model-writtenrun 9186 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Peter Eisentraut opened the thread after seeing a misleading “Success” error on a corrupted database, proposing explicit SLRU short-I/O diagnostics and a separate return-check cleanup; he explained that the existing reporting structure lacked the byte counts needed for more detailed messages #1.
- Shawn Debnath suggested describing the operations as partial reads or writes and observed that SlruPhysicalWritePage already substitutes ENOSPC when a failed write leaves errno zero #2.
- Michael Paquier recommended avoiding errcode_for_file_access() when errno is zero, keeping diagnostic wording consistent for translators, and covering the omitted CloseTransientFile checks; he also considered generating error strings at failure time but questioned whether that complexity was justified #3.
- Peter's return-check cleanup landed separately as 7e9a4c5c3d, expanding the five SLRU changes from #1 to 45 checks across 19 files and using != 0 instead of < 0.
- Peter returned with the remaining SLRU fix, separating errno-based errors from short-I/O messages, and added a pg_checksums patch distinguishing read errors and short writes #4.
- Michael accepted the simpler SLRU diagnostics and proposed nesting pg_checksums read-error classification inside the short-read check for consistency with other code; he also noted the related PostgreSQL 11 issue but questioned backpatching that far #5.
- The SLRU patch from #4 landed unchanged as 396e4afdbc, while its pg_checksums patch landed with Michael's proposed restructuring as 1d7a6e3eb4; Peter confirmed both were committed to master and PostgreSQL 12 #6.
- After the fixes landed, Bruce Momjian proposed adding an explicit request to report any future “Success” error as a bug, but no implementation followed in this thread #7.
- TAKATSUKA Haruka later reported repeated recovery failures on newly generated standbys and proposed bounded retries for short reads #8; Peter argued that retries would require widespread changes and were unlikely to help serious storage problems #9.
- The thread ended with all submitted changes landed and no retry patch pending: TAKATSUKA Haruka accepted Peter's explanation, identified DRBD as the standby storage, and offered to report reproducible conditions if found #10.