Hi Richard,
Thanks for working on this! The problem analysis is spot-on. In an
outer join scenario, null-extended rows have NULLs not only in the indexed
column but also across other columns of the nullable table. Once those other
columns are removed from GROUP BY, the null-extended rows and the table's
actual NULL row get mistakenly merged into the same group, leading to wrong
query results. This is indeed a subtle bug introduced by commit bd10ec529,
and the fix looks very solid.
Just a very minor thought on the implementation: while nullable_relids is
allocated in the planner context and will be reclaimed anyway, perhaps we
could consider adding a bms_free(nullable_relids) before exiting
remove_useless_groupby_columns(), just to stay consistent with how other
temporary bitmapsets are cleaned up in that function?It is not a big deal at
all, so feel free to ignore it if you prefer keeping it as-is.
Regards
