How to do Multithreading in Oracle PlSql using a Single Session

Question

Hi All,
I have a complex but interesting requirement.
1) I have a table with 90k rows.
2) My PlSql Package picks up all the 90k rows one by one, does some processing/transformation on it
3) if the row satisfies some business rules, then it gets inserted into an application table. If the row is not eligible, then it inserts the row into a Error table.
4) After processing all the 90k records, If my Error Table Count reaches 10% of the input count (90k), then all inserts (i.e. both application table inserts and
error table inserts) must be rolled back.
5) Right now, we are achieving this logic in a single session but there are far many performance issues, because processing has to go through row by row.
Is it possible to implement the above logic as "multi threading within PlSql" whereby i can run Step 2 in 10 parallel slaves and if the total error count of all the 10 slaves exceeds 10% of the input data, then there should be a rollback for all 10 slaves.
+A sample testcase is given below. Issue is: rollback statement in p1 is not taking any effect (because plsql spawns sessions to execute p2 which autocommits itself on graceful exit). But is there any other way by which i can make the rollback to be effective+
create table t1 (a number);
create table j1(b number);
create or replace procedure p1
as
l_cnt number;
i number;
l_job number;
begin
     for i in 1 .. 5 loop
        dbms_job.submit( l_job, 'p2('||i||');' );       
    end loop;
    commit;
    loop
        dbms_lock.sleep(30);
        select count(*) into l_cnt from j1;
        exit when (l_cnt = 5);
    end loop;
   rollback;
end;
create or replace procedure p2(i number)
as
begin
insert into t1 values(i);
insert into j1 values(1);
end;
/

Answer 1

Basically your current 'algorithm' can be described as 'slow-by-slow processing' (co Tom Kyte)
Relational databases are about sets not about records.
Consequently your approach must be qualified as evil.
Instead of fixing the problem (and getting rid of slow-by-slow processing), you propose to add even more evil by adding multi-threading to the mix.
The issue you describe is not an issue at all, it is just making clear you are heading for disaster.
An approach that works quite nicely is
- add an extra status column to the table to be processed.
- UPDATE the status column of all records that do not qualify. By UPDATE I mean the UPDATE statement, not record by record processing
- (BULK)-INSERT the remaining records, for which the flag is NULL.
The simple guidelines for application development are:
Only do it in PL/SQL when you can't do it in straight SQL
Only do it in Java when you can't do it in PL/SQL
Only do it in 3GL when you can't do it in Java.
Your 'strategy' is doomed to fail.
Get rid of it.
Before it is too late, and your heart is pumping like mad, because your problems are increasing, and increasing, and increasing, and you have to work 7 x 24 to deal with 'performance problems'
You have already demonstrated your approach doesn't work.
Don't try to make it worse.
Sybrand Bakker
Senior Oracle DBA
Experts: those who do read documentation