model-writtenrun 4257 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Aug - Nov 2021 · 6 entries
- Noriyoshi Shinoda opened the thread with a patch raising failed huge-page allocation diagnostics to WARNING, because the default huge_pages=try could silently fall back to ordinary memory #1.
- Julien Rouhaud requested SQL-visible status that would survive log rotation, while Fujii Masao recommended LOG-level ereport messages; Shinoda combined those ideas with a using_huge_pages GUC #2 #3 #4.
- Kyotaro Horiguchi raised misleading postgres -C behavior and overly verbose diagnostics, so Shinoda removed the GUC pending related discussion elsewhere and shortened the fallback messages with input from Fujii and Tom Lane #5 #6 #7 #9 #13.
- Concern about noisy messages under the default configuration led Shinoda to condition logging on the kernel's huge-page count, but Fujii found that silence could falsely imply success, prompting a redesign that reported allocation with or without huge pages #18 #19 #22.
- Justin Pryzby identified a translation problem in the assembled message, and Masahiko Sawada found misleading success tracking and initdb noise; revisions adopted complete messages, corrected allocation tracking, and restricted logging to postmaster startup #23 #26 #24 #29 #31.
- Sawada's discussion of diagnosing insufficient huge pages led to logging both success and fallback under huge_pages=try, with the requested allocation size on fallback; Shinoda also floated changing shared_memory_size_in_huge_pages but explicitly left that out #32 #33.
Aug 2022 · 1 entry
- After a long period without sufficient reviewer interest, the CommitFest entry was returned and Shinoda withdrew the logging patch in August 2022 #34 #35.
Nov 2022 · 2 entries
- Discussion resumed with Thomas Munro proposing an opt-in verbose mode and John Naylor suggesting fallback even for huge_pages=on, but Andres Freund and Tom Lane defended the existing hard-failure semantics of on #36 #38 #40 #41 #42.
- Justin proposed persistent status through an internal GUC, Michael Paquier supported that direction, and Justin implemented effective_huge_pages in January 2023 before renaming it huge_pages_active and reopening the entry #39 #44 #45 #48.
Jan - Jul 2023 · 9 entries
- Álvaro Herrera found that the new GUC failed to report success for huge_pages=on, while Nathan Bossart highlighted misleading postgres -C output; Justin fixed the success condition and adopted an unknown initial state, also correcting Windows detection #49 #50 #51 #52 #54.
- Nathan questioned using a string for three fixed states, prompting Justin to provide an enum alternative alongside revised documentation and unknown/true/false values #55 #56.
- Stephen Frost and Nathan preferred a boolean system-information function to avoid the unknown state, and Justin implemented pg_huge_pages_active() despite his own preference for a GUC #59 #62 #66.
- Michael discovered that the function's global boolean would not reach Windows EXEC_BACKEND children, and Justin concluded that the GUC machinery already provided the required propagation without additional infrastructure #74 #75 #76.
- Justin returned to the earlier GUC approach as huge_pages_status, reused the existing enum with on/off/unknown, removed GUC_RUNTIME_COMPUTED, and reported Windows testing #78.
- Michael requested clearer unknown-state documentation and broader EXEC_BACKEND checking; Justin then fixed the System V shared-memory path and added an assertion ensuring that a running instance never reports unknown #79 #81.
- Kyotaro accepted a GUC with an unknown state, and Nathan ultimately endorsed the revised patch and confirmed that read/write_nondefault_variables already handled EXEC_BACKEND propagation #83 #86 #88.
- Michael requested comments explaining the allocation checks and proposed TAP coverage; Justin clarified the assertion, and Michael supplied the final revision with Windows status assignment moved after allocation checks, two TAP checks, and editorial cleanup #87 #90 #91 #92.
- Michael committed the final attachment unchanged as a14354cac0, crediting Justin as author, and closed the thread by reporting additional EXEC_BACKEND testing on Windows and macOS; the status GUC landed, while the original additional startup logging remained withdrawn #92 #93 #35.