AWS ML Engineer Associate (MLA-C02) Study Guide & Cheat Sheet

AWS Certified Machine Learning Engineer - Associate (MLA-C02) Study Guide

Exam domains

Who it’s for: ML engineers who build, deploy, and operate both traditional ML and generative AI solutions on AWS. AWS recommends at least 1 year with SageMaker AI and Amazon Bedrock plus 1 year in a related role such as backend developer, DevOps developer, data engineer, or data scientist.

Study & test-day tips

• Map every scenario to the managed option first: Amazon Bedrock for foundation models and SageMaker AI built-in tooling for traditional ML, whenever the question rewards lower operational overhead.
• Reach for the lightest generative AI fix that works: prompt engineering first, RAG when answers need your own or current data, fine-tuning when the model must learn a task or style, and domain adaptation on unlabeled text (continued pre-training) when it must learn a domain's vocabulary.
• Know RAG end to end: chunking and metadata at ingestion, the same embedding model for documents and queries, a vector store that fits the workload, then hybrid search, metadata filters, and reranking to improve retrieval.
• Evaluate generative output with the right method: reference-based metrics such as BLEU, ROUGE, and BERTScore when you have ground truth, LLM-as-a-judge for open-ended quality at scale, and human review when judgment or brand risk matters.
• Know endpoint trade-offs cold: real-time for low-latency steady traffic, serverless for intermittent traffic, asynchronous for large payloads or long inference, and Batch Transform for offline scoring of whole datasets.
• Match the metric to the problem: precision, recall, F1, and AUC for imbalanced classification, accuracy only for balanced classes, and RMSE or MAE for regression. Watch for questions that penalize false positives versus false negatives.
• For drift, capture endpoint traffic, compare it with a baseline built from training data in a scheduled job, and let a CloudWatch alarm or EventBridge rule start retraining; model quality checks also need ground truth labels. SageMaker Model Monitor, which packaged this, is closed to new customers.
• Treat prompts and agents like code: version prompts in Amazon Bedrock Prompt Management, roll out agents through versioned deployments, and test prompt changes in the CI/CD pipeline before promoting them.
• Default to least privilege and layered safeguards: scope IAM roles tightly, encrypt with KMS, isolate with VPCs and PrivateLink, prefer IAM or short-term credentials for Amazon Bedrock, and add Amazon Bedrock Guardrails for harmful content and sensitive data.
• Watch generative AI cost drivers: model choice, input and output tokens, embedding volume, and vector storage. Batch inference, prompt caching, smaller or distilled models, and Provisioned Throughput each fit a different traffic pattern.

Cheat sheet

Data Preparation and Feature Engineering

• Amazon S3 as the central data lake; AWS Lake Formation for fine-grained access governance.
• AWS Glue, Glue DataBrew, and Spark on Amazon EMR for ETL; AWS Glue Data Quality and DataBrew to validate data.
• SageMaker Data Wrangler for visual feature engineering; SageMaker Feature Store for online and offline feature reuse.
• SageMaker Ground Truth for labeling and human-in-the-loop annotation (no longer open to new customers; existing customers can keep using it).
• Handle class imbalance (resampling, SMOTE, class weights) and clean data by detecting outliers, imputing missing values, and deduplicating.

Streaming and Ingestion

• Kinesis Data Streams: real-time, custom consumers, replayable shards.
• Amazon Data Firehose: near-real-time delivery to S3, Redshift, and OpenSearch with transforms.
• Amazon MSK for managed Apache Kafka; Amazon Managed Service for Apache Flink for stream processing.
• Choose formats by access pattern: columnar Parquet or ORC for analytics, JSON or CSV for interchange.

Vector Stores, Embeddings, and RAG Data

• Vector stores on AWS: Amazon OpenSearch Service, Amazon RDS for PostgreSQL with pgvector, and Amazon S3 Vectors.
• Embedding models on Amazon Bedrock turn text and images into vectors; embed documents and queries with the same model.
• Amazon Bedrock knowledge bases chunk documents with standard, hierarchical, or semantic chunking, or no chunking, and can attach metadata for filtering.
• Mask or redact sensitive data before training or indexing, for example with Amazon Comprehend PII detection or Amazon Macie discovery.
• Fine-tuning uses labeled prompt and response pairs; continued pre-training uses unlabeled domain text; distillation trains a smaller model on a larger model's responses.

