ne 27. 10. 2019 v 19:47 odesílatel Tom Lane <tgl@sss.pgh.pa.us> napsal:
At pgconf.eu, someone whose name I've forgotten pointed out to me that this doesn't work:
regression=# select (row(1, 2.0)).f1; ERROR: could not identify column "f1" in record data type LINE 1: select (row(1, 2.0)).f1; ^
The fields of an anonymous rowtype are certainly named f1, f2, etc, so it seems like this *should* work. A related case is
regression=# select (row(1, 2.0)).*; ERROR: record type has not been registered
Admittedly, these probably aren't terribly useful cases in practice, but it's unfortunate that they don't work as one would expect. So I propose the attached patch to make them work.
The underlying reason for both of these failures is that RowExpr doesn't carry a typmod, so if it's of type RECORD then get_expr_result_type doesn't know how to find a tupdesc for it. The minimum-code solution is to teach get_expr_result_type to build a tupdesc directly from the RowExpr, and that seems to be necessary for complicated cases like
select (r).f1 from (select row(1, 2.0) as r) ss;
In an earlier version of the patch I chose to add in some fast-path logic in ParseComplexProjection and ExpandRowReference, so as to make the really simple cases shown above a bit less inefficient. But on second thought, these are such corner cases that it doesn't seem worth carrying extra code for them. The cases that are more likely to arise in practice are like that last example, and we can't optimize that in the parser. (The planner will optimize FieldSelect-from-RowExpr after flattening subqueries, which is probably as much as we really need to do here.)
I don't feel a need to back-patch this, but I would like to push it into HEAD.
some times I hit this limit, an can be nice more consistent behave of composite types.
It's new feature - and there is not a reason for back-patching
