STEP 1 · Study

MLA-C02 StudyMachine Learning Engineer – Associate

Read the key points for each exam domain, then use “Practice this domain” to answer questions from that domain only. Domains with a larger weight appear more often on the exam.

Data Preparation for ML and AI

28%

Learning goals

Ingest data from sources such as S3, EFS, FSx, RDS and DynamoDB, and from streaming with Kinesis, Flink and Kafka, in a format suited to the purpose (Parquet, ORC, CSV, JSON)
Transform data and create features with Glue, DataBrew, Spark on EMR, Data Wrangler and Feature Store, choosing appropriately among scaling, binning, log transforms and encoding
Prepare RAG documents with embeddings, chunking and metadata extraction, and choose vector stores such as OpenSearch, pgvector and S3 Vectors
Validate quality with Glue Data Quality and DataBrew, label with Ground Truth, and address class imbalance with Clarify bias metrics and resampling

Key points

Formats: when analytics reads only some columns, use columnar Parquet or ORC (good compression, less data read). For row-wise appends and logs, use JSON or CSV. Some SageMaker AI built-in algorithms read fast in Pipe mode with RecordIO-protobuf.
Storage: S3 is the default home for training data. For training that reads many small files at high speed, use FSx for Lustre (integrated with S3); for sharing across multiple instances, use EFS. For large uploads from distant locations, use S3 Transfer Acceleration.
Streaming: ingest with Kinesis Data Streams, transform with Managed Service for Apache Flink or Lambda, and deliver to S3 and elsewhere with Data Firehose. For existing Kafka assets, use MSK.
Transformation tools: for code-free visualization and transformation, Data Wrangler or DataBrew; for large-scale ETL, Glue (Spark) or EMR. Use Feature Store (online and offline) so created features share the same definitions in both training and inference.
Feature engineering: standardization (mean 0, variance 1) and normalization (0–1), log transforms for skewed distributions, binning numeric ranges, one-hot encoding for categories (unordered) and label encoding (ordered). Impute missing values, handle outliers and remove duplicates.
Generative AI data: convert text and images to vectors with embedding models. RAG retrieval accuracy changes with chunking (fixed-size, hierarchical, semantic) and metadata. Also prepare training data for fine-tuning, continued pre-training and distillation with consistent formats and quality.
Personal information: mask, redact or anonymize before using data for training or RAG. Find sensitive data in S3 with Macie, and meet PII, PHI and data residency requirements.
Quality and bias: validate missing values, ranges and uniqueness with Glue Data Quality or DataBrew rules. Label with Ground Truth (combining people and automated labeling). Measure bias with Clarify pre-training metrics (class imbalance CI, difference in proportions of labels DPL and others), and address it with oversampling, undersampling, synthetic data (such as SMOTE) and stratified splits.

Workplace uses

SalesJoin CRM opportunity data with web access logs in Glue and register features such as “visits in the last 30 days” in Feature Store. Use the same features for training the win-prediction model and for real-time inference on the sales screen.
General AffairsChunk internal policy PDFs by heading, attach department and revision date as metadata, and store them in a vector store, so old versions of policies do not show up in search.
AccountingIn expense fraud detection, fraud examples make up only 0.5% of the total. Confirm class imbalance with Clarify, oversample only the training data, and evaluate on test data with the original ratio.

Exam objectives (from the exam guide)

1.1 データを収集して保存する。
1.2 データ変換、特徴量エンジニアリング、前処理を実行する。
1.3 データ品質を検証し、バイアスを管理する。

Learning resources

Learn in detail (3 topics with mini quizzes) →Practice this domain (84 questions) →Next domain ↓

