Black Friday, three pods, one restart
Picture an AI shopping assistant whose MCP server runs on three pods behind a round-robin load balancer. It's 9:02 on Black Friday. Under the legacy protocol, each shopper's session (their negotiated capabilities, their cart, the paused "which size?" question) lives in one pod's memory, so the load balancer has to pin every shopper to their pod.
At 9:04 the busiest pod runs out of memory and restarts. Several thousand shoppers lose their sessions in the middle of checkout.
Under 2026-07-28 the same incident is a non-event. Every request says who's asking and what it supports, carts are handles stored in a shared database, and the "which size?" question travels inside the request. The next request goes to a healthy pod, which answers it without knowing the old one ever existed.