Parallel By ANY

Previous posts in this category introduced a way to perform DML in a SELECT using the following technology:

  • Table Functions
  • Autonomous Transactions

The following concepts were included:

  • A REF CURSOR as a parameter
  • Row by row (aka slow by slow) processing of the REF CURSOR
  • Raw parallel query processing to try to overcome the slow by slow processing

This post makes the parallel processing less raw by specifying that one instance of the function will run on each parallel process.

To keep this demo simple I’ll let Oracle decide how many parallel processes to use. This wont break anything because all I’m after is extra processing power.

SQL> -- Create the table
SQL> CREATE TABLE demo
 2   ( col1 NUMBER,
 3     col2 NUMBER );
Table created.

SQL> -- create 1000 demo records
SQL> BEGIN
 2     FOR counter IN 1..1000 LOOP
 3       INSERT INTO demo
 4       VALUES(counter,
 5              counter);
 6     END LOOP;
 7   END;
 8 /
PL/SQL procedure successfully completed.

SQL> COMMIT;
Commit complete.

SQL> -- create a return type for the function we will write later
SQL> CREATE OR REPLACE TYPE v_n_t AS TABLE OF NUMBER;
 2 /
Type created.

SQL> CREATE OR REPLACE FUNCTION delete_demo ( p_curs SYS_REFCURSOR )
 2                     RETURN v_n_t
 3                     --
 4                     -- Here is the parallel clause. The BY ANY clause tells Oracle to
 5                     -- simply use whatever parallelism it sees fit
 6                     --
 7                    PARALLEL_ENABLE ( PARTITION p_curs BY ANY ) AS
 8                    --
 9                    PRAGMA AUTONOMOUS_TRANSACTION;
 10                   v_ret_val v_n_t := v_n_t();
 11                   v_number NUMBER;
 12 BEGIN
 13   LOOP
 14     FETCH p_curs INTO v_number;
 15     EXIT WHEN p_curs%NOTFOUND;
 16     v_ret_val.EXTEND;
 17     DELETE demo
 18     WHERE col1 = v_number;
 19     v_ret_val(1) := SQL%ROWCOUNT;
 20   END LOOP;
 21   COMMIT;
 22   RETURN(v_ret_val);
 23 END;
 24 /
Function created.

SQL> ALTER SESSION FORCE PARALLEL QUERY;
Session altered.

SQL> SELECT COUNT(*)
 2     FROM TABLE(delete_demo(CURSOR(SELECT col1
 3                                     FROM demo)))
 4 /

COUNT(*)
----------
      1000

SQL> ALTER SESSION DISABLE PARALLEL QUERY;
Session altered.

SQL> SELECT COUNT(*)
 2     FROM demo
 3 /

COUNT(*)
----------
         0

SQL> SELECT statistic,
 2          value
 3     FROM v$pq_sysstat
 4    WHERE value > 0
 5   ORDER BY value DESC;

STATISTIC            VALUE
-------------------- -----
Local Msgs Recv'd       30
Local Msgs Sent         30
Servers Idle            12
Servers Started         12
Servers Highwater        4
Server Sessions          4
Queries Initiated        1
DFO Trees                1

Note the much more efficient use of PQ processing. Only 30 messages!

The next post will how parallelism can be controlled to provide raw power and functionality. It’s Here.

Thanks for reading!

REF Cursors, Row By Row and Parallel Query

In the first post in this category I introduced a way to perform DML in a SELECT using the following technology:

  • Table Functions
  • Autonomous Transactions

In this post I’ll add the following:

  • A REF CURSOR as a parameter
  • Row by row (aka slow by slow) processing of the REF CURSOR
  • Raw parallel query processing to try to overcome the slow by slow processing

Lets get to the code shall we…

SQL> -- Create the table
SQL> CREATE TABLE demo
  2 ( col1 NUMBER,
  3   col2 NUMBER );
Table created.

SQL> -- create 1000 demo records
SQL> BEGIN
  2    FOR counter IN 1..1000 LOOP
  3      INSERT INTO demo
  4      VALUES(counter,
  5             counter);
  6    END LOOP;
  7  END;
  8 /
PL/SQL procedure successfully completed.

SQL> COMMIT;
Commit complete.

SQL> -- create a return type for the function we will write later
SQL> CREATE OR REPLACE TYPE v_n_t AS TABLE OF NUMBER;
  2 /
Type created.

