SagaMaker Pipelines

Generative AI systems evolve continuously. Models are refined with new data, prompt strategies change, safety controls are tightened, and cost optimizations are introduced. Each change can alter model behavior in subtle, sometimes undesirable ways. In production environments, these changes accumulate quickly, increasing the risk of difficult-to-detect regressions.

In these systems, failures are rarely obvious during training. Regressions often surface only after deployment, when models are exposed to real user inputs, long-tail prompts, or sustained traffic. Managing this rate of change manually does not scale. Informal processes make it difficult to reproduce behavior, understand what changed, or recover quickly when problems occur.

SageMaker Pipelines exists to address this challenge by providing execution-level control over how model changes are produced, evaluated, and proposed for production use.

SageMaker Pipelines for managing the execution lifecycle

SageMaker Pipelines manages the execution lifecycle of model changes. It orchestrates and automates the workflow steps for producing, evaluating, and making promotion decisions for new model versions. This distinction is central to understanding its role in production GenAI systems.

Pipelines do not store long-term model state or serve inference traffic. Instead, they coordinate execution steps like running customization jobs, evaluation steps, and policy checks, and then emitting outcomes indicating whether a model version should advance or be blocked. These outcomes are recorded in SageMaker Model Registry, which acts as the system of record for model versions and their approval status.

Foundational components of SageMaker Pipelines

SageMaker Pipelines is composed of a small set of ...