model-writtenrun 9195 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Sean reported an OpenBSD 6.6 compilation failure caused by undeclared FD_SETSIZE in vacuumdb in #1; Tom Lane and Andres Freund traced it to the removed sys/select.h inclusion, and Tom reproduced the same failure in reindexdb on OpenBSD 6.4 in #2, #3, and #5.
- Andres Freund proposed centralizing the limit behind ParallelSlotsMax() in #4, while the missing buildfarm coverage prompted requests for modern OpenBSD machines in #5 and an offer from Mikael Kjellström in #18.
- Andres Freund identified a separate safety problem in #8: FD_SETSIZE limits descriptor values rather than connection counts, so the existing job limit could permit invalid FD_SET() operations; Álvaro Herrera reproduced a buffer-overflow abort in #12.
- The discussion settled on checking actual socket descriptors instead of predicting a safe connection count in #10 and #14, while Tom Lane noted the Windows portability obstacle to simply switching to poll() in #9.
- Michael Paquier first posted the ParallelSlotsMax() compilation fix in #15, then expanded it into a two-patch submission in #16 with a separate runtime check in ParallelSlotsGetIdle().
- Michael Paquier committed the compilation-fix portion of #16 as 56f8f9624b, with minor comment and formatting changes, and left runtime error handling under discussion in #20.
- Andres Freund's review in #21 called for removing the ineffective argument-limit check, validating sockets earlier in ParallelSlotsSetup(), and backpatching because the existing behavior could write outside the descriptor set; Michael implemented that revised remainder in #22, removing the recently added ParallelSlotsMax() wrapper.
- Andres Freund rejected applying the same diagnostic wording to pgbench because its option describes clients rather than jobs in #23, while Álvaro Herrera requested translatable logging and a suggested usable job count in #24.
- Michael Paquier and Álvaro Herrera refined the proposed count in #26 and #27, and Tom Lane established the POSIX lowest-unused-descriptor basis for the hint in #30; Michael's updated patch in #32 added pg_log_fatal(), the hint, and explanatory comments while dropping the pgbench change.
- Andres Freund challenged the hint's i - 1 calculation in #33, and Tom Lane supplied simpler wording without a select() reference in #35; those final diagnostic corrections were incorporated without changing the socket-validation approach.
- Michael Paquier committed the revised #32 as 71d84efba7 for vacuumdb and reindexdb, with the vacuumdb-only branch adaptation represented by 63fc3b1240, and announced completion through PostgreSQL 9.5 in #37.
- The original reporter confirmed a successful OpenBSD ARM64 build in #38; a later discussion in #41 recorded a separate Windows follow-up for FD_SETSIZE declaration order and a missing newline on older branches, and Álvaro Herrera requested consistent cross-branch commit messages.
- By the thread's end, the submitted fixes had landed and attention had shifted to preventing recurrence: the reporter announced the OpenBSD-current scorpionfly buildfarm animal in #43, and Mikael Kjellström confirmed another OpenBSD animal covering all active branches in #47.