedrock to perform RAG queries.
  • Convert the data into vectors. Store the data in an Amazon Neptune database. Connect the database to Amazon Bedrock. Call the Amazon Bedrock API to perform RAG queries.
  • Fine-tune an existing LLM by using an AutoML job in Amazon SageMaker. Configure the S3 bucket as a data source for the AutoML job. Deploy the LLM to a SageMaker endpoint. Use the endpoint to perform RAG queries.
  • Create a knowledge base for Amazon Bedrock. Configure a data source that references the S3 bucket. Use the Amazon Bedrock API to perform RAG queries.
Answer: D

The short version

D — Bedrock knowledge bases provide RAG over S3 with least ops. It ingests csv and docx documents and retrieves context for Bedrock models automatically.

Key concepts in this question

  • Retrieval Augmented Generation: Grounding LLM answers by retrieving relevant private documents at query time, avoiding full retraining while improving factuality.
  • Bedrock knowledge bases: Managed RAG that connects an S3 data source, handles parsing and chunking, creates vector embeddings, stores them in a vector store, and exposes Retrieve and RetrieveAndGenerate APIs.
  • Operational overhead: Includes document parsing, embedding pipelines, vector index management, and sync on S3 updates, all of which Bedrock automates.
  • Fine-tuning versus retrieval: Fine-tuning or SageMaker Pipelines retraining changes model weights, which is heavier and unnecessary when the need is document-grounded Q and A.

Why D is correct

Creating a Bedrock knowledge base pointed at the S3 bucket meets the RAG need without building embedding or orchestration code. Bedrock syncs csv and docx content, vectorizes it behind the scenes, and lets the application call the Bedrock API to retrieve relevant chunks and generate grounded answers with the chosen open-source model. This avoids custom pipelines with least overhead.

Why the others are wrong

  • A. SageMaker Pipelines generating a new model is retraining, not retrieval, and adds training plus hosting effort without a vector retrieval layer.
  • B. Manual vectorization plus Neptune requires designing embeddings, index schema, and sync logic, far more ops than a managed knowledge base.
  • C. AutoML fine-tuning plus a SageMaker endpoint bakes documents into weights and needs training and endpoint management, not lightweight RAG querying.

MLA-C02 exam tip

When you see RAG plus S3 documents plus least overhead, choose Bedrock knowledge bases, when you see bake knowledge into weights, think fine-tuning.

5A company plans to deploy an ML model for production inference on an Amazon SageMaker endpoint. The average inference payload size will vary from 100 MB to 300 MB. Inference requests must be processed in 60 minutes or less. Which SageMaker inference option will meet these requirements?
  • Serverless inference
  • Asynchronous inference
  • Real-time inference
  • Batch transform
Answer: B

The short version

B — SageMaker asynchronous inference handles very large payloads with relaxed latency. It queues 100 to 300 MB requests and processes them within the 60-minute budget, unlike real-time or serverless options.

Key concepts in this question

  • SageMaker inference options: Real-time for low-latency small payloads, serverless for bursty small traffic, asynchronous for large payloads with queueing, batch transform for offline bulk jobs.
  • Payload limits: Real-time and serverless enforce small synchronous limits around a few MB, while asynchronous supports very large payloads up to GB scale via S3-backed request and response handling.
  • Latency tolerance: 60 minutes allows queueing, autoscaling, and retries, which fits asynchronous semantics with S3 input and output plus notifications.
  • Production endpoint requirement: The model must stay deployed for on-demand invocation, which rules out pure offline batch transform workflows.

Why B is correct

Average payloads of 100 to 300 MB far exceed synchronous real-time and serverless limits, so those options would reject the requests. Asynchronous inference accepts the large payload via S3, queues it, processes it on a persistent endpoint, and returns results within the generous 60-minute window. It preserves a production endpoint for per-request large inference.

Why the others are wrong

  • A. Serverless inference suits intermittent small-payload traffic and enforces small payload and timeout limits, so it cannot take 100 to 300 MB.
  • C. Real-time inference requires small synchronous payloads and low latency, so huge payloads would fail or time out.
  • D. Batch transform runs offline bulk jobs, not a responsive production endpoint for individual large inference requests.

MLA-C02 exam tip

Memorize the split by size and latency, small plus instant equals real-time, large plus tolerant equals async, bulk offline equals batch transform.

6An ML engineer notices class imbalance in an image classification training job. What should the ML engineer do to resolve this issue?
  • Reduce the size of the dataset.
  • Transform some of the images in the dataset.
  • Apply random oversampling on the dataset.
  • Apply random data splitting on the dataset.
Answer: C

The short version

C — Random oversampling rebalances image classes by duplicating minority examples. It corrects learning bias without discarding valuable majority data or merely augmenting pixels.

Key concepts in this question

  • Class imbalance: When one class dominates, the classifier optimizes for the majority and underlearns minority boundaries, hurting recall on rare classes.
  • Random oversampling: Replicates minority-class samples at random until classes balance, forcing the optimizer to see rare patterns more often.
  • Resampling family: Oversampling adds minority copies, undersampling removes majority rows, and synthetic methods create new variants, all aimed at balanced loss signals.
  • Augmentation versus rebalancing: Simple image transforms improve robustness but do not by themselves fix severe count skew unless applied in a balancing strategy.

