1.2.1. Data → Model → Deploy → Monitor: The Four-Stage Loop
💡 First Principle: Each stage of the ML lifecycle has a primary question it answers, and understanding these questions tells you which AWS services are relevant and which exam domain is being tested.
Stage 1: Data Preparation (Domain 1 — 28%) answers "What do we learn from?" This stage handles ingestion from sources like S3, Kinesis, and databases; transformation through Glue, DataBrew, and Data Wrangler; feature engineering; and data quality validation. The critical output is clean, properly formatted, bias-checked data ready for training.
Stage 2: Model Development (Domain 2 — 26%) answers "What patterns exist in the data?" This stage covers algorithm selection (built-in algorithms, foundation models, AI services), training (epochs, batches, distributed training), hyperparameter tuning with AMT, and performance evaluation with metrics like F1, RMSE, and ROC-AUC.
Stage 3: Deployment & Orchestration (Domain 3 — 22%) answers "How do we serve predictions at scale?" This stage handles endpoint selection (real-time, serverless, async, batch), infrastructure provisioning (compute, containers, auto scaling), CI/CD pipelines (CodePipeline, SageMaker Pipelines), and deployment strategies (blue/green, canary).
Stage 4: Monitoring & Security (Domain 4 — 24%) answers "Is it still working correctly and securely?" This stage covers data drift detection, model performance monitoring (Model Monitor), infrastructure optimization (CloudWatch, Cost Explorer), and security (IAM, VPC, KMS). When monitoring detects degradation, it triggers the loop back to Stage 1.
⚠️ Exam Trap: Questions often present a scenario spanning multiple stages and ask which service handles a specific stage. A question about "a model's predictions becoming less accurate over time" is testing Stage 4 (monitoring/drift), not Stage 2 (model retraining)—even though retraining might ultimately be needed. Always identify which stage the question is asking about first.
Reflection Question: A real-time fraud detection model starts flagging legitimate transactions as fraudulent after a holiday shopping season. Which lifecycle stage detected the problem? Which stage fixes it? Which AWS services are involved in each?
