On Sun, Apr 3, 2016 at 4:26 PM, Tom Lane <t...@sss.pgh.pa.us> wrote: > Robert Haas <robertmh...@gmail.com> writes: > > Jim, can you re-review this? > > I'm not Jim, but I have a question: what's the motivation for the > Fortran-order traversal of the result (down rows before across columns)? > It seems less than intuitive to do it that way. Perhaps there's a good > reason, but I do not see any defense of this choice in the thread. >
If I am understanding you correctly, it does work the way you find intuitive: all results from the first row are executed before any in the second row, so SELECT a, b UNION ALL SELECT c, d will execute the queries in order: a, b, c, d as is shown in the changes to the sgml and the test cases. Did you get the impression of Fortran-ordering from the phrase "top-to-bottom, left-to-right order" in the sgml patch? If so, would calling it "rows first" or something else be more clear? Or am I misunderstanding you and you find the order a, c, b, d more intuitive? I also note that the patch seems to be missing resetting gexec_flag > in some error exit paths, possibly allowing the \gexec to be applied > to the next query unexpectedly. It should clear that in all the same > places where gfname or gset_prefix get cleared. > Will do! I'm only seeing one place where those two vars are deallocated and nulled, and that's at the tail end of SendQuery. Were you expecting more than just that?
