model-writtenrun 9358 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Oct 2018 - Mar 2019 · 15 entries
- Surafel Temesgen proposed multi-row INSERT dumps for faster, smaller transfers to other database systems, initially using a separate function and fixed batches of 100 with options for both pg_dump and pg_dumpall #1.
- Tom Lane questioned the error-isolation and memory tradeoffs, and Michael Paquier questioned cross-database compatibility, while Stephen Frost defended improving the existing INSERT export facility and suggested --inserts=100 syntax #2 #5 #6.
- Surafel adopted that syntax, then generalized the count after Álvaro Herrera objected to hardcoding 100; the generalized patch also dropped the earlier pg_dumpall changes without explaining that omission #7 #8 #10.
- After a rebase request and several weeks' delay, Surafel refreshed the patch for current master #11 #12.
- Fabien Coelho tested the refreshed patch and requested a separate option to avoid optional-argument parsing problems, stricter integer validation, tests and removal of duplicated INSERT code; Surafel introduced --insert-multi and validation but initially retained the separate implementation #13 #14.
- David Rowley demonstrated broken output for zero-column tables, prompting a fallback to one DEFAULT VALUES statement per row and a clearer explanation of that exception #17 #18 #19 #20.
- David supplied a replacement prototype using the existing INSERT function and independent statement counters, arguing that FETCH must remain fixed at 100 regardless of output batch size to avoid excessive round trips or memory use; he also proposed the name --rows-per-insert #21.
- Surafel integrated that approach, but combining --inserts with --rows-per-insert produced malformed SQL, and his first correction prohibited the combination; David argued that the options should cooperate, leading Surafel to remove the conflict and use a default count of one #23 #24 #26 #31 #32.
- Álvaro suggested batching zero-column rows with SELECT UNION ALL, but David argued for retaining standard DEFAULT VALUES syntax for portability, and Álvaro accepted that reasoning #33 #35 #41.
- David took over substantial implementation cleanup, restored his simpler batching state, repaired documentation and option ordering, and made --rows-per-insert imply --inserts #36.
- Fabien showed how one dump_inserts integer could represent both COPY selection and the INSERT row count without making command-line ordering significant; David implemented that simplification and then added TAP coverage #37 #39 #40 #42.
- Fabien pressed for strict numeric parsing and boundary tests; Surafel restored strtol validation and then fixed narrowing from long to int after David's review #44 #46 #50 #51 #52.
- David resolved a collision with the new extra-float-digits option, added zero-column test coverage, corrected regex escaping and clarified the documentation's failure semantics #56.
- Fabien reported successful builds and manual boundary-case tests, then requested a single validation error and one output line per row; David implemented those changes and restored ERANGE checking, while leaving further requested test expansion for others #57 #60.
- Álvaro committed #60 as 7e413a0f82, with light code cleanup and stronger tests covering incomplete final batches and multiple zero-column rows; he identified combined --rows-per-insert and --on-conflict-do-nothing testing as a remaining opportunity #61.
Jun 2019 · 2 entries
- Three months later, Peter Eisentraut noticed that pg_dumpall could not forward the new option, and Fabien supplied the missing forwarding and documentation patch #63 #65.
- Álvaro applied David's off-list ordering corrections and committed #65 as a193cbec11, leaving both the main feature and the pg_dumpall follow-up landed at the thread's close #66.