1.2. The ML Lifecycle on AWS

💡 First Principle: Every ML system follows the same four-stage lifecycle—Data, Model, Deploy, Monitor—and these stages form a loop, not a line. Understanding this loop is the single most important mental model for the exam, because every question maps to a specific stage and the transitions between stages.

What happens when you think of ML as a linear process—collect data, train model, deploy, done? You end up with models that rot in production. Real-world data changes constantly: customer behavior shifts seasonally, new product categories appear, fraud patterns evolve. Without the monitoring stage feeding back into data preparation, your model becomes stale within months. The exam tests whether you understand this cyclical nature by presenting scenarios where the correct answer involves triggering retraining based on monitoring signals.

Think of the ML lifecycle like the water cycle. Data evaporates (gets ingested and transformed), forms clouds (model training), falls as rain (deployment serves predictions), and flows back to the ocean (monitoring captures new data that feeds the next cycle). If any stage breaks, the whole system degrades.

⚠️ Common Misconception: Candidates often treat the lifecycle as linear — build once, deploy, move on. The exam presents scenarios where the correct answer is to retrain or adjust the data pipeline, not to tweak the deployed model. If a question describes declining prediction accuracy over time, the answer almost always involves the Monitor → Data feedback loop, not manual model adjustment.