Skip to content- Sources: primary, reproduction, discussion
- Summary: The post, dated 2026-09-13, traces a 40ms stop-the-world pause to garbage collector metadata the kernel had paged out to swap, so the collector read its own bookkeeping from disk with every P halted. The author measures this in a mock run under induced memory pressure, on a Hetzner box running kernel 6.8 with MGLRU enabled, rather than in an observed production incident, and reports a median pause of about 51us against a 40ms worst case with the metadata on NVMe. The bookkeeping lives outside the heap and is reused rather than freed, which is what lets it go cold enough for the kernel to age it out while the process stays busy, and a 2026-09-14 update measures Go 1.26 Green Tea against the same case and reports a negligible change.
- Why it matters: Go GC bookkeeping sits outside the heap and is reused rather than freed, so a host under memory pressure can turn a routine pause into disk latency paid with every P halted.