model-writtenrun 7605 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jun - Aug 2020 · 3 entries
- Bharath Rupireddy opened the thread with #1, proposing a core enable_connectioncache GUC to close postgres_fdw connections at transaction end; Vignesh found configuration-placement issues and a regression-test failure in #2 and #3.
- David G. Johnston proposed an explicit disconnect function in #4, while Masahiko Sawada argued for keeping policy inside postgres_fdw through a server option in #6; Bruce Momjian also suggested remote idle-session timeouts in #7.
- Bharath explored recovery from remotely terminated connections in the separate retry prototype #14, but Ashutosh objected to its libpq changes and shared-connection assumptions in #17, and Bharath later asked readers to ignore the prototype after build problems in #26.
Sep - Dec 2020 · 6 entries
- After several months, Alexey Kondratov revived the original connection-lifetime problem with the extension-local GUC prototype #29, and Bharath expanded it into disconnect functions, a GUC, a server option, and tests in #35.
- Alexey's reviews in #36 and #38 simplified connection selection, required errors for unknown servers, and clarified transaction tracking and option precedence, producing #42.
- Craig Ringer suggested integrating cleanup with DISCARD ALL in #39, but Alexey argued in #41 that a targeted disconnect function remained useful without discarding unrelated session state.
- Fujii Masao requested a way to inspect cached connections before disconnecting them in #43, prompting Bharath to add an array-returning postgres_fdw_get_connections() in #45.
- Review exposed problems with invalidated connections and disappearing catalog entries: Fujii requested validity information and cached server OIDs in #46 and #60, while Bharath moved the underlying connection-leak fix into a separate discussion in #62.
- Alexey questioned allowing disconnect inside explicit transactions in #66, but Bharath chose to preserve in-use connections and report warnings with partial success in #67; the function-only revision #68 implemented these rules.
Jan - Feb 2021 · 20 entries
- Bharath restored the complete series and extension upgrade handling in #73, then repaired missing SQL files, documentation builds, and unstable regression output through #77.
- Fujii's review in #78 led to proper extension upgrade packaging and record-based listing output in #82, replacing the textual array interface.
- Fujii extracted cached server-OID storage into #83, Bharath tested it in #84, and it landed unchanged as 5e5f4fcd89.
- Further review in #83 drove materialized listing results, NULL server names for dropped servers, and protection against duplicate callback registration, incorporated in #87.
- Fujii extracted the listing function in #90, Bharath stabilized its username-dependent test in #91, and that version landed unchanged as 708d165ddb.
- Zhijie Hou noticed that NULL server names were undocumented in #92; Bharath supplied wording in #97, and Fujii's attached patch #100 landed unchanged as ee79a548e7.
- The remaining disconnect and policy work continued in #98, and Fujii's review led Bharath to abandon hash-table destruction in #102, avoiding its callback-management complexity.
- The disconnect API became separate postgres_fdw_disconnect(text) and postgres_fdw_disconnect_all() functions in #107, followed by safe handling of dropped servers and additional tests in #109.
- After OID-based matching arrived in #118, Fujii demonstrated in #120 that multiple user mappings could create several connections for one server, requiring removal of the early scan exit in #128.
- The permission discussion concluded that disconnecting another user's connections within the same shared session was acceptable in #126, and Fujii prepared the final comments, documentation, and tests in #129.
- The disconnect functions landed from #129 as 411ae64997, with only a documentation typo corrected.
- Tom Lane's buildfarm report in #132 exposed nondeterministic warning order; Fujii proposed temporarily suppressing warnings in #133 and committed that discussion-derived fix as 6adc5376dc.
- A second buildfarm failure reported in #136 required regress_ prefixes on test roles, and Bharath's patch #137 landed unchanged as 0c3fc09fe3.
- With the functions committed, Bharath resent the remaining GUC and server-option patches in #140.
- Tom reported CLOBBER_CACHE_ALWAYS failures in #141 and raised safety and possible-revert concerns in #143, but Fujii explained in #144 that invalidation-time disconnection predated the new functions and preserved connections in use.
- Bharath proposed removing immediate invalidation-callback closure in #146, but Fujii rejected that direction in #147 because it would revive connection leaks without resolving the test behavior.
- Bharath then combined API simplification with transaction-based test fixes in #159; Fujii preserved the valid column and dropped-server visibility, extracted revised tests in #160, and committed them unchanged as f77717b298.
- The follow-up discussion in #166 left the existing listing interface intact, so neither the callback-removal proposal nor the API-removal proposal landed, and no revert occurred.
- Fujii's option reviews in #167 and #169 clarified the value of local connection control despite remote timeouts and required explicitly specified server settings to override the GUC; Bharath implemented matching names and tri-state precedence in #170.
- After the rebase in #171, Fujii simplified the proposal in #172 to a server-only cached Boolean, deferring the GUC because its behavior was awkward for already-open unused connections and unlike other FDW options.
Apr 2021 · 1 entry
- Bharath made the final documentation adjustment in #173, which landed unchanged as b1be3074ac; Fujii's last message #175 confirms that the server option was committed, completing the retained feature set while leaving the GUC deferred.