DBMS_OUTPUT to a table with timestamp - plsql

Question

I am deleting a massive table and want to delete the table in batches. I am deleting records older than 467 days. I want to insert the dbms_output status in the following procedure to be written to a table that has two columns such as the number of records to be deleted with timestamp columns:
CREATE OR REPLACE PROCEDURE delete_tab (tablename IN VARCHAR2, nrows IN NUMBER ) IS
sSQL1 VARCHAR2(2000);
sSQL2 VARCHAR2(2000);
nCount NUMBER;
BEGIN
DBMS_OUTPUT.enable (100000);
nCount := 0;
sSQL1:='delete from '|| tablename ||
' where ROWNUM < ' || nrows || ' and where cast(time_stamp as date) < sysdate - 467';
sSQL2:='select count(ROWID) from ' || tablename ||
' where cast(time_stamp as date) < sysdate - 467';
EXECUTE IMMEDIATE sSQL2 INTO nCount;
LOOP
EXECUTE IMMEDIATE sSQL1;
nCount := nCount-nrows;
DBMS_OUTPUT.PUT_LINE('Existing records to be deleted: ' || to_char(nCount));
commit;
EXIT WHEN nCount = 0;
END LOOP;
END delete_tab;
/
Let me know how I can add an insert statement within the block to write the progress.

Answer 1

Besides the missing Insert there are a couple other minor issues and a major issue with this code.
ssql1:='delete from '|| tablename ||
' where ROWNUM < ' || nrows || ' and where cast(time_stamp as date) < sysdate - 467';
If the above were valid then when executed it nrows-1 (not nrows as I suspect you are thinking) every time as long as that many rows still exist. However, it is not valid; the where clause is invalid. The proper format for the where clause is
"Where <condition> and <condition> ..."; do not repeat where.
ssql2:='select count(ROWID) from ' || tablename ||
' where cast(time_stamp as date) < sysdate - 467';
There is no reason to count the rowids, every row has exactly 1. ROWID is pseudo column telling Oracle basically where on the disk the row is located. You have probably heard that it is the fastest way to retrieve a specific - that is true - but it is not true selecting rowid itself is faster, it will if anything, be slower. It would require a full table scan and calculation of its value for every row in the table. The optimizer may notice what is happening and change this to select count() but why hope for that just start with count().
Now we get to the worst issue. Unless the total number of rows to delete is an exact multiple of the parameter nrows then the procedure becomes a never ending loop until it throws the exception ORA-22054: underflow error.
This is because result as the only exit condition is "exit when ncount = 0". Suppose you have 1002 rows to delete and nrows is 1000. The first iteration deletes 999 then reduces ncount to 2 (1002-1000). Then the second iteration deletes another 999 and reduces ncount to -998 (2-1000). The procedure continues iterating until eventually ncount gets so small it cannot be held any longer (something like -1*10^39 - 1).
You could change the exit condition to "exit when nCount < 1". But even that is not necessary, no need for the code to calculate the exit condition at all. Let Oracle tell it. The delete statement returns SQL%rowcount that contains the number of rows processed by the last DML statement. If no are processed it returns 0. So the exist condition becomes a simple "exit when sql%rowcount = 0". Lesson: always be careful with equal 0 conditions.
Taking all this into account the procedure becomes. (Also changed variable names as I do not like ssql1, ssql2 and reusing a variable for 2 things. When using such I get confused as to which is which, esp on large procedures.):
create or replace procedure delete_tab
( tablename in varchar2
, nrows in number
)
is
delete_sql varchar2(2000);
count_sql varchar2(2000);
deleted_count number;
expect_delete number;
begin
dbms_output.enable (100000);
count_sql := 'select count(*) from ' || tablename ||
' where time_stamp < current_timestamp - 467';
dbms_output.put_line('COUNT_SQL: ' || count_sql
);
execute immediate count_sql into expect_delete;
dbms_output.put_line('Existing records to be deleted: ' || to_char(expect_delete));
delete_sql := 'delete from '|| tablename ||
' where rownum <= ' || nrows ||
' and cast(time_stamp as date) < sysdate - 467';
dbms_output.put_line('DELETE_SQL: '
|| delete_sql
);
deleted_count := 0;
loop
execute immediate delete_sql;
exit when sql%rowcount = 0;
deleted_count := deleted_count+sql%rowcount;
commit;
end loop;
insert into delete_log( table_name, date_deleted, rows_deleted)
values (tablename, sysdate, deleted_count) ;
commit; -- commit the deletes in final loop and insert
end delete_tab;
See demo (includes example of danger with ncount = 0 and procedure has additional trace information). BTW in a production run my commit interval (nRows) would something like 500000 or larger.