model-writtenrun 6371 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Alexander Lakhin reported reproducible relmapper checksum failures during concurrent VACUUM workloads across supported PostgreSQL versions and supplied a reproducer in #1.
- Heikki Linnakangas suspected non-atomic concurrent file access and posted shared-lock protection for load_relmap_file() in #2, while asking whether the same assumptions endangered pg_control.
- Thomas Munro reproduced the failure on ext4, supported read/write locking, and identified the need to lock relmap_redo() as well in #3; Alexander confirmed that the attached patch eliminated the failure with fsync disabled in #5.
- Tom Lane questioned whether the behavior violated kernel guarantees in #4 and #7, while Michael Paquier advocated defensive locking and distinguished existing control-file locking in #6; Thomas supplied a standalone concurrent-I/O reproducer in #9, which Alexander also reproduced in #10.
- Heikki concluded that relmapper reads needed locking regardless of the disputed guarantees and separately proposed redundant control-file copies against torn writes in #11; Thomas supported redundancy with a synchronization barrier in #12.
- Heikki committed #2 as b6d8d2073f, changing only the atomicity commentary, and announced the landing while inviting separate control-file work in #13.
- Thomas pointed out that the landing still omitted relmap_redo() locking in #14, prompting 9b8ed0f52b to lock the WAL replay writer and correct its comment; Heikki confirmed the fix in #15.
- Thomas ended the thread with an experimental control-file patch and corruption-recovery tests in #16, writing both copies with a synchronization barrier instead of alternating timestamped copies; the relmapper fixes had landed, while this separate experiment remained open for discussion with copy selection and broader redundancy unresolved.