User-Defined Functions

• Functions return a value, and procedures do not return a value.
• Procedures must be invoked using a statement. Functions can be invoked in nearly any context, such as SELECT, FROM, and WHERE clauses, expressions, and expressions.
• Functions have settings, and procedures do not.

​

Overview

• An argument has a mode and a type.
  • CockroachDB supports the IN (default), OUT, and INOUT argument modes. For an example, see .
  • The type can be a built-in type, or type, or implicit record type. A type can have a DEFAULT value.
• The return type can be a built-in , user-defined or type, , , PL/pgSQL type, implicit record type, , or VOID.
  • Preceding a type with SETOF indicates that a set, or multiple rows, may be returned. For an example, see .
  • VOID indicates that there is no return type and NULL will always be returned.
• The indicates whether the function has side effects. VOLATILE and NOT LEAKPROOF are the default.
  • Annotate a function with side effects with VOLATILE. This also prevents the from pre-evaluating the function.
  • A STABLE or IMMUTABLE function does not mutate data. You cannot create a STABLE or IMMUTABLE function that executes a mutation ( INSERT, UPSERT, UPDATE, DELETE ) statement.
  • LEAKPROOF indicates that a function has no side effects and that it communicates nothing that depends on its arguments besides the return value (i.e., it cannot throw an error that depends on the value of its arguments). You must precede LEAKPROOF with IMMUTABLE, and only IMMUTABLE can be set to LEAKPROOF. NOT LEAKPROOF is allowed with any other volatility.
  • Non- VOLATILE functions can be optimized through inlining. For more information, see Create an inlined UDF.
• LANGUAGE specifies the language of the function body. CockroachDB supports the languages SQL and .
• The function body:
  • Can reference arguments by name or by their ordinal in the function definition with the syntax $1.
  • Can be enclosed in a single line with single quotes '' or multiple lines with $$.
  • Can reference tables.
  • Can reference only the SELECT statement.

​

Statement statistics

​

Examples

​

Create a UDF

CREATE FUNCTION add(a INT, b INT) RETURNS INT IMMUTABLE LEAKPROOF LANGUAGE SQL AS 'SELECT a + b';

• name: add
• arguments: a of type INT, b of type INT
• return type: INT
• volatility: IMMUTABLE LEAKPROOF
• language: SQL
• function body: 'SELECT a + b'

CREATE FUNCTION add(a INT, b INT) RETURNS INT IMMUTABLE LEAKPROOF LANGUAGE SQL AS 'SELECT $1 + $2';

CREATE FUNCTION add(a INT, b INT) RETURNS INT LANGUAGE SQL AS $$
  SELECT a + b;
$$;

​

View a UDF definition

SHOW CREATE FUNCTION add;

  function_name |                 create_statement
----------------+---------------------------------------------------
 add            | CREATE FUNCTION public.add(IN a INT8, IN b INT8)
                |     RETURNS INT8
                |     IMMUTABLE
                |     LEAKPROOF
                |     CALLED ON NULL INPUT
                |     LANGUAGE SQL
                |     AS $$
                |     SELECT a + b;
                | $$
(1 row)

​

Invoke a UDF

SELECT add(3,5) as sum;

  sum
-------
    8
(1 row)

​

Call a UDF from a view

​

Create a UDF using PL/pgSQL

CREATE FUNCTION fib(n int) RETURNS INT AS $$
    DECLARE
        tmp INT;
        a INT := 0;
        b INT := 1;
        i INT := 2;
    BEGIN
        IF n < 0 THEN
            RAISE EXCEPTION 'n must be non-negative';
        END IF;
        IF n = 0 THEN RETURN 0; END IF;
        IF n = 1 THEN RETURN 1; END IF;
        LOOP
            IF i > n THEN EXIT; END IF;
            tmp := a + b;
            a := b;
            b := tmp;
            i := i + 1;
        END LOOP;
        RETURN b;
    END
  $$ LANGUAGE PLpgSQL;

SELECT fib(8);

  fib
-------
   21

​

Create an inlined UDF

• It is labeled as IMMUTABLE, STABLE, or LEAKPROOF (i.e., non- VOLATILE ).
• It has a single statement.
• It is not a .
• Its arguments are only variable or constant expressions.
• It is not a .

1. Create tables a and b:
   CREATE TABLE a (
     a INT
   );
   
   CREATE TABLE b (
     b INT PRIMARY KEY
   );
   
2. Insert a value (10) into 1000 rows in a and 1 row in b:
   INSERT INTO a SELECT 10 FROM generate_series(1, 1000);
   INSERT INTO b VALUES (10);
   
3. Create a VOLATILE function foo_v() and a STABLE function foo_s():
   CREATE FUNCTION foo_v(x INT) RETURNS INT VOLATILE LANGUAGE SQL AS $$
     SELECT b FROM b WHERE b = x
   $$;
   
   CREATE FUNCTION foo_s(x INT) RETURNS INT STABLE LANGUAGE SQL AS $$
     SELECT b FROM b WHERE b = x
   $$;
   
   Each function returns a specified value from table b.
4. View the query plan when foo_v() (the VOLATILE function) is used in a selection query to retrieve equal values from table a:
   EXPLAIN ANALYZE SELECT foo_v(a) FROM a WHERE a = 10;
   
