Re: sqlsmith: ERROR: XX000: bogus varno: 2

On Mon, Dec 20, 2021 at 1:13 PM Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> The reason the regression tests fail if I only patch ruleutils is
> that psql \d on a partitioned table invokes
> ... pg_get_expr(c.relpartbound, c.oid) FROM pg_catalog.pg_class c
> and evidently relpartbound does contain precisely these node types.

Right. I'm not surprised that relpartbound uses those node types. I
*am* surprised that pg_get_expr() is expected to be able to handle
them. IOW, they ARE node trees, consonant with the fact that the
column type is pg_node_tree, but they're NOT expressions.

If we're going to have a policy that all node types stored in the
catalog should be supported by expression_tree_walker even if they're
not actually expressions, we ought to have a rather explicit comment
about that in the comments for expression_tree_walker, because
otherwise somebody might easily make this same mistake again.
Alternatively, maybe pg_get_expr() should just fail and tell you that
this is not an expression, and if you want to see what's in that
column, you should use the SQL-callable functions specifically
provided for that purpose (pg_get_partkeydef, I think). I don't know
why it should be legitimate for pg_get_expr() to just assume that any
random node tree it gets handed must be an expression without doing
any sanity checking.

--
Robert Haas
EDB: http://www.enterprisedb.com