ML Model and Foundation Model (FM) Development

24%

Learning goals

Choose built-in algorithms (XGBoost, Linear Learner, k-means, RCF, DeepAR, BlazingText and others), AI services, Bedrock FMs and JumpStart suited to the problem, based on interpretability, performance, latency and cost
Make training faster, cheaper and better with script mode, automatic model tuning (AMT), early stopping, distributed training, Spot training, regularization and ensembles
Choose appropriately among prompt engineering, RAG, fine-tuning and distillation, and optimize retrieval components (embeddings, reranking)
Analyze performance with metrics such as confusion matrices, F1, AUC and RMSE, reproducible experiments with MLflow, shadow testing, Clarify and Debugger, and ROUGE, BLEU, BERTScore, LLM-as-a-judge and human evaluation

Key points

Algorithms: XGBoost (gradient boosting) or Linear Learner for tabular classification and regression, k-means for clustering, Random Cut Forest for anomaly detection, DeepAR for time-series forecasting, BlazingText for text classification and word vectors, and image classification and object detection for images. Use linear models or shallow trees when explanation is needed, and gradient boosting or deep learning when performance matters most.
Build or use: transcription, translation, OCR and image analysis are often covered by AI services (Transcribe, Translate, Textract, Rekognition, Comprehend). Text generation and summarization use Bedrock FMs. Predictions that can only be solved with your own data are trained in SageMaker AI.
Training elements: epochs (how many passes over the whole dataset), steps, batch size and learning rate. Too large a learning rate makes loss oscillate or diverge; too small makes convergence slow.
Script mode: pass your own training script to SageMaker AI containers for TensorFlow, PyTorch, scikit-learn and others. It takes less effort than bringing your own container (BYOC).
Tuning: use AMT strategies such as random search, Bayesian optimization and Hyperband to maximize the objective metric. Early stopping halts unpromising trials. Shorten training time with distributed training (data parallel, model parallel), and cut costs with Managed Spot Training (save checkpoints).
Countering overfitting: regularization (L1, L2, dropout, weight decay), early stopping, adding or augmenting data, and reducing features. For underfitting, increase features or model capacity. For catastrophic forgetting, where fine-tuning loses original capabilities, lower the learning rate, mix in original data or use methods such as LoRA. Improve performance with ensembles (bagging, boosting, stacking).
Evaluation: for classification, the confusion matrix, precision (to reduce false positives), recall (to reduce misses), F1, and the ROC curve and AUC. For regression, RMSE and MAE. Compare against a baseline (a simple model or the current approach). Shadow testing sends production traffic to the new model too and compares without affecting users.
Generative AI evaluation: ROUGE for summarization, BLEU for translation, BERTScore or embedding similarity for semantic similarity, and LLM-as-a-judge or human evaluation for quality scoring. Bedrock evaluations can assess models and RAG (retrieval accuracy, answer faithfulness). Keep experiment conditions and results reproducible with MLflow and Prompt Management.

Workplace uses

SalesFor churn prediction, XGBoost is accurate, but the sales director asked for the reasons. Use Clarify's SHAP to show, per customer, the factors that pushed up churn probability (such as an increase in inquiries).
General AffairsAn internal FAQ bot ranks retrieved documents poorly and answers wrongly often. Try changing the embedding model and adding reranking, and compare retrieval accuracy and answer faithfulness with Bedrock's RAG evaluation before switching.
AccountingIn automatic account classification of invoices, misclassification (precision) causes more rework than misses (recall). Adjust the threshold to prioritize precision and route low-confidence items to staff.

Exam objectives (from the exam guide)

2.1 ML と AI のソリューションに適したモデリングアプローチを選択する。
2.2 ML と AI のソリューションのモデルをトレーニング、ファインチューニング、カスタマイズする。
2.3 ML と AI のシステムのパフォーマンスを分析および評価する。

Learning resources

Learn in detail (3 topics with mini quizzes) →Practice this domain (72 questions) →Next domain ↓

