model-writtenrun 5553 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Aug 2021 · 4 entries
- Julien Rouhaud opened the thread with a patch marking direct and indirect GUC variables PGDLLIMPORT after encountering repeated Windows extension-build failures, and asked whether PostgreSQL 14 could receive it #1.
- Michael Paquier favored targeted additions on the development branch, while Tom Lane questioned treating every GUC as an extension interface; Julien responded with a proof of concept that instead hid previously unexported GUC variables on GCC builds #2 #6 #8.
- Robert Haas favored direct access and platform parity, while Alvaro Herrera and Chapman Flack explored typed getters and change-observer APIs; Peter Eisentraut argued that any required API should also serve core code, and no replacement API emerged from this discussion #17 #19 #20 #33.
- Tom clarified that he would accept automatic exposure without additional source annotations, prompting discussion of header rewriting and export machinery; Julien's Windows experiments and Craig Ringer's explanation showed the awkwardness and limitations of extension-side import workarounds #44 #49 #53 #57.
Feb - May 2022 · 13 entries
- After a gap of several months, Andres Freund reported that the hidden-variable patch failed CI and needed enforcement infrastructure, while Julien declined to invest further without agreement on the intended policy #58 #59.
- The discussion converged on marking globals already declared in headers while leaving currently private variables private: Robert proposed that simpler scope and Tom accepted it as the practical route to Windows parity #71 #73 #74.
- After Andres requested automated checking, Robert supplied a new two-part implementation consisting of a deliberately simple Perl helper and generated PGDLLIMPORT markings across header globals #76 #77.
- Andres reviewed the generated patch, questioned frontend and shared-library markings, and proposed broader function-visibility work; John Naylor explained the existing frontend macro behavior, while Robert ultimately deferred function annotations and hidden visibility to separate work #78 #82 #83 #84.
- Robert's revised series improved declaration detection by stripping comments and trailing whitespace, regenerated the header changes, and proposed landing immediately after feature freeze to reduce conflicts with other patches #84.
- Andrew Dunstan approved the helper subject to adding descriptive and copyright comments, and the release discussion accepted a post-freeze landing #85 #88 #90.
- Robert landed the revised series as the helper commit 80900d4690 and the broad header-marking commit 8ec569479f, adding the requested script comments and updating the generated markings for intervening source changes, then announced completion #84 #91.
- Andres later found that the helper had missed the comma-separated signal-mask declarations, raised remaining globals declared only in C files, and noticed that bbsink operation tables unnecessarily had external linkage #94.
- Michael and Robert agreed to split the signal declarations into the conventional style rather than complicate the parser, and left moving C-file-only declarations into headers outside the agreed scope #95 #96.
- Michael attached a follow-up marking the three signal-mask variables, documenting script arguments, and restoring a marker on __pg_log_level that he believed had been accidentally removed; it landed unchanged as 5edeb57428 #97 #98.
- The separate bbsink issue raised by Andres and accepted by Michael and Robert produced b5f44225b8, which made the operation tables and private target helper functions static without an attached patchset #94 #95 #96.
- Andres objected to marking the frontend-only __pg_log_level variable, Tom confirmed that its earlier removal was intentional, and Michael's final message agreed to undo that part of his fix; 8d33412665 subsequently made the one-line removal #99 #100 #101.
- The supplied commits complete the header-marking series and its immediate corrections; broader visibility machinery, exposing C-file-only globals, and Michael's final suggestion to remove __pg_log_level from the public header remained outside the completed work #84 #96 #101.