model-writtenrun 7578 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Takamichi Osumi proposed making archive recovery fail when it encounters WAL generated with wal_level=minimal, replacing a warning that allowed potentially incomplete recovery, and initially requested backpatching in #1.
- Sergei Kornilov established that the problematic case required hot_standby=off because hot standby already rejected such WAL, while Kyotaro Horiguchi requested automated reproduction; Osumi added the TAP test in #6 following #3 and #2.
- Laurenz Albe tested the patch and requested a clearer fatal error and actionable hint in #7; Osumi incorporated that feedback and covered both standby and ordinary archive recovery in #8.
- Fujii Masao warned that the stronger safeguard removed a way to salvage data from an otherwise unstartable database and identified an unreachable hot-standby error in #10; Laurenz favored stopping unsafe recovery and leaving any corruption-override setting to separate work in #13.
- Osumi removed the redundant error and reorganized the surrounding conditions in #17, but after the discussion resumed in March, David Steele requested keeping the patch focused by undoing that restructuring in #21.
- Fujii accepted proceeding while proposing a recovery-to-an-earlier-point hint in #22, and Kyotaro examined the recovery scenarios and supported the safeguard in #23 and #24; Osumi combined the feedback in #27.
- Fujii's review in #28 led Osumi to remove unnecessary test assertions and explicitly wait for WAL archival in #30; Fujii then consolidated duplicate recovery tests and removed redundant configuration checks in #32.
- Kyotaro suggested identifying a recovery boundary by LSN in #34, but Fujii discovered that a failed replay persists minimal wal_level in pg_control, causing retries to fail before checking recovery targets, and proposed removing the misleading hint and documenting the risk in #37.
- An intermittent coverage-instrumentation failure delayed confidence in the tests, but suggestions from Takayuki Tsunakawa and Kyotaro in #35 and #36 helped Osumi establish that unpatched HEAD also reproduced it in #41 and #42.
- Fujii posted the final patch in #43, removing the recovery-target hint and documenting that changing wal_level to minimal invalidates earlier backups for archive recovery; proposed additional wording was left for possible later improvement after #44, #45 and #47.
- Fujii committed #43 unchanged as 9de9294b0c and announced the landing in #50, retaining Osumi's author credit and crediting Laurenz, Kyotaro, David and Fujii for review.
- Windows buildfarm failures exposed missing TCP/IP replication authorization in the new test; Fujii diagnosed the missing allows_streaming initialization option and attached a one-line fix in #51, while Tom Lane independently reported the Windows failure in #52.
- Fujii committed that fix unchanged as 8ee9b662da and confirmed it in #53; Tom's message reported the failure rather than reviewing the fix, so it does not establish reviewer credit for this commit.
- The thread ended with Osumi thanking the participants in #55, with the safeguard and its Windows test correction committed and no remaining attached patch awaiting review.