There are four levels of constraint: column-level, row-level, table-level and schema-level.
A table-level could, for example, involve a target table other than the source table on which it was declared but only gets checked when the source table changes. In theory a schema-level constraint would be checked for every change in every table in the schema but in practise the optimizer would be able to detect changes in a more granular way; consequently, if your DBMS had support for schema-level constraint then you wouldn’t find much use for table-level constraints in practise.
No current SQL product supports schema-level constraints i.e. CREATE ASSERTION. Apparently Rdb did support it when it was looked after by DEC but that is no longer the case. — UPDATE: in a private message I was advised that Sybase’s SQL Anywhere supports CREATE ASSERTION but with serious errors that allow such constraints sometimes to be violated!
The only SQL-like product I’ve used that currently supports subqueries in CHECK constraints, which enables table-level constraints, is the Access database engine (ACE, Jet, whatever). It has issues, though. First, there is not support for SQL-92 functionality (or equivalent) to defer constraint checking. Second, table-level constraints are checked for each row affected, rather than when the statement completes as required by the SQL-92 Standard. Needless to say, the workaround is very clunky e.g. drop the constraint and in doing so lock the table, execute the update, recreate the constraint. Schema-level constraints, arguably achievable by adding the same constraint to all the tables it involves, is virtually unworkable.
Possibly for these reasons, the Access Team have never publicized its CHECK constraint functionality at all beyond the initial announcements for Jet 4.0 (it remains missing from the Access Help, for example). All that said, for intra-table constraints (e.g. a sequenced key in a valid-state ‘history’ temporal table) the functionality works well, especially when you consider that Access only got trigger-like functionality (not SQL based, though) last year.
SQL of course has UNIQUE constraints and referential integrity constraints that are of course table-level but these are special cases. Therefore, all constraints you will encounter ‘in the wild’ will be either colum- or row-level.
Do be aware with MySQL that, although using CHECK() in SQL DDL will parse without error, it will have no effect. How users can tolerate a SQL product with no CHECK constraints at all is beyond me! PostgreSQL has a excellent constraints model, hint hint 🙂
So why are inter-table constraints so rarelt supported? One reason must be due to historical circumstances. As @gbn correctly identifies (under the title Concurrency), the Sybase/SQL Server family of SQL implementations is based on a model that cannot cope with inter-table constraint checking and that’s not something that is likely to ever change.
Consider looking at this the other way around: if you were creating a SQL product today, would you include CREATE ASSERTION? If you did, you would certainly have to also implement DEFERRABLE constraints (even though multiple assignment is arguably the better model). But you would be able to draw on a lot more research and experience if you went down the route of building a ‘traditional’ optimizer. And perhaps you’d find there is no commercial demand for schema-level constraints (if MySQL can get anyway without CHECK constraints…) If PostgreSQL doesn’t do it, I don’t think anyone ever will.
I think the real show stopper is that most industrial-strength products have already developed trigger functionality that allows users to write ‘constraints’ of arbitrary complexity (plus can a lot more e.g. send an email to tell something happened). Sure, they are procedural rather than declarative, the coder has to do a lot of extra work that the system would take care of with true constraints, and the performance tends to be not so great. But the fact is they existing in real products today and do provide a ‘get out of jail free card’ card for vendors. Why should they bother implementing worthy features if customers are not banging the table for them?
As regards academic/teaching langauges, as @Damien_The_Unbeliever correctly identifies, a Tutorial D CONSTRAINT are always ‘schema’-level, hence allow for global constraints of arbitrary conplexity by definition. If you are looking to design your own DBMS(!!) with this kind of functionality, you should consider implementing the D specification while using an existing SQL DBMS for storage, as Dataphor have done.
A question has been bothering me: given that existing ‘industrial strength’ SQL DBMSs support triggers, why don’t they simply map declarative CREATE ASSERTION to a trigger under the covers? I’ve long suspected the answer is because they know that performance would be appalling given their legacy technology.
A more satisfying answer is provided in Applied Mathematics for Database Professionals  By Lex de Haan, Toon Koppelaars, chapter 11. They define various execution models to use when using triggers to enforce multi-tuple constraints. The most sophisticated (though still highly doable) model, which they call EM6, involves the following steps:
- Translate the formal specification into a constraint validation query.
- Develop code to maintain transition effects.
- Devise transition effect (TE) queries that ensure the constraint validation query is only run when necessary [e.g. Can I limit the check to only the updated rows? Can a DELETE ever violate this constraint? Are there only specific columns that an UPDATE must involve to require the constraint to be checked? etc]
- Discover a means to optimize the constraint validation query by having the TE query provide values that can be used in the validation query.
- Devise and add a serialization strategy to the data integrity (DI) code. [i.e. solve the concurrecy problem where a transaction cannot read the ‘bad’ data another transaction is writing].
They then assert (no pun intended!):
Because we believe that it is not possible for a DBMS vendor to
program an algorithm that accepts an arbitrarily complex predicate and
then computes efficient transition effect (TE) queries, a minimal
validation query, and optimal serialization code to implement
execution model EM6, we should not expect full support for multi-tuple
constraints—in a practical, usable and acceptable way—from these
vendors in the future. The best we can hope for is that database
researchers first come up with more common classes of constraints and
develop convenient shorthands for these. The DBMS vendors, in their
turn, should then provide us with new declarative constructs,
consistent with these shorthands, to state these common classes of
constraints easily to the DBMS. Given such a common class declaration,
the DBMS vendor should be able to program an algorithm that provides
us with an EM6-like execution model under the covers to implement the
constraint.
One such common class of database constraint is a foreign key, which is already widely implemented, of course.