Re: Make PG's "NOT NULL"s and attnotnull ("is_nullable") conform to SQL-2011

Vitaly Burovoy <vitaly.burovoy@gmail.com> writes:
> On 4/27/16, Alvaro Herrera <alvherre@2ndquadrant.com> wrote:
>> Point 2 is where things differ from what I remember; my (possibly
>> flawed) understanding was that there's no difference between those
>> things.  Many (maybe all) of the things from this point on are probably
>> fallout from that one change.

> It is just mentioning that CHECK constraints have influence on
> nullability characteristic, but it differs from NNC.
> NNC creates CHECK constraint, but not vice versa. You can create
> several CHECK "col IS NOT NULL" constraints, but only one NNC (several
> ones by inheritance only?). And DROP NOT NULL should drop only those
> CHECK that is linked with NNC (and inherited), but no more (full
> explanation is in my initial letter).

This seems to me to be a most curious reading of the standard.
SQL:2011 11.4 <column definition> syntax rule 17a says

         If a <column constraint definition> is specified that contains
         the <column constraint> NOT NULL, then it is equivalent to the
         following <table constraint definition>:

                CND CHECK ( C IS NOT NULL ) CA

As a rule, when the SQL spec says "equivalent", they do not mean "it's
sort of like this", they mean the effects are indistinguishable.  In
particular, I see nothing whatsoever saying that you're not allowed to
write more than one per column.

So I don't like the proposal to add an attnotnullid column to
pg_attribute.  What we'd talked about earlier was converting attnotnull
into, effectively, a hint flag saying that there's at least one NOT NULL
constraint attached to the column.