Tom Lane wrote:
> "Andrew Dunstan" <andrew@dunslane.net> writes:
> >> I object to creating gratuitous incompatibilities with the SQL standard,
> >> which will obstruct legitimate features down the road. The SQL standard
> >> says it is <schema>.<constraint>.
>
> > Is there a case for enforcing uniqueness on constraint names, then?
>
> Other than "SQL92 says so"? Very little. This seems to me to be a
> design error in the spec. Per-table constraint names are easier to
> work with --- if they're global across a schema then you have a serious
> problem avoiding collisions.
I assume that SQL99 and later don't specify anything different than what
SQL92 calls for in this regard?
Without any meaningful guidance from the spec, the best we can do is
support per-table constraint names and provide optional (via a GUC
variable) support for SQL92-compliant constraint names. Let the DBA
decide which (if not both) is best for his situation. Inasmuch as
one of our "selling points" is our compliance with the SQL spec, I see
little reason to entirely avoid compliance with the spec on this issue --
just make it possible to do something else when/if necessary.
The two approaches aren't necessarily mutually exclusive (though SQL99
compliance on constraint names would obviously make it unnecessary to
specify a tablename along with a constraint name), so I see little
problem here. But the current arrangement is obviously untenable,
because it allows you to create a situation (multiple constraints by
the same name) that you can't reasonably extricate yourself from.
--
Kevin Brown kevin@sysexperts.com
