AWS ML Engineer MLA-C02 Study Guide and Exam Tips

AWS is in the middle of swapping this exam over. The MLA-C01 is in use until September 28, 2026. The MLA-C02 is in use from September 29, 2026. Registration for the C02 opened on September 1, so both are bookable right now, and which one you should sit depends only on your date.

Sitting before September 28? Take the C01. It is still the live exam, everything in this guide's domain sections still applies, and a pass is a full three-year certification. You do not lose it when the version changes.

Sitting after that, or still deciding? You are taking the C02. Study the same core ML engineering material, then add the generative AI scope covered below. There is no way to prepare for the C01 and sit it in time if you are starting from scratch this week.

The AWS Certified Machine Learning Engineer Associate is, at heart, a SageMaker exam. That sounds reductive, but it's the single most useful framing for anyone starting to study. Most scored questions touch SageMaker in some form: pipelines, training jobs, feature stores, model registry, endpoint deployment. If you walk in without deep SageMaker fluency, the other AWS services won't save you. What the C02 adds is a second pillar next to it, Amazon Bedrock, and the exam now expects you to be fluent in both.

Exam Format

Both versions score on a scale of 100 to 1,000, and both need a 720 to pass. Scoring is compensatory, so you don't have to clear each domain individually. You only have to pass overall.

The MLA-C01 has 65 questions and a 130-minute time limit. Fifteen of the 65 are unscored pretest items you can't identify, so treat every question like it counts. It costs $150 USD.

The MLA-C02 is launching as a beta. The beta runs 85 questions in 170 minutes and costs $75. AWS has not published a general availability date or the price it will settle at, so check the exam page before you book. Beta exams also don't return a result at the end of the session the way a standard AWS exam does.

The question types changed, and this one catches people. The C01 used four: multiple-choice, multiple-response, ordering, and matching. The C02 exam guide lists two, multiple-choice and multiple-response. If you have been drilling ordering and matching for the C02, you can stop.

AWS recommends at least one year of hands-on experience for both. For the C02 it is more specific: a year with Amazon SageMaker AI, Amazon Bedrock, and other AWS services for ML engineering, plus experience with both traditional ML and generative AI.

What Changed from the MLA-C01

Amazon Bedrock appears nowhere in the C01 task statements. Neither do foundation models, RAG, fine-tuning, or agentic workflows. Every one of them is in the C02 task statements, and one of the four domains is named after them. Generative AI went from absent to graded.

The domain weights moved to make room:

• Data Preparation for ML and AI — 28%, unchanged from the C01
• ML Model and Foundation Model (FM) Development — 24%, down from 26% and renamed
• Deployment and Orchestration of ML and AI Workflows — 24%, up from 22%
• Operating, Monitoring, and Securing ML and AI Solutions — 24%, unchanged in weight, renamed from monitoring and maintenance

The new generative AI content is spread across all four domains rather than quarantined in one. In data preparation you configure vector databases and prepare documents for RAG. In model development you select foundation models from Bedrock, pick a RAG architecture pattern, and identify fine-tuning strategies. In deployment you stand up Bedrock knowledge bases backed by a vector database and deploy agentic workflow infrastructure. In operations you monitor token usage and embedding costs, run Bedrock evaluations, and apply Bedrock Guardrails.

The in-scope service list changed too. Amazon Bedrock AgentCore is new. Amazon SageMaker is now listed as Amazon SageMaker AI. Amazon Inspector, AWS CodeCommit, and AWS CodeConnections were added. Several services dropped off: Amazon Augmented AI, Amazon Fraud Detector, Amazon Kendra, Amazon Q, Amazon Mechanical Turk, and the Lookout family. If your study material still drills those, it is built for the C01.

The Four Domains

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

The largest domain, and the one where candidates underperform most, because they put their study time into model training instead. Data preparation here means getting data from source systems into a format you can train on, at scale, reliably. On the C02 that includes preparing data for foundation models as well as SageMaker algorithms.

