AWS Certified Machine Learning Engineer - Associate

About the AWS Certified Machine Learning Engineer - Associate (MLA-C02) exam

The exam at a glance

Questions

85 questions in the MLA-C02 beta, multiple choice and multiple response; the guide sets 65 (50 scored, 15 unscored) at general availability

Time

170 minutes in the beta

Pass mark

720 on a scale of 100 to 1,000

Exam fee

US$75 at beta pricing (MLA-C01 was US$150); varies by country

Valid for

3 years

Where you sit it

Pearson VUE test centre or online proctored; English only during the beta

Prerequisites

None required; AWS recommends at least 1 year with SageMaker AI, Amazon Bedrock and other AWS ML services

As published in the vendor’s exam guide at the time of writing. Fees vary by country and change; the vendor’s page is the source of truth.

What is on the exam, domain by domain

1. Data Preparation for ML and AI

   The largest domain. Ingesting and storing data (S3, streaming with Kinesis and Kafka, formats such as Parquet), transforming it with Glue, Data Wrangler and Spark, feature engineering and Feature Store, vector stores and embeddings, preparing documents for RAG and data for fine-tuning, and checking quality and bias with SageMaker Clarify.

   28% of the paper

2. ML Model and Foundation Model (FM) Development

   Choosing between AI services, built-in algorithms, custom models and foundation models in Amazon Bedrock; training and tuning in SageMaker AI; customising FMs and retrieval; and evaluating models, including drift baselines, shadow tests and RAG and GenAI evaluation.

   24% of the paper

3. Deployment and Orchestration of ML and AI Workflows

   Real-time, serverless, asynchronous and batch inference, endpoints in a VPC and auto scaling, containers, Bedrock knowledge bases and agents, and CI/CD with CodePipeline and SageMaker Pipelines, Model Registry and prompt management.

   24% of the paper

4. Operating, Monitoring, and Securing ML and AI Solutions

   Monitoring models, data and agents in production with Model Monitor and CloudWatch, controlling inference and infrastructure cost, and securing ML systems with IAM, VPC isolation, encryption, logging and Amazon Bedrock Guardrails.

   24% of the paper

How to prepare

1. 1.Read the MLA-C02 exam guide, its in-scope and out-of-scope service lists and AWS's comparison with MLA-C01. Study to the task statements, not to everything SageMaker AI can do.
2. 2.Build the core loop once in your own account: data in S3, a SageMaker AI training job, a model in the Model Registry, a real-time endpoint with auto scaling, and Model Monitor on its traffic.
3. 3.Do the same for generative AI: call a model in Amazon Bedrock, build a knowledge base over your own documents, add a guardrail, and run a Bedrock model evaluation. MLA-C02 asks about these as often as traditional ML.
4. 4.Learn the trade-offs that decide most questions: which inference option fits the traffic, which data store fits the access pattern, fine-tuning against RAG against prompting, and the cheapest compute that meets a requirement.
5. 5.Two timed practice papers above 80% before you book. The beta is 85 questions in 170 minutes, so pacing matters.

Sample questions

Three questions in the exam’s style, written by Square 1 against the published domains. They are ours, not the vendor's. The answer and the reasoning are under each one; the practice papers on this page draw from a bank written the same way, with an explanation on every question.

1. Question 1 · Data Preparation for ML and AI

   An online retailer wants shoppers to type descriptions such as 'blue waterproof hiking jacket' and get matching product photos. The photos have no captions or tags. Which solution meets this requirement?

   1. AExtract text from the photos with Amazon Textract and embed it with Titan Text Embeddings V2
   2. BEmbed the photos and the typed queries with Amazon Titan Multimodal Embeddings G1
   3. CEmbed the photos with Titan Multimodal Embeddings G1 and the queries with Titan Text Embeddings V2
   4. DLabel each photo with Amazon Comprehend and embed the resulting labels as text

   Answer: B. Titan Multimodal Embeddings G1 places text and images in one shared vector space, so a text query can be compared directly with photo vectors. Embedding queries with a different model puts them in a different space, so the comparison is meaningless. Product photos hold little text for Textract, and Comprehend analyses text, not images.

2. Question 2 · Deployment and Orchestration of ML and AI Workflows

   An internal tool calls a small CPU-based text classifier a few dozen times a day at random times. Users accept a short delay on the first request after an idle period. The team wants the lowest cost and no instance management. Which option should the team choose?

   1. AA real-time endpoint on the smallest instance
   2. BAsynchronous Inference with an Amazon SNS topic
   3. CServerless Inference with a suitable memory size
   4. DA batch transform job every 15 minutes

   Answer: C. Serverless Inference scales to zero when idle, bills only for use and needs no instance choice, which suits intermittent traffic that tolerates cold starts. A real-time endpoint bills for its instance around the clock even when no requests arrive.

3. Question 3 · Operating, Monitoring, and Securing ML and AI Solutions

   A team is concerned that someone could insert mislabelled records into the S3 prefix that feeds its weekly retraining pipeline. Which combination of controls best mitigates this data poisoning risk?

   1. ALimit writes to the ingestion role, enable versioning and log S3 data events
   2. BEncrypt the prefix with SSE-S3 and enable S3 Transfer Acceleration for uploads
   3. CScan the training container with Amazon Inspector before each retraining run
   4. DAdd a guardrail with contextual grounding checks to the deployed endpoint

   Answer: A. Restricting who can write, keeping prior versions and recording object-level activity limit and expose tampering with training data. Image scanning finds package CVEs, not bad data, and contextual grounding checks apply to generated answers, not training records.

Questions people ask before they book

Should I sit MLA-C01 or MLA-C02?

In English, MLA-C02. AWS ended MLA-C01 in English on 28 September 2026 and opened the MLA-C02 beta on 29 September. MLA-C01 stays available in Japanese, Korean and Simplified Chinese until MLA-C02 reaches general availability.

How much does it cost and how long is it valid?

US$75 at the MLA-C02 beta price, against US$150 for MLA-C01, varying by country. The certification is valid for three years, whichever version you pass.

Do I need the AI Practitioner or Cloud Practitioner first?

No, neither is required. They help if you are new to AWS or to AI, but people who already train or deploy models on AWS usually go straight to the ML Engineer Associate.

Which languages can I take it in?

The MLA-C02 beta is in English only. AWS says Japanese, Korean and Simplified Chinese will follow at general availability, and until then MLA-C01 remains available in those three languages.

How long should I study?

Eight to twelve weeks part-time for most people who already use AWS for data or ML work, longer if SageMaker AI and Amazon Bedrock are new to you.