model-writtenrun 7161 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Aug - Sep 2022 · 3 entries
- Alvaro Herrera proposed cataloguing NOT NULL as CHECK constraints while retaining attnotnull and avoiding redundant constraints for primary keys in #1; Laurenz Albe and Amit Langote exposed inheritance requirements that shaped the first attachment #6 #7 #9.
- The initial CHECK-based patches required substantial repairs to parsing, inheritance, primary-key propagation, and constraint removal, while binary upgrade remained unresolved; Zhihong Yu and Amit reviewed concrete implementation problems #12 #13 #14.
- Peter Eisentraut argued that explicit NOT NULL constraints must remain distinguishable from arbitrary CHECK expressions, and Tom Lane supported a separate constraint type to avoid expression overhead; Robert Haas also raised display and execution costs #15 #19 #26.
Feb - Apr 2023 · 8 entries
- After several months, Alvaro replaced CHECK emulation with contype='n' and separated diagnostic and psql display patches, then removed those peripheral pieces and matched inherited constraints by column rather than by name #27 #31 #32.
- Peter requested cleaner tests and information-schema integration, supplying separate patches for the latter; Alvaro then added NO INHERIT so parents could remain constrained without forcing the same restriction on children #33 #34 #35.
- Justin Pryzby's review drove API, resource, and test corrections, while Alvaro added pg_dump support and attempted to repair binary upgrade #38 #39 #40.
- The first feature landing, e056c557ae, came from #40 with additional inheritance-loop and validation changes, leaving system-catalog entries, NOT VALID, and DEFERRABLE support outside its scope.
- Andres Freund traced intermittent parallel-test failures to a shared table name and committed the unattached rename as 728015a470 #43 #45.
- Andres also reported sepgsql fallout, and Tom accepted revised catalog-access logging in 76c111a7f1, while possible missing hook calls remained a follow-up concern #47 #48 #49.
- Cross-version upgrade failures and Tom's discovery that omitted NOT NULL decorations forced expensive restore scans led to agreement to revert; 9ce04b50e1 removed the feature and its immediate fallout without an attached revert patch #52 #54 #55.
- Peter's information-schema proposal #34 prompted Alvaro's faster catalog-join implementation #109, while tests and domain limitations received separate discussion #113 #114; Alvaro confirmed moving the revised patch to thread #48404 in #115, and that thread's #541921 supplied the final version for 3af7217942.
Jun - Nov 2023 · 10 entries
- The next attempt introduced temporary NOT NULL constraints that pg_dump would drop after primary-key creation, plus relcache handling for invalid partitioned primary keys, to preserve upgrade performance #57.
- Peter's review supplied documentation fixes and challenged obsolete comments and CREATE TABLE LIKE behavior; Alvaro corrected LIKE propagation but deliberately deferred NOT VALID because it required more dump and upgrade work #60 #61 #62.
- The unchanged parallel_schedule comment was extracted from #63 into c8e43c22be following Peter's request, and the next version rebased and separated the relcache prerequisite #60 #64.
- Dean Rasheed's review exposed duplicate-constraint and naming problems, while discussion with Vik Fearing led to accepting redundant NOT NULL declarations; the revised series also settled on a separate psql footer, which Robert accepted #65 #66 #74 #76 #78 #79.
- Dean's inheritance tests found missing and undroppable constraints derived from primary keys; an unposted dependency-link experiment proved troublesome, so Alvaro returned to coninhcount and conislocal accounting #86 #91 #92 #97.
- Further review distinguished checking existing constraints during ALTER TABLE INHERIT from creating new ones, while Alvaro defended catalog entries as necessary for planner dependency invalidation #98 #99.
- Alvaro rewrote primary-key-derived constraint recursion through ATPrepCmd to handle grandchildren, integrated Peter's documentation and display fixups, and completed extensive cleanup before announcing the second landing #101 #102 #103 #104.
- The second landing immediately exposed a locale-sensitive test ordering issue, and Peter noted that domain constraints were absent; Alvaro explained that domain support had deliberately been separated from table support #105 #106 #107.
- Alexander Lakhin reported a partition-attachment crash and tested the proposed fix; d0ec2ddbe0 came from #110 with a further correction to the tuple-existence condition and added regression cases #108 #111 #112.
- Alexander's deep-inheritance crash reproducers led directly to b0f7dd915b, which added explicit stack-depth checks to four recursive functions without an attached patch #116 #117.
Jan - Feb 2024 · 4 entries
- Andrew Bille reported that parent and child tables both having primary keys could fail pg_upgrade, exposing a collision involving inherited temporary NOT NULL constraints #118.
- Alexander then reported a multiple-inheritance self-update error, Michael Paquier identified a missing command-counter increment, and Alvaro's proposed fix also changed partition-attachment behavior and revealed a separate UNIQUE-versus-primary-key mismatch #119 #120 #122 #124 #125.
- After considering jian he's alternative CCI placement, Alvaro committed the two parts of #125 as c3709100be and cee8db3f68, expanding tests and tightening the constraint-type lookup #126 #127.
- The restore repair d9f686a72e came from #560372 in thread #49465, with further pg_dump logic changes, allowing inherited constraints to take over temporary NO INHERIT constraints and preserving inherited definitions during restore; Alvaro announced it here as the fix for Andrew's report #118 #128.
Apr - May 2024 · 9 entries
- Alexander found a missing-lock assertion in that repair, prompting descendant locking and restrictions against directly changing NO INHERIT status; further testing exposed invalid NO INHERIT behavior on partitioned tables #129 #130 #131 #132.
- The two patches in #134 landed as d45597f72f and 13daa33fa5, with light test and comment edits, restricting direct inheritance-status changes and rejecting partitioned-table NO INHERIT constraints.
- Alexander questioned why CREATE TABLE LIKE still copied NOT NULL with EXCLUDING CONSTRAINTS, and accepted Alvaro's explanation that this preserved existing behavior and the standard's special treatment of nullability #135 #136 #137.
- Kyotaro Horiguchi's wording report #138 led to the first patch of #139, committed unchanged in 21ac38f498 alongside version-comment corrections reported by Justin Pryzby in thread #49465.
- The remaining patches in #139 tried to repair ordinary dump/restore by refactoring pg_dump and limiting temporary constraints and NO INHERIT conversions to binary upgrade, but Robert raised concerns about the behavior's maintainability #140.
- Alvaro found that suppressing temporary constraints also slowed ordinary primary-key restoration, while their inheritance interactions needed more invasive fixes than the release schedule allowed, so he abandoned that attempted repair and chose another revert #141.
- The revert attached in #142 became 6f8bb7c1e9, with light adjustments preserving unrelated code modernization and domain documentation.
- After the revert, Robert argued for separating cataloguing from changes to primary-key-implied nullability, and Alvaro agreed to consider preserving existing persistent NOT NULL behavior in a future attempt #144 #146 #147 #148.
- Bruce removed the combined release-note item in #149, but Alvaro clarified that domain-constraint work remained; the thread ends with Peter questioning that inconsistency and Bruce awaiting a decision about the retained work and its release notes #151 #152 #153.