The services you need:

• AWS Glue for ETL jobs, crawlers, and the Data Catalog. Know when Glue is the right choice versus running Spark on EMR. Glue handles structured and semi-structured data from S3, RDS, and Redshift well; EMR is for when you need more control over the Spark cluster.
• SageMaker Feature Store for feature management. The exam tests whether you understand why Feature Store exists: it keeps training and inference features consistent, which prevents train-serve skew. It also enables feature reuse across teams and models.
• SageMaker Data Wrangler for visual data preparation. Know its capabilities but also its limits; it's a UI tool for data exploration and transformation, not a replacement for Glue on large-scale ETL.
• Data formats and storage. Parquet versus CSV versus RecordIO. When to use each. SageMaker's built-in algorithms often expect RecordIO or CSV; understanding format conversion is testable.

Feature engineering questions show up here too. Expect questions about handling missing values, encoding categorical variables, scaling numerical features, and when to apply specific transformations. The exam frames these as AWS-service questions, not statistics questions: which SageMaker processing job or Glue transform accomplishes the goal?

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

Model development on this exam is less about ML theory and more about which SageMaker tools handle each stage of the training workflow. You should know:

• Built-in algorithms. XGBoost, Linear Learner, BlazingText, k-NN, LightGBM, and others. The exam tests when to select each one. XGBoost for tabular data with complex feature interactions. Linear Learner for regression and binary classification on large datasets. BlazingText for text classification and Word2Vec. Know the input formats each algorithm expects.
• SageMaker Training Jobs. How to configure instance types, distributed training with data parallelism versus model parallelism, and spot instance training for cost optimization. The cost angle matters; several questions test whether you can pick the approach that balances performance against spend.
• Hyperparameter tuning. SageMaker Automatic Model Tuning runs Bayesian optimization by default. Know how to set objective metrics, define hyperparameter ranges, and interpret tuning job results.
• Model evaluation. Accuracy, precision, recall, F1, AUC-ROC. The exam won't ask you to calculate these from scratch, but it will give you a confusion matrix or metric set and ask which model performs better for a given use case. An imbalanced fraud detection dataset where accuracy is 99% but recall is 12% is a classic setup.

SageMaker Experiments and Model Registry appear in this domain. Model Registry is how you version, catalog, and approve models before they move to production. Expect at least a few questions on the approval workflow.

On the C02 this domain also covers foundation models, which is where the rename comes from. You need to evaluate and select an FM from Bedrock against task requirements, choose a RAG architecture pattern for a given use case, and identify a fine-tuning strategy for a pre-trained model. Know when to leave a model alone and prompt it, when to fine-tune, and when to reach for retrieval instead. Model evaluation widens here too: Bedrock evaluations and Bedrock Prompt Management sit alongside the classic precision and recall questions.

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

Domain 3 is about getting a trained model into production and keeping the pipeline that built it running. Key areas:

• Endpoint types. Real-time endpoints for synchronous inference, batch transform for large offline prediction jobs, asynchronous endpoints for payloads that take longer to process, and serverless endpoints for intermittent traffic. The exam gives you a traffic pattern and expects you to pick the right endpoint type.
• SageMaker Pipelines. The CI/CD backbone for ML on AWS. You define steps (processing, training, evaluation, registration, deployment) as a directed acyclic graph. Expect questions on step dependencies, conditional logic, and how Pipelines integrates with Model Registry.
• Multi-model and multi-container endpoints. When you have dozens of models serving different customer segments, multi-model endpoints let you host them on shared infrastructure. Know the tradeoffs: cold start latency on less-frequently-called models versus cost savings from consolidation.
• Infrastructure as code. CloudFormation and CDK for deploying ML infrastructure. Step Functions for orchestrating workflows that extend beyond SageMaker.
• Generative AI deployment, new on the C02. Creating and managing Bedrock knowledge bases backed by a vector database, deploying agentic workflow infrastructure, and choosing an FM deployment option such as Bedrock Custom Model Import. Amazon Bedrock AgentCore is now an in-scope service.

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

The second-largest domain. It catches candidates who studied training and deployment but skipped the operational side.

