model-writtenrun 9308 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Andres Freund proposed removing WITH OIDS and storing catalog OIDs as ordinary columns to simplify storage infrastructure, catalog access, and debugging, posting an initial working prototype in #1; replacing the global OID counter with sequences remained a separate project.
- Tom Lane opposed exposing catalog OIDs through SELECT * and proposed a hidden-column flag, while also recommending complete removal of user-table OID support instead of pg_dump compatibility translation in #6; Andres accepted dropping that translation but favored making catalog OIDs visible in #7.
- Robert Haas and Stephen Frost argued that visible catalog OIDs would be less confusing and that tools should already select explicit columns in #9 and #11, while Andres explained the on-disk incompatibility preventing direct upgrades of WITH OIDS tables in #14.
- After several weeks, Andres posted #16 with conflict resolutions, passing regression tests, pg_dump warnings instead of OID preservation, and documentation changes, retaining harmless WITHOUT OIDS syntax and planning to reject incompatible tables during pg_upgrade.
- Andres replaced implicit executor-side OID assignment with an explicit pg_nextoid() function in #17, avoiding catalog column defaults because of relcache complications and moving bootstrap OID assignment into genbki.pl.
- Noah Misch questioned whether the replacement sequence-based pgbench workload exercised the required locks sufficiently in #18; Andres later reported that both workloads detected weakened LWLocks on PowerPC in #23.
- John Naylor found that ordinary OID columns broke reformat_dat_file.pl and supplied #20, then revised it in #22 after Andres requested preserving OIDs first in the formatted catalog data in #21.
- Andres folded those fixes into #24, completed explicit OID attribute handling in GetSysCacheOid(), added pg_upgrade rejection checks, restored pageinspect support for old tuple headers, and polished tests and documentation; John supplied the remaining bootstrap documentation update in #25.
- The hidden-column discussion produced a separate prototype in #28 that skipped flagged attributes during SELECT-star expansion while retaining them in whole-row values; this experiment was not included in the OID-removal commit.
- Andres committed the main implementation and John's documentation contribution as 578b229718, with final changes preserving pg_language OIDs during replacement and removing duplicate BKI OID syntax, then announced the landing in #30.
- A buildfarm failure exposed missed sepgsql references to the removed system OID attribute, and Thomas Munro supplied and compile-tested #31; Andres applied it unchanged as 937e4e5099 and confirmed that in #32.
- Andreas Karlsson supplied a correction for stray quotation marks in three historical release notes in #33.
- Andrew Dunstan reported failures involving large objects after upgrading older clusters in #34; Andres traced them to copying the now-incompatible pg_largeobject_metadata table and proposed reconstructing its metadata through SQL, including ACLs, in #36.
- Andrew tested the upgrade repair and recommended applying it in #37, identifying only a remaining dump-order difference for large objects; Andres's initial repair explicitly left the pre-9.0 reconstruction path unresolved in #36.
- Andres landed the upgrade repair with comment-only adjustments as 12a53c732c and discussed the unresolved ordering question in #40, then applied the release-note correction unchanged as 54bb22f66a, acknowledged in #41.
- Andrew accommodated large-object ordering differences by allowing additional fuzz in buildfarm upgrade comparisons in #42.
- Months later, Noah demonstrated that the replacement concurrency test missed the original xlc 13.1.3 failure and posted #43, using concurrent enum creation to generate many OIDs; it landed unchanged as 8a0cbb8852 with a PostgreSQL 12 backport.
- Andres subsequently suggested direct pg_nextoid() calls as another test strategy in #44, but Noah explained the tradeoffs in #45 and pointed to the already-landed commit in #47; the thread ended with Andres's acknowledgment in #48, with the main change and repair patches committed and the hidden-column experiment unlanded.