So it turns out there's another case we have to check for, as reported
on IRC by Yaroslav Schekin (though I've simplified his query a little):
WITH RECURSIVE
x(a) AS ((VALUES ('a'),('b'))
UNION ALL
(WITH
z AS /*NOT*/ MATERIALIZED (SELECT a FROM x)
SELECT z.a || z1.a AS a FROM z CROSS JOIN z AS z1
WHERE length(z.a || z1.a) < 5))
SELECT * FROM x;
Here, uncommenting that NOT actually changes the result, from 22 rows to
4 rows, because we end up generating multiple worktable scans and the
recursion logic is not set up to handle that.
So what I think we need to do here is to forbid inlining if (a) the
refcount is greater than 1 and (b) the CTE in question contains,
recursively anywhere inside its rtable or the rtables of any of its
nested CTEs, a "self_reference" RTE. We only need to make this check if
we're somewhere (possibly nested) inside a recursive query, but I don't
know how practical it will be to test that condition at the point we do
inlining; doing it unconditionally will (I think) be harmless because
self_reference will not be set.
--
Andrew (irc:RhodiumToad)
