model-writtenrun 8270 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jun - Jul 2019 · 6 entries
- Dave Cramer proposed binary pgoutput transmission to reduce conversion costs and support change-data-capture clients using a core plugin available on managed services #1 #3.
- Andres Freund supported binary output provided built-in logical replication also consumed it, while warning that embedded type OIDs make user-defined arrays and composites difficult to transfer between databases #4 #6.
- The discussion distinguished portable binary network representations from internal storage formats, and Petr Jelinek pointed Dave toward pglogical's implementation #10 #12.
- A proposed expansion into nullability and other schema metadata met objections about complexity and pgoutput's purpose, so Dave limited the immediate work to binary transmission #16 #19 #27.
- Petr recommended subscriber-selected binary output through START_REPLICATION options, and Dave's first patch implemented sending and receiving binary values while retaining tuple conversion in worker.c #30 #31.
- Following Tomas Vondra's recommendation to use a dedicated boolean catalog column, Dave added CREATE/ALTER SUBSCRIPTION support and then supplied a documentation delta #33 #34 #35.
Oct - Dec 2019 · 3 entries
- After several months, Dmitry Dolgov questioned unnecessary relation-ID API changes and found that GetSubscription failed to load the binary flag; Dave restored the insert API and supplied the missing assignment, alongside a declaration-order fix prompted by Thomas Munro #37 #38 #39 #40 #41 #45.
- Dave also introduced publisher/subscriber architecture checks, with Alvaro Herrera recommending failure rather than silent fallback, and subsequently resent the complete accumulated series for testing #42 #43 #45 #48.
- December and February refreshes kept the feature alive across upstream changes and removed a catalog-version bump that repeatedly prevented application #50 #52 #53.
Jan - Apr 2020 · 1 entry
- Petr's review of cursor initialization and oversized fixed arrays led Dave to allocate tuple storage dynamically and combine binary and changed flags into a single format array #54 #56 #58.
Jul 2020 · 9 entries
- Following another long gap, Dave expressed frustration with repeated rebases, then refreshed the series; Daniel Gustafsson's compiler review and advice from Alvaro Herrera led to consolidating eight incremental patches into implementation and documentation patches #60 #61 #62 #63 #68.
- Daniel supplied psql display support and SQL and TAP tests, reviewed format initialization and diagnostics, and tested Dave's revisions, while noticing that the TAP file had been omitted during resubmission #69 #70 #71.
- Tom Lane and Petr explained that architecture compatibility checks belonged to internal-format transfer rather than send/receive functions; Dave removed much of that logic, and Daniel removed the remnants, restored the missing TAP test, and produced a cleaned single patch #72 #74 #76 #77.
- Tom questioned the misleading base-types terminology, prompting Dave's final naming revision, while Andres and Tom debated whether array and record receive functions should retain OID checks for useful error reporting #79 #80 #82 #84 #85 #86.
- Tom committed Dave's final submission as 9de77b5453, adding pg_dump support, worker restart handling, input validation, and other local corrections while documenting cross-version limitations and initially retaining restrictions on user-defined arrays and composites #80 #89.
- In the same announcement, Tom posted a separate follow-up that permitted binary transfer of user-defined arrays and composites by rejecting embedded type-OID mismatches only when both OIDs were in the manually assigned range #89.
- Peter Geoghegan reported a buildfarm invalid read, and Tom traced it to an older updated-columns bug exposed by the newly variable-sized arrays; d5daae47db corrected column mapping and dropped-column handling, added assertions and regression coverage, and was backpatched #90 #91.
- At the thread's last exchange, Petr explained that unchanged markers accommodate unavailable TOAST values rather than reliably identify every unchanged column, leaving Tom's proposed improvement to column-trigger semantics and transmission efficiency unresolved #92 #93.
- After the last recorded message, Tom committed the array/record follow-up from #89 unchanged as 670c0a1d47, completing all attached work represented by the three candidate commits.