3.1. Selecting and Deploying Models
💡 First Principle: A model deployment is defined by four orthogonal choices — which model, which version, which deployment type, and how much capacity — and they trade off independently. Most "which deployment fits this scenario" questions are really asking you to match one of these axes to a stated constraint (latency, residency, cost, or burstiness).
The planning domain's biggest single concept is the deployment-type decision, because it drives cost, latency, availability, and data residency all at once — and the exam loves scenarios where two options both "work" but only one matches the constraint. Before any of that, though, you select a model from the catalog that fits the capability (modality, context window, reasoning strength) and a version that fits your stability needs.
⚠️ Common Misconception: "Serverless and managed/provisioned deployment are the same thing with different billing." They differ in what they guarantee. Pay-as-you-go standard is best-effort capacity billed per token; provisioned reserves dedicated capacity billed hourly whether you use it or not. The guarantee — not just the invoice — is the real difference.
