model-writtenrun 3175 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Fujii Masao reported that summarize_wal could be enabled by configuration reload while wal_level was minimal and proposed a GUC check hook to reject that configuration in #1.
- Nathan Bossart, Jelte Fennema-Nio, and Tom Lane challenged the hook's dependence on another GUC's value and discussed a more general consistency-check mechanism in #2, #4, and #7.
- Robert Haas proposed preventing unsafe summary generation inside the summarizer in #8, and Fujii Masao identified the additional risk that old minimal-level WAL could still be summarized after restarting with a higher wal_level in #9.
- Robert Haas replaced the hook with WAL-record metadata and fast-forward processing in #10, arguing that checkpoint records must identify their original wal_level and that preventing corrupted incremental backups justified changing WAL and control-file formats after beta2.
- Jakub Wartak's off-list testing exposed broken incremental backup chains across timeline switches, prompting Robert Haas to add wal_level to XLOG_END_OF_RECOVERY in #11; Robert also explained that skipping an initial partial checkpoint cycle was safe and that making summarize_wal startup-only would not solve the historical-WAL problem.
- Fujii Masao demonstrated that shutdown checkpoints could retain the previous wal_level across a restart in #12, leading Robert Haas to add XLOG_PARAMETER_CHANGE handling in #25.
- Fujii Masao identified unused shared-memory state in #27, and Robert Haas removed it and added a regression test proving that incremental backups fail across an intervening minimal-level period in #29.
- Discussion of the original GUC hook continued: Nathan Bossart explained why this particular combination could work in #35, but Robert Haas rejected its fragile assumptions and unnecessary restrictions on processing older safe WAL in #38, with Nathan agreeing in #39.
- Fujii Masao requested documentation for advancing summarizer positions without writing files in #32 and accepted landing only the fast-forward solution in #40; Robert Haas incorporated those documentation changes in #41, leaving the proposed GUC hook and additional user-facing diagnostics out of the fix.
- Fujii Masao approved the final patch with a wording correction in #42, and Robert Haas committed #41 as 402b586d0a with that correction and comment wrapping, confirming the PostgreSQL 17 backpatch in #43.
- The thread ended with Fujii Masao's thanks in #44; the WAL-safety fix had landed, with no further bug report or replacement patch pending in the thread.