model-writtenrun 3613 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Andres Freund opened the thread with an emergency-autovacuum test design using prepared locks and pg_resetwal, but identified excessive pg_subtrans initialization, a 4GB failsafe threshold, and CLOG alignment as obstacles to routine testing in #1.
- Justin Pryzby suggested sparse relation files in #2, but Andres rejected that approach because of portability concerns and VACUUM potentially allocating the space in #4.
- Andres argued that failsafe checks should not depend on indexes or dead tuples in #5 and proposed periodic checks in #7; Peter Geoghegan agreed to decouple them from the vacuum strategy in #10.
- The initial testing also exposed autovacuum log percentages above 100% after truncation in #1; Peter accepted the diagnosis in #3 and fixed the denominator to use the original table size in fbe9b80610, without posting a matching patchset.
- Peter posted periodic heap-scan failsafe checks in #11; Masahiko Sawada identified the remaining small-table exclusion in #15, and Peter removed it while committing the revised patch as c242baa4a8, confirmed in #16.
- After Anastasia Lubennikova requested the test work in #17, Andres shared his WIP script in #18; Masahiko revived it in #19 with pg_resetwal followed by single-user-mode checks to avoid concurrent autovacuum.
- Masahiko abandoned that sequence in #20 and introduced a test extension that advances nextXid while the server runs, restoring prepared locks and avoiding initialization of every intervening transaction-status page.
- During that work, Masahiko reported a launcher race with autovacuum disabled and only one worker: a replacement launcher can start before the previous worker exits and then leave without launching another worker, as described in #20.
- Heikki Linnakangas revived the work for SLRU testing in #22 and supplied #23, replacing direct XID assignment with helpers that exercise real allocation near important boundaries and adding warning-limit, stop-limit, and repeated-wraparound tests.
- Masahiko found incorrect Meson registration and resource problems in #24, then posted #30 with optional execution, background transactions instead of prepared locks, and explicit checks that failsafe mode triggered.
- John Naylor reported that the background-session API had changed in #31; Masahiko adapted the tests in #32, and John confirmed that they passed in #34.
- Tom Lane rejected running such expensive tests by default in #35, while Daniel Gustafsson requested the established PG_TEST_EXTRA mechanism in #36; Masahiko adopted that gate in #37, retaining all three test scenarios.
- Daniel found that the attempted environment-based timeout extension took effect too late in #38, Michael Paquier requested documentation in #39, and Noah Misch proposed scaling the configured timeout in #41; these reviews led to the separate BackgroundPsql timeout prerequisite and revised tests in #42.
- Daniel requested API documentation fixes in #43, and vignesh C reviewed argument handling, module ordering, and includes in #44; Masahiko incorporated these changes, including xid8 interfaces, in #48, while retaining epoch-and-XID progress output.
- Masahiko discovered an existing false claim that BackgroundPsql timeouts could be changed after construction in #46; Daniel confirmed the limitation in #47, prompting a separate backpatchable documentation fix in #48 and final wording in #50.
- The final series in #50 landed as the unchanged documentation component 5abff2aaa1, the unchanged timeout-option component 334f512f45, and the test module with light Perl edits e255b646a1; Masahiko confirmed the push in #51.
- Masahiko enabled the optional suite on buildfarm animal perentie in #52, providing regular execution despite its exclusion from default test runs.
- Peter Eisentraut reported that Meson installed the test module into production installations in #53; Masahiko supplied #54, Peter verified the behavior in #55, and the patch landed unchanged as 1aa67a5ea6.
- Alexander Lakhin reproduced the previously described autovacuum launcher race as a failure of the landed emergency test in #57 and also reported that the expensive suite ran twice in some buildfarm configurations; this thread records no subsequent fix for either issue.
- Alexander later investigated slow shutdowns and suggested a larger timeout plus PostgreSQL 17 coverage in #58; in the final reply #59, Masahiko attributed that failure to overlapping personal workloads, committed to avoiding the overlap, and enabled PostgreSQL 17 testing, leaving the patches landed but the earlier launcher-race report unresolved here.