Обсуждение: Domain checks not always working when used in compound type

Hello All,

Today I was experimenting with domain and compound types.

-- Creating a domain type with NOT NULL constraint and at least 2 
characters.
CREATE DOMAIN domain_a AS VARCHAR(10) NOT NULL
   CHECK (LENGTH(TRIM(value)) > 2);

-- Using the domain within a compound type. The NOT NULL constraint 
should be
-- inherited from the domain, therefore not repeated here for column a.
CREATE TYPE compound_ab AS (
   a    domain_a,
   b    varchar(10)
);

-- Creating a table using the compound type.
CREATE TABLE test1 (
   name  VARCHAR(25) NOT NULL UNIQUE,
   ab    compound_ab,
   id    INTEGER GENERATED ALWAYS AS IDENTITY
);

-- Insert into table without providing a value for ab, thus (ab).a is 
NULL, which shouldn't be allowed.
INSERT INTO test1 VALUES ('person1');
INSERT 0 1

-- Select proves that (ab).a is NULL - violating the NOT NULL constraint.
SELECT *, (ab).a FROM test1;
   name   | a
---------+---
  person1 | ¤

¤ is set as a marker for NULL values in order to make it distinguishable 
from an empty string.

-- Directly inserting a NULL gives an error, so the constraint is in place.
INSERT INTO test1 (name, ab.a, ab.b)
   VALUES ('person2', NULL, 'something');
ERROR 23502: domain domain_a doesn't allow NULL values

INSERT INTO test1
   VALUES ('person3', '(,something)');
ERROR 23502: domain domain_a doesn't allow NULL values

-- But it can be circumvented by only feeding ab.b, but not ab.a.
INSERT INTO test1 (name, ab.b)
   VALUES ('person3', 'something');

SELECT *, (ab).a FROM test1;
   name   |      ab      | id | a
---------+--------------+----+---
  person1 | ¤            |  1 | ¤
  person3 | (,something) |  2 | ¤

The tests have been carried out with PostgreSQL 16.1.

Is the behaviour as expected? I think it isn't.

Regards,

Holger

-- 
Holger Jakobs, Bergisch Gladbach, Germany, Tel. +49-178-9759012
Hint: Only my mail server is entitled to send mails from my mail address. The SPF regulation is strict.
Mailing lists must not use the original from address when forwarding!

Holger Jakobs <holger@jakobs.com> writes:
> -- Using the domain within a compound type. The NOT NULL constraint 
> should be
> -- inherited from the domain, therefore not repeated here for column a.
> CREATE TYPE compound_ab AS (
>    a    domain_a,
>    b    varchar(10)
> );

You are assuming in effect that a simple NULL value for a composite
type is the same thing as ROW(NULL, NULL).  They are not quite the
same, and one way in which they are not is that we don't consider
field-level constraints when deciding if a simple NULL value is legal
for the composite -- it always is.  Thus

regression=# select null::compound_ab;
 compound_ab 
-------------
 
(1 row)

regression=# select row(null, null)::compound_ab;
ERROR:  domain domain_a does not allow null values

The SQL spec itself is pretty schizophrenic about whether ROW(NULL, NULL)
is equivalent to bare NULL.  This is how we've chosen to interpret it.
I'll freely admit that there's some implementation considerations
involved in that choice, but we're unlikely to revisit it.

If you don't want things to work like this, you could attach a NOT
NULL constraint to the test1.ab column (as well as having the domain
constraint).

            regards, tom lane

Am 29.12.23 um 20:49 schrieb Tom Lane:
> Holger Jakobs <holger@jakobs.com> writes:
>> -- Using the domain within a compound type. The NOT NULL constraint
>> should be
>> -- inherited from the domain, therefore not repeated here for column a.
>> CREATE TYPE compound_ab AS (
>>     a    domain_a,
>>     b    varchar(10)
>> );
> You are assuming in effect that a simple NULL value for a composite
> type is the same thing as ROW(NULL, NULL).  They are not quite the
> same, and one way in which they are not is that we don't consider
> field-level constraints when deciding if a simple NULL value is legal
> for the composite -- it always is.  Thus
>
> regression=# select null::compound_ab;
>   compound_ab
> -------------
>   
> (1 row)
>
> regression=# select row(null, null)::compound_ab;
> ERROR:  domain domain_a does not allow null values
>
> The SQL spec itself is pretty schizophrenic about whether ROW(NULL, NULL)
> is equivalent to bare NULL.  This is how we've chosen to interpret it.
> I'll freely admit that there's some implementation considerations
> involved in that choice, but we're unlikely to revisit it.
>
> If you don't want things to work like this, you could attach a NOT
> NULL constraint to the test1.ab column (as well as having the domain
> constraint).
>
>             regards, tom lane