Model Development and Training

• SageMaker AI built-in algorithms, script mode with supported frameworks, and SageMaker JumpStart models.
• Automatic model tuning (AMT) for hyperparameters; early stopping and distributed training to cut training time; managed spot training to cut cost.
• Prevent overfitting, underfitting, and catastrophic forgetting when you train or fine-tune.
• Track experiments with MLflow on SageMaker AI and diagnose convergence with TensorBoard on SageMaker and CloudWatch; compute bias metrics (CI, DPL) and SHAP explanations in your own processing code. SageMaker Clarify and Debugger are no longer open to new customers.
• Metrics: accuracy, precision, recall, F1, and AUC for classification; RMSE and MAE for regression.

Foundation Models and Evaluation

• Select FMs on Amazon Bedrock by task, modality, quality, latency, and cost.
• Customize on Amazon Bedrock with supervised fine-tuning, reinforcement fine-tuning, or model distillation; bring outside models in with Custom Model Import.
• RAG with Amazon Bedrock knowledge bases: managed ingestion and retrieval, metadata filtering, and reranking for relevance.
• Amazon Bedrock evaluations support automatic metrics, human review, and LLM-as-a-judge; know BLEU, ROUGE, and BERTScore.
• Use AWS AI services when they fit: Textract for documents, Rekognition for images and video, Comprehend for text, Transcribe for speech.

Deployment and Inference

• Real-time endpoints for low-latency steady traffic; serverless endpoints for intermittent traffic with no idle cost.
• Asynchronous endpoints for large payloads and long inference; Batch Transform for offline scoring.
• Multi-model and multi-container endpoints to host many models and cut cost.
• FM options: on-demand Amazon Bedrock inference, cross-Region inference profiles for bursts, Provisioned Throughput for reserved capacity, or SageMaker AI endpoints for self-hosted models.
• Scale GPU workloads with instance choice and auto scaling on the right metric.

Agents and MLOps

• Amazon Bedrock AgentCore to deploy and operate agents with runtime, memory, gateway, identity, and observability. Amazon Bedrock Agents, now called Agents Classic, is closed to new customers.
• Amazon Bedrock Prompt Management to version prompts; versioned agent deployments for controlled rollout.
• SageMaker Pipelines and Model Registry for ML workflows and versioning; Step Functions and EventBridge to orchestrate.
• CodePipeline, CodeBuild, and CodeDeploy for CI/CD, with automated tests and rollback.
• Keep knowledge bases fresh by syncing data sources on a schedule or when source data changes.

Monitoring, Cost, and Security

• Drift detection: enable data capture on endpoints, compare live data with a training baseline in scheduled jobs (AWS documents MLflow and Evidently AI based solutions), and alert through CloudWatch or Amazon SNS. SageMaker Model Monitor is no longer open to new customers.
• Amazon CloudWatch, including generative AI observability, plus AWS X-Ray and AgentCore Observability to trace agents, tool failures, and latency.
• Control FM cost with model choice, batch inference, prompt caching where supported, and tracking of token, embedding, and vector storage spend.
• IAM least privilege, CloudTrail and AWS Config for audit, VPC isolation, PrivateLink, and KMS encryption.
• Amazon Bedrock Guardrails: content filters, denied topics, sensitive information filters, contextual grounding checks, and Automated Reasoning checks. Scan container images and code with Amazon Inspector; CodeGuru Reviewer stopped accepting new repositories on November 7, 2025.

Glossary

Amazon S3

Object storage service that serves as the central data lake for ML datasets, models, and artifacts.

AWS Glue

Serverless ETL service for discovering, cataloging, and transforming data at scale.

AWS Glue DataBrew

Visual data preparation tool for cleaning and normalizing data without writing code.

Amazon EMR

Managed big data platform running Apache Spark, Hadoop, and related frameworks for large-scale processing.

Amazon Kinesis Data Streams

Real-time streaming data service with replayable shards for custom consumer applications.

SageMaker Data Wrangler

Visual tool for importing, exploring, and engineering features from data for ML.

SageMaker Feature Store

Repository for storing, sharing, and serving ML features with online and offline stores.

SageMaker Ground Truth

Data labeling service with human and automated workflows to create training datasets. No longer open to new customers.

