Reclaiming CPU for free with Go's Profile Guided Optimization

Cloudflare’s Observability team outlines how they deployed Go 1.20’s Profile Guided Optimization (PGO) for the globally deployed wshim service to reduce CPU usage. They gathered representative CPU profiles from high-load Tier 1 datacenters via pprof, merged them, built with the -pgo flag, and measured ~3.5% immediate reduction (~97 cores) and ~71 cores saved on a 7-day average, validating the efficacy of PGO. The post also proposes next steps including automating profile collection, refining a two-step deployment approach, improving sampling strategies, and exploring BOLT/LTO.