AWS Certified Machine Learning Engineer – Associate Cheat Sheet (MLA-C01)
Every in-scope SageMaker feature, MLOps workflow, monitoring and security decision the MLA-C01 exam actually tests — organised by the four exam domains with runbooks and decision tables
MLA-C01 (English) is no longer available — it retired on September 28, 2026. MLA-C02 has replaced it, and this material has not yet been updated for it.
See the full timelineAbout This Cheat Sheet
- 14 pages
- Every in-scope MLA-C01 service
- For ~1 year of hands-on SageMaker
- 5 reference architectures
- 25 answer-pattern scenarios
- Updated April 2026
Exam domains
Share of the scored MLA-C01 content
- Data Preparation for ML28%
- ML Model Development26%
- Deployment & Orchestration of ML Workflows22%
- ML Solution Monitoring, Maintenance & Security24%
Services covered
SageMaker data prep
- Data Wrangler
- Processing
- Ground Truth
- Feature Store
- Clarify
SageMaker build & train
- Studio
- Unified Studio
- Canvas
- JumpStart
- Autopilot
- Experiments
- Debugger/Profiler
- HyperPod
SageMaker deploy & MLOps
- Pipelines
- Model Registry
- Projects
- Lineage Tracking
- MLflow
- Model Monitor
- Neo
- Inference Recommender
Data preparation services
- S3
- Glue
- Glue DataBrew
- Glue Data Quality
- Lake Formation
- Athena
- Redshift
- EMR
- Kinesis
- Kinesis Video Streams
- Data Firehose
- Managed Service for Apache Flink
- OpenSearch
- QuickSight
- DataSync
- Storage Gateway
Purpose-built AI services
- Bedrock
- Comprehend
- Comprehend Medical
- Rekognition
- Textract
- Transcribe
- Polly
- Translate
- Lex
- Kendra
- Personalize
- Fraud Detector
- Lookout for Equipment
- Lookout for Metrics
- Lookout for Vision
- HealthLake
- A2I
- Amazon Q
- CodeGuru
- DevOps Guru
- Amazon Forecast
Show 50 more servicesShow fewer
Compute
- EC2
- Lambda
- Batch
- SAR
- ECS
- EKS
- ECR
- Auto Scaling
- Compute Optimizer
Orchestration & integration
- Step Functions
- MWAA
- EventBridge
- SNS
- SQS
- API Gateway
Storage
- EBS
- EFS
- FSx
- S3 Glacier
Databases for ML apps
- RDS
- DynamoDB
- DocumentDB
- ElastiCache
- Neptune
Security
- IAM
- KMS
- Secrets Manager
- Macie
- CloudTrail
- Config
- Organizations
Networking
- VPC
- Direct Connect
- CloudFront
Observability, cost & governance
- CloudWatch
- CloudWatch Logs
- X-Ray
- Chatbot
- Systems Manager
- Trusted Advisor
- Billing
- Budgets
- Cost Explorer
- Service Catalog
IaC & CI/CD
- CloudFormation
- CDK
- CodePipeline
- CodeBuild
- CodeDeploy
- CodeArtifact
Feature details
- S3
- Lifecycle · Object Lock · Access Points
- Glue Data Quality
- DQDL
- Lake Formation
- FGAC
- Redshift
- Redshift ML
- OpenSearch
- k-NN · Neural plugin
- Ground Truth
- Mechanical Turk
- Feature Store
- Offline + online
- Clarify
- Bias · FM eval · Explainability + SHAP
- Training
- Built-in algorithms · Script Mode · BYO containers · SMDDP · SMP · Managed Spot · Hyperparameter Tuning
- Inference
- Real-time · Serverless · Async · Batch Transform · Multi-Model / Multi-Container · Production Variants · Endpoint Auto Scaling
- SageMaker security
- Execution roles · Network isolation · VPC mode
- Pipelines
- Full step reference (all 11 step types)
- Model Registry
- Approvals
- Projects
- Service Catalog
- MLflow
- Managed tracking server
- Model Monitor
- Four types: Data Quality / Model Quality / Bias Drift / Feature Attribution Drift
- EC2
- GPU · Trainium · Inferentia · Capacity Blocks
- FSx
- FSx for Lustre / NetApp / OpenZFS / Windows
- Amazon Forecast
- Legacy note
- Organizations
- SCPs
- VPC
- Gateway/interface endpoints
Concepts covered
- ML fundamentals
- Supervised / unsupervised / self-supervised / RL · Bias vs variance · Splits and cross-validation · Feature engineering · Common algorithms · Neural-network families
