model-writtenrun 7664 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
May - Jun 2016 · 1 entry
- Craig Ringer proposed asynchronous libpq batching in #1, using queued commands and application-controlled synchronization to reduce round trips, with documentation and tests; per-command application annotations were explicitly deferred.
Aug - Nov 2016 · 4 entries
- Michael Paquier's initial review found a queue-cleanup crash, portability problems, and unsuitable library error handling in #11; Craig fixed these and moved testing into a TAP-driven module in #13, retaining the embedded queue because simple_list's allocation behavior was unsuitable for libpq.
- Daniel Verite identified a mismatch between documented queue counting and its boolean implementation in #14, and Michael proposed consistent API names, consolidated status, and additional error checks in #22.
- Shay Rojansky and Craig clarified transaction boundaries and error recovery in #23, #25, and #26: clients could insert synchronization points between individual queries without leaving batch mode, while automatic transaction inference and a separate flush-control API were left out.
- Craig explained that psql integration required substantial asynchronous control-loop work and invited others to adopt the patch in #31; Aya Iwata revived it with a rebase in #34, and Vaishnavi Prabakaran took over implementation revisions and split code from tests in #35.
Feb - Apr 2017 · 6 entries
- Vaishnavi applied Michael's follow-up review in #37, while Daniel introduced a separate pgbench companion in #39 to exercise the API and measure batching benefits.
- The two series informed each other: Daniel's queue-count and error-message findings were fixed in #41, and Vaishnavi's review of aborted-state handling and return-value checks produced Daniel's revised pgbench patch in #42.
- Daniel demonstrated that single-row mode failed during batch result processing and proposed a per-command flag in #43; Vaishnavi instead removed premature input parsing in #47 and incorporated that alternative into #48.
- Craig and Michael supported selecting single-row mode immediately before retrieving each result in #52 and #53, while Daniel requested clearer documentation and consistent large-object errors in #56 and #57; these changes and expanded tests appeared in #58.
- Repeated COPY-test failures exposed disagreement over intentionally provoking protocol errors in #59, #60, and #64; Vaishnavi temporarily withdrew the tests in #66, hoping to replace them with psql-based coverage.
- Andres Freund rejected landing this much code without tests in #69 and rejected expanding the scope to psql as a prerequisite in #71; Vaishnavi restored revised tests and adjusted API names and error handling in #72, but Andres deferred the feature beyond PostgreSQL 10 in #73.
Jun 2017 · 2 entries
- Daniel revived the benchmarking phase by rebasing libpq and updating pgbench together in #78; Andres reported doubled local throughput in #79, then traced unnecessary send and receive calls in #81.
- Andres supplied a syscall-reduction delta in #85, but warned that unrestricted buffering could consume excessive memory in #88; Vaishnavi converted the idea into threshold-based flushing in #90.
Aug - Oct 2017 · 1 entry
- Further review in #91 led to removing the linear-time queue-count API and tightening state handling in #92; Craig objected that the proposed CommitQueue name confused synchronization with database commits in #93, leading to SendQueue in #97.
Nov 2017 - Jan 2018 · 1 entry
- After small compatibility updates in #99 and #100, Kyotaro Horiguchi questioned the separate queue-processing API, COPY recovery, and queue structure in #101; Daniel reiterated that psql integration should remain separate in #102.
Aug - Sep 2019 · 1 entry
- Following a long gap, Nikhil Sontakke rebased the series and restored portable timing tests in #105; Amit Kapila noted outstanding Horiguchi feedback in #106, and Álvaro Herrera supplied another rebase in #107, followed by further API, error-handling, and transactional-test review from Andres in #108.
Mar 2021 · 9 entries
- Development continued in the split thread thread #42876, and the archive/commitfest mismatch was explained in #112; Álvaro brought the substantially revised pipeline API, tests, and pgbench integration back here in #114.
- Álvaro resolved successive Windows build failures across #115, #117, #118, and #119 by fixing dependency discovery, renaming the test module, and adding PostgreSQL support libraries and Winsock linkage.
- After Zhihong Yu's documentation and test-header review in #120, Álvaro added pipeline-mode PQsendQuery support using the extended protocol in #121, while noting that Windows CI had built but not run the test program.
- Álvaro substantially simplified query tracking in #122 by routing ordinary and pipelined commands through one command queue and removing connection-level last_query state.
- Tom Lane identified an input-parser path that could repeatedly process the same asynchronous message in #123; Álvaro revised parsing, renamed queue helpers, incorporated David Johnston's documentation suggestions, and expanded describe-command tests in #124.
- The final submission #125 split libpq from pgbench, settled on PQpipelineSync, and added pgbench tests and benchmark results; Justin Pryzby then identified a disabled assertion in #126, prompting Álvaro's correction in #127.
- Álvaro committed the libpq implementation from #125 as acb7e4eb6b, preserving its code exactly while transferring the pgbench documentation to the companion commit.
- Daniel's pgbench integration from #125 landed as 9aa491abbf with the transferred documentation, corrected assertion, and small test refinements, including prepared-query coverage.
- Both feature components were merged by the thread's end: Andres welcomed the landing in #128, and Matthieu Garrigues reported post-merge tests running up to 15 percent faster than his previously tested version in #129; psql integration remained outside this work.