AWS Certified Machine Learning Engineer – Associate (MLA-C02)

Learning plan

About 80 hours of focused study, split into 5 phases. Pick your weekly pace and the timeline adjusts.

Estimated finish: 8 weeks

What the exam covers

Source: official exam guide. Weights can change, so check before you start.

Before you start

• An AWS account with SageMaker AI and Bedrock model access enabled (new accounts get Free Tier credits; watch endpoint costs)
• Working Python plus pandas, and comfort reading SQL and JSON
• Core ML vocabulary (train/validation/test splits, overfitting, common algorithms like XGBoost and k-means). AWS recommends about 1 year with SageMaker AI and Bedrock

1. Weeks 1–2 · 20 h

   Data preparation for ML and RAG (largest domain)

   • Pick the right ingestion and storage path (S3, Kinesis Data Streams/Firehose, Glue, EMR) and file format (Parquet, ORC, CSV, RecordIO) for a given volume and latency
   • Engineer features with SageMaker Data Wrangler and Glue DataBrew, then store them in SageMaker Feature Store (online vs offline store)
   • Detect pre-training bias with SageMaker Clarify (class imbalance, DPL) and fix it with resampling or synthetic data
   • New in C02: chunk documents for RAG, pick an embedding model, and configure a vector store (OpenSearch Service, RDS/Aurora with pgvector, S3 Vectors)
   • Mask or redact PII before training or fine-tuning (Comprehend PII detection, Glue sensitive-data detection, Macie)

   Hands-on: Load a public tabular dataset (for example UCI Adult) into S3, run a SageMaker Clarify pre-training bias report on a sensitive attribute, then build a Data Wrangler flow that one-hot encodes categoricals and exports to Feature Store.

   Use: AWS Skill Builder – MLA-C02 exam prep · MLA-C01 vs MLA-C02 comparison (official) · SageMaker Clarify processing jobs

2. Weeks 3–4 · 20 h

   Model and foundation model development

   • Choose between SageMaker built-in algorithms (XGBoost, Linear Learner, BlazingText, DeepAR), JumpStart, Bedrock FMs and AI services for a scenario, including cost and latency trade-offs
   • Run a SageMaker training job with Automatic Model Tuning (Bayesian vs random search) and use early stopping, distributed training and Spot training
   • Read evaluation metrics (RMSE, AUC-ROC, F1, confusion matrix) and debug with SageMaker Debugger; explain results with Clarify SHAP values
   • Pick a customisation route (prompt engineering, RAG, Bedrock fine-tuning, continued pre-training, distillation) and evaluate FM output with BLEU, ROUGE, BERTScore, LLM-as-a-judge and Bedrock Evaluations
   • Version models in SageMaker Model Registry with approval status

   Hands-on: Train the built-in XGBoost algorithm on your prepared data with a 10-job hyperparameter tuning run, register the best model in Model Registry, then run a Bedrock automatic evaluation comparing two FMs on a small summarisation prompt set.

   Use: Amazon SageMaker AI Developer Guide · Amazon Bedrock model evaluation · amazon-sagemaker-examples (GitHub)

3. Weeks 5–6 · 18 h

   Deployment, orchestration and agents

   • Choose the endpoint type (real-time, serverless, asynchronous, batch transform, multi-model) from traffic and payload requirements, and configure auto scaling on InvocationsPerInstance
   • Use deployment guardrails (blue/green, canary, linear traffic shifting) and shadow tests for safe rollouts
   • Build a SageMaker Pipelines workflow and trigger it from CodePipeline or EventBridge; know when Step Functions or Managed Airflow (MWAA) fits better
   • New in C02: create a Bedrock Knowledge Base, manage prompts with Bedrock Prompt Management, import an outside model with Custom Model Import, and deploy an agent (Bedrock Agents or AgentCore) with versioned aliases

   Hands-on: Build a 3-step SageMaker Pipeline (processing, training, conditional register) and deploy the approved model to a serverless endpoint. Then create a Bedrock Knowledge Base over a few PDFs in S3 and query it with RetrieveAndGenerate. Delete endpoints when done.

   Use: SageMaker Pipelines · Bedrock Knowledge Bases · AWS Agentic AI Demonstrated (Skill Builder)

4. Weeks 6–7 · 12 h

   Monitoring, cost and security

   • Configure SageMaker Model Monitor for data quality, model quality, bias drift and feature attribution drift, and alarm on it in CloudWatch
   • Track FM and agent cost drivers (tokens, provisioned throughput vs on-demand, embedding and vector storage) and right-size with Inference Recommender, Savings Plans and Spot
   • Lock down ML resources with least-privilege IAM roles, VPC-only training and endpoints, KMS encryption, and network isolation
   • Apply Bedrock Guardrails and choose between Bedrock API keys and IAM credentials; scan pipeline images with Amazon Inspector

   Use: Amazon Bedrock Guardrails · AWS MLOps Demonstrated (Skill Builder)

5. Week 8 · 10 h

   Practice exams and gap review

   • Take the official practice question set in Skill Builder, then one full timed practice exam under the 170-minute beta limit
   • Tag every miss by domain and task (for example 3.2 or 4.3) and re-read that section of the exam guide
   • Drill the C02-only content (RAG, agents, FM cost and Guardrails) because most third-party question banks were written for C01

   Use: AWS Skill Builder – MLA-C02 exam prep · Tutorials Dojo MLA practice exams

You're ready to book when…

• You can pick the right SageMaker endpoint type (real-time, serverless, async, batch) for any traffic pattern and justify it
• You can name which Model Monitor job catches data drift vs bias drift vs feature attribution drift
• You can explain the C02 additions (vector stores, Knowledge Bases, agent deployment, FM cost controls) without notes
• You score 75% or higher on two different full-length practice exams

Exam tips

• Beta takers get one attempt only: if you fail the C02 beta you must wait for the standard version to retake (AWS beta policy).
• Out of scope per the exam guide is designing full end-to-end ML architectures, so prefer the answer that operationalises an existing design with the most managed service.
• The C02 guide drops SageMaker Neo edge optimisation and bring-your-own-container, so don't over-invest in them even if older courses cover them.