model-writtenrun 5531 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Apr - May 2022 ยท 16 entries
- Tom Lane opened the investigation after wrasse began intermittently choosing bitmap scans instead of the expected index-only scans, suspecting tenk1's relallvisible and recent statistics or VACUUM changes #1.
- Peter Geoghegan proposed waiting for a sufficiently recent vacuum horizon #9, while Tom favored a SQL-accessible helper #13; Andres Freund warned that horizons can move backward #16, and Tom rejected disabling index-only scans because their plan shape was the purpose of the test #19.
- Noah Misch enabled autovacuum logging #22, which let Tom rule out autovacuum workers during the failing interval #23; Tom then committed temporary relallvisible probes directly from the discussion as 5bb2b6abc8.
- Andres posted a separate patch exposing VACUUM VERBOSE diagnostics at DEBUG1 #27, and Peter identified inconsistent message levels and an autovacuum assertion needing adjustment #28 #29; this attachment has no matching landing among the candidates.
- The probes showed relallvisible remaining zero throughout index creation #30, and Noah supplied a reproducer #44; Tom finally reproduced the failure after changing his script to run initdb on every iteration and decided further temporary logging commits were unnecessary #50.
- Tom's instrumentation found a vacuum cutoff older than tenk1's inserting transaction #57 and traced it to the logical replication launcher #61, while Andres independently reproduced the same class of failure by delaying the autovacuum launcher #58.
- Tom identified the faulty assumption that databaseId zero uniquely identifies walsenders #62, and Andres proposed an explicit flag to preserve hot-standby feedback's global effect #64; Tom subsequently clarified that both launchers could incorrectly hold back user-table horizons #71.
- The investigation also exposed misleading VACUUM log wording: Tom supplied a replacement expressing cutoff age rather than apparent horizon advancement #59, and Peter accepted it #68.
- Tom posted the walsender flag implementation, including a workaround for the restored-xmin assertion #78, and Andres explained why standby feedback must affect horizons across databases #79.
- Peter committed the discussion-derived wording correction as 357c8455e6, also standardizing XID capitalization and relocating related instrumentation.
- Tom committed the walsender patch as 6fea65508a, adding a database-less-client restriction and completed comments to #78, then told Noah wrasse could resume #80; the commit deliberately deferred backpatching while the fix aged in development.
- Broader VACUUM instrumentation and freezing ideas remained separate: Peter proposed calculating FreezeLimit from NextXID #77, Andres distinguished that from cutoff logging #81, and Peter favored additionally reporting the all-visible result #82.
- Alvaro Herrera followed the landed walsender fix with a delta that copied only selected xmin-related flags #83; Tom identified that its OR operation incorrectly preserved existing recipient flags and proposed masked replacement plus a clearer macro name #84.
- Masahiko Sawada supported selective copying and explained the original assertion's purpose #85; Alvaro's revised patch adopted masked replacement and PROC_XMIN_FLAGS #86, after which Masahiko questioned whether the mask should include every xmin-related flag #87.
- In the final message, Alvaro supplied stable-branch adaptations and deferred adding further flag bits to a separate development-only change because older branches interpret the masks differently #88; later that day, 8d061acd12 landed the unchanged development patch from #86 with backports through PostgreSQL 13.
- The supplied history therefore ends with the horizon fix, log-wording correction and flag-copying cleanup landed, while the DEBUG1 attachment #27, broader instrumentation ideas #82 and additional flag bits #88 remain outside those commits; no removal of the temporary probes is recorded.