MLA-C02 Domains Explained

The short version

MLA-C02 has four domains: data preparation (28%), model and foundation model development (24%), deployment and orchestration (24%), and operating, monitoring, and securing (24%). The weightings come from AWS’s MLA-C02 exam guide, read on October 18, 2026. The guide itself says its list is not comprehensive, so treat the topics below as the core, not the whole.

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

What it covers. Three tasks: collect and store data, transform it and engineer features, and validate quality and manage bias.

In practice.

  • Pulling data from sources such as S3, RDS, DynamoDB, and OpenSearch, and choosing storage by cost, performance, and compliance.
  • Streaming ingestion with Kinesis, Apache Flink, or Apache Kafka, and picking formats such as Parquet, JSON, CSV, or ORC for the access pattern.
  • Transforming with AWS Glue, DataBrew, Spark on EMR, or SageMaker Data Wrangler, and managing features in SageMaker Feature Store.
  • Feature engineering: scaling, normalization, binning, log transforms.
  • Newer for this version: vector databases (OpenSearch Service, RDS with pgvector, S3), embedding models, tokenization, chunking documents for RAG, masking and anonymizing data, and preparing data for fine-tuning.
  • Data quality, labeling, outliers, missing values, deduplication, bias, and class imbalance.

Typical traps.

  • Choosing a batch tool when the question needs state across a stream.
  • Mixing up similar data-prep tools. Know what each one is for.
  • Picking a vector store without reading the requirements: joins with relational data, keyword matching, query volume, and cost point to different answers.

What to practice. Build a small pipeline from raw files to a feature set. Try one RAG data prep run, with chunking and metadata.

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

What it covers. Choose approaches, train and customize, then analyze and evaluate.

In practice.

  • Selecting Bedrock foundation models and comparing custom models, managed services, pre-trained models, and FMs on interpretability, latency, and cost.
  • Choosing RAG patterns and applying AWS AI services such as Textract, Rekognition, Comprehend, and Transcribe.
  • Using SageMaker AI built-in algorithms and script mode, hyperparameter tuning, early stopping, and distributed training.
  • Preventing overfitting, underfitting, and catastrophic forgetting.
  • Fine-tuning and prompt engineering as customization.
  • Evaluation: baselines, drift detection, shadow variants, explainability, human review, LLM-as-a-judge, and metrics such as BLEU, ROUGE, and BERTScore.

Typical traps.

  • Using accuracy on a heavily imbalanced dataset, when precision, recall, or a precision-recall curve tells the real story.
  • Reaching for fine-tuning when prompting or retrieval would do, or the reverse.
  • Treating a managed AI service as a modeling job. If a service already solves the problem, the exam often wants it.

What to practice. Tune one model, read the metrics, and write down why you chose them. Compare prompting, RAG, and fine-tuning for a single use case on paper.

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

What it covers. Deployment infrastructure, provisioning, and CI/CD.

In practice.

  • Choosing compute, deployment targets, and inference style: real-time or batch, multi-model or multi-container.
  • Deploying models built outside AWS, and configuring FM hosting.
  • Deploying agents, agent state, and agentic workflow infrastructure.
  • Containers, SageMaker AI endpoints in a VPC, deploying through the SDK, CLI, or Boto3, and picking metrics for auto scaling.
  • Bedrock knowledge bases, retrieval pipelines, and GPU scaling.
  • Rollout and rollback strategies, CodePipeline and related services, retraining mechanisms, model registries, prompt management, and testing in pipelines.

Typical traps.

  • Missing the one constraint that decides: quota limits, latency, GPU needs, or a private network.
  • Mixing up deployment strategies. Know how canary, linear, all-at-once, and rolling differ in capacity and risk.
  • Confusing which scaling metric belongs to which endpoint type.

What to practice. Deploy a model, put it in a VPC, attach auto scaling, and trigger a rollout through a pipeline. Then roll it back.

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

What it covers. Monitoring models in production, managing cost, and security.

In practice.

  • Monitoring performance and data drift, A/B testing, and watching agent behavior.
  • Choosing inference instance families, using CloudWatch, X-Ray, and dashboards, and setting up purchasing options and cost quotas.
  • Watching FM cost drivers such as token use, embedding computation, and vector storage.
  • Least-privilege IAM, auditing with CloudTrail and Config, VPCs, subnets, and security groups.
  • Choosing credential types for FM access and applying safeguards such as Bedrock Guardrails.

Typical traps.

  • Assuming logging is on by default. Some request-level activity must be switched on deliberately.
  • Choosing a broad IAM policy for convenience when the question asks for least privilege.
  • Confusing a monitoring tool for infrastructure with one for model quality.

What to practice. Set up a monitoring baseline and an alarm, read an IAM policy line by line, and review which API calls end up in your logs.

How the domains connect

The exam likes scenarios that cross domains. A single question may start with data in a stream, ask about a model choice, and finish with a deployment constraint. Practice by following one project through all four domains rather than studying each in isolation.

More guides for MLA-C02

Next step

Take the sample questions to see which domain feels weakest, then plan your weeks with the study plan. Our MLA-C02 practice tests are built around these same four domains, so you can focus where it counts.

Frequently asked questions

Which domain should I study first?

Start with data preparation. It carries the highest weighting, and later domains build on it.

Is the weighting by number of questions?

AWS gives weightings as a share of scored content, so heavier domains have more scored questions, but the exact counts are not published.

Does MLA-C02 test code?

The exam guide mentions programmatic deployment with the SageMaker AI SDK, CLI, and Boto3. The exam uses multiple-choice and multiple-response questions only, so you answer by choosing, not by writing code.

Try these practice exams