model-writtenrun 5771 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- The thread began with Noah Misch’s announcement of fdd965d074, which made RelationBuildDesc() retry after relevant invalidations to prevent concurrent index builds from producing indexes that silently miss rows; the fix was backpatched through PostgreSQL 9.6 #1.
- Months later, Tomas Vondra reported that CLOBBER_CACHE_ALWAYS builds could loop indefinitely during initdb, with the avocet buildfarm animal stuck for 104 days, and traced the loop to invalidations during relation descriptor construction #2.
- Andres Freund and Michael Paquier discussed whether the imminent minor releases needed intervention, but Tomas found the failure confined to CLOBBER_CACHE_ALWAYS builds, and Noah judged emergency action unnecessary because the regression had already shipped in November #3 #4 #5 #6 #7 #8.
- Noah initially questioned whether the debug-discard handling covered initdb, then discovered that PostgreSQL 13 and earlier had an additional invalidation call omitted from the backpatch; his attached fix passed true to InvalidateSystemCachesExtended() at that call site #8 #10.
- The prolonged buildfarm hang also exposed a monitoring problem: Michael discussed timeout configuration, while Tom Lane explained that wait_timeout terminated clients without reporting the failure and advocated fixing reporting and making the timeout finite by default #9 #11.
- Tomas confirmed that the patch fixed PostgreSQL 10 and 13 in the thread’s last message #12; Noah subsequently committed its unchanged code with revised explanatory wording and Tomas’s review credit as 1b8462bf1d, covering supported branches 10 through 13, while the buildfarm reporting concern remained unresolved in this thread.