Disable all constraints and table checks while restoring a dump

So you look up other tables in a CHECK constraint.

CHECK constraints are supposed to run IMMUTABLE checks. What passes OK for a row at one time should pass OK at any time. That's how CHECK constraints are defined in the SQL standard. That's also the reason for this restriction in the manual:

Currently, CHECK expressions cannot contain subqueries nor refer to variables other than columns of the current row.

Still, expressions in CHECK constraints are allowed to use functions, even user-defined functions. Those should be IMMUTABLE, but Postgres does not currently enforce this. According to this related discussion on pgsql-hackers, one reason is to allow references to the current time, which is not IMMUTABLE by nature.

But you are looking up rows of another table, which is completely in violation of how CHECK constraints are supposed to work. I am not surprised that pg_dump fails to provide for this.

Move your check in another table to a trigger (which is the right tool), and it should work with modern versions of Postgres.

PostgreSQL 9.2 or later

While the above is true for any version of Postgres, several tools have been introduced with Postgres 9.2 to help with your situation:

pg_dump option --exclude-table-data

A simple solution would be to dump the db without data for the violating table with:

--exclude-table-data=my_schema.my_tbl

Then append just the data for this table at the end of the dump with:

--data-only --table=my_schema.my_tbl

But complications with other constraints on the same table might ensue. There is an even better solution:

Solution: NOT VALID

Up to Postgres 9.1, the NOT VALID modifier was only available for FK constraints. This was extended to CHECK constraints in Postgres 9.2. The manual:

If the constraint is marked NOT VALID, the potentially-lengthy initial check to verify that all rows in the table satisfy the constraint is skipped. The constraint will still be enforced against subsequent inserts or updates [...]

A plain Postgres dump file consists of three "sections":

• pre_data
• data
• post-data

^{Postgres 9.2 also introduced an option to dump sections separately with -- section=sectionname, but that's not helping with the problem at hand.}

Here is where it gets interesting. The manual:

Post-data items include definitions of indexes, triggers, rules, and constraints other than validated check constraints. Pre-data items include all other data definition items.

Bold emphasis mine.
You can change the offending CHECK constraint to NOT VALID, which moves the constraint to the post-data section. Drop and recreate:

ALTER TABLE a
  DROP CONSTRAINT a_constr_1
, ADD  CONSTRAINT a_constr_1 CHECK (fail_if_b_empty()) NOT VALID;

A single statement is fastest and rules out race conditions with concurrent transactions. (Two commands in a single transaction would work, too.)

This should solve your problem. You can even leave the constraint in that state, since that better reflects what it actually does: check new rows, but give no guarantees for existing data. There is nothing wrong with a NOT VALID check constraint. If you prefer, you can validate it later:

ALTER TABLE a VALIDATE CONSTRAINT a_constr_1;

But then you are back to the status quo ante.

It seems that this is due to the way pg_dump creates the dump. Looking at the actual dump I saw that the CHECK constraint was present in the dump file using the syntax that's part of the CREATE TABLE command:

CREATE TABLE a (
    i integer NOT NULL,
    CONSTRAINT a_constr_1 CHECK (fail_if_b_empty())
);      

This creates the failure upon restoration of the database as the check is put in place before either table a or table b have any data in them. If however, the dump file is edited and the CHECK is added using the following syntax instead, at the end of the dump file:

ALTER TABLE a ADD CONSTRAINT a_constr_1 CHECK (fail_if_b_empty()); 

... then there is no problem in the restoration.

The exact same logic can be implemented using a TRIGGER as in the following script:

CREATE OR REPLACE FUNCTION fail_if_b_empty (
    ) RETURNS BOOLEAN AS $$
    SELECT EXISTS (SELECT 1 FROM b)
$$ LANGUAGE SQL;

DROP TABLE IF EXISTS a;

CREATE TABLE IF NOT EXISTS a (
    i   INTEGER   NOT NULL
);

INSERT INTO a(i) VALUES (0),(1);

CREATE TABLE IF NOT EXISTS b (
    i  INTEGER NOT NULL
);

INSERT INTO b(i) VALUES (0);

CREATE TRIGGER tr1 AFTER INSERT OR UPDATE ON a
FOR EACH ROW
EXECUTE PROCEDURE fail_if_b_empty();  

In this case however, pg_dump creates (by default) the trigger at the end of the dump file (and not in the CREATE TABLE statement as in the case of a check) and so the restoration succeeds.