SageMaker Training Plans — Full Feature Documentation

SageMaker's new GPU capacity reservation feature was launching at re:Invent 2024: no existing content, no prior model, no reference architecture to follow.

Build complete documentation from scratch for a re:Invent launch — one of the most compressed, high-stakes delivery windows in the AWS calendar.

Determined the full content scope: concept introduction, API reference, IAM taxonomy, status lifecycle, checkpointing guidance, and quota management — then designed the information architecture to organize it. Identified that the feature required two distinct user mental models — plan creators vs. plan users — and structured the entire surface around that split. Built the checkpointing section after determining that users needed to understand the failure mode beyond the API. Created the lifecycle and workflow diagrams. The engineering team authored code samples; everything else — scope, structure, organization, diagrams, and written content — is mine. Independent work under a tight deadline, with a lead SDE as technical POC and an SDM overseeing the launch.

Feature shipped at re:Invent 2024 with complete documentation. The API shape was changed based on documentation feedback. The two-persona structure became the organizing principle for the entire feature surface. The lifecycle and workflow diagrams were selected by the solutions architects authoring the launch blog, who used them as the primary illustration of the mental model.