- Metrics
- Binary / multiclass / regression / ranking / forecast / NLP / clustering · Imbalanced-data guidance
- Encryption patterns
- At-rest + in-transit · VPC mode + network isolation
- Reference architectures
- End-to-end SageMaker training + deployment · Feature Store-backed real-time inference · Multi-tenant Multi-Model Endpoint · Streaming anomaly detection · RAG on Bedrock
- Playbooks
- Data prep + feature engineering playbook (cleaning, encoding, scaling, text/image/time-series, leakage traps) · SageMaker Pipelines step reference
- Runbooks
- Training OOM · Endpoint latency miss · Drift detection · Runaway cost
- Exam practice
- 25-scenario answer patterns · 12 common pitfalls · 23-row quick-pick decision table
Built from the official MLA-C01 exam guide; every service on its in-scope list is audited present.
What's Inside
ML Vocabulary & Lifecycle
The ML fundamentals the exam assumes: learning types, data splits, feature engineering, algorithms and the metrics to judge them.
- Supervised / unsupervised / self-supervised / reinforcement learning
- Bias-variance tradeoff (with underfit / overfit fixes)
- Training / validation / test splits; cross-validation
- Feature engineering: scaling, encoding, imputation, binning, interactions, embeddings
- Algorithms: Linear / Logistic, Trees, Random Forest, XGBoost, LightGBM, K-Means, DBSCAN
- Algorithms: SVM, PCA, RCF, Factorization Machines, DeepAR, BlazingText
- Neural-network families: MLP, CNN, RNN / LSTM / GRU, Transformer, GAN, Diffusion
- Metrics cheat table: binary / multiclass / regression / ranking / forecast / NLP / clustering
- Imbalanced-data guidance
Data Preparation
Getting training data ready: AWS data services for storage, processing and streaming, plus the SageMaker prep, labeling and bias tools.
- S3 (Lifecycle + Object Lock + Access Points + SSE-KMS)
- Glue + Glue DataBrew + Glue Data Quality DQDL; Lake Formation fine-grained access
- Athena, Redshift + Redshift ML, EMR + EMR Serverless
- Streaming: Kinesis + Kinesis Video Streams + Data Firehose + Managed Service for Apache Flink
- OpenSearch Service with k-NN + Neural plugin; QuickSight; DataSync + Storage Gateway
- SageMaker-native prep: Data Wrangler visual prep, Processing jobs
- Ground Truth labeling (with Mechanical Turk workforce)
- Feature Store with online + offline stores for point-in-time-correct training
- SageMaker Clarify: pre-training + post-training bias, SHAP, FM evaluation
Amazon SageMaker Core
SageMaker end to end: the studios, built-in algorithms, training at scale, every inference option and the security controls.
- Studio + Unified Studio + Canvas + JumpStart + Autopilot
- Built-in algorithms: XGBoost, Linear Learner, K-Means, PCA, KNN, RCF, Factorization Machines
- Built-in: IP Insights, BlazingText, Object2Vec, DeepAR, Seq2Seq, Neural Topic Model, LDA
- Built-in vision: Image Classification, Object Detection, Semantic Segmentation
- Custom containers + Script mode; distributed training (SMDDP + SMP); Managed Spot + S3 checkpoints
- Hyperparameter Tuning (Bayesian / Random / Grid / Hyperband); Experiments; Debugger + Profiler
- HyperPod for FM-scale training; Production Variants with A/B + shadow
- Inference: Real-time, Serverless Inference, Asynchronous Inference, Batch Transform
- Multi-Model + Multi-Container endpoints, Neo compilation, Inference Recommender
- Security: execution roles, network isolation, VPC mode, KMS, inter-container encryption
MLOps — Pipelines & Registry
Automating the ML lifecycle: SageMaker Pipelines step by step, model approval and deployment, tracking and orchestration choices.