                                               info
   --------------------------------------------------------------------------------------------
     planning time: 2ms
     execution time: 77ms
     distribution: local
     vectorized: true
     rows read from KV: 1,000 (39 KiB, 1 gRPC calls)
     cumulative time spent in KV: 330µs
     maximum memory usage: 80 KiB
     network usage: 0 B (0 messages)
     sql cpu time: 75ms
     estimated RUs consumed: 0
   
     • render
     │
     └── • filter
         │ nodes: n1
         │ actual row count: 1,000
         │ sql cpu time: 75ms
         │ estimated row count: 1,000
         │ filter: a = 10
         │
         └── • scan
               nodes: n1
               actual row count: 1,000
               KV time: 330µs
               KV contention time: 0µs
               KV rows read: 1,000
               KV bytes read: 39 KiB
               KV gRPC calls: 1
               estimated max memory allocated: 60 KiB
               sql cpu time: 87µs
               estimated row count: 1,000 (100% of the table; stats collected 19 seconds ago)
               table: a@a_pkey
               spans: FULL SCAN
   (33 rows)
   
   The query takes 77ms to execute because the function is invoked for each row scanned in table a.
5. View the query plan when using foo_s() (the STABLE function) instead:
   EXPLAIN ANALYZE SELECT foo_s(a) FROM a WHERE a = 10;
   
                                                 info
   ------------------------------------------------------------------------------------------------
     planning time: 5ms
     execution time: 4ms
     distribution: local
     vectorized: true
     rows read from KV: 1,001 (39 KiB, 2 gRPC calls)
     cumulative time spent in KV: 832µs
     maximum memory usage: 420 KiB
     network usage: 0 B (0 messages)
     sql cpu time: 3ms
     estimated RUs consumed: 0
   
     • render
     │
     └── • merge join (left outer)
         │ nodes: n1
         │ actual row count: 1,000
         │ estimated max memory allocated: 340 KiB
         │ estimated max sql temp disk usage: 0 B
         │ sql cpu time: 3ms
         │ estimated row count: 1,000
         │ equality: (a) = (b)
         │ right cols are key
         │
         ├── • filter
         │   │ nodes: n1
         │   │ actual row count: 1,000
         │   │ sql cpu time: 5µs
         │   │ estimated row count: 1,000
         │   │ filter: a = 10
         │   │
         │   └── • scan
         │         nodes: n1
         │         actual row count: 1,000
         │         KV time: 722µs
         │         KV contention time: 0µs
         │         KV rows read: 1,000
         │         KV bytes read: 39 KiB
         │         KV gRPC calls: 1
         │         estimated max memory allocated: 60 KiB
         │         sql cpu time: 202µs
         │         estimated row count: 1,000 (100% of the table; stats collected 42 seconds ago)
         │         table: a@a_pkey
         │         spans: FULL SCAN
         │
         └── • scan
               nodes: n1
               actual row count: 1
               KV time: 110µs
               KV contention time: 0µs
               KV rows read: 1
               KV bytes read: 30 B
               KV gRPC calls: 1
               estimated max memory allocated: 20 KiB
               sql cpu time: 11µs
               estimated row count: 1 (100% of the table; stats collected 42 seconds ago)
               table: b@b_pkey
               spans: FULL SCAN
   (57 rows)
   
   The query takes only 4ms to execute because the function is inlined and transformed to a with an equality comparison (a) = (b), which has much less overhead than invoking a function for each row scanned in table a.

​

Video Demo

​

Known limitations

• User-defined functions are not supported in partial index predicates.
• Views cannot reference a UDF that contains mutation statements ( INSERT, UPDATE, UPSERT, DELETE ).
• A RECORD -returning UDF cannot be created without a RETURN statement in the root block, which would restrict the wildcard type to a concrete one.
• User-defined functions are not currently supported in:
  • Expressions (column, index, constraint) in tables.
  • Partial index predicates.
• User-defined functions cannot call themselves recursively.
• The setval function cannot be resolved when used inside UDF bodies.
• Casting subqueries to in UDFs is not supported.
• Routines cannot be invoked with named arguments, e.g., SELECT foo(a => 1, b => 2); or SELECT foo(b:= 1, a:= 2);.
• Routines cannot be created if they reference temporary tables.
• Routines cannot be created with unnamed INOUT parameters. For example, CREATE PROCEDURE p(INOUT INT) AS $$ BEGIN NULL; END; $$ LANGUAGE PLpgSQL;.
• Routines cannot be created if they return fewer columns than declared. For example, CREATE FUNCTION f(OUT sum INT, INOUT a INT, INOUT b INT) LANGUAGE SQL AS $$ SELECT (a + b, b); $$;.
• Routines cannot be created with an OUT parameter of type RECORD.
• DDL statements (e.g., CREATE TABLE, CREATE INDEX ) are not allowed within UDFs or stored procedures.
• Polymorphic types cannot be cast to other types (e.g., TEXT ) within routine parameters.
• Routine parameters and return types cannot be declared using the ANYENUM polymorphic type, which is able to match any type.
• cannot be collected for statements executed inside UDFs or stored procedures. You can request statement diagnostics for the top-level invocation of the function or procedure, and the resulting trace includes spans for each statement executed. However, there is no way to target statements executed inside the function or procedure with a statement diagnostics request.
• Statements within routines do not currently respect hint injections. The workaround is to modify the inline hints directly in the body by replacing the routine.

​

See also