SRE Is Weird, Down the Stack
summary
Infrastructure SRE work uncovers millions of lines of opaque firmware that break real systems: expect 30–50M lines of invisible firmware, mysterious 100 ms SMM pauses from 16‑bit handlers that freeze all cores, and 35 minute firmware flashes caused by dumb microcontroller I/O. Workarounds are common, upstream fixes rare, and cloud shifts (cgroups v2) can suddenly raise CPU by ~30% for Go binaries. Practical steps: demand source/open BMCs, benchmark across machine types, track QPS-per-dollar, and share hardware/cloud test suites (Open Compute, contests).
AI-written summary · about