model-writtenrun 7376 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Zhijie Hou reported that postgres_fdw batch inserts could exceed the protocol's 65535-parameter limit and proposed automatically reducing batch size according to the number of columns, with implementation and documentation patches in #1.
- Bharath Rupireddy requested a shared macro in error messages, clearer documentation, and a combined patch in #2, then requested a runtime guard against zero parameters and markup corrections in #4; Hou incorporated these changes in #3 and #6.
- Bharath found that the patch could dereference a NULL modification state during regression tests and supplied the guard in #8; discussion with Tomas Vondra established that EXPLAIN without ANALYZE leaves that state unset, and #10 refined the comments.
- Tomas simplified batch-size limiting to a division and Min expression and separated the documentation into its own paragraph in #12, which Bharath reviewed in #13.
- Tomas committed cb92703384 from #12, adding a 70000-row regression test and making the documented limit precise; in #14 he explained that rejecting oversized options alone would not handle varying column counts.
- Bharath measured a substantial regression-test slowdown in #15, and Tom Lane confirmed increased buildfarm runtime in #17; Tomas identified tuple-descriptor bookkeeping as the cause and pointed to a separate performance fix in #18 and #19.
- After applying that performance fix, Tomas reported that valgrind still took over an hour instead of about five minutes in #22, prompting Tom to propose keeping the test only long enough for slow buildfarm machines to exercise it before removal in #23.
- Alvaro Herrera suggested testing many columns with fewer rows in #24; Tomas supplied a 1500-column, 50-row replacement in #25, but reported that its roughly thirty-minute valgrind runtime remained excessive.
- Bharath independently supplied a 1600-column, 41-row alternative in #26 and compared both replacements under CLOBBER_CACHE_ALWAYS in #28 and #29; neither proposed replacement landed.
- Two years later, Andres Freund reported that the retained test exhausted memory on a 32-bit address-sanitizer build and still consumed nearly an hour on a valgrind machine in #30; Tom supported removal from all affected branches in #31, and Tomas agreed to perform it in #32.
- Tomas removed the expensive test and its expected output in 087a933b21, confirming removal from every affected branch in #33; the batch-size fix remained in place and the replacement-test proposals were left unused.