model-writtenrun 685 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jun 2025 · 2 entries
- Paul A Jungwirth opened the thread in #1 by revisiting trigger behavior for temporal leftovers: conference participants agreed to avoid SPI, while Paul defended nested INSERT row and statement triggers using the SQL standard and comparisons with DB2 and MariaDB.
- The first series in #2 implemented leftovers through direct executor inserts and subtraction SRFs, alongside documentation, temporal foreign-key actions, and trigger metadata; after a summer rebase, #4 restored PERIOD DDL based on generated range columns.
Aug 2025 - Apr 2026 · 22 entries
- Peter Eisentraut's documentation review in #8 asked for sections within the existing DDL and DML chapters, concrete table examples, and removal of descriptions of unimplemented features; Paul applied that restructuring in #9.
- Paul added isolation tests in #10, explaining that READ COMMITTED can miss concurrently inserted leftovers and recommending an explicit preceding SELECT FOR UPDATE instead of adding automatic rescans or locking costs.
- The first two documentation patches from #10 became e4d8a2af07, with Peter adjusting examples, emphasis, and diagram padding as described in #11.
- Peter questioned the subtraction functions' SRF interface in #13, and Paul explained in #18 that it supports polymorphic range results and possible future user-defined types, with multiranges deliberately returning at most one row.
- Chao Li's reviews in #14, #15, and #19 found documentation mistakes and implementation cleanup opportunities; his standalone typo patch #17 was folded into the remaining series, while #21 also removed obsolete foreign-key helper parameters and fixed trigger-data null-check ordering.
- The subtraction-helper component of #21 landed as 5eed8ce50c with a catalog-version bump, and Paul resent the remaining eight patches in #23 after Peter's commitment report in #22.
- Peter's detailed review in #24 exposed incorrect early query augmentation and challenged row counts, privileges, constructor lookup, and missing deparsing; Paul's #25 moved overlap conditions into rewriting, excluded leftovers from command counts, added domain and decoding tests, and introduced a constructor helper.
- Paul removed PERIOD DDL from the active series in #26 because it repeatedly caused rebase conflicts and was lower priority than completing temporal DML.
- Peter proposed storing constructor OIDs directly in pg_range in #27; Paul's testing and naming feedback in #28 led to #29, which Paul and Kirill Reshke reviewed in #31 and #30.
- The revised catalog patch #29 landed as c257ba8397 with only a catalog-version addition, replacing fragile constructor discovery with explicit catalog links.
- Paul resumed the main series in #34 using the new catalog fields, added OLD/NEW RETURNING coverage in #36, and received successful partition-related testing from Kirill in #39 after Kirill resolved a stale-build problem.
- Peter's review in #40 prompted parser, error-location, and test improvements in #41; Paul defended returning only directly changed rows and retaining NULL as an unbounded range endpoint, since an UNBOUNDED keyword caused substantial grammar conflicts.
- Peter accepted the RETURNING explanation and supplied a small fixup in #43, while requesting clearer concurrency diagrams, tests of the anomalies themselves, and fewer slow isolation permutations; Paul incorporated those changes in #44.
- The constructor helper and main temporal DML components of #44 landed together as 8e72d914c5, with light code cleanup and edited documentation.
- The isolation-test component of #44 followed as b6ccd30d8f, preserving the submitted tests and diagram with minor prose edits; Peter announced in #46 that this was the release's feature cutoff.
- Post-commit review immediately raised further issues: jian he questioned tracking of the implicitly updated range column and proposed coercion cleanup in #45, while SATYANARAYANA NARLAPURAM demonstrated overlapping leftovers after a BEFORE UPDATE trigger widened the range in #47.
- SATYANARAYANA NARLAPURAM then reported an INSTEAD OF view-trigger crash in #48 and supplied rejection patches in #49; jian he moved that rejection into RewriteQuery in #50.
- Paul changed the view-trigger approach in #54, arguing that INSTEAD OF triggers should remain usable and take responsibility for leftovers themselves, so the executor should skip automatic leftover insertion.
- For the separate BEFORE UPDATE overlap issue, Paul proposed runtime rejection in #55, saving the application-time value before triggers and checking afterward whether they changed it.
- Tom Lane supplied a Coverity-driven SRF protocol fix in #56, Paul added function statistics tracking in #57, and the combined patch landed as 207cb2abcb with only a comment-position adjustment.
- After that SRF landing, Tom and Paul discussed possible infinite loops and interrupt checks in #58, #59, and #60, concluding that current range cardinalities and insertion-path checks did not require another change.
- jian he supplied more view-trigger test feedback in #61 and tested the BEFORE UPDATE restriction with text ranges in #62; Paul moved the view-trigger repair to a dedicated thread in #63, leaving its later outcome outside this thread.
Jun - Jul 2026 · 3 entries
- Paul revisited the BEFORE UPDATE restriction in #64, explaining that the text-range result followed C collation ordering and simplifying the implementation with ExecInitForPortionOf and validation immediately after triggers.
- After comparing MariaDB and DB2, Paul offered #65 as a different resolution: preserve existing trigger behavior and test both overlapping results without a temporal key and rejection by a WITHOUT OVERLAPS primary key.
- The test-only alternative #65 landed unchanged as 5454fe4a6e, and Peter closed the release open item in #68; temporal foreign-key actions, trigger exposure of FOR PORTION OF bounds, and PERIOD DDL remained deferred beyond the completed core feature.