- SageMaker Pipelines steps: ProcessingStep, TrainingStep, TuningStep, ModelStep, TransformStep, ConditionStep, CallbackStep
- More steps: LambdaStep, EMRStep, ClarifyCheckStep, QualityCheckStep, AutoMLStep, FailStep, NotebookJobStep
- SageMaker Model Registry: versioning + approval gates + EventBridge → CodePipeline cross-account deploy
- SageMaker Projects with Service Catalog templates
- SageMaker Experiments + Lineage Tracking + managed MLflow tracking server
- MWAA for Airflow-native DAGs; Step Functions for heterogeneous durable orchestration
- EventBridge for training-job-state-change + endpoint alarms + registry approvals
Monitoring, Drift & Explainability
Keeping deployed models healthy: the four Model Monitor types, explainability, endpoint metrics and autoscaling.
- Model Monitor types: Data Quality, Model Quality, Bias Drift, Feature Attribution Drift
- Data Quality checks missing values, types, distributions
- Model Quality (accuracy / precision / recall) needs ground-truth capture
- DataCaptureConfig on endpoints and baseline jobs
- SageMaker Clarify: explainability at inference time (SHAP) + bias reports pre + post training
- CloudWatch endpoint metrics: Invocations, ModelLatency, OverheadLatency, InvocationsPerInstance
- CPU/GPU/MemoryUtilization plus custom metrics; X-Ray distributed tracing
- Endpoint Auto Scaling: target tracking on InvocationsPerInstance or custom metric
- Scale-to-zero only via Serverless Inference
Purpose-Built AI Services
The managed AI services that can replace a custom model, with the features that decide which one fits.
- Bedrock managed FMs: Nova + Titan + Claude + Llama + Mistral + Cohere + AI21 + Stability
- Bedrock Guardrails + Knowledge Bases + Agents + Prompt Management + Model Evaluation
- Comprehend + Comprehend Medical (HIPAA); Rekognition (images + video + custom labels)
- Textract (forms + tables + lending + invoice); Transcribe (ASR, diarisation, Call Analytics, medical)
- Polly (TTS, neural voices); Translate (NMT + Active Custom); Lex (chatbots)
- Kendra (enterprise semantic search); Personalize (recommendations); Fraud Detector
- Lookout for Equipment / Metrics / Vision; HealthLake (FHIR + NLP)
- Augmented AI (A2I) human review; Amazon Q (Business + Developer + in QuickSight)
- CodeGuru + DevOps Guru
- Amazon Forecast legacy note: SageMaker Canvas / DeepAR / Chronos for modern forecasting
Compute, Containers & Storage
Picking compute, containers and storage for training and inference, from accelerated instances to shared file systems.
- EC2 instance families for ML: p5 / p4d / p4de NVIDIA, g5 / g6
- trn2 / trn1 Trainium, inf2 / inf1 Inferentia; Capacity Blocks for ML
- Lambda for glue and async preprocessing; Batch job queues + array jobs
- Serverless Application Repository; ECR images with Enhanced scanning
- ECS + Fargate, EKS + Karpenter; Auto Scaling + Compute Optimizer
- Storage: S3 + S3 Glacier, EBS gp3 / io2, EFS for shared notebooks
- FSx for Lustre for distributed training scratch (S3 linked)
- FSx for NetApp ONTAP / OpenZFS / Windows File Server for enterprise shared storage
- Storage Gateway for hybrid seeds
Databases for ML Apps
Where ML applications keep source data, features, state and caches, database by database.