Amazon Bedrock

Fully managed service for building generative AI applications with foundation models from Amazon and other providers through one API.

Foundation model (FM)

Large model pretrained on broad data that can be adapted to many tasks through prompting, retrieval, or customization.

Embedding model

Model that converts text or images into numeric vectors whose distances reflect semantic similarity.

Vector store

Database that indexes embeddings for similarity search, such as OpenSearch Service, Amazon RDS for PostgreSQL with pgvector, or Amazon S3 Vectors.

Retrieval Augmented Generation (RAG)

Pattern that retrieves relevant content at query time and passes it to the model so answers are grounded in your data.

Chunking

Splitting documents into smaller passages before embedding so retrieval returns focused, relevant context.

Reranking

Second-stage scoring that reorders retrieved passages by relevance before they reach the model.

Amazon Bedrock knowledge bases

Managed RAG capability that ingests data sources, chunks and embeds content into a vector store, and retrieves context for prompts.

Fine-tuning

Further training a pretrained model on labeled examples so it performs a specific task or style better.

Continued pre-training

Further training a foundation model on unlabeled domain text to deepen its knowledge of that domain, which SageMaker JumpStart offers as domain adaptation fine-tuning.

Model distillation

Training a smaller student model on responses from a larger teacher model to get similar quality at lower cost and latency.

Catastrophic forgetting

Loss of previously learned general capabilities when a model is trained too narrowly on new data.

Amazon Bedrock Custom Model Import

Feature for importing a customized open-source model, built outside Amazon Bedrock, into Bedrock for inference.

Provisioned Throughput

Amazon Bedrock purchase option that reserves model capacity for consistent throughput.

Cross-Region inference

Amazon Bedrock inference profiles that route requests across AWS Regions to absorb bursts of traffic.

Amazon Bedrock Agents

Capability for building agents that use a foundation model to plan steps, call actions, and query knowledge bases. Now called Amazon Bedrock Agents Classic and closed to new customers; AWS points new agent work to Amazon Bedrock AgentCore.

Amazon Bedrock AgentCore

Services for deploying and operating agents at scale, including runtime, memory, gateway, identity, and observability.

Amazon Bedrock Prompt Management

Feature for creating, versioning, and testing prompts so applications can use a specific prompt version.

Amazon Bedrock Guardrails

Configurable safeguards that filter harmful content, block denied topics, filter sensitive information, and check responses for grounding.

LLM-as-a-judge

Evaluation method that uses a model to score another model's outputs against criteria such as correctness or helpfulness.

BLEU, ROUGE, and BERTScore

Reference-based NLP metrics: BLEU measures n-gram precision, ROUGE measures n-gram recall and overlap, and BERTScore compares meaning using embeddings.

Amazon Bedrock API keys

Credentials for calling Amazon Bedrock, available as short-term keys or long-term keys; IAM credentials remain an alternative.

Automatic model tuning (AMT)

SageMaker AI hyperparameter optimization that searches parameter combinations to maximize an objective metric.

SageMaker Clarify

Service that detects bias in data and models and explains predictions with SHAP. No longer open to new customers; AWS now documents computing the same bias metrics with pandas and scikit-learn and explanations with the SHAP library.

Real-time endpoint

Persistent SageMaker AI endpoint serving low-latency, synchronous predictions with auto scaling.

Serverless endpoint

SageMaker AI inference option that scales to zero, suited to intermittent or spiky traffic.

Asynchronous endpoint

SageMaker AI endpoint that queues requests for large payloads and long-running inference.

Batch Transform

SageMaker AI feature for offline, high-throughput inference over an entire dataset.

Data capture

SageMaker AI endpoint setting that saves inference requests and responses to Amazon S3, the input for drift monitoring and retraining.

SageMaker Pipelines

Purpose-built service for building, automating, and managing ML workflows.

SageMaker Model Monitor

Service that monitors deployed models for data quality, model quality, and drift against a baseline. No longer open to new customers; AWS documents drift monitoring solutions built on data capture, scheduled jobs, MLflow, and Evidently AI.

Amazon Inspector

Vulnerability management service that scans workloads such as container images for software vulnerabilities.

AWS KMS

Key Management Service for creating and controlling the encryption keys that protect data.