ExamPark
Course menuMLA-C02

Exams / MLA-C02 / Study guide

AWS Certified Machine Learning Engineer – Associate (MLA-C02) study guide

What the MLA-C02 exam tests, domain by domain, and how to tell the right answer from the nearly right one. Each topic links to fuller study notes and free practice questions.

Last checked against the official exam guide and AWS documentation · How we write our content

About the MLA-C02 exam

The AWS Certified Machine Learning Engineer – Associate exam checks whether you can take machine learning and generative AI workloads from data to production on AWS: prepare data, train and customize models, deploy them, automate the workflow and keep it monitored and secure. MLA-C02 replaces MLA-C01 and covers both traditional ML with Amazon SageMaker AI and generative AI with Amazon Bedrock. AWS aims it at people with at least a year of hands-on ML engineering on AWS and a year in a related role such as backend developer, DevOps engineer, data engineer or data scientist.
Questions 65: 50 scored and 15 unscored, which aren’t marked
Time 130 minutes in our practice exams; confirm the current length on the official page
Question types Multiple choice (one answer from four) and multiple response (two or more from five or more)
Score Scaled from 100 to 1,000; 720 to pass
Scoring model Compensatory: you need to pass overall, not in each domain
Guessing An unanswered question counts as wrong, and there is no penalty for a wrong answer, so answer everything
Status In beta now; the standard exam starts on January 14, 2027

Check the official certification page for the current price, length, languages and booking options before you book. Beta results arrive later than standard results.

What the exam really tests

Questions describe an ML team’s situation and ask for the AWS feature or setting that solves it with the least effort. A few decisions come up again and again:

  • Managed AI service, Bedrock or SageMaker AI? A pretrained service (Textract, Comprehend, Rekognition) fits a common task; Bedrock fits generative AI with foundation models; SageMaker AI fits training and hosting your own models.
  • Which inference option? Real-time, serverless, asynchronous or batch, chosen by latency, payload size and traffic pattern.
  • Prompt, retrieve or tune? Prompt engineering first, retrieval-augmented generation (RAG) for private or changing knowledge, fine-tuning or distillation for style, format or cost.
  • How do you know it still works? Model Monitor for drift, Clarify for bias and explanations, CloudWatch for Bedrock tokens and latency, Guardrails for safe output.
  • Who and what can reach the data? IAM least privilege, KMS keys, VPC endpoints and network isolation.

How to read a question

  1. Note the data (size, format, labels, sensitivity) and the traffic pattern (steady, spiky, offline).
  2. Note the qualifier: least operational overhead, lowest cost, lowest latency, no code changes.
  3. Rule out answers that build what a managed feature already does, such as custom drift scripts instead of Model Monitor.
  4. Between the rest, pick the one that meets every stated requirement, not just the main one.

How the domains add up

Data Preparation for ML and AI is the largest domain at 28%. Development, Deployment and Operations each weigh 24%, so the exam is balanced: you can’t pass on modeling knowledge alone. Generative AI appears in every domain, from preparing documents for a knowledge base to monitoring token use.

Domain 1: Data Preparation for ML and AI 28%

The largest domain covers getting data ready for both classic ML and generative AI: ingesting and storing data, streaming, vector stores, SageMaker Feature Store, transforming data with Glue, EMR and Data Wrangler, feature engineering, document preparation and embeddings for RAG, masking sensitive data, preparing data for fine-tuning and distillation, data quality and labeling, and detecting bias and class imbalance.

Retrieval-augmented generation with an Amazon Bedrock knowledge base When the knowledge base syncs, documents in S3 are split into chunks, turned into vectors by an embeddings model, and stored in a vector store. At question time, the question is used to retrieve the most similar chunks from the vector store, and the foundation model answers using the question plus those chunks, citing its sources. Knowledge base sync S3 documents source files Embed chunks embeddings model Vector store e.g. OpenSearch At question time User question Retrieve top chunks similarity search Foundation model question + chunks Answer with sources
Retrieval-augmented generation in two phases. When the knowledge base syncs, it splits documents into chunks, turns each chunk into a vector with an embeddings model, and stores the vectors. At question time it retrieves the closest chunks and passes them to the model with the prompt, so answers come from your documents and can cite them, with no fine-tuning.

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

This domain covers choosing and building models: picking foundation models in Bedrock, choosing SageMaker AI algorithms or AWS AI services, RAG retrieval tuning, training jobs, hyperparameter tuning, distributed training, fixing overfitting and underfitting, prompt engineering and fine-tuning, experiment tracking with MLflow, evaluating ML and generative models, and explaining predictions.

