model-writtenrun 7460 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Dilip Kumar reported that SP-GiST insertion repeatedly split an oversized tuple until memory exhaustion, providing a reproducer using INCLUDE columns in #1.
- Pavel Borisov identified that included values cannot be shortened by prefix stripping in #2 and proposed rejecting oversized tuples whenever INCLUDE columns exist in #4.
- Pavel corrected an attribute-count error in #5, and Dilip's successful test and comment review in #6 and #8 led to the explanatory wording finalized in #9.
- Tom Lane rejected that restriction in #11 because it disabled valid prefix-shortening cases, proposed checking whether insertion actually makes progress, and discovered that buffer locks also prevented query cancellation.
- Alvaro Herrera supplied an interrupt-macro split and a prospective btree caller in #12, then noted in #13 that the btree change might already be obsolete.
- Tom's cancellation prototype in #14 released buffers before servicing interrupts and used the existing insertion-retry mechanism; his revision in #15 incorporated Alvaro's macro and added an assertion guarding against callers that still block interrupts.
- Tom replaced the original INCLUDE restriction with a progress check in #16, allowing ten iterations without tuple-size reduction to accommodate alignment and operator-class corner cases, and included regression tests alongside the cancellation changes.
- Alvaro confirmed cancellation worked and challenged redundant interrupt processing in #17, then verified the loop fix and regression test in #18 while suggesting a larger grace period for compatibility.
- Tom accepted moving interrupt processing outside the locked loop in #20 and split the combined work into three patches in #21, retaining the ten-cycle limit because extra iterations could bloat the index.
- Tom added documentation of the progress requirement in #22, and Dilip confirmed that the revised fix resolved his original failure in #23.
- Pavel proposed renaming or inverting longValuesOK in #24 and #25, but Tom declined an unnecessary operator-class API change in #27.
- Robert Haas investigated external operator classes while discussing backpatch compatibility in #26; Tom concluded in #28 that ten cycles were generous and the external classes examined did not enable longValuesOK.
- The macro component of #21 landed with comment and whitespace edits as e47f93f981, and its cancellation component landed unchanged as eb7a6b9229, both with backports.
- The loop fix from #21 and documentation from #22 landed with backports as c3c35a733c, after Tom strengthened progress detection to require improvement over the smallest tuple size previously reached.
- By the final message, all three SP-GiST-related changes had landed, but the proposed btree caller remained outside the commits: Tom explained in #29 that it incorrectly assumed ProcessInterrupts() could not return and could consequently reach an assertion failure.