model-writtenrun 8322 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Nov 2018 - Jan 2019 · 5 entries
- David Steele proposed removing exclusive backup mode in PostgreSQL 12 because a crash during backup could leave backup_label requiring unavailable WAL and prevent automatic restart #1.
- Robert Haas opposed removing a widely used interface after a short deprecation period, while Peter Eisentraut explained that maintaining a database connection during filesystem copying complicated existing shell scripts #4 #12.
- Steele posted the initial removal patch, deleting exclusive backup state and shutdown handling, simplifying the SQL API and documentation, and replacing a logical-decoding test's filesystem hot backup with a stopped-server copy #28.
- Eisentraut challenged changing the meaning of existing SQL calls, and Haas proposed retaining the exclusive arguments but rejecting requests unless explicitly false so old scripts would fail visibly; Stephen Frost accepted that approach #45 #49 #67.
- Without consensus for PostgreSQL 12, Steele deferred removal toward PostgreSQL 13 while participants advocated stronger deprecation documentation #53 #70 #71.
Feb 2019 · 4 entries
- The February 2019 discussion exposed additional migration barriers: Christophe Pettus warned about blocked upgrades, Andres Freund proposed a wrapper to maintain the connection while running a copy command, and Laurenz Albe highlighted backup products limited to separate pre-backup and post-backup hooks #78 #80 #83.
- Albe suggested retaining WAL during exclusive backups, but Frost and Steele objected to disk exhaustion and longer crash recovery; Haas subsequently questioned whether even PostgreSQL 13 removal had sufficient support #84 #86 #93 #107.
- Fujii Masao proposed a backup_label.pending file with explicit restore-time renaming, but Steele and Albe warned that unchanged restore procedures or a missing recovery.signal could silently recover from the wrong checkpoint #120 #121 #161.
- The discussion shifted toward documenting failure scenarios, and Steele opened a separate documentation effort while rejecting routine client warnings that might encourage users to suppress useful messages #153 #157 #166 #173.
Jun - Jul 2020 · 8 entries
- After more than a year, Steele returned with a rebased patch that incorporated the SQL compatibility safeguard and cited a working shell implementation of nonexclusive backups; returning backup metadata at backup start remained a possible extension rather than part of the patch #174.
- Albe, Bruce Momjian, and Haas continued to oppose outright removal, while Frost argued that the restart hazard and maintenance burden justified removing or repairing the deprecated interface #181 #184 #185 #191.
- Magnus Hagander proposed disconnectable nonexclusive backups that return the label and a unique cookie at start, track outstanding backups, and accept the cookie at stop, aiming to support separate backup hooks without retaining exclusive mode #187 #193.
- Haas warned about losing automatic cleanup on disconnect, and Steele explored tracking the oldest starting LSN and a backup count to manage forced full-page writes, while Hagander challenged duplicate-stop and overlapping-backup cases #191 #195 #197 #209.
- Steele considered retaining one shared-memory backup slot but remained concerned that users would put backup_label in the live data directory and that removing shutdown cleanup could worsen the consequences #214.
- Jehan-Guillaume de Rorthais and Hagander proposed exposing manual backups through progress monitoring and administrative functions, with possible timeout-based cleanup, but no implementation was attached #210 #212 #215.
- Steele withdrew the patch on July 10, 2020, calling the approach a dead end and promising a new thread for a design incorporating the discussion, leaving both removal and the alternative API unresolved here #218.
- The eventual removal landed in April 2022 as 39969e2a1e, crediting David Steele and Nathan Bossart, but its matching final patch is #504432 in thread #45155, so the landing is attributed to that later thread.