>>>>> "PG" == PG Bug reporting form <noreply@postgresql.org> writes:
PG> fetch backward 9 from "mycursor";
PG> commit;
PG> The result set is wrong (9, 8, 7, 6, 5, 4, 3, 2, 1).
PG> It seems that selection predicate is ignored.
PG> This bug has been reproduced with PostgreSQL 9.6 and 10.0
I wonder if we have a contender here for the oldest reported bug in PG
history; while I haven't tested anything older than 9.5, the incorrect
logic seems to date back to the introduction of subqueries in
6.something.
Here is a simpler test case:
begin;
declare foo cursor for select * from generate_series(1,3) i where i <> all (values (2));
fetch all from foo; -- returns the expected 2 rows
fetch backward all from foo; -- assertion failure, or incorrect result
The problem is that the scan direction is being set to "backward" in the
EState, and as a result the subquery evaluation is run in the backward
direction too, which obviously doesn't do anything sensible. The
assertion failure is from the tuplestore code complaining about doing a
backward fetch on a tuplestore not initialized for backward access.
I'm really not sure yet what the correct fix is, though.
--
Andrew (irc:RhodiumToad)