• SageMaker Model Monitor. Detects data drift, model quality degradation, bias drift, and feature attribution drift. Know the four monitoring types and when to configure each. Data drift is the most commonly tested: your production input distribution shifts away from your training distribution, and model performance degrades even though nothing in the code changed.
• CloudWatch integration. Metrics, alarms, and logs for SageMaker endpoints. Invocation count, model latency, CPU/GPU utilization. The exam tests whether you can set up alerts that trigger retraining when performance drops below a threshold.
• Security. IAM roles for SageMaker execution, VPC configurations for training jobs and endpoints, encryption at rest and in transit with KMS, and network isolation. VPC endpoint questions are common: how to keep SageMaker traffic off the public internet.
• Cost optimization. Spot instances for training, auto-scaling for endpoints, choosing the right instance family. The exam presents scenarios where you need to reduce cost without breaking SLA commitments.
• AI-specific operations, new on the C02. Token usage, embedding computation costs, and vector database storage are all cost levers you are expected to know. On the monitoring side, Bedrock evaluations and Bedrock AgentCore Observability. On the security side, Bedrock Guardrails for responsible AI policy, and picking the right credential type for FM access.

What to Study First

Start with SageMaker. Specifically, go through the SageMaker Developer Guide sections on training jobs, built-in algorithms, and endpoints. Then move to Pipelines and Model Monitor. Those four components cover roughly half the exam.

After SageMaker, pick up Glue and Feature Store for the data preparation domain. Then fill in Step Functions, CloudFormation, IAM policies, and KMS for the orchestration and security questions. For the C02, add Bedrock as a second block of roughly the same size as Pipelines and Model Monitor.

A reasonable timeline for someone with existing ML knowledge and some AWS experience:

• Weeks 1-2: SageMaker core: training jobs, built-in algorithms, hyperparameter tuning, model evaluation. Run at least one training job end-to-end in the console or SDK.
• Weeks 3-4: Data preparation. Glue ETL, Feature Store, data formats, processing jobs. Feature engineering patterns. For the C02, add vector databases and document preparation for RAG.
• Weeks 5-6: Deployment and orchestration. Endpoint types, Pipelines, Model Registry, Step Functions. Set up a simple pipeline that trains, evaluates, and conditionally deploys.
• Week 7: C02 only. Bedrock end to end. Pick a model, build a knowledge base over a vector store, put a guardrail on it, and run an evaluation. An afternoon of this beats a week of reading about it.
• Weeks 8-9: Monitoring and security. Model Monitor setup, CloudWatch alarms, IAM roles, VPC configuration. Take full-length practice exams and review every wrong answer.

If you already work with SageMaker daily, compress this to 4-5 weeks. If you're coming from a data science background with minimal AWS experience, budget 10 weeks and spend extra time on the infrastructure and security topics. If you use Bedrock at work already, the C02 costs you almost no extra preparation over the C01.

Common Mistakes

Three patterns show up repeatedly in post-exam reports from candidates who failed:

Underestimating Domain 1. Data preparation is 28% of the exam, but many candidates treat it as the "easy" domain and allocate the least study time. Glue ETL and Feature Store questions require specific service knowledge, not general data engineering intuition.

Studying ML theory instead of AWS implementation. You don't need to derive backpropagation. You need to know that SageMaker's XGBoost implementation expects CSV or LibSVM input, supports Pipe mode for streaming large datasets, and can run distributed training across multiple instances. The exam tests AWS service knowledge, not textbook ML.

Ignoring Model Monitor. Candidates who work primarily on model development often skip monitoring entirely. Four types of monitoring, each with specific configuration requirements, across 24% of the exam. Don't skip it.

Test Day Strategy

Both versions give you about two minutes per question: 65 questions in 130 minutes on the C01, 85 in 170 on the C02 beta. That sounds generous until you hit a multi-response question with six options and a detailed scenario. Flag long questions and come back to them. These exams are not adaptive, so you can review and change answers before you submit.

Read the scenario carefully before looking at the answer choices. AWS exam questions often include details that rule out two options immediately. A question that mentions "intermittent traffic with unpredictable spikes" is steering you toward serverless or asynchronous endpoints, not real-time. A question about "retraining when data distribution changes" is pointing at Model Monitor with CloudWatch alarms, not a cron job.

When two answers both seem correct, look for the one that's more operationally sound. AWS consistently favors managed services over custom solutions, automation over manual processes, and least-privilege security over convenience. If one answer uses a native SageMaker feature and the other involves stitching together Lambda functions and custom scripts, the SageMaker answer is almost always right.