model-writtenrun 50 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- David Steele opened the thread with #1, proposing that pg_backup_stop() return a consistent pg_control copy carrying a flag that prevents recovery without backup_label, while existing backup tools could continue copying the disk file.
- Michael Paquier questioned the SQL API extension's maintenance cost and suggested separating replication-protocol protection from the optional SQL feature in #2 and #4; David first adjusted archive ordering in #3, then split the work and extracted a shared helper in #5.
- David fixed bytea alignment problems in #6 and maintained the series through upstream documentation and catalog changes in #7, #8, #10, and #11, without changing its central design.
- Haibo Yan asked whether the proposal should detect an incorrect backup_label in #12; David limited this work to missing labels in #14, and Haibo subsequently reviewed the code and suggested simpler buffer initialization in #15.
- Michael requested clearer documentation of flag reset and crash behavior in #16, supplied a shared-memory/disk consistency assertion in #18, and identified that the SQL path must copy pg_control before finishing the backup; David integrated these changes and reported crash-restart testing in #19.
- Michael advocated concentrating on replication-protocol backups in #20, and David agreed in #21 that this useful subset could proceed while the SQL extension remained a later possibility.
- Corey Huinker reported a joint review with Surya Poondla, passing tests, and a minor ereport() style issue in #22; David applied that correction in #23.
- After nearly two years without sufficient traction, David withdrew the series in #25, noting that torn pg_control copies remained possible and needed attention.
- David revived the work in #26, dropping the SQL extension and concentrating on the backup-label flag and consistent control-file copies for replication-protocol backups.
- Solai v reported successful pg_basebackup restore and missing-label tests in #27, and Andy Pogrebnoi reviewed the narrowed patch positively in #29.
- Shihao Zhong reviewed and tested the narrowed patch, supplied dedicated primary-and-standby TAP coverage, and identified contradictory advice to remove backup_label in #31; David preserved that work and added conditional hints, documentation, and further tests in #32.
- Andrey Borodin demonstrated that standby checkpoint replay legitimately makes shared-memory control data differ from disk in #33, and Shihao confirmed the assertion problem and warned against presenting pg_resetwal as a recovery remedy in #35.
- David removed the invalid assertion and strengthened the corruption warnings in #36, after which Shihao marked the patch ready in #37; they reserved stronger pg_resetwal handling for separate work in #38 and #39.
- Following another rebase in #40, David split the work in #41 into a backpatchable torn-control-file bug fix and a separate backup-label flag feature.
- Heikki Linnakangas reviewed the fix positively but preferred reading the on-disk control file for consistency with existing SQL functions in #42; David adopted that approach in #45, replacing the shared-memory helper with a locked disk read and CRC validation.
- Heikki also challenged documentation that suggested modifying pg_control or restoring an isolated backup_label in #43; David removed that advice and recommended a complete fresh restore in #46, leaving additional startup hardening for separate work.
- To support the fix on older branches, Michael backpatched the existing sendFileWithContent() length-parameter prerequisite as 0cfd56254a and d2c1e18de9, following the dependency identified in #45 and reporting completion in #47.
- Michael committed the torn-copy portion of #46 as 4a92c8c29b, retaining its executable changes while shortening a comment, and announced backpatches through PostgreSQL 15 in #47; PostgreSQL 14 was excluded because its approaching end of life did not justify the risk.
- The new stack buffer was followed by a Solaris/SPARC buildfarm crash, prompting Michael to propose heap allocation in #48; David reviewed the workaround positively in #49 while warning that it might mask an underlying defect.
- Meanwhile, Heikki explained in #50 why the old snapshot workflow did not justify retaining the hint to remove backup_label; David agreed in #51 to prepare a hint-removal patch before continuing the flag feature.
- A second SPARC failure reported in #52 occurred during backup cancellation, and David argued in #53 that the enlarged stack frame could expose an alignment or unwinding problem and recommended applying the heap-allocation workaround.
- Michael committed #48 unchanged as c97931c16d and announced the palloc0()/pfree() workaround in #56.
- By the thread's end, Michael reported that billbug had returned to green in #57, while Marcel Hofstetter enabled future core collection and noted another Solaris/SPARC machine had remained healthy in #58; the underlying platform failure was still unexplained.
- The torn-copy fix and its SPARC workaround had landed, but the backup-label flag portion of #46 remained pending, with the hint-removal follow-up promised in #51 and a possible future revival of the SQL extension described in #46.