What we actually change in production — not a newsletter.
Before we touch the rest of the system, we leave one real, reproducible path — with rollback in a single command. Every other service copies that pattern.
SLOs, error budgets and burn-rate alerting. On-call stops being a tax when every page justifies itself.
An AI gateway is not a chat box. It is keys, quotas, traces, and a switch for when the data cannot leave for a vendor.
State splitting, modules that encapsulate decisions, policy over the plan, and credentials that do not outlive the job.
Requests and limits from data, three scaling loops that ignore each other, probes that do not self-harm, and an attributable bill.
Real inventory, networking first, verified replication, and a cutover with abort criteria written in the cold. No Saturday war room.