MLA-C02 study guide: domains, format and a 4-week plan

MLA-C02 is the updated AWS Certified Machine Learning Engineer – Associate exam. It validates that you can build, deploy, operate and secure both traditional ML models and foundation model workloads on AWS, including RAG and agentic workflows, and it is aimed at engineers with at least a year of hands-on SageMaker AI and Amazon Bedrock experience. It replaces the English MLA-C01 from 29 September 2026 and launches as a beta: 85 multiple-choice and multiple-response questions in 170 minutes, in English, at a Pearson VUE test centre or online.

MLA-C02 at a glance

Details as of September 2026 — confirm on the official exam page before booking.

How the switch from MLA-C01 works

AWS states that the last day to take MLA-C01 in English is 28 September 2026. Beta registration for MLA-C02 opened on 1 September 2026 and beta delivery begins on 29 September 2026. MLA-C01 stays available in Japanese, Korean and Simplified Chinese until MLA-C02 reaches general availability, and the general availability dates were still listed as to be confirmed at the time of writing.

The two versions share the same four-domain structure, and the weights barely move: model development drops from 26% to 24% and deployment rises from 22% to 24%. The real change is inside the tasks. The comparison section at the end of the C02 exam guide lists a large block of new content on vector databases, embeddings, RAG document preparation, foundation model selection and customisation, agents, generative AI evaluation and token-level cost monitoring. A few MLA-C01 items were removed, including edge optimisation with SageMaker Neo and bring-your-own-container. If you studied for MLA-C01, treat the generative AI material as new study, not revision.

Who this exam is for

The exam guide asks for at least one year of experience using SageMaker AI, Amazon Bedrock and other AWS services for ML engineering, plus a year in a related role such as backend developer, DevOps engineer, data engineer or data scientist, with experience of both traditional ML and generative AI. It also expects comfort with CI/CD, infrastructure as code, data formats and AWS security basics.

There is no prerequisite certification. The guide is explicit that designing full end-to-end ML architectures and setting ML strategy are out of scope: this is an implementer's exam. If you want the generative AI application angle at professional depth, see the [AIP-C01 study guide](/blog/aip-c01-study-guide).

What the exam covers

Domain 1: Data Preparation for ML and AI (28%)

The largest domain. It covers ingesting and storing data (S3, EFS, databases, streaming with Kinesis or Kafka, and file formats such as Parquet and CSV chosen by access pattern), transforming it with AWS Glue, DataBrew, Spark on EMR or Data Wrangler, and managing features in SageMaker Feature Store. New for C02 are embeddings, vector database configuration, chunking and metadata extraction for RAG, masking and anonymisation, and preparing data for fine-tuning and distillation. It ends with data quality, labelling, class imbalance and bias in the data.

Domain 2: ML Model and Foundation Model (FM) Development (24%)

This domain asks you to choose an approach (a built-in algorithm, a managed AI service, a Bedrock foundation model, fine-tuning or RAG) against interpretability, latency and cost. It covers SageMaker AI training: script mode, automatic model tuning, early stopping, distributed training, and preventing overfitting, underfitting and catastrophic forgetting. Evaluation now spans classic metrics and drift baselines alongside BLEU, ROUGE, BERTScore, human evaluation and LLM-as-a-judge.

Domain 3: Deployment and Orchestration of ML and AI Workflows (24%)

Expect questions on choosing real-time, serverless, asynchronous or batch inference; multi-model and multi-container endpoints; importing models built elsewhere; and deploying agents and Bedrock knowledge bases. The CI/CD half covers the AWS developer tools, configuring training and inference jobs, Model Registry and MLflow for versioning, automated retraining, prompt management and refreshing a knowledge base on a schedule. Know which rollout strategy gives a safe rollback.

Domain 4: Operating, Monitoring, and Securing ML and AI Solutions (24%)

This domain covers watching models in production (data and concept drift, A/B tests, agent failures such as tool errors or truncated streaming), controlling cost (instance families, purchasing options, token and embedding costs) and securing the workload (least-privilege IAM, VPC isolation for endpoints, CloudTrail and Config for audit, image scanning in pipelines, Bedrock Guardrails and choosing the right credential type for Bedrock).

