MLA-C02 Study Guide: One ML Lifecycle, Two Tracks (Classic Models and Foundation Models)
The AWS Certified Machine Learning Engineer – Associate exam used to be about one kind of model: you prepare data, train it on SageMaker, deploy it to an endpoint and watch it drift. MLA-C02 keeps all of that and lays a second track over it: foundation models on Amazon Bedrock, with retrieval, agents, guardrails and evaluation. The lifecycle stages are the same; the tools and the traps at each stage differ. This guide walks the lifecycle once and shows both tracks side by side, so you learn where they meet and where they part.
Study guide · AWS Certified Machine Learning Engineer – Associate (MLA-C02)

Where the exam stands
MLA-C02 is in beta at the time of writing (in English only, since 29 September 2026). AWS has announced general availability in all languages on 14 January 2027, when MLA-C01 retires. The official exam guide lists 65 questions (50 scored, 15 unscored) and a minimum passing score of 720 on a scaled score of 100–1,000; it does not state an exam time. The beta runs longer, 85 questions in 170 minutes, so check the official exam page for the version you book. The four domains keep MLA-C01's structure: Data Preparation for ML and AI (28%), ML Model and Foundation Model (FM) Development (24%), Deployment and Orchestration of ML and AI Workflows (24%), and Operating, Monitoring, and Securing ML and AI Solutions (24%).
Of the 469 MLA-C02 questions in this question bank, 80 are new for MLA-C02, exactly 20 per domain, and almost all of those 80 are about generative AI. Across the whole bank about one question in five touches generative AI. SageMaker still dominates: it is tagged on far more questions than any other service. Plan your time accordingly: generative AI is the new part, not the main part.
The two tracks at a glance
| Stage | Classic ML (mostly SageMaker AI) | Foundation models (mostly Bedrock) |
|---|---|---|
| Data | Columnar files in S3, Glue and Data Wrangler, Feature Store, Clarify bias checks, Ground Truth labels | Parsing, chunking, embeddings, a vector store, metadata for filtering |
| Build | Built-in algorithm, script or container; training jobs; automatic model tuning | Choose a model; prompt, RAG, fine-tune or distill, lightest first |
| Evaluate | Precision, recall, F1, AUC, RMSE; overfitting checks | Bedrock evaluations (automatic, human, LLM-as-a-judge), RAG retrieval and answer quality |
| Deploy | Real-time, serverless, asynchronous or batch inference; Model Registry; Pipelines | On-demand, cross-Region or provisioned inference; knowledge bases; agents on AgentCore |
| Operate | Model Monitor, endpoint auto scaling, IAM, VPC, KMS, lineage | Guardrails, invocation logs with masking, private endpoints, token cost |
Data: where most of the weight is
Domain 1 is the largest. On the classic track, know when to use each preparation tool: Glue for serverless batch ETL, Data Wrangler (now part of SageMaker Canvas) for visual exploration that exports to a pipeline, Clarify for bias before training, Ground Truth for labels, and Feature Store to avoid training-serving skew (the online store serves low-latency lookups at inference time, the offline store keeps history for training). Know the techniques too: scaling, one-hot versus label encoding, missing values, and class imbalance. Time-series data is split by time, never shuffled.
On the foundation-model track, data preparation means getting documents ready for retrieval. Scanned PDFs need OCR before they can be chunked. The chunking strategy (fixed size, hierarchical, semantic) changes what the model sees. Metadata stored with each chunk lets a query filter by department or date.
Going deeper: the embedding dimension contract
An embedding model always returns a vector of a fixed length, and a vector index stores vectors of one declared length. Similarity search compares a query vector against stored vectors position by position, so the two must come from the same model with the same dimension. If you switch embedding models, or change the output size of the same model, every existing vector becomes incomparable with new queries; writes into the old index fail on a dimension mismatch. The fix is never to edit the field: create a new index with the new dimension and re-ingest every document. Questions that describe "after we upgraded the embedding model, ingestion fails" or "results became nonsense" are testing this.
Build and evaluate: pick the metric before the model
The recurring skill in Domain 2 is choosing the metric that matches the cost of being wrong. Missing fraud is expensive: watch recall. Flagging good customers is expensive: watch precision. Training accuracy far above validation accuracy is overfitting (add regularization, dropout, early stopping or more data); poor scores on both is underfitting (a more expressive model or better features). Managed Spot Training cuts cost if the job checkpoints.
