PostgreSQL’s Table-Valued Functions

Table-valued functions are an awesome thing. Many databases support them in one way or another and so does PostgreSQL. In PostgreSQL, (almost) everything is a table. For instance, we can write:
 
 
 
 
 
 
 
 

CREATE OR REPLACE FUNCTION 
    f_1 (v1 INTEGER, v2 OUT INTEGER)
AS $$
BEGIN
    v2 := v1;
END
$$ LANGUAGE plpgsql;

… and believe it or not, this is a table! We can write:

select * from f_1(1);

And the above will return:

+----+
| v2 |
+----+
|  1 |
+----+

It’s kind of intuitive if you think about it. We’re just pushing out a single record with a single column. If we wanted two columns, we could’ve written:

CREATE OR REPLACE FUNCTION 
    f_2 (v1 INTEGER, v2 OUT INTEGER, v3 OUT INTEGER)
AS $$
BEGIN
    v2 := v1;
    v3 := v1 + 1;
END
$$ LANGUAGE plpgsql;

… and then:

select * from f_2(1);

And the above will return:

+----+----+
| v2 | v3 |
+----+----+
|  1 |  2 |
+----+----+

That’s useful, but those are just single records. What if we wanted to produce a whole table? It’s easy, just change your functions to actually return TABLE types, instead of using OUT parameters:

CREATE OR REPLACE FUNCTION f_3 (v1 INTEGER)
RETURNS TABLE(v2 INTEGER, v3 INTEGER)
AS $$
BEGIN
    RETURN QUERY
    SELECT *
    FROM (
        VALUES(v1, v1 + 1), 
              (v1 * 2, (v1 + 1) * 2)
    ) t(a, b);
END
$$ LANGUAGE plpgsql;

When selecting from the above very useful function, we’ll get a table like so:

select * from f_3(1);

And the above will return:

+----+----+
| v2 | v3 |
+----+----+
|  1 |  2 |
|  2 |  4 |
+----+----+

And we can LATERAL join that function to other tables if we want:

select *
from book, lateral f_3(book.id)

… which might yield, for example:

+----+--------------+----+----+
| id | title        | v2 | v3 |
+----+--------------+----+----+
|  1 | 1984         |  1 |  2 |
|  1 | 1984         |  2 |  4 |
|  2 | Animal Farm  |  2 |  4 |
|  2 | Animal Farm  |  4 |  6 |
+----+--------------+----+----+

In fact, it appears that the keyword LATERAL is optional in this case, at least for PostgreSQL.

Table-valued functions are very powerful!

Discovering table-valued functions

From jOOQ’s schema reverse-engineering perspective, things might get a bit tricky as can be seen in this Stack Overflow question. PostgreSQL deals with OUT parameters in a very similar way as with TABLE return types. This can be seen in the following query against the INFORMATION_SCHEMA:

SELECT r.routine_name, r.data_type, p.parameter_name, p.data_type
FROM   information_schema.routines r
JOIN   information_schema.parameters p
USING (specific_catalog, specific_schema, specific_name);

… and the output:

routine_name | data_type | parameter_name | data_type
-------------+-----------+----------------+----------
f_1          | integer   | v1             | integer
f_1          | integer   | v2             | integer
f_2          | record    | v1             | integer
f_2          | record    | v2             | integer
f_2          | record    | v3             | integer
f_3          | record    | v1             | integer
f_3          | record    | v2             | integer
f_3          | record    | v3             | integer

As you can see, the output is really indistinguishable from that perspective. Luckily, we can also join the pg_catalog.pg_proc table, which contains the relevant flag to indicate if a function returns a set or not:

SELECT   r.routine_name, 
         r.data_type, 
         p.parameter_name, 
         p.data_type, 
         pg_p.proretset
FROM     information_schema.routines r
JOIN     information_schema.parameters p
USING   (specific_catalog, specific_schema, specific_name)
JOIN     pg_namespace pg_n
ON       r.specific_schema = pg_n.nspname
JOIN     pg_proc pg_p
ON       pg_p.pronamespace = pg_n.oid
AND      pg_p.proname = r.routine_name
ORDER BY routine_name, parameter_name;

Now, we’re getting:

routine_name | data_type | parameter_name | data_type | proretset
-------------+-----------+----------------+-----------+----------
f_1          | integer   | v1             | integer   | f
f_1          | integer   | v2             | integer   | f
f_2          | record    | v1             | integer   | f
f_2          | record    | v2             | integer   | f
f_2          | record    | v3             | integer   | f
f_3          | record    | v1             | integer   | t
f_3          | record    | v2             | integer   | t
f_3          | record    | v3             | integer   | t

We can see that f_3 is the only function actually returning a set of record, unlike f_1 and f_2, which only return a single record.

Now, remove all those parameters that are not OUT parameters, and you have your table type:

SELECT   r.routine_name, 
         p.parameter_name,
         p.data_type,
         row_number() OVER (
           PARTITION BY r.specific_name 
           ORDER BY p.ordinal_position
         ) AS ordinal_position
FROM     information_schema.routines r
JOIN     information_schema.parameters p
USING   (specific_catalog, specific_schema, specific_name)
JOIN     pg_namespace pg_n
ON       r.specific_schema = pg_n.nspname
JOIN     pg_proc pg_p
ON       pg_p.pronamespace = pg_n.oid
AND      pg_p.proname = r.routine_name
WHERE    pg_p.proretset
AND      p.parameter_mode = 'OUT'
ORDER BY routine_name, parameter_name;

Which will give us:

routine_name | parameter_name | data_type | position |
-------------+----------------+-----------+----------+
f_3          | v2             | integer   |        1 |
f_3          | v3             | integer   |        2 |

How to run such queries in jOOQ?

Once the above code is generated, you can easily call the table-valued function in any jOOQ query. Consider again the BOOK example (in SQL):

select *
from book, lateral f_3(book.id)

… and with jOOQ:

DSL.using(configuration)
   .select()
   .from(BOOK, lateral(F_3.call(BOOK.ID)))
   .fetch();

The returned records then contain values for:

record.getValue(F_3.V2);
record.getValue(F_3.V3);

All that typesafety is only available in the upcoming jOOQ 3.5, for free! (SQL Server, Oracle, and HSQLDB table-valued functions are already supported!)