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

This domain covers getting models into production and automating the lifecycle: real-time, serverless, asynchronous and batch inference, multi-model endpoints, deploying foundation models with Bedrock, agents and AgentCore, endpoint scaling, containers and VPCs, Bedrock knowledge bases, SageMaker Pipelines and CI/CD, the Model Registry, and safe rollouts with deployment guardrails.

Choosing a SageMaker AI inference option A decision tree. If you need an answer in seconds and traffic is steady, use a real-time endpoint; if traffic is spiky or idle, use serverless inference. If you do not need an answer in seconds and want to score a whole dataset, use Batch Transform; otherwise use asynchronous inference for large or slow requests. Need an answer in seconds? yes no Steady traffic? Score a full dataset? yes no yes no Real-time endpoint steady traffic Serverless inference spiky or idle Batch Transform offline scoring Async inference big or slow jobs
How to pick a SageMaker AI inference option. Real-time endpoints suit steady traffic that needs low latency. Serverless inference scales to zero for spiky or idle traffic but can have cold starts. Batch Transform scores a whole dataset with no endpoint left running, and asynchronous inference queues large payloads or long jobs (up to an hour) and can scale to zero between them.

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

This domain covers running ML and AI safely once it’s live: monitoring data and model drift with Model Monitor, observing foundation models and agents in CloudWatch, controlling inference cost and capacity, IAM least privilege, network isolation and encryption, Bedrock Guardrails and responsible AI, and auditing with CloudTrail.

A 6-week study plan

This plan assumes about an hour on weekdays and a longer, hands-on session at the weekend. If you already train and deploy models on SageMaker AI and build with Bedrock, four weeks may be enough.

Week Focus Do this
1 Baseline and data Take the free practice exam in timed mode. Then data formats and sources for training, Kinesis, Firehose and MSK, Glue and Data Wrangler, Feature Store, feature engineering and data quality.
2 Data for AI and responsible data Vector stores, chunking and embeddings for knowledge bases, PII masking with Comprehend and Macie, data for fine-tuning and distillation, Clarify bias metrics and class imbalance.
3 Development Choosing algorithms, AI services and foundation models, training jobs (script mode, containers, Spot), tuning, distributed training, overfitting, prompt engineering and fine-tuning, MLflow, evaluation and SHAP.
4 Deployment and MLOps Inference options, multi-model endpoints, Bedrock throughput and Custom Model Import, agents and AgentCore, auto scaling, VPC hosting, knowledge base syncs, Pipelines, Model Registry and deployment guardrails.
5 Operations and security Model Monitor (data quality, model quality, bias and attribution drift), Bedrock metrics and agent traces, cost and Inferentia, IAM for SageMaker AI and Bedrock, network isolation, KMS, Guardrails, CloudTrail.
6 Exam practice Full timed practice exams. Review every wrong answer and rebuild the weakest pieces by hand.

How to use practice questions well

  • Learn the matching pairs. “Large payloads, minutes per request” → asynchronous inference; “many small rarely used models” → multi-model endpoint; “right passage ranks low” → reranking.
  • Read the wrong options. They are usually real features that solve a neighboring problem.
  • Don’t memorize questions. The real exam won’t reuse them.
  • Aim for 80% or more across a few full timed practice exams before you book.

Get hands-on

In a personal account with a budget alert set: train a built-in XGBoost model on a small CSV in SageMaker AI, deploy it to a serverless endpoint and call it; create a Bedrock knowledge base from a few documents in S3 and test retrieval; attach a guardrail and watch it block a denied topic. Delete endpoints, vector stores and notebooks afterward, because they bill while they run.

Frequently asked questions

How much math do I need?

Very little. You need to know what metrics mean (precision, recall, F1, AUC, RMSE), what overfitting looks like and how common algorithms behave, not how to derive them. The exam is about engineering choices on AWS.

How is MLA-C02 different from MLA-C01?

MLA-C02 adds much more generative AI: Bedrock model choice, knowledge bases, agents, guardrails, fine-tuning and distillation, and monitoring foundation models, alongside the SageMaker AI content. See our AWS exam changes page for dates.

Should I take the AI Practitioner exam first?

It isn’t required. AI Practitioner (AIF-C01) is a foundational exam about concepts; this one expects you to build and operate solutions. If generative AI is new to you, AIF-C01 is a gentle start.

Are the practice questions the real exam questions?

No. Every question on this site is original, written from the public exam guide and checked against AWS documentation. Using real exam questions breaks the AWS Certification Program Agreement.

Test yourself

The free MLA-C02 practice exam has 65 questions with an explanation for every option. No sign-up needed.

Take the free practice examReview questions untimed

Want to be sure you're ready? The 5 full-length practice sets add 325 new questions, none of them in the free exam. One payment of $14.99, no subscription. See the practice sets