- RDS: relational source for training data + Redshift ML via SQL
- DynamoDB: online feature serving + session state + inference logging
- DynamoDB Streams to Lambda for retraining
- DocumentDB: JSON training data + metadata catalogs with vector search
- ElastiCache (Redis OSS / Valkey / Memcached): semantic cache in front of Bedrock
- ElastiCache also as hot-feature cache and for prompt outputs
- Neptune: graph DB for entity features + Neptune Analytics for GraphRAG
Security, Networking & Governance
Locking down ML workloads: scoped roles, encryption, audit and compliance rules, and private networking for SageMaker.
- IAM execution roles scoped by resource ARN + conditions
- Permission boundaries for ML teams; Organizations SCPs pinning Region + denying unapproved services
- KMS CMKs across S3 artefacts / EBS / endpoints / Feature Store offline
- Secrets Manager: DB + third-party API rotation; Macie for PII in training data
- CloudTrail org trail; Config managed rules + Conformance Packs
- Rules: sagemaker-endpoint-configuration-kms-key-configured, sagemaker-notebook-no-direct-internet-access
- VPC with private subnets + flow logs; Direct Connect for on-prem hybrid
- Gateway endpoints for S3 / DDB; PrivateLink interface endpoints for SageMaker API + runtime
- More PrivateLink endpoints: ECR + STS + CloudWatch + Glue + Athena + KMS + Secrets Manager
- CloudFront + API Gateway + WAF + authorizers for inference fronting
Observability, Cost & DevOps
Operating ML systems: monitoring and alerting, cost control, and the IaC and CI/CD tooling around SageMaker.
- CloudWatch metrics + Logs + Logs Insights + subscription filters → Firehose → OpenSearch
- X-Ray distributed tracing; Chatbot → Slack / Teams / Chime
- Systems Manager: Parameter Store + Patch Manager + Change Manager + Incident Manager
- Trusted Advisor security + cost checks
- AWS Billing + Budgets with actions + Cost Explorer
- Service Catalog for approved SageMaker templates; CloudFormation + CDK for IaC
- CI/CD: CodePipeline + CodeBuild + CodeDeploy + CodeArtifact
- Cost patterns: Managed Spot training with checkpoints, Serverless / Async inference
- Multi-Model / Multi-Container endpoints, Inference Recommender
- Graviton / Inferentia / Trainium, FastFile mode, delete unused endpoints
Reference Architectures
Five full walkthroughs that combine SageMaker and other AWS services into production ML systems, from training pipelines to a RAG assistant.
- End-to-end SageMaker training + deployment with Pipelines + Clarify + Condition + Model Registry
- Deployment via EventBridge-triggered CodeDeploy blue/green + Model Monitor
- Feature Store-backed real-time inference with online + offline reconciliation
- Multi-tenant inference on a Multi-Model Endpoint with per-tenant CloudWatch + Auto Scaling
- Streaming anomaly detection with Kinesis + Managed Flink + SageMaker RCF endpoint
- Anomaly alerts via EventBridge + nightly retraining Step Function
- RAG assistant using Bedrock Knowledge Bases with Guardrails
- RAG evaluation: Bedrock Model Evaluation + SageMaker Clarify FM evaluation
Scenarios, Pitfalls, Runbooks & Decision Table
25 scenario → answer mappings, 12 common pitfalls, four runbooks and a 23-row quick-pick decision table.
