Why LLM inference splits prefill and decode, and how that leaked user code at Cloudflare

为什么大模型推理都在做 PD 分离，而它又如何漏出了别人的代码

Prefill-decode disaggregation speeds up inference by splitting the two stages, but Cloudflare's GLM-5.3 deployment leaked cache across tenants through a request-handling bug. Its postmortem says 155 long-context replies mixed in other users' conversations or code: over-limit requests were replaced with a one-token placeholder yet still crossed nodes, and decode nodes read stale cache that had not been overwritten.

Why it matters: Cloudflare's fix adds transfer-count checks and delayed memory reclamation, folding data overwrite and page ownership into safety checks when nodes exit abnormally.