conclusion that it's really a bad idea for plpgsql to not associate
variables with output columns --- that is, I think we should make
RETURNS TABLE columns semantically just the same as OUT parameters.
Here are some reasons:
1. It's ludicrous to argue that "standards compliance" requires the
behavior-as-submitted. plpgsql is not specified by the SQL standard.
2. Not having the parameter names available means that you don't have
access to their types either, which is a big problem for polymorphic
functions. Read the last couple paragraphs of section 38.3.1:
http://developer.postgresql.org/pgdocs/postgres/plpgsql-declarations.html#PLPGSQL-DECLARATION-ALIASES
as well as the following 38.3.2. How would you do those things with
a polymorphic TABLE column?
3. Not treating the parameters as assignable variables makes RETURN NEXT
nearly worthless in a TABLE function. Since they're not assignable,
you can't use the parameterless form of RETURN NEXT (which'd return
the current values of the variables). The only alternative available
is to return a record or row variable; but there's no convenient way
to declare such a variable, since after all the whole point here is
that the function's output rowtype is anonymous.
4. It's a whole lot easier to explain things if we can just say that
OUT parameters and TABLE parameters work alike. This is especially
true when they actually *are* alike for all the other available PLs.
If we insist on the current definition then we are eventually going to
need to kluge up some solutions to #2 and #3, which seems like make-work
to me when we already have smooth solutions to these problems for
OUT parameters.
Comments?
For the archives, here is the patch as I currently have it (with the
no-plpgsql-variables behavior). But unless I hear a good argument
to the contrary, I'm going to change that part before committing.
regards, tom lane
No comments:
Post a Comment