Why C is correct

The stem identifies class imbalance as the root cause in an image classification job, so the direct remedy is to rebalance class frequencies. Random oversampling increases minority representation, giving the loss function and gradients sufficient signal from underrepresented classes. It preserves majority data while correcting bias.

Why the others are wrong

  • A. Reducing dataset size discards information and can worsen imbalance or starve the model of training examples.
  • B. Transforming some images is generic augmentation, it improves invariance but does not systematically equalize class counts.
  • D. Random data splitting only partitions train and test sets, it does not change class proportions or loss weighting.

MLA-C02 exam tip

Imbalance equals resample or reweight, oversample minority or undersample majority, splitting and shrinking never fix skew.

7A company receives daily .csv files about customer interactions with its ML model. The company stores the files in Amazon S3 and uses the files to retrain the model. An ML engineer needs to implement a solution to mask credit card numbers in the files before the model is retrained. Which solution will meet this requirement with the LEAST development effort?
  • Create a discovery job in Amazon Macie. Configure the job to find and mask sensitive data.
  • Create Apache Spark code to run on an AWS Glue job. Use the Sensitive Data Detection functionality in AWS Glue to find and mask sensitive data.
  • Create Apache Spark code to run on an AWS Glue job. Program the code to perform a regex operation to find and mask sensitive data.
  • Create Apache Spark code to run on an Amazon EC2 instance. Program the code to perform an operation to find and mask sensitive data.
Answer: B

The short version

B — Glue Sensitive Data Detection on a Glue Spark job masks cards with least code. It natively scans S3 csv files and redacts PII without hand-built regex or servers to manage.

Key concepts in this question

  • PII masking before retraining: Credit card numbers in daily csv files must be detected and masked or tokenized so retraining data stays compliant.
  • Glue Sensitive Data Detection: Built-in classifiers that identify PII such as card numbers in S3 datasets inside Glue ETL, with actions to redact or hash.
  • S3 plus Spark fit: Glue Spark jobs scale across many daily files, integrate directly with S3 as source and sink, and run serverless.
  • Least development effort: Prefers managed detection over custom regex logic and over provisioning EC2 or separate discovery workflows.

Why B is correct

A Glue Spark job with Sensitive Data Detection reads the S3 csv files, applies maintained PII patterns for card numbers, and writes masked output for retraining with minimal custom code. Native S3 integration avoids moving data or managing clusters, and built-in detection avoids writing and testing fragile regular expressions for varied card formats. It stays fully managed and minimizes development.

Why the others are wrong

  • A. Macie discovery finds sensitive data and can aid governance, but masking csv content for a retraining pipeline is better done inline in Glue ETL.
  • C. Hand-coded regex on Glue works but requires building and maintaining patterns, increasing effort versus built-in detectors.
  • D. Spark on EC2 adds infrastructure provisioning and maintenance plus custom logic, the highest effort path.

MLA-C02 exam tip

When you see mask PII in S3 for analytics with least effort, choose Glue Sensitive Data Detection, save regex for custom edge cases.

8A company needs to extract entities from a PDF document to build a classifier model. Which solution will extract and store the entities in the LEAST amount of time?
  • Use Amazon Comprehend to extract the entities. Store the output in Amazon S3.
  • Use an open source AI optical character recognition (OCR) tool on Amazon SageMaker to extract the entities. Store the output in Amazon S3.
  • Use Amazon Textract to extract the entities. Use Amazon Comprehend to convert the entities to text. Store the output in Amazon S3.
  • Use Amazon Textract integrated with Amazon Augmented AI (Amazon A2I) to extract the entities. Store the output in Amazon S3.
Answer: A

The short version

Answer A is correct because Amazon Comprehend now handles semi-structured documents natively with built-in OCR and NLP in one call, so a PDF can go straight to entity extraction and then to S3 with no separate preprocessing stage.

Key concepts in this question

The question asks for LEAST amount of time to extract entities from a PDF for a classifier. Amazon Comprehend provides entity detection including custom entity recognition. Amazon Textract provides OCR and layout extraction but does not by itself do NLP entity extraction. Amazon Augmented AI adds human review, which adds time. Self-managed OCR on SageMaker adds build and maintenance time.

Why A is correct

A uses one managed service to do both text extraction and entity analysis internally. AWS documents that Comprehend can classify and extract entities from PDF, Word and image files directly without first extracting text, with support for real-time and batch jobs. That removes extra hops and polling, so latency is minimal.

Why the others are wrong

B is wrong because running an open source OCR tool on SageMaker means provisioning, tuning and maintaining models and endpoints yourself, which is far slower than a single managed API. C is wrong because it adds an unnecessary two-stage pipeline; you pay for Textract time plus Comprehend time plus orchestration. D is wrong because Textract plus A2I is designed for human-in-the-loop verification of low-confidence extraction, which improves accuracy for sensitive workflows but directly increases time and cost.