For foundation models, climb the customization ladder from the bottom: prompt engineering first, RAG when answers need current or citable facts, fine-tuning when you need a consistent behavior or format, heavier options only when those fail. A custom model trained on structured, labeled data still beats an LLM for high-volume, low-latency numeric predictions.
Beginner tip: if precision and recall blur together, make up one confusion matrix and compute them by hand: 1,000 transactions, 20 frauds, the model flags 30, of which 15 are real frauds. Precision is 15 of 30 (50%), recall is 15 of 20 (75%). Then change one number, say the model flags 60 and catches 18, and recompute. Ten minutes of this makes every "which metric should the team optimize" question readable.
Deploy: the traffic shape picks the option
For SageMaker endpoints, read the traffic description. Steady, latency-sensitive requests: real-time endpoint. Long idle periods and tolerance for cold starts: serverless inference. Large payloads or minutes of processing: asynchronous inference, which can scale to zero. A nightly scoring of the whole table: batch transform, with no endpoint at all. Many small models with uneven traffic: a multi-model endpoint. Production variants let a challenger take a small share of traffic before it replaces the champion.
On Bedrock the same reading applies: on-demand for variable traffic, cross-Region inference profiles to absorb bursts (a geographic profile when data must stay in one geography), Provisioned Throughput for steady, forecastable volume, and batch inference for large offline jobs with a deadline rather than a latency target. One frequent trap: creating a knowledge base with CloudFormation or CDK does not load any documents. Nothing is searchable until an ingestion job has run, and it must run again after every content change.
Batch transform, with no endpoint at all
Asynchronous inference (can scale to zero)
Serverless inference
A multi-model endpoint
A real-time endpoint for steady, latency-sensitive requests
Operate: four kinds of drift, and guardrails that are not IAM
SageMaker Model Monitor checks four things against a baseline: data quality (input distribution), model quality (predictions against ground truth), bias drift and feature attribution drift. Model Monitor is no longer open to new customers (existing customers keep it, and AWS points to open-source monitoring with CloudWatch as the replacement); the exam guide speaks of drift detection pipelines, and the same four checks still apply. A drift event can go through EventBridge to start a retraining pipeline, but the new model should still pass the same evaluation and approval gate in the Model Registry.
Going deeper: why model quality monitoring always lags
Data quality monitoring needs only the inputs, which the endpoint captures the moment a request arrives, so it can raise an alert within one monitoring run. Model quality monitoring compares predictions with the true outcome, and the true outcome often arrives days or weeks later: whether a loan defaulted, whether a customer churned. Until those labels are uploaded and merged with captured predictions, there is nothing to score. That is why a question that asks for the earliest warning points to data quality monitoring, and a question that asks whether the model has actually become worse points to model quality monitoring.
For generative AI, Bedrock Guardrails filter prompts and responses (denied topics, content filters, masking of personal data), but they do not decide which APIs an application or agent may call; IAM roles and tool-level policies do. Production applications call Bedrock with role credentials, privately through an interface VPC endpoint when traffic must not cross the internet.
A five-week order
- Week 1: S3 formats, Glue, Feature Store, feature engineering, Clarify and data splits.
- Week 2: training jobs, tuning, metrics, overfitting, Model Registry.
- Week 3: endpoint types, Pipelines and Step Functions, CI/CD with approval gates.
- Week 4: the generative AI track end to end: RAG data preparation, model choice and customization, Bedrock inference options, guardrails.
- Week 5: monitoring, security, cost, then timed mixed sets.
Practise
Practise in the question bank: every question explains each option, right and wrong. Start with the domain you just studied, then take a timed mock exam.
- MLA-C02 exam page · All 529 MLA-C02 questions · Timed mock exam · 35-day study plan
- By domain: Data Preparation for ML and AI · ML Model and Foundation Model (FM) Development · Deployment and Orchestration of ML and AI Workflows · Operating, Monitoring, and Securing ML and AI Solutions
Exam facts as published for MLA-C02 at the time of writing. The exam is in beta until general availability, and Bedrock features change often; check the official AWS exam guide and documentation.