MLA-C02 Cheat Sheet: Key Facts to Review

How to use this sheet

This is a review aid in our own words, not a full syllabus and not a list of exam questions. It does not cover everything on the exam. AWS’s exam guide says its own list is not comprehensive. Use it after you have studied, to spot gaps. Facts below were read on the AWS pages on October 22, 2026.

Exam facts at a glance

Item	Detail
Level	Associate
Scoring	Scaled 100 to 1,000; minimum passing score 720 (exam guide)
Scored questions	50 scored plus 15 unscored per the exam guide; beta listed as 85 questions in 170 minutes
Types	Multiple choice and multiple response
Guessing	No penalty; blanks count as wrong
Scoring model	Compensatory: pass the exam overall, not each section
Domains	28% data, 24% model/FM development, 24% deployment, 24% operations and security
Validity	3 years
Retake	14 calendar days after a fail

Domain 1: Data preparation (28%)

Key concepts

• Pick storage and ingestion by access pattern: batch files, streams, relational, or search data.
• Columnar formats such as Parquet suit analytics reads. CSV and JSON are simple but heavier to scan.
• Feature engineering: scaling, normalization, binning, encoding, log transforms.
• New in this version: embeddings, vector stores, chunking and metadata for RAG, masking and anonymizing data.
• Data quality: missing values, outliers, deduplication, bias, class imbalance.

Easily confused pairs

• Kinesis Data Streams is for custom stream consumers and replay. Data Firehose is for delivering streaming data to a destination with little code.
• Glue DataBrew is for visual data preparation. SageMaker Data Wrangler is for preparation inside the ML workflow.
• Feature Store keeps reusable features. A data lake holds raw data.
• Macie finds sensitive data in S3. Comprehend finds entities and sentiment in text.

Decision rules

• If the scenario needs state across events in a stream, think a stream processor, not a batch job.
• If the data has relations and vectors together, think a relational database with a vector extension. If it needs search plus vectors, think a search service.
• If one class is rare, avoid plain accuracy. Think resampling, weighting, or better metrics.

Domain 2: Model and FM development (24%)

Key concepts

• Overfitting: great on training data, weak on new data. Underfitting: weak on both.
• Hyperparameter tuning, early stopping, distributed training.
• Prompting, RAG, and fine-tuning are three levels of customization.
• Metrics: precision, recall, F1, AUC for classification; RMSE for regression; BLEU, ROUGE, and BERTScore for text.
• Evaluation also includes human review and LLM-as-a-judge approaches.

Easily confused pairs

• Precision is “of what I flagged, how much was right”. Recall is “of what exists, how much did I find”.
• RAG adds fresh knowledge at question time. Fine-tuning changes model behavior or style.
• Random Cut Forest is for anomaly detection. k-means is for clustering.
• Managed AI services (Textract, Rekognition, Transcribe) versus training your own model.

Decision rules

• If a managed service already solves the task, the answer often uses it.
• If the knowledge changes often, think RAG before fine-tuning.
• If the cost of a miss is high, favor recall. If false alarms are costly, favor precision.

Domain 3: Deployment and orchestration (24%)

Key concepts

• Inference styles: real-time, asynchronous, batch, and serverless. Match them to latency, payload size, and traffic.
• Multi-model endpoints host many models on shared resources. Multi-container endpoints combine containers.
• Rollouts: all-at-once, canary, linear, and rolling differ in risk and capacity.
• Model registry stores versions and approval status. Pipelines automate build, test, and release.
• Bedrock knowledge bases, agents, and GPU scaling are in scope.

Decision rules

• If traffic is spiky or rare, think serverless or asynchronous. If it is steady and latency-sensitive, think real-time with auto scaling.
• If private access is required, think VPC settings and endpoints.
• If a quota limits spare capacity during a deployment, check which rollout strategy needs extra capacity.
• If something must roll back quickly, favor strategies with a ready previous version.

Domain 4: Operating, monitoring, and securing (24%)

Key concepts

• Drift: data drift (inputs change) versus model quality drift (predictions degrade). Monitoring needs a baseline.
• CloudWatch for metrics and alarms, CloudTrail for API activity, X-Ray for tracing, Config for resource state.
• Cost levers: instance choice, purchasing options, token use, embeddings, vector storage.
• Security: least-privilege IAM, encryption at rest and in transit, VPC isolation, Bedrock Guardrails.

Easily confused pairs

• CloudWatch shows how things perform. CloudTrail shows who did what.
• Management events are logged by default in CloudTrail. Data-plane activity such as model invocation generally needs to be enabled on purpose.
• Guardrails filter content and protect sensitive data. IAM controls who can call the model.

Decision rules

• If the question says “least privilege”, pick the narrowest policy that works.
• If it says “audit”, think CloudTrail. If it says “alert”, think CloudWatch alarms.
• If it mentions model quality in production, think model monitoring, not only instance metrics.

Common traps

1. Choosing a good answer that ignores the deciding requirement (private network, quota, latency, cost).
2. Using accuracy on imbalanced data.
3. Treating fine-tuning as the default when prompting or retrieval would do.
4. Forgetting that multiple-response questions need every correct option.
5. Relying on notes from MLA-C01 only, which miss RAG, embeddings, agents, and Bedrock topics.

Last-day checklist

• Re-read the confused pairs above aloud.
• Redo the questions you missed in practice, and say why the closest wrong option is wrong.
• Skim the four domain lists in the exam guide for anything you cannot explain.
• Confirm your appointment, ID, and the rules for test center or online proctoring.
• Sleep. Plan to answer every question; blanks count as wrong.

More guides for MLA-C02

• MLA-C02 Exam Format, Scoring and Retakes
• MLA-C02 Study Plan: 4-Week and 8-Week
• Is the MLA-C02 Exam Hard?
• MLA-C02 Prerequisites and Skills Check
• MLA-C02 Renewal and Retakes Explained
• Is MLA-C02 Worth It? An Honest Look
• AIF-C01 vs MLA-C02: Which First?

Next step

Use the domains explained page for detail behind any row, and the exam overview for the full picture. For a self-test, try the sample questions, then our MLA-C02 practice tests.