model-writtenrun 5678 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Oct 2020 · 2 entries
- Amit Langote opened the thread with a report that moving a referenced row between partitions could incorrectly delete referencing rows through ON DELETE CASCADE, and proposed rejecting that movement in #1.
- David G. Johnston argued that the update should work normally in #2, and Tomas Vondra advocated a complete development-branch fix with potentially simpler backbranch treatment in #4, which Amit accepted in #5.
Nov 2020 - Apr 2021 · 6 entries
- Amit's replacement series in #6 created foreign-key triggers on partitioned parents and linked child triggers to them, then suppressed inherited DELETE actions during row movement in favor of a root-table UPDATE event whose tuples were initially kept in a tuplestore.
- Arne Roland argued for similarly transparent behavior for user-defined triggers in #8, #11, and #14, but Amit kept this fix limited to foreign-key enforcement and deferred broader changes to the documented trigger behavior to a separate discussion in #17.
- Sawada demonstrated that the original guard still allowed unwanted SET NULL and SET DEFAULT actions in #19, prompting the broader restrictive patch in #20; his self-reference counterexample in #23 then helped establish that such a backbranch restriction would reject valid operations.
- Rahila Syed reported successful tests of the functional fix and requested documentation in #21 and #25, leading Amit to document the unsupported non-root ancestor case in #27, while the restrictive backbranch approach was set aside in #24.
- Sawada's infrastructure review in #28 identified missing comments, unnecessary helper functions, a tuple-memory leak, and client compatibility issues, which Amit addressed in #29.
- Amit also discovered that query-lifetime tuple storage could not support deferred foreign keys in #29; the series remained unresolved through the April rebase in #31 and the July postponement in #33.
Sep - Oct 2021 · 2 entries
- Work resumed in September with #36, including a separate crash reproducer and a proposal to commit the infrastructure independently; #38 then introduced transaction-lifetime tuplestores and deferred-constraint tests while incorporating Zhihong Yu's diagnostic suggestions from #37.
- Following an off-list discussion with Robert, Amit replaced tuple storage with CTIDs and source/destination partition OIDs in #39, avoiding excessive memory consumption and fetching the required tuples when triggers execute.
Dec 2021 - Jan 2022 · 7 entries
- After the December rebase in #40, Herrera committed its infrastructure portion as f4566345cf, correcting pg_dump compatibility and a psql version check, and reposted the unchanged enforcement remainder in #41.
- Herrera reopened the possibility of backpatching in #41, and Amit supplied PostgreSQL 13 and 14 infrastructure ports in #44, while noting that existing foreign keys would lack the newly required parent triggers.
- Julien Rouhaud restored CI coverage by resending the pending development patch in #46, and Amit followed with complete branch-specific series plus the HEAD patch in #48.
- ABI analysis in #49 exposed risks from changed executor and trigger interfaces, and Zhihong Yu and Herrera favored leaving older branches alone in #50 and #51.
- Tom Lane requested adjacent trigger-event flag bits in #53, and Herrera incorporated that arrangement with formatting and readability changes in #63, while questioning the new ExecInsert outputs and cross-partition detection API.
- Amit explained why tuple routing required the ExecInsert outputs and replaced inferred cross-partition status with an explicit boolean in #64, folding Herrera's changes into a full revision that was resent unchanged in #65.
- Herrera questioned whether an insertion canceled by a BEFORE trigger could crash the new code in #66, and Amit explained in #67 that the unset destination relation prevents the foreign-key helper from being called.
Mar 2022 · 3 entries
- In March, Herrera adapted the patch to extensive executor changes and moved ancestor-relation management into execMain.c in #68; Zhihong Yu's review in #69 led to comment refinements in #70 and Amit's final delta in #71.
- Herrera merged #70 with #71 and committed ba9a7e3921, with only an additional comment wording change, announcing completion in #72.
- The thread ends with Amit's thanks in #73: both development-branch changes are committed, backports remain unpursued because of compatibility costs, and row movement involving independently referenced non-root ancestors retains the documented restriction.