model-writtenrun 8487 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Zhenglong Li, posting as Keyerror Smart, reported that JsonExpr's missing strictness handling could produce non-NULL results for unmatched outer-join rows and proposed treating every JsonExpr as non-strict in #1; Greg Burd independently reproduced the issue and endorsed that conservative approach in #2.
- Responding to Ethan Tran's performance review, Zhenglong replaced the blanket rule with selective PASSING and behavior-expression checks in #5, explaining that disabling context-only STRICT wrapper inlining could cause roughly fourfold slowdowns in a backpatched release.
- Ethan demonstrated that RETURNING a NOT NULL domain could turn a NULL input into an empty array in #6; Zhenglong responded in #7 by restricting acceptable behaviors to NULL or ERROR and adding domain tests, while declining plan-dependency machinery as unsuitable for backpatching.
- Ethan confirmed that the revised rule preserved his benchmark's inlining and suggested accepting UNKNOWN ON ERROR in #8; Zhenglong incorporated that case and a regression test in #9, while deferring constant-NULL-default refinements and rebasing around conflicting regression additions.
- Tomas Vondra identified that a domain error also violates a STRICT wrapper's required NULL result in #10, and detailed the optimization losses caused by rejecting constant PASSING values and behaviors in #11.
- Zhenglong adopted Tomas's domain guard in #12, corrected the earlier test that expected unsafe domain inlining, and initially retained the conservative behavior restrictions for backpatch simplicity.
- Amit Langote explained in #13 that use_json_coercion could reflect stale domain constraints and proposed rejecting every domain return type; his replacement implementation inspected PASSING and behavior expressions for Vars, PlaceHolderVars, or Params, restoring safe constant-input inlining.
- Tomas reported in #14 that the separately landed FULL JOIN fix introduced another walker needing equivalent JsonExpr handling, and accepted backpatching the more precise approach subject to its correctness in #15.
- Zhenglong integrated Amit's implementation and credited him as co-author in #16, adding tests for restored constant-input inlining and rejected unconstrained-domain inlining; he left equivalent FULL JOIN walker handling for a subsequent addition.
- Amit supplied a corrected delta against #12 in #17 because his earlier JsonBehavior handling stopped traversal too soon and could miss a non-strict DEFAULT subquery; he also recommended identical checks across branches and clarified that the domain guard could precede the executor coercion changes.
- At the end of #17, no commit was linked to this thread, Amit explicitly had not reviewed #16, and integrating the subquery correction and extending the FULL JOIN walker remained outstanding.