model-writtenrun 2794 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jan 2022 · 2 entries
- Jelte Fennema-Nio proposed nonblocking query cancellation using libpq's existing connection machinery, gaining encryption support while explicitly giving up signal-handler safety, and included cancellation tests in #1.
- Andres Freund identified a duplicate Windows export ordinal in #2, and Jelte corrected the build and tests while adding a separate test-module README patch in #3.
Mar - Apr 2022 · 3 entries
- Jacob Champion tested GSS cancellation and questioned the expanded API in #4; Tom Lane favored an encrypted blocking interface in #5, while Andres and Robert Haas defended nonblocking support for backend and event-loop users in #6 and #7.
- The discussion established that encryption could not preserve PQcancel's signal-safety promise, and Jelte tried a PGconn-based interface that repurposed PQrequestCancel, added PQrequestCancelStart and PQconnectComplete, and documented the behavior in #12.
- Jelte then pinned cancellation to the original server address, added nonblocking postgres_fdw cancellation and tests, and replaced timing sleeps with a monitoring connection across #14, #15, and #17.
Jun 2022 · 1 entry
- Tom committed the expanded libpq_pipeline README from #17 as f40346ff0b, but rejected changing the old API's behavior, mixing two kinds of PGconn, and breaking enum ABI compatibility in #20.
Sep - Nov 2022 · 1 entry
- The attachment in #21 accidentally repeated the old implementation without the committed README; Jacob caught the mismatch in #22, and #23 finally supplied the separate PGcancelConn API while preserving existing functions.
Jan - Apr 2023 · 3 entries
- After rebasing for libpq error-reporting changes, Jelte substantially reorganized the implementation in #27: private address arrays and copied host data replaced connection-state special cases, and EOF handling and caller migration became separate components.
- Denis Laxalde tested the interface through Psycopg, supplied documentation fixes in #40, and argued that blocking cancellation should consume an existing reusable PGcancelConn in #42; Jelte adopted that design in #43.
- Denis supplied missed caller updates in #44, and Jelte incorporated them while dropping the committed address prerequisite in #45; subsequent rebases eventually removed the already-applied pgindent component in #48.
Nov 2023 - Apr 2024 · 16 entries
- After further rebasing and a long review gap, Thomas Munro's spelling review in #50 prompted a standalone cleanup, while #52 abandoned the proposed -2 EOF convention in favor of existing errno-based EOF detection.
- Álvaro Herrera committed the spelling component of #54, with one additional correction, as 8c9da1441d, then proposed moving cancellation code into fe-cancel.c in #55.
- Jelte implemented the file split and fixed Makefile, macOS, and Windows build issues through #61; Álvaro committed that component as 6d4565a05f, renaming and relocating the integer-parsing helper.
- Álvaro supplied incremental cleanup patches in #64, Jelte incorporated them in #65, and the host-release extraction and internal-function exports landed as 53747f7222 and 774bcffe4a.
- Further API review in #66 and #67 led to explicit PQcancelStart and PQcancelBlocking operations and substantially expanded documentation in #69.
- Denis rechecked the feature and supplied documentation corrections in #72; Jelte added missing connection-state documentation in #73 and subsequently renamed object creation to PQcancelCreate while retaining a distinct wrapper type for compiler checking in #76.
- The preparatory components of #76 landed as 095493a377 for connection-status documentation, 319e9e53f3 for legacy cancellation tests, and 4dec98c2af for moving pg_cancel; the test commit additionally introduced PID-based monitoring and connection copying.
- Álvaro posted the remaining feature and caller patches in #77, and Jelte supplied final comment revisions through #81; the encrypted blocking and nonblocking API landed as 61461a300c, chiefly with documentation edits described in #82.
- Buildfarm failures reported in #83 and #86 exposed cancellation-test timing races, and Jelte's idle-before-active synchronization patch in #87 landed with helper cleanup as 1ee910ce43.
- Jacob demonstrated that cancellation could still arrive between Bind and Execute in #89; the test was temporarily disabled, and Jelte proposed switching it to simple-query protocol and restoring it in #90.
- Álvaro instead restored the test using the PgSleep wait event to synchronize extended-query execution in 6b3678d347, replacing the workaround proposed in #90.
- The dblink, isolationtester, and fe_utils portions of #90 landed as 66ab9371a2, while Álvaro retained postgres_fdw for more review in #93; psql and pgbench remained outside the migration because their signal-handler usage required larger refactoring.
- Noah Misch proposed a reusable backend helper in #94, Jelte implemented it in #95, and Álvaro revised its error handling in #96; after dropping unsafe previous-error tracking, it landed as 2466d6654f.
- The new helper exposed a macOS translation-linking failure in #101, so Álvaro removed its translation wrappers in #102, noting that contrib lacked translation support.
- The postgres_fdw test could time out while executing RESET itself, so Jelte proposed transaction-local statement_timeout in #104; following debate over probabilistic test coverage, that inline fix landed as ecf741cfae.
- Tom's C++ headerscheck report in #112 prompted Álvaro's const-correctness patch in #113, which Jelte approved in #114 and which landed unchanged as b8b37e41ba.
Jun - Oct 2024 · 7 entries
- Alexander Lakhin reproduced both ignored cancellation requests and incorrectly attributed timeout errors in #116; Jelte proposed preserving timeout indicators in #117 and added a test-timing workaround after Alexander identified extra remote queries in #118.
- Noah found that PQcancelCreate marked the wrong connection bad on allocation failure in #120; Jelte added the exact one-line correction to #121, and it landed as 6d2ac55491, leaving the two earlier proposals pending in #122.
- Alexander traced a separate indefinite wait to Cygwin socket polling in #123; Álvaro split the cancellation regression into a platform-skippable file in #125, then committed the corrected split as d329a515f4.
- Alexander supplied logging instrumentation in #130 showing that a Windows socket-write error could be hidden while the cancellation path waited for a response to a packet never sent.
- Tom revisited the pending test workaround in #133 and adapted #122 to self-join a table without remote estimates, initialize the remote transaction before timing, and allow 100ms; this landed as 8749d850f9, as explained in #134.
- The separate timeout-indicator proposal remained disputed: Tom questioned disabling-timeout semantics in #133, while Jelte defended preserving an already-triggered timeout in #135 and #137.
- Alexander reproduced another ignored-cancel window between cursor creation and FETCH in #138, and later failures reported in #139 showed that the longer test timeout had not eliminated the race.
Nov - Dec 2024 · 4 entries
- Tom confirmed the idle-session cancellation race and proposed exponentially spaced repeat requests within the existing 30-second cleanup deadline in #142, preferring retries over letting an idle cancellation kill the next query.
- Jelte endorsed the retry approach and suggested clearer naming in #143, then accepted that the retry belonged in postgres_fdw rather than libpqsrv_cancel because the helper does not wait for the running query's result, as resolved in #144 and #145.
- Tom committed the retry patch with naming changes and a reduced 10ms test timeout as 89962bfef6, announcing its development-branch and PostgreSQL 17 landing in the final message #146.
- At the end of the thread, the new API, backend helper, caller migrations, and postgres_fdw retry fix had landed, while the timeout-indicator proposal and Windows lost-write investigation remained unresolved in the supplied discussion, as distinguished in #143.