MLA-C02 exam tip

When the stem says PDF or scanned image plus LEAST time or LEAST effort for entities or classification, choose Comprehend direct IDP over Textract plus Comprehend. Reserve Textract for when you need raw text, tables or forms, and reserve A2I for when the stem explicitly asks for human review or audit.

9A company shares Amazon SageMaker Studio notebooks that are accessible through a VPN. The company must enforce access controls to prevent malicious actors from exploiting presigned URLs to access the notebooks. Which solution will meet these requirements?
  • Set up Studio client IP validation by using the aws:sourceIp IAM policy condition.
  • Set up Studio client VPC validation by using the aws:sourceVpc IAM policy condition.
  • Set up Studio client role endpoint validation by using the aws:PrimaryTag IAM policy condition.
  • Set up Studio client user endpoint validation by using the aws:PrincipalTag IAM policy condition.
Answer: A

The short version

A — Client IP validation with aws sourceIp blocks stolen presigned URLs. It binds Studio access to the VPN egress range so URLs replayed elsewhere are denied.

Key concepts in this question

  • SageMaker Studio presigned URLs: Pre-authenticated notebook URLs that must be constrained so interception or sharing does not grant broad access.
  • aws sourceIp condition: IAM policy key that restricts requests to an allowed IPv4 or IPv6 range, ideal for VPN-anchored access with known egress addresses.
  • VPC and tag conditions: aws sourceVpc and principal tags constrain origin network or identity attributes, but they do not directly validate the client address tied to VPN use.
  • Defense against replay: Even if a URL leaks, an IP condition causes IAM to deny calls from unexpected networks, containing the blast radius.

Why A is correct

Notebooks are reachable through a VPN, so legitimate Studio traffic presents a predictable source address range. Adding an aws sourceIp condition to the Studio access policy requires that presigned URL calls originate from that range, and malicious replay from outside the VPN fails the condition. This directly enforces the access-control goal with a simple IAM guardrail and no endpoint redesign. VPC checks help elsewhere, but source IP is the settled VPN control.

Why the others are wrong

  • B. aws sourceVpc validates VPC origin, useful for endpoint architectures, but it does not anchor the check to the VPN client address as required here.
  • C. aws PrimaryTag style endpoint validation focuses on role tags, not on where the presigned URL is being replayed from.
  • D. PrincipalTag validation constrains user attributes, which helps authorization granularity but does not stop off-network URL reuse.

MLA-C02 exam tip

Presigned URL abuse plus VPN equals sourceIp condition, think where is the call coming from, not just who the caller is.

10Case study - An ML engineer is developing a fraud detection model on AWS. The training dataset includes transaction logs, customer profiles, and tables from an on-premises MySQL database. The transaction logs and customer profiles are stored in Amazon S3. The dataset has a class imbalance that affects the learning of the model's algorithm. Additionally, many of the features have interdependencies. The algorithm is not capturing all the desired underlying patterns in the data. After the data is aggregated, the ML engineer must implement a solution to automatically detect anomalies in the data and to visualize the result. Which solution will meet these requirements?
  • Use Amazon Athena to automatically detect the anomalies and to visualize the result.
  • Use Amazon Redshift Spectrum to automatically detect the anomalies. Use Amazon QuickSight to visualize the result.
  • Use Amazon SageMaker Data Wrangler to automatically detect the anomalies and to visualize the result.
  • Use AWS Batch to automatically detect the anomalies. Use Amazon QuickSight to visualize the result.
Answer: C

The short version

C — SageMaker Data Wrangler auto-detects anomalies and visualizes them in one place. Its built-in analyses profile aggregated fraud data and surface outliers without separate BI tooling.

Key concepts in this question

  • Post-aggregation quality checks: After joining S3 and MySQL sources, the engineer must spot skewed, missing, or anomalous values before training.
  • Data Wrangler analyses: One-click data quality, bias, and anomaly summaries with histograms, scatter plots, and outlier reports inside Studio.
  • Automatic anomaly signal: Statistical profiling flags distributional shifts and rare values, accelerating fraud feature review.
  • Single-tool visualization: Built-in charts satisfy the visualize the result need without wiring a separate dashboard service.

Why C is correct

Data Wrangler is purpose-built for exactly this post-aggregation exploration step. It connects to the unified dataset, runs built-in analyses that quantify skew and flag anomalies, and renders visuals in the same workflow where transforms are authored. That meets both halves of the requirement, automatic detection plus visualization, in one managed notebook flow. Other options need extra visualization plumbing, while Wrangler offers few-click visuals.

Why the others are wrong

  • A. Athena queries data with SQL but does not automatically detect anomalies or render ML-oriented visual profiles by itself.
  • B. Redshift Spectrum plus QuickSight splits detection and visualization across two services, increasing setup versus a single Wrangler flow.
  • D. AWS Batch runs container jobs, it provides no built-in anomaly detection or charting for tabular ML data.

MLA-C02 exam tip

When one option both profiles anomalies and draws charts inside Studio, choose Data Wrangler over Athena plus QuickSight.

Want the full bank of 106 questions for AWS Certified Machine Learning Engineer - Associate (MLA-C01)? See all practice exams.