MLA-C02 AWS Machine Learning Engineer – Associate (new)
- Vendor
- AWS
- Level
- Associate
- Questions
- 65
- Time
- 130 min
- Domains
- 4
- In the bank
- 529
Practice 529 questionsTake a timed mock examDay-by-day study plan
New exam version: MLA-C02 is in beta (English since 2026-09-29) and becomes generally available on 2027-01-14, replacing MLA-C01. It keeps MLA-C01's four domains and adds generative AI: Amazon Bedrock, RAG and vector stores, agents, guardrails and foundation model evaluation and cost. This bank shares the still-valid MLA-C01 questions and adds MLA-C02-only questions. MLA-C01 exam page.
Guides for this exam
The exam at a glance
| Level | Associate |
|---|---|
| Questions | 65 |
| Time | 130 minutes (about 2 minutes per question) |
| Passing mark | 720 on a scaled score of 100–1,000 |
Domains and weights
| Domain | Weight | Practice questions here |
|---|---|---|
| D1 · Data Preparation for ML and AI | 28% | 140 |
| D2 · ML Model and Foundation Model (FM) Development | 24% | 130 |
| D3 · Deployment and Orchestration of ML and AI Workflows | 24% | 115 |
| D4 · Operating, Monitoring, and Securing ML and AI Solutions | 24% | 144 |
62 of the 529 practice questions (12%) ask for more than one answer.
D1 · Data Preparation for ML and AI 28%
Data preparation questions describe a dataset and a problem that shows up later: strong validation scores but weak production results, skewed scaling, a leaky evaluation split, sensitive fields, or a vector index that rejects documents. The job is to trace the symptom back to how the data was built and fix it at that stage. Distractors often treat the symptom with access controls or more compute.
What to watch for
- When offline scores are much better than live results, check what the training data could see. Features joined from a later time, or near-duplicates spread across splits, inflate evaluation.
- Choose a scaling method with the value distribution in mind. Under some methods, a few extreme values squeeze most of a feature into a narrow range.
- For RAG data, the vector index field must match the embedding model's output size. If the model or its dimension changes, plan for a new index and a fresh ingestion run.
Tasks in the official exam guide (3)
D2 · ML Model and Foundation Model (FM) Development 24%
Classic models and foundation models share this domain. A scenario sets a goal (adapt a large model on limited hardware, satisfy a regulator, avoid one costly kind of error, label images in a certain way) and asks which technique, metric or approach fits. Good answers respect the stated constraint; weak ones chase overall accuracy or set compliance aside.
What to watch for
- Map the costly mistake to a metric. If false alarms hurt most, one metric matters; if missed cases hurt most, its counterpart does. Plain accuracy rarely settles the choice.
- When a large model cannot be fully fine-tuned on the hardware available, look at parameter-efficient methods that train small adapters on a frozen, possibly quantized, base.
- Strong overall results can hide a weak customer segment, and regulators may require explanations of single decisions. Check segment metrics and interpretability before picking a model on accuracy alone.
Tasks in the official exam guide (3)
D3 · Deployment and Orchestration of ML and AI Workflows 24%
These questions follow a model or FM workflow from build to serving: pulling images across accounts, shifting traffic safely, chaining build, approval and release stages, scaling an asynchronous endpoint, or waiting on a long Amazon Bedrock job. Each asks for the mechanism that does this with the least custom code. Options that add polling scripts, manual steps or unrelated networking changes are usually wrong.
What to watch for
- Cross-account access to an ECR image needs permission on both sides: a policy on the repository in the owner account and permissions on the role that pulls it.
- Know how traffic-shifting configurations differ: one moves a small slice and then the rest, another moves in several equal steps. Both can roll back automatically.
- For a job that runs for hours, prefer a Step Functions integration that waits for completion and handles failure. For asynchronous endpoints that scale to zero, check what starts the first instance.
Tasks in the official exam guide (3)
D4 · Operating, Monitoring, and Securing ML and AI Solutions 24%
Operations questions take a running ML or generative AI system and ask how to keep it healthy, affordable and locked down. Typical stems mention throttled model calls discovered too late, a job denied access to its data, costs that must be split by project, or a cheaper model that might be good enough. Correct answers rely on metrics, alarms, roles and tags rather than manual checks.
What to watch for
- A SageMaker job runs as its execution role, so the permissions of the person who launched it do not carry over. Data and image access belong on that role.
- For endpoint health, watch errors and latency. For Amazon Bedrock, the service's CloudWatch throttling metrics can drive an alarm that warns the team before users report failures.
- One tagging standard can serve both cost allocation and access control. A smaller model is a valid saving once it meets the quality bar on your own labeled test set.
Tasks in the official exam guide (3)
A study order that works
The 35-day study plan for MLA-C02 takes the domains in this order. Each day has a short reading, key terms and a quiz from the bank.
- Days 1–9: D1 · Data Preparation for ML and AI
- Days 10–18: D2 · ML Model and Foundation Model (FM) Development
- Days 19–26: D3 · Deployment and Orchestration of ML and AI Workflows
- Days 27–34: D4 · Operating, Monitoring, and Securing ML and AI Solutions
- Day 35: Final Day: Full Practice Exam and Exam-Day Checklist
Recent changes to these questions
- : 13 answer choice sets changed, 8 answer keys corrected, 13 explanations rewritten
- : 60 questions added
- : 9 explanations rewritten
- : 7 answer choice sets changed, 7 answer keys corrected, 7 explanations rewritten
- : 21 explanations rewritten
About these questions
Every question is original, written from the public exam guide and checked against the vendor's own documentation. None comes from a real exam. Each one has an explanation of why the answer is right and why the other options are wrong. How the questions are written.