Create very big table

Question

Hi ,
call_fact table contain about 300 milion tables.
exceptions table contain about 150 milion tables.
Both tables have an uptodate statistics.
The machine have 8 CPUs
The statment already run 48 hours.
Can one suggest a faster way to do it ?
create table abc parallel
as
select /*+ parallel(t,32) */ *
from STARQ.CALL_FACT t
where rowid NOT IN (select /*+ parallel(ex,32) */ row_id
from starq.exceptions ex );
The plan is:
Plan
CREATE TABLE STATEMENT ALL_ROWSCost: 1,337,556,446,040                                                        
15 PX COORDINATOR                                                   
14 PX SEND QC (RANDOM) PARALLEL_TO_SERIAL SYS.:TQ30001 :Q3001Cost: 26,234 Bytes: 43,994,469,792 Cardinality: 282,015,832                                              
13 LOAD AS SELECT PARALLEL_COMBINED_WITH_PARENT :Q3001                                        
12 BUFFER SORT PARALLEL_COMBINED_WITH_CHILD :Q3001                                   
11 PX RECEIVE PARALLEL_COMBINED_WITH_PARENT :Q3001Cost: 26,234 Bytes: 43,994,469,792 Cardinality: 282,015,832                               
10 PX SEND ROUND-ROBIN PARALLEL_FROM_SERIAL SYS.:TQ30000 Cost: 26,234 Bytes: 43,994,469,792 Cardinality: 282,015,832                          
9 FILTER                     
4 PX COORDINATOR                
3 PX SEND QC (RANDOM) PARALLEL_TO_SERIAL SYS.:TQ20000 :Q2000Cost: 26,234 Bytes: 43,994,469,792 Cardinality: 282,015,832           
2 PX BLOCK ITERATOR PARALLEL_COMBINED_WITH_CHILD :Q2000Cost: 26,234 Bytes: 43,994,469,792 Cardinality: 282,015,832 Partition #: 10 Partitions accessed #1 - #46     
1 TABLE ACCESS FULL TABLE PARALLEL_COMBINED_WITH_PARENT STARQ.CALL_FACT :Q2000Cost: 26,234 Bytes: 43,994,469,792 Cardinality: 282,015,832 Partition #: 10 Partitions accessed #1 - #46
8 PX COORDINATOR                
7 PX SEND QC (RANDOM) PARALLEL_TO_SERIAL SYS.:TQ10000 :Q1000Cost: 4,743 Bytes: 10 Cardinality: 1           
6 PX BLOCK ITERATOR PARALLEL_COMBINED_WITH_CHILD :Q1000Cost: 4,743 Bytes: 10 Cardinality: 1      
5 TABLE ACCESS FULL TABLE PARALLEL_COMBINED_WITH_PARENT STARQ.EXCEPTIONS :Q1000Cost: 4,743 Bytes: 10 Cardinality: 1

Answer 1

> When in doubt, I use exists. Here it is clear to me that exists will be faster
If the row_id column is declared not null, this is not true: exactly the same path is chosen as can be seen below.
select /* with primary key */
  from call_fact t
where rowid not in
       ( select row_id
           from exceptions ex
call     count       cpu    elapsed       disk      query    current        rows
Parse        1      0.00       0.00          0          0          0           0
Execute      1      0.00       0.00          0          0          0           0
Fetch     1001      0.46       0.46          0      32467          0       15000
total     1003      0.46       0.47          0      32467          0       15000
Misses in library cache during parse: 1
Optimizer mode: ALL_ROWS
Parsing user id: 61 
Rows     Row Source Operation
  15000  NESTED LOOPS ANTI (cr=32467 pr=0 pw=0 time=600105 us)
  30000   TABLE ACCESS FULL CALL_FACT (cr=1466 pr=0 pw=0 time=120050 us)
  15000   INDEX UNIQUE SCAN EX_PK (cr=31001 pr=0 pw=0 time=297574 us)(object id 64376)
select /* with primary key */
  from call_fact t
where not exists
       ( select 'same rowid'
           from exceptions ex
          where ex.row_id = t.rowid
call     count       cpu    elapsed       disk      query    current        rows
Parse        1      0.00       0.00          0          0          0           0
Execute      1      0.00       0.00          0          0          0           0
Fetch     1001      0.51       0.46          0      32467          0       15000
total     1003      0.51       0.47          0      32467          0       15000
Misses in library cache during parse: 1
Optimizer mode: ALL_ROWS
Parsing user id: 61 
Rows     Row Source Operation
  15000  NESTED LOOPS ANTI (cr=32467 pr=0 pw=0 time=585099 us)
  30000   TABLE ACCESS FULL CALL_FACT (cr=1466 pr=0 pw=0 time=120048 us)
  15000   INDEX UNIQUE SCAN EX_PK (cr=31001 pr=0 pw=0 time=298198 us)(object id 64376)
********************************************************************************Note that the tables, scaled down to 30,000 and 15,000 rows, are created like this:
SQL> create table call_fact (col1, col2)
  2  as
  3   select level
  4        , lpad('*',100,'*')
  5     from dual
  6  connect by level <= 30000
  7  /
Tabel is aangemaakt.
SQL> create table exceptions (row_id, col)
  2  as
  3  select rowid
  4       , lpad('*',100,'*')
  5    from call_fact
  6   where mod(col1,2) = 0
  7  /
Tabel is aangemaakt.
SQL> alter table exceptions add constraint ex_pk primary key (row_id)
  2  /
Tabel is gewijzigd.
SQL> exec dbms_stats.gather_table_stats(user,'call_fact')
PL/SQL-procedure is geslaagd.
SQL> exec dbms_stats.gather_table_stats(user,'exceptions',cascade=>true)
PL/SQL-procedure is geslaagd.Without declaring row_id not null, I've test exists to be definitely much faster, as the not in variant cannot do an antijoin anymore.
Regards,
Rob.