model-writtenrun 7459 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Apr 2021 · 3 entries
- Fujii Masao reported that the default compute_query_id=off silently left pg_stat_statements empty, while Julien Rouhaud explained that enabling calculation unconditionally would impose overhead and interfere with alternative identifier providers #1 #2.
- Magnus Hagander proposed defaulting to auto so consumers could request identifiers without imposing the cost on everyone, and Alvaro Herrera suggested separating the configured setting from runtime enablement #4 #24.
- A parallel design discussion considered explicit provider selection: Andres Freund advocated loading a provider library through a GUC, Tom Lane suggested a dedicated hook, and Julien cautioned against dynamically switching identifier algorithms #27 #28 #31.
May 2021 · 13 entries
- After a two-week lull and Michael Paquier's request for a resolution before beta1, Julien posted the first off/on/auto patch; Magnus Hagander and Fujii Masao objected to its direct rewriting of auto to on and requested a separate runtime flag #32 #34 #36 #38.
- Pavel Stehule proposed diagnosing missing query identifiers when users queried pg_stat_statements, and Julien added a warning and regression tests in the second version; Maciek Sakrejda subsequently reported successful testing #54 #62 #91.
- Fujii clarified that SHOW should retain auto while a separate flag recorded extension demand, including transfer to EXEC_BACKEND children, and Julien accepted that arrangement #84 #85.
- Kyotaro Horiguchi posted a competing proof of concept with a string query_id_provider setting and dynamically loaded provider functions, but Julien questioned how multiple consumers and competing providers would interact; this implementation was not adopted #86 #88 #93 #94.
- Bruce Momjian declined to commit the auto interface because he found its semantics confusing, while compatibility concerns and Tom Lane's objection to unconditional calculation overhead kept auto viable; Alvaro then took responsibility for committing it #95 #108 #114 #116 #118.
- Alvaro's third version preserved auto using a separate flag and enablement helpers, and Fujii found that starting with on and later reloading auto could wrongly disable calculation despite pg_stat_statements being loaded #122 #129.
- Tom independently posted CRC and XOR hashing experiments that reduced measured overhead but raised portability and hash-quality concerns; neither experiment became part of the configuration fix #131.
- Discussion also considered exposing the active provider through a function or read-only setting, but Bruce proposed inspecting pg_stat_activity.query_id for now and postponing a provider-reporting GUC until extension registration was designed #136 #139 #140 #141.
- Alvaro's fourth version fixed the on-to-auto transition, inlined the enablement check and revised documentation, while dropping the warning because existing statistics could suppress it and a useful diagnostic could be developed separately #142.
- Alvaro also questioned errors from custom providers when core calculation was enabled, while Julien warned that removing safeguards could mix incompatible identifiers; these concerns did not produce another attached revision #142 #144.
- Justin Pryzby reviewed the fourth version and requested wording fixes, an assertion and release-note changes; Alvaro incorporated those light edits and committed Julien's patch as cafde58b33, explicitly leaving the warning for separate work #143 #147.
- Bruce subsequently refined the release notes, and the authorship discussion confirmed Julien's credit with Alvaro describing his contribution as janitorial #148 #150 #151 #153.
- The thread ended with the main auto-setting fix landed and the open item closed by agreement; Alvaro planned no further work here, leaving the warning, provider introspection and hashing experiments outside the completed fix #156 #157 #158.