Deployment and Orchestration of ML and AI Workflows

24%

Learning goals

Choose by requirements among real-time, serverless, asynchronous and batch transform inference, multi-model and multi-container endpoints, ECS, EKS and Lambda, and edge optimization with Neo
Choose Bedrock FM deployment options (on-demand, Provisioned Throughput, Custom Model Import), RAG retrieval settings (hybrid search, reranking) and the agent foundation (AgentCore)
Create resources automatically with CloudFormation and CDK, and configure ECR containers, endpoints inside a VPC, and metric-based auto scaling
Automate training, evaluation, registration, deployment, retraining and rollback with SageMaker Pipelines, Step Functions, MWAA and EventBridge plus CodePipeline, CodeBuild and CodeDeploy

Key points

Inference options: for consistently low latency, a real-time endpoint; for sparse traffic where you do not want to pay for idle time, serverless inference (with cold starts); for large inputs or long processing, asynchronous inference (input and output in S3, can scale to 0 while waiting); for bulk processing, batch transform.
Many models: for many similar models invoked sparsely, a multi-model endpoint (loads and shares models on demand). To run different containers such as preprocessing and inference on one endpoint, a multi-container endpoint (direct invocation or a serial inference pipeline).
Edge: compile models for specific hardware with SageMaker Neo to make on-device inference faster and smaller.
FM deployment: Bedrock on-demand (per-token billing), Provisioned Throughput (reserving a fixed throughput) and Custom Model Import (using open models fine-tuned elsewhere in Bedrock). You can also host on your own GPU instances with SageMaker AI endpoints or JumpStart.
Infrastructure as code: CloudFormation (templates) and CDK (definitions in programming languages). Store containers in ECR. For endpoints inside a VPC, specify subnets and security groups, and access S3 and others through VPC endpoints.
Auto scaling: target tracking with metrics such as InvocationsPerInstance (invocations per instance). For increases and decreases at fixed times, scheduled scaling. For GPU workloads, also watch GPU utilization and queue length.
Orchestration: SageMaker Pipelines for ML steps (processing → training → evaluation → conditional branch → model registration), Step Functions for business flows spanning AWS services, and MWAA for existing Airflow assets. Start them on schedules or events (arrival of new data) with EventBridge.
CI/CD: CodePipeline (flow), CodeBuild (build and test), CodeDeploy (deployment) and CodeConnections (connecting to GitHub and others). Release to production only model versions approved in Model Registry. SageMaker AI deployment guardrails provide blue/green (all at once, canary, linear) with automatic rollback on CloudWatch alarms. Version prompts and agents too, and test them before release.

Workplace uses

SalesReturn a discount-rate suggestion on the quote screen within 100 ms with a real-time endpoint. Release new models via canary starting at 10% of traffic, with automatic rollback on an error-rate alarm.
General AffairsOn the 1st of every month, retrain the anomaly detection model on the previous month's entry/exit logs. Start SageMaker Pipelines with an EventBridge schedule, and register to Model Registry only when evaluation metrics exceed the threshold.
AccountingScore 200,000 transactions in bulk at month-end. Instead of an always-on endpoint, use batch transform to process input from S3 and output results to S3.

Exam objectives (from the exam guide)

3.1 ML と AI のモデルタイプのデプロイインフラストラクチャを管理する。
3.2 既存のアーキテクチャと要件に基づいて ML と AI のワークロードのリソースをプロビジョニングして設定する。
3.3 MLOps と AI のワークロードのための、自動オーケストレーション、継続的インテグレーションおよび継続的デリバリー (CI/CD) パイプラインを実装する。

Learning resources

Learn in detail (3 topics with mini quizzes) →Practice this domain (72 questions) →Next domain ↓

Operating, Monitoring, and Securing ML and AI Solutions

24%

Learning goals

Monitor production models and agents with Model Monitor (data quality, model quality, bias drift, feature attribution drift), A/B testing, CloudWatch generative AI observability and AgentCore Observability
Investigate latency and errors with CloudWatch, X-Ray, Logs Insights and dashboards, and right-size instances with Inference Recommender and Compute Optimizer
Manage ML and generative AI costs with Cost Explorer, Budgets and tags, Spot and Savings Plans, and optimization of token usage and vector store costs
Protect ML and AI with least-privilege IAM, Role Manager, bucket policies, VPCs and security groups, encryption, CloudTrail and Config, Bedrock credentials (API keys, IAM) and guardrails