A 4-week study plan

Domain 1 carries the most weight; the other three are equal. Split the time accordingly, and give extra hours to the generative AI tasks if you come from MLA-C01.

Week 1: Domain 1. Work through ingestion, formats and transformation services, then spend a full session on RAG data preparation: chunking, metadata and embeddings. Finish with bias, imbalance and data quality, and a Domain 1 mixed paper.

Week 2: Domain 2. Revise SageMaker AI training options and tuning, then foundation model selection and customisation. Build a comparison table of when to prompt, retrieve or fine-tune. Practise evaluation metrics until you can say what each one measures. Take a Domain 2 paper.

Week 3: Domain 3. Cover inference options and when each fits, then the pipeline and versioning tooling. Deploy a small endpoint and a knowledge base if you can: seeing the configuration makes the options stick. Take a mixed paper and topic papers on anything below your target.

Week 4: Domain 4, then full mocks. Spend two or three days on monitoring, cost and security. Then sit a full-length mock at beta length, 85 questions in 170 minutes, review every explanation, and sit a second mock after fixing your weakest topics.

Where candidates lose marks

• Studying from MLA-C01 material alone. Much of the generative AI scope is new in C02. Check any resource against the C02 exam guide before relying on it.
• Choosing the wrong inference option. Real-time, serverless, asynchronous and batch inference each fit a specific combination of payload size, latency and traffic pattern. Many questions turn on one detail such as an intermittent workload or a large payload.
• Confusing data drift with concept drift. One is a change in inputs, the other a change in the relationship between inputs and outcome. They call for different monitoring and different responses.
• Mixing up customisation methods. Fine-tuning, continued pre-training, distillation and RAG solve different problems and need differently shaped data. Know what training data each expects.
• Multiple-response scoring. Credit is given only when every correct option is chosen. Read how many answers are required.
• Learning product names instead of capabilities. SageMaker AI features that older MLA material leans on, such as SageMaker Clarify and Model Monitor, are no longer open to new customers, and the C02 guide describes skills such as drift detection and bias metrics rather than naming those tools. If you understand the capability (bias measurement, drift baselines, human review) you will recognise the right answer however it is presented.

Practising for MLA-C02 on CertifyAIx

CertifyAIx has 901 original MLA-C02 questions written against the C02 exam guide, covering every domain and explained option by option. See [how CertifyAIx questions are written](/blog/how-certifyaix-practice-questions-are-written) for our method.

• 3 full-length mock exams, 85 questions in 170 minutes, matching the beta
• 38 mixed domain papers and 42 topic papers of up to 25 questions each (60 minutes per 25)
• Explanations for every option, with pointers to the relevant AWS documentation
• Progress tracked by domain and topic

Start with the [MLA-C02 practice questions](/certifications/aws-machine-learning-engineer-associate-mla-c02).

Frequently asked questions

Can I still take MLA-C01?

AWS lists 28 September 2026 as the last day to take MLA-C01 in English. From 29 September 2026 the English exam is MLA-C02, starting with the beta. MLA-C01 remains available in Japanese, Korean and Simplified Chinese until MLA-C02 reaches general availability.

How difficult is MLA-C02?

It is an associate exam, but a broad one: it spans classic ML engineering and foundation model workloads. AWS recommends a year of hands-on SageMaker AI and Bedrock experience, and the scenarios assume you have configured these services yourself.

How long should I prepare for MLA-C02?

With the recommended experience, four to eight weeks is a realistic range. If you have worked mainly with traditional ML, allow more time for the RAG, agent and evaluation tasks, which make up a large share of the new content.

Why is the beta longer and cheaper than the standard exam?

AWS uses beta exams to check how new questions perform before they go into the standard exam, which is why the beta runs to 85 questions against the 65 in the exam guide. AWS prices the MLA-C02 beta at $75, half the $150 fee for MLA-C01.

Are CertifyAIx questions real MLA-C02 exam questions?

No. Every CertifyAIx question is original and written against the public exam guide. We never use content from a live or beta exam.

CertifyAIx is independent and not affiliated with Amazon Web Services. AWS Certified Machine Learning Engineer – Associate is a trademark of its owner. All CertifyAIx practice questions are original.