model-writtenrun 201 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Zsolt Parragi opened the thread with #1, showing that autovacuum leaders updated their own balanced cost limits without publishing them to parallel workers, and proposed propagation after each cost-limit update with a regression test.
- Daniel Gustafsson favored fixing the code over weakening the documentation but questioned the timing-dependent filler-table test in #2; Bharath Rupireddy proposed centralizing propagation in #3, then accepted the smaller change and integration into the existing test suite in #6.
- Zsolt moved the test into the existing suite in #7, and Masahiko Sawada identified a remaining race where the second worker could finish too soon in #8; Zsolt added a dedicated holding injection point in #9.
- Daniel and Bharath supported the dedicated injection point in #10 and #11, but CI and local instability led Zsolt to suppress unrelated autovacuum activity through test thresholds in #14.
- Daniel committed #14 as 4af0528a0e, preserving its code and tests while editing the commit prose, backpatched the fix to PostgreSQL 19, and confirmed completion in #16.
- After that landing, Nikolay Samokhvalov reported a separate gap in #17: a leader waiting for parallel workers neither processed configuration reloads nor noticed cost rebalancing; he supplied an independent reproducer and a callback-based fix with periodic polling.
- Masahiko confirmed the fix but considered a generic callback excessive in #18, prompting Nikolay to move the timed wait into vacuumparallel.c with an atomic completed-index counter in #19.
- Bharath rejected duplicated wait machinery and recurring polling wakeups in #20, proposing instead to refresh costs inside the existing worker-finish wait and wake balanced autovacuum workers when their count changes.
- Daniel verified Bharath's alternative with the reproducer and check-world in #21, and Zsolt integrated tests and simplified the refresh helper in #22.
- Manu tested both approaches, demonstrated that reload and rebalance tests failed when their respective fixes were removed, measured the avoided polling wakeups, and supplied an additional worker-departure test in #23.
- Masahiko proposed targeting only waiting leaders through a condition variable in #24; Manu's instrumentation in #25 showed many unnecessary latch calls but a small practical cost, and Zsolt and Daniel favored the less invasive solution near release in #26 and #27.
- Manu reported no measurable CPU penalty in #28, while noting a small unexplained throughput difference; Bharath then separated the production fix from the tests in #29 so the fix could land without delaying PostgreSQL 19.
- Manu repeatedly tested the split series in #30 but found that unrelated worker starts could mask the worker-departure path, leaving that additional coverage without a deterministic test.
- Masahiko accepted the latch approach and requested a shorter helper comment in #31, which Bharath incorporated in #32; Daniel endorsed the production fix and requested test cleanups in #33.
- Following Masahiko's suggestion in #34, Zsolt revised the tests in #35 to let injection points disable leader participation, changed polling failures to TAP assertions, and removed a redundant standalone reload test while leaving the production fix unchanged.
- Masahiko prepared and tested the production-only patch in #39, then committed it unchanged modulo rebase as 42e96cf2fe with Bharath credited as author, backpatched it to PostgreSQL 19, and announced the push in #41.
- The tests remained separate: Masahiko supplied a delta in #42 that replaced three injection points with one worker-start hold and DEBUG2 propagation logging, allowing the leader to finish all indexes before testing updates during its wait.
- Zsolt approved and folded that delta into the remaining test patch in #43, crediting Nikolay as author and Masahiko as co-author; at the thread's end both production fixes had landed, while the revised regression tests remained submitted for review and CI.