Key points

Drift: when the distribution of input data changes from training time (data drift), accuracy drops. Model Monitor regularly compares production data against a baseline built from training data and sends violations to CloudWatch. If ground truth labels arrive later, model quality monitoring can also detect accuracy drops.
A/B testing: put multiple production variants on one endpoint, split traffic by weight and compare metrics. Shadow testing compares without returning results to users.
Generative AI monitoring: track token counts, latency and errors with CloudWatch generative AI observability, and agent steps, tool calls and failures with AgentCore Observability. Record inputs and outputs with Bedrock model invocation logging.
Troubleshooting: CloudWatch metrics and alarms, searching logs with Logs Insights, request paths and per-segment latency with X-Ray, and function performance with Lambda Insights.
Choosing instances: general purpose, compute optimized, memory optimized, GPU, inference-oriented (Inferentia) and training-oriented (Trainium). Load-test with Inference Recommender to choose instance types and counts that meet latency and cost requirements.
Cost: analyze with Cost Explorer, set budgets and alerts with Budgets, and aggregate by project with cost allocation tags. Use Managed Spot for training and SageMaker AI Savings Plans for steady usage. For generative AI, cut costs with token reduction (shorter prompts, caching, smaller models), batch inference and fewer embedding recomputations.
Capacity issues: address throttling with service quotas, sudden load with auto scaling settings or Provisioned Throughput, and Lambda with provisioned concurrency.
Security: least privilege with IAM roles (create persona-based roles with Role Manager), S3 bucket policies, isolation with VPC subnets and security groups, and private traffic through VPC endpoints. KMS at rest and TLS in transit. Record actions with CloudTrail and evaluate configurations with Config. Check container images for vulnerabilities with Inspector. Bedrock has API keys and IAM credentials; temporary IAM credentials are the norm in production. Protect sensitive information with guardrails.

Workplace uses

SalesAfter a new product launch, the input distribution of the demand forecasting model changed. Notify on Model Monitor data quality violations with a CloudWatch alarm and start the retraining pipeline.
General AffairsToken usage of the internal assistant grows every month. Look at token counts per department in CloudWatch, shorten the long system prompt, and use prompt caching for the repeated leading portion.
AccountingRun training jobs using payroll data in VPC subnets with no internet access, reach S3 through a VPC endpoint, and encrypt data and models with KMS keys. Record who started each job with CloudTrail.

Exam objectives (from the exam guide)

4.1 ML と AI のモデルの推論とパフォーマンスをモニタリングする。
4.2 ML と AI のインフラストラクチャのコストとパフォーマンスを最適化および管理する。
4.3 ML と AI のワークロードとモデルエンドポイントのセキュリティを確保する。

Learning resources

Learn in detail (3 topics with mini quizzes) →Practice this domain (73 questions) →Try the mock exam →

When you finish reading

Practice all domains (301 questions)Mock exam (65 questions · 130 min)MLA-C02 home

Notes (unofficial, original questions, no pass guarantee)

The questions on this site are original practice questions written by this site based on publicly available skills outlines and exam guides. They are not actual exam questions.
Amazon Web Services, AWS and related marks are trademarks of Amazon.com, Inc. or its affiliates. This site is not affiliated with or endorsed by AWS.
This content is a study aid and does not guarantee a pass. Exam content, format and prerequisites may change without notice; always check the latest information on the official page before taking the exam.
Your accuracy on the mock exam cannot be converted to the official (scaled) score and does not predict a pass or fail.
Progress is saved only in this browser (localStorage) unless you sign in with Google to sync it. You can delete it from the Cookie & storage settings page (Japanese).