>>>>> "Andrew" == Andrew Gierth <andrew@tao11.riddles.org.uk> writes:
Andrew> So what's happening here is that there's an equivalence class
Andrew> containing members
Andrew> "greatest(pg_catalog.pg_stat_user_tables.idx_scan,
Andrew> public.pg_stat_user_tables.idx_scan)" and "0" (as an integer
Andrew> constant), and somehow a pathkey for this eclass is becoming
Andrew> attached to the query going to the remote for statistics
Andrew> purposes.
Andrew> It seems obviously wrong that a constant pathkey with no actual
Andrew> reference to the foreign table should be being pushed down, so
Andrew> so far I suspect that get_useful_pathkeys_for_relation isn't
Andrew> being selective enough about what is "useful". In this context
Andrew> I find it suspicious that find_em_expr_for_rel will return an
Andrew> expr with no vars as being "for" every rel, since it's just
Andrew> looking for a subset.
So this looks to me like an oversight in aa09cd242 (CCing rhaas and
Ashutosh accordingly), which changed find_em_expr_for_rel from using
bms_equal to bms_is_subset without considering the degenerate case of
members with no relids at all. I propose to simply add a !bms_is_empty
condition there; anyone have any better idea?
--
Andrew (irc:RhodiumToad)
