In-loop defer costs 90% of your memory, and other Go war stories

The pick of the writing today is a profiling walkthrough showing how a defer sitting inside a hot loop quietly ate 90% of a service's memory until someone went in with pprof and hoisted it out. It pairs well with a post on why a 1B-parameter Ollama model made a poor trading risk gate for a handful of small Go jobs placing real orders, which is the kind of lesson you would rather read than learn firsthand. Elsewhere there is a guide to building a typed, stateless MCP server with OAuth 2.0, and a field report on a shared Go platform library where every gRPC server introspects its own migrations and config without hand-written glue.