CREATE TABLE test1 (
   name  VARCHAR(25) NOT NULL UNIQUE,
   ab    compound_ab NOT NULL,
   id    INTEGER GENERATED ALWAYS AS IDENTITY
);

INSERT INTO test1 VALUES ('person1');
INSERT 0 1
ERROR 23502: NULL value in column "ab" of relation "test1" violates NOT 
NULL constraint

That one is fine now. Thanks for the explanation.


OTOH, the second problem persists.

INSERT INTO test1 (name, ab.b)
   VALUES ('person3', 'something');
INSERT 0 1

SELECT *, (ab).a FROM test1;
   name   |      ab      | id | a
---------+--------------+----+---
  person3 | (,something) |  2 | ¤

Can something be done about this one?

Regards,

Holger


-- 

Holger Jakobs, Bergisch Gladbach, Tel. +49-178-9759012

On Sat, Dec 30, 2023 at 3:56 AM Holger Jakobs <holger@jakobs.com> wrote:
>
> Am 29.12.23 um 20:49 schrieb Tom Lane:
> > Holger Jakobs <holger@jakobs.com> writes:
> >> -- Using the domain within a compound type. The NOT NULL constraint
> >> should be
> >> -- inherited from the domain, therefore not repeated here for column a.
> >> CREATE TYPE compound_ab AS (
> >>     a    domain_a,
> >>     b    varchar(10)
> >> );
> > You are assuming in effect that a simple NULL value for a composite
> > type is the same thing as ROW(NULL, NULL).  They are not quite the
> > same, and one way in which they are not is that we don't consider
> > field-level constraints when deciding if a simple NULL value is legal
> > for the composite -- it always is.  Thus
> >
> > regression=# select null::compound_ab;
> >   compound_ab
> > -------------
> >
> > (1 row)
> >
> > regression=# select row(null, null)::compound_ab;
> > ERROR:  domain domain_a does not allow null values
> >
> > The SQL spec itself is pretty schizophrenic about whether ROW(NULL, NULL)
> > is equivalent to bare NULL.  This is how we've chosen to interpret it.
> > I'll freely admit that there's some implementation considerations
> > involved in that choice, but we're unlikely to revisit it.
> >
> > If you don't want things to work like this, you could attach a NOT
> > NULL constraint to the test1.ab column (as well as having the domain
> > constraint).
> >
> >                       regards, tom lane
>
> CREATE TABLE test1 (
>    name  VARCHAR(25) NOT NULL UNIQUE,
>    ab    compound_ab NOT NULL,
>    id    INTEGER GENERATED ALWAYS AS IDENTITY
> );
>
> INSERT INTO test1 VALUES ('person1');
> INSERT 0 1
> ERROR 23502: NULL value in column "ab" of relation "test1" violates NOT
> NULL constraint
>
> That one is fine now. Thanks for the explanation.
>
>
> OTOH, the second problem persists.
>
> INSERT INTO test1 (name, ab.b)
>    VALUES ('person3', 'something');
> INSERT 0 1
>
> SELECT *, (ab).a FROM test1;
>    name   |      ab      | id | a
> ---------+--------------+----+---
>   person3 | (,something) |  2 | ¤
>
> Can something be done about this one?
>

CREATE DOMAIN domain_a AS VARCHAR(10) NOT NULL CHECK (LENGTH(TRIM(value)) > 2);

CREATE TYPE compound_ab AS (a domain_a,b varchar(10));

CREATE TABLE test1 (
name  VARCHAR(25) NOT NULL UNIQUE,
ab    compound_ab,
id    INTEGER GENERATED ALWAYS AS IDENTITY,
CONSTRAINT compound_ab_a CHECK ((ab).a IS NOT NULL),
CONSTRAINT compound_ab_b CHECK ((ab).b IS NOT NULL));

This should be fine.