This module establishes the foundational mindset that "performance lives in the pipeline." Learners will discover that a large language model (LLM) application is a multi-stage system where overall speed is dictated by the slowest component. They will learn to deconstruct a complex Retrieval-Augmented Generation (RAG) architecture, trace a user request through it, and use system diagrams to form an evidence-based hypothesis about the primary performance bottleneck.

In this module, learners move from hypothesis to evidence. They will learn to use system logging and profiling data to quantify the precise latency contribution of each stage in an LLM pipeline. The focus is on designing small, reversible, and hypothesis-driven experiments to prove or disprove their initial findings and distinguish a performance bottleneck's root cause from its symptoms.

This module bridges the gap between a working prototype and a resilient, production-ready service. Learners will design and manage declarative deployments using Helm and Kubernetes, package a multi-component RAG stack, and implement Horizontal Pod Autoscaling (HPA) for dynamic, cost-efficient scaling. They will also master the critical operational skills of performing controlled, zero-downtime rollouts and rapid rollbacks.