- Inference: Serverless inference, Batch Transform, Async for large payloads, Multi-Model endpoints, Inference Recommender
- Training: Managed Spot training, HyperPod FM training, JumpStart + LoRA, forecasting
- MLOps: Model Registry approvals, Model Monitor, Experiments + Lineage, MWAA vs Pipelines
- Data: Macie PII, Feature Store, Ground Truth, A2I review, Clarify SHAP, VPC isolation
- AI services: Lookout for Vision, Lookout for Metrics, Personalize, Comprehend Medical, Bedrock + KB
- Pitfalls: Managed Spot checkpointing, endpoint cost, File vs Pipe vs FastFile
- Pitfalls: Feature Store consistency, Model Quality needs labels, drift baselines
- Pitfalls: IAM execution role scope, VPC-only internet, Bedrock vs JumpStart
- Pitfalls: Edge Manager deprecation; Clarify needs facet column; Inference Recommender needs model package
- Runbooks: training OOM, endpoint latency SLO miss, drift detected, runaway cost
Why This Cheat Sheet Helps
MLA-C01 is the ML engineering exam — it tests the practical choices: which inference option (Real-time / Serverless / Async / Batch / Multi-Model), which orchestration (Pipelines / MWAA / Step Functions), which drift monitor (Data Quality / Model Quality / Bias / Feature Attribution), when to reach for SageMaker JumpStart vs Bedrock, how to wire Model Registry approvals into CI/CD, how to right-size an endpoint without guessing. This cheat sheet puts those decisions side by side with the exact AWS vocabulary examiners use.
It assumes you already understand ML fundamentals (bias / variance / metrics / feature engineering) and basic Python. Use it in the final weeks to map the exam's surface area, reinforce the SageMaker + AI services matrix, and rehearse the four operational runbooks the exam loves: OOM training, endpoint latency miss, drift detection, and runaway cost.
How to Use It
Skim the whole sheet once to see how the four domains map to sections. Then hammer practice exams — for each wrong answer, find the relevant section and study the tables and decision trees. Pay extra attention to sections 2 (Data Preparation), 3 (SageMaker core), 4 (MLOps), 5 (Monitoring + Clarify), 14 (Scenario → Answer Patterns), 15 (Pitfalls), 16 (Data prep + leakage traps), 17 (Pipelines step reference), 18 (Runbooks), and 19 (Quick-Pick decision table).
In the final week, walk through the four domain weights against your confidence: Data Preparation (28%) → sections 2, 16; ML Model Development (26%) → sections 1, 3, 6; Deployment & Orchestration (22%) → sections 3–4, 7–8, 13; Monitoring, Maintenance & Security (24%) → sections 5, 10–12, 18. Pair with CloudNinja's free MLA-C01 practice exam to surface gaps.
Frequently Asked Questions
Is this AWS Machine Learning Engineer Associate cheat sheet free?
Yes, completely free with no signup required. Download the PDF directly from CloudNinja and use it as a study reference for the MLA-C01 exam.
How much experience do I need before MLA-C01?
AWS recommends around one year of hands-on SageMaker experience plus general software engineering. You should be comfortable with Python, at least one ML framework (scikit-learn, PyTorch, TensorFlow, or XGBoost), and the basics of model training + hosting on SageMaker. If you do not already hold AI Practitioner (AIF-C01) or have production ML experience on AWS, build that foundation first.
Did AWS retire the old ML Specialty (MLS-C01)?
Yes. AWS retired Machine Learning – Specialty (MLS-C01) and split the coverage into the AI Practitioner (AIF-C01, foundational), Machine Learning Engineer – Associate (MLA-C01) for operational ML + MLOps, and the Generative AI Developer – Professional (AIP-C01) for production GenAI development. MLA-C01 is the current associate-tier ML exam for engineers.
Is this cheat sheet updated for the current MLA-C01 exam?
Yes, it is built directly from the current MLA-C01 exam guide (machine-learning-engineer-associate-01) and audited against the full in-scope services list. It reflects current services and recent changes — SageMaker Unified Studio, SageMaker HyperPod for FM training, managed MLflow tracking server, Inference Recommender, Feature Store online + offline, Model Monitor with Bias + Feature Attribution drift, Amazon Bedrock (Guardrails + Knowledge Bases + Agents + Prompt Management), Amazon Q, SageMaker Edge Manager deprecation guidance toward Greengrass + Neo, and Amazon Forecast as legacy (exam prefers SageMaker Canvas / DeepAR / Chronos).
Keep Studying
Recommended courses
Affiliate links — if you enrol through them, CloudNinja may earn a commission at no extra cost to you.