model-writtenrun 881 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- SATYANARAYANA NARLAPURAM reported that FOR PORTION OF left dependent stored generated values stale on the updated row, although leftover inserts recomputed them correctly, and proposed adding the range attribute during ExecInitGenerated in #1.
- jian he expanded the fix in #2 to initialize child executor state earlier, account for column-specific triggers and other dependency consumers, and preserve inherited-table leftovers in their child tables.
- SATYANARAYANA NARLAPURAM confirmed the generated-column and inheritance fixes in #3 but reported disappearing leftovers when INSERT triggers rejected them; jian he and Paul A Jungwirth explained in #4 and #5 that honoring those trigger cancellations was expected behavior.
- Paul A Jungwirth took over revisions in #5, cleaned up the executor branches, and questioned duplicating dependency logic; jian he proposed centralizing it in ExecGetUpdatedCols in #6, which Paul incorporated in #7 with corrected partition attribute mapping and broader tests.
- jian he's suspected crash in #8 disappeared after a clean rebuild in #10, while his test improvements and unused-include correction were incorporated alongside a rebase in #11.
- After Paul's rebase in #12, Peter Eisentraut found an injection-point test crash in #13; Paul fixed the premature NULL perminfo dereference in #14.
- Chao Li demonstrated that multiple inheritance still failed in #16 and supplied an additional attribute-mapping fix with UPDATE and DELETE tests in #17, requesting separation of inheritance work from dependency tracking.
- Chao Li then demonstrated unintended mutation of the permission bitmap in #18; Paul incorporated the correction and reorganized the work into dependency-tracking and partition/inheritance patches in #20.
- Chao Li's allocation review in #21 led Paul to avoid redundant bitmap copies and use the current memory context in #22, which Chao approved in #23; Paul declined jian he's subsequent query-context optimization because its allocation behavior would be inconsistent in #24 and #25.
- Nathan Bossart reported a separate crash when the range column itself was virtual generated in #26; Paul proposed rejecting both stored and virtual generated range columns during parse analysis in #27, while noting that the validation phase would eventually need reconsideration.
- Paul moved the generated-range prohibition to a separate discussion in #28, leaving future PERIOD support and later validation as follow-up concerns.
- After discussing permissions with Peter Eisentraut, Paul replaced the executor approach in #29 with registration of the range column in RTEPermissionInfo->updatedCols, simultaneously enforcing UPDATE permission and fixing generated-column and trigger dependencies; INSERT permission for historical leftovers remained unnecessary, and inheritance fixes were redirected to their own thread.
- jian he approved the replacement and requested terminology corrections in #30; Peter committed the attachment unchanged apart from whitespace and rebasing as 7ac030d5b1, announcing it in #31.
- Peter separately implemented jian he's terminology suggestion as f0aedc7cb0 and explained in the thread's final message #32 that one test covered both generated-column kinds, so its comment should omit 'stored'; the main dependency fix was landed, while the separate prohibition and inheritance work were no longer being pursued here.
- The supplied later commit e994f956e4 landed the prohibition originating in #27 on July 6, moving validation into the planner after view rewriting and adding documentation plus view and SQL-function regression coverage.