"David G. Johnston" <david.g.johnston@gmail.com> writes:
> On Monday, August 20, 2018, Wu Ivy <ivywuyzl@gmail.com> wrote:
>> Why are SELECT query never marked nullable?
> Basically the nullability property is used by the planner for optimization
> during the joining of physical tables. As soon as you try outputting
> columns the ability to enforce not null goes away because of, in
> particular, outer joins. While some changes could maybe be made the
> cost-benefit to do so doesn't seem favorable.
A further thought on this is that really it's a historical accident that
the elements of tuple descriptors are exactly pg_attribute rows. There
are a *whole lot* of fields in pg_attribute that aren't especially
relevant to tuple sets generated on-the-fly within a query, and typically
won't get filled with anything except default values. The only fields
that really mean a lot for a dynamic tuple set are the data type and
values derived from that, and in some usages the column name.
[ wanders away wondering if it'd be worth our time to design a new,
more compact TupleDesc struct without the meaningless fields ... ]
regards, tom lane