model-writtenrun 8737 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Victor Yegorov reported that bt_metap() failed on an out-of-range integer and suspected block-number formatting; Peter Geoghegan identified the oldest_xact column's incorrect SQL type as the likely cause #1 #2.
- Victor found that VACUUM temporarily cleared the failure, then followed Alvaro Herrera's diagnostic advice and supplied a backtrace; Peter explained that oldest_xact uniquely combined unsigned formatting with an int4 declaration #3 #4 #6 #7.
- Peter submitted a master-only patch using bigint for block and level fields, xid for oldest_xact, and float8 for the tuple count, with an attribute-count check requiring users of older declarations to update pageinspect; he left analogous block-number input changes outside its scope #5.
- Peter committed that patch with expanded comments and revised commit prose as 691e8b2e18, announcing the fix in #8.
- Andres Freund reopened the stable-branch question because affected indexes would remain impossible to inspect for years, proposing that oversized transaction IDs display wrapped values instead of causing errors #10.
- Peter considered a NOTICE and HINT warning about unreliable values, but Andres proposed simply changing unsigned formatting to signed formatting, matching the existing behavior of root block fields; Peter accepted the workaround #11 #12 #13.
- After clarifying the affected release range, they settled on applying the workaround only to PostgreSQL 11 and 12, which introduced the oldest_xact field #14 #15.
- Peter announced the stable-branch workaround as landed in both releases, represented by c788115b5e; the thread ended with corrected types on master and potentially negative values replacing errors on the affected stable branches #16.