model-writtenrun 5959 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Masahiko Sawada reported that parallel vacuum workers lacked PROC_IN_VACUUM and therefore held back oldest xmin, with an analogous PROC_IN_SAFE_IC problem for concurrent index operations; his initial patch copied the leader's flags during generic parallel-worker startup #1.
- Alvaro Herrera questioned the effects of copying other process flags, and Sawada clarified that autovacuum did not support parallel vacuum #4 #6; Peter Geoghegan separately called for better visibility into processes blocking vacuum, but this thread produced no implementation of that broader proposal #5.
- Michael Paquier requested documentation of the autovacuum restriction, while Greg Nancarrow requested tests #7 #8; Sawada explained that stable tests lacked worker suspension controls and added a vacuum-worker assertion and comment instead #9 #10.
- Herrera argued that flags must be installed before the worker exposes its restored snapshot and proposed doing this in ProcArrayInstallRestoredXmin(), with exclusive locking when flags need copying #11 #12; Sawada adopted that integration and added an index-worker assertion #14.
- Matthias van de Meent tested the revised patch and confirmed that it fixed blocked xmin advancement and a parallel-worker horizon-retreat case, while identifying broader xmin soundness issues as separate work that should not delay this fix #15.
- Amit Kapila questioned the conditional lock upgrade and whether flag copying belonged in snapshot restoration #16; Andres Freund argued that unconditional exclusive locking was inexpensive in this context and proposed asserting that copied flags belong to an allowed set #20.
- After discussion of returning to a separate helper, Herrera favored installing xmin and flags together to avoid exposing xmin without its exemption flags #26; Sawada retained that design, switched to unconditional exclusive locking, and introduced PROC_COPYABLE_FLAGS #28.
- Kapila manually reproduced the vacuum blockage using debugger-controlled workers, confirmed that the patch allowed oldest xmin to advance, and posted the final development-branch patch with minor comment edits #29.
- Sawada confirmed similar manual testing with pgbench and VACUUM VERBOSE and supplied PostgreSQL 13 and 14 backport patches #30; no stable automated test emerged, and Kapila planned a final check before committing #31.
- Kapila committed #29 unchanged as 0f0cfb4940 on November 19, crediting Sawada as author and backpatching through PostgreSQL 13.
- John Naylor marked the work committed in the CommitFest application, and Kapila acknowledged the update, leaving this fix complete with no remaining patch under review in the thread #32 #33.