SQL> -- now for the function
SQL> /*
SQL> Key points
SQL>    1) The parameter is a REF cursor - how cool is that?
SQL>    2) We loop through the records one by one aka SLOW BY SLOW...
SQL>    3) The function deletes the rows it queries
SQL>    4) This is madness
SQL>    5) Yes, it is madness but I'm leading up to something
SQL>    6) When will you make your point
SQL>    7) 3 or 4 more posts; I promise
SQL>    8) OK. This better be good...
SQL> */
SQL> CREATE OR REPLACE FUNCTION delete_demo ( p_curs SYS_REFCURSOR )
  2     RETURN v_n_t AS
  3     PRAGMA AUTONOMOUS_TRANSACTION;
  4     v_ret_val v_n_t := v_n_t();
  5     v_number NUMBER;
  6  BEGIN
  7    LOOP
  8      FETCH p_curs INTO v_number;
  9      EXIT WHEN p_curs%NOTFOUND;
 10      v_ret_val.EXTEND;
 11      DELETE demo
 12      WHERE col1 = v_number;
 13      v_ret_val(1) := SQL%ROWCOUNT;
 14    END LOOP;
 15    COMMIT;
 16    RETURN(v_ret_val);
 17 END;
 18 /
Function created.

SQL> -- One way to overcome the slow by slow processing is with MORE power!
SQL> -- This can be provided by Parallel Query because we are using a SELECT here...
SQL> ALTER SESSION FORCE PARALLEL QUERY;
Session altered.

SQL> SELECT COUNT(*)
  2    FROM TABLE(delete_demo(CURSOR(SELECT col1
  3                                    FROM demo)))
  4 /

COUNT(*)
----------
      1000

SQL> ALTER SESSION DISABLE PARALLEL QUERY;
Session altered.

SQL> SELECT COUNT(*)
  2    FROM demo
  3 /

COUNT(*)
----------
         0

SQL> -- Check the parallel query server usage
SQL> -- Please note the number of message sent and received
SQL> -- because that become important in later posts
SQL> SELECT statistic,
  2         value
  3    FROM v$pq_sysstat
  4   WHERE value > 0
  5 ORDER BY value DESC;

STATISTIC           VALUE
------------------ ------
Local Msgs Sent     22000
Local Msgs Recv'd   22000
Server Sessions      4000
DFO Trees            1000
Queries Initiated    1000
Servers Idle           12
Servers Started        12
Servers Highwater       4

At this point you may/should be thinking this is a lot of work just to delete records from a table. And you would be right.

It would be far simpler to just set the parallel degree for the table and run the DELETE directly against it.

And if you want to process the rows deleted just add a RETURNING clause to the DELETE statement.

And looping through the rows one by one will take forever!

Those are all valid arguments. And this series of posts is not intended to debunk any of them.

It’s intended to show Table Function as an option for processing large amounts of data especially when extra processing such as running totals must be added to the workload.

Over the next few articles I’ll expand on these topics. The next post introduces Parallelism for Table Functions.

Thanks for reading!

Explicit Cursors For Better Program Control

Arrgh! I just spent several days working through a bug in a large (1000+ line) unfamiliar package that was raising unexpected NO DATA FOUND errors.

After working through several calls to the procedure and monitoring the results I isolated the problem to code very similar to the following.

CREATE OR REPLACE PROCEDURE abc AS
  v_a VARCHAR2(1);

    PROCEDURE def IS
      v_b VARCHAR2(1);
    BEGIN
      SELECT NULL
        INTO v_b
        FROM another_table
       WHERE 1 = 2;
   END;

BEGIN

  -- get row from a_table
   BEGIN
     SELECT NULL
       INTO v_a
       FROM a_table
      WHERE 1 = 2;
   EXCEPTION
     WHEN NO_DATA_FOUND THEN
     -- if row not found then run def procedure
     def;
   END;

END;
/

A simple piece of code right? Nope. The call to the DEF procedure in the EXCEPTION handler was raising the NO DATA FOUND error which was not handled.

This caused the calling code to fail.

This brings up one of the things I don’t like to use in PL/SQL – implicit cursors. I agree they are simpler to type than explicit ones but when their associated EXCEPTION handlers are used for program flow it can quickly get unruly. I much prefer explicit cursors such as in this example:

CREATE OR REPLACE PROCEDURE abc AS

  v_a VARCHAR2(1);

  CURSOR curs_a IS
  SELECT NULL
    FROM a_table
   WHERE 1 = 2;

  CURSOR c_another IS
  SELECT NULL
    FROM another_table
   WHERE 1 = 2;

    PROCEDURE def IS
      v_b VARCHAR2(1);
    BEGIN
      OPEN c_another;
      FETCH c_another INTO v_b;
      CLOSE c_another;
    END;

BEGIN

  -- self documenting!?
  -- easy to follow!
  OPEN curs_a;
  FETCH curs_a INTO v_a;
  IF curs_a%NOTFOUND THEN
    def;
  END IF;
  CLOSE curs_a;

END;

Alright that’s enough ranting for today.

Thanks for reading.