model-writtenrun 9047 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Andres Freund opened the thread with a PostgreSQL 12 JIT tuple-deforming regression and a twelve-patch series that also repaired hash-comparison JIT compilation, exposed per-expression JIT information through EXPLAIN, and added regression tests #1.
- The proposed EXPLAIN design passed ExprState into expression-display functions, distinguished projection from ordinary output, exposed aggregate phases and hash keys, and made generated function names reproducible for testing; aggregate memory reporting remained future work #1.
- Andres committed the tuple-descriptor fix from the initial series as 97e971ee05 and the hash-comparator fix as ac88807f9b, preserving their code while leaving out tests dependent on the unfinished EXPLAIN infrastructure, then announced both landings #2.
- Maciek Sakrejda proposed sibling JIT groups and a Projects boolean to avoid changing existing field shapes, while Robert Haas questioned omitting unnecessary deforming fields only in text output; Andres defended explicit nulls as useful evidence that deforming was unnecessary #3 #4 #5.
- Andres posted the eight-patch remainder with a rebase and minor conflict resolution, explaining that already-applied fixes had been removed; he also proposed suppressing Output for nodes that do not project #6.
- Maciek and Andres explored making expression fields consistently structured groups, accepting a compatibility break rather than conditionally changing their types or requiring consumers to associate separately named JIT fields #7 #8.
- Robert opposed conditional Output-to-Project renaming and conditional grouping, but accepted consistently omitting Output for non-projecting nodes; Andres favored extensible expression groups, while Maciek advocated a more consistent structured format and suggested version metadata #9 #10 #11 #12.
- Tom Lane later supported expression subnodes but rejected conditional Project naming and format-version metadata, suggested text-only suppression or an all-columns indication for non-projecting output, and questioned whether detailed JIT information belonged in text output #13.
- Andres accepted the objections to Project renaming, welcomed extensible expression subnodes, and said he could live without per-expression JIT details in text output, although the representation of projection status remained unresolved #14.
- Robert emphasized that mechanically interpreting expression strings was a larger tooling limitation and insisted that every output format should expose the same information; Tom allowed that an explicit detail option could address readability but still questioned the layout #15 #16 #17.
- Maciek withdrew the version-number proposal after Tom's explanation and highlighted the need for consistent structured output and better EXPLAIN schema documentation #18.
- The thread ended with Robert proposing indented JIT fields beneath each text expression and consistent inclusion or omission of null fields across formats; the two performance fixes had landed, while the remaining EXPLAIN and testing series still awaited a revised implementation resolving these design questions #19.