In a forum I frequent a question was raised regarding an index not being used. The poster is using Oracle 11.1.0.7, which limits the index access paths Oracle can use; a concatenated index is created on the table (which has columns x,y,z,t,q and w, for lack of better names) on columns x, y, and z. Let’s look at what 11.2.0.3 will do with such a situation then explain what may be going on with 11.1.0.7 and, as a result, why the index isn’t being used.
The original problem states that the table is new, meaning it has been created and has no data in it (possibly a daily occurrence) and at least one index is created on this table during the create table process. Of course having no data in the table means that even if you do generate statistics they will be essentially useless. Let’s set this up with 11.2.0.3, using a single-column index (since at the time this example was created the existence of the concatenated index wasn’t known) and see what happens:
SQL>
SQL> --
SQL> -- Create an empty table with no indexes
SQL> --
SQL> create table emp2 as select * from emp where 0=1;
Table created.
SQL>
SQL> --
SQL> -- Create an index on the empty table
SQL> --
SQL> create index empno_idx on emp2(empno);
Index created.
SQL>
SQL> --
SQL> -- Populate the table (will insert 1,800,000 rows)
SQL> --
SQL>
SQL> begin
2 for i in 1..100000 loop
3 insert into emp2
4 select * from emp;
5
6 update emp2
7 set empno = empno+i, mgr=mgr+i, updated='YES'
8 where updated is null;
9
10 commit;
11 end loop;
12
13 commit;
14 end;
15 /
PL/SQL procedure successfully completed.
SQL>
SQL> --
SQL> -- Select from emp2 and see if any indexes are used
SQL> --
SQL> set autotrace on
SQL>
SQL> select *
2 from emp2
3 where empno = (select min(empno)+1 from emp);
EMPNO ENAME JOB MGR HIREDATE SAL COMM DEPTNO UPD
---------- ---------- --------- ---------- --------- ---------- ---------- ---------- ---
7370 SMITH CLERK 7903 17-DEC-80 800 20 YES
Execution Plan
----------------------------------------------------------
Plan hash value: 1300931333
-----------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
-----------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | 90 | 4 (0)| 00:00:01 |
| 1 | TABLE ACCESS BY INDEX ROWID| EMP2 | 1 | 90 | 1 (0)| 00:00:01 |
|* 2 | INDEX RANGE SCAN | EMPNO_IDX | 82203 | | 1 (0)| 00:00:01 |
| 3 | SORT AGGREGATE | | 1 | 13 | | |
| 4 | TABLE ACCESS FULL | EMP | 18 | 234 | 3 (0)| 00:00:01 |
-----------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("EMPNO"= (SELECT MIN("EMPNO")+1 FROM "EMP" "EMP"))
Note
-----
- dynamic sampling used for this statement (level=2)
Statistics
----------------------------------------------------------
7 recursive calls
1 db block gets
859 consistent gets
0 physical reads
0 redo size
1094 bytes sent via SQL*Net to client
523 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
1 rows processed
SQL>
SQL> set autotrace off
SQL>
SQL> --
SQL> -- Generate statistics
SQL> --
SQL> exec dbms_stats.gather_table_stats('GRIBNAUT', 'EMP2', estimate_percent=>100, cascade=>true)
PL/SQL procedure successfully completed.
SQL>
SQL> --
SQL> -- Try the query again
SQL> --
SQL> set autotrace on
SQL>
SQL> select *
2 from emp2
3 where empno = (select min(empno)+1 from emp);
EMPNO ENAME JOB MGR HIREDATE SAL COMM DEPTNO UPD
---------- ---------- --------- ---------- --------- ---------- ---------- ---------- ---
7370 SMITH CLERK 7903 17-DEC-80 800 20 YES
Execution Plan
----------------------------------------------------------
Plan hash value: 1300931333
-----------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
-----------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 18 | 738 | 22 (0)| 00:00:01 |
| 1 | TABLE ACCESS BY INDEX ROWID| EMP2 | 18 | 738 | 19 (0)| 00:00:01 |
|* 2 | INDEX RANGE SCAN | EMPNO_IDX | 18 | | 3 (0)| 00:00:01 |
| 3 | SORT AGGREGATE | | 1 | 13 | | |
| 4 | TABLE ACCESS FULL | EMP | 18 | 234 | 3 (0)| 00:00:01 |
-----------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("EMPNO"= (SELECT MIN("EMPNO")+1 FROM "EMP" "EMP"))
Note
-----
- dynamic sampling used for this statement (level=2)
Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
12 consistent gets
0 physical reads
0 redo size
1094 bytes sent via SQL*Net to client
523 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
1 rows processed
SQL>
SQL> set autotrace off
SQL>
Even with no ‘formal’ statistics Oracle 11.2.0.3 can use the CBO to generate access paths to use the index because of dynamic sampling. Notice that when statistics are gathered the cost estimates change to reflect those more accurate statistics. Knowing that a concatenated index was created we drop the original single-column index and replace it with a concatenated index where the column of interest is not the leading column:
SQL> --
SQL> -- Drop the empno index and replace it
SQL> --
SQL> drop index empno_idx;
Index dropped.
SQL> create index ename_empno_idx on emp2(ename, empno);
Index created.
SQL>
SQL> --
SQL> -- Try the query again
SQL> --
SQL> set autotrace on
SQL>
SQL> select *
2 from emp2
3 where empno = (select min(empno)+1 from emp);
EMPNO ENAME JOB MGR HIREDATE SAL COMM DEPTNO UPD
---------- ---------- --------- ---------- --------- ---------- ---------- ---------- ---
7370 SMITH CLERK 7903 17-DEC-80 800 20 YES
Execution Plan
----------------------------------------------------------
Plan hash value: 2131058526
-----------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
-----------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 18 | 738 | 26 (0)| 00:00:01 |
| 1 | TABLE ACCESS BY INDEX ROWID| EMP2 | 18 | 738 | 23 (0)| 00:00:01 |
|* 2 | INDEX SKIP SCAN | ENAME_EMPNO_IDX | 18 | | 20 (0)| 00:00:01 |
| 3 | SORT AGGREGATE | | 1 | 13 | | |
| 4 | TABLE ACCESS FULL | EMP | 18 | 234 | 3 (0)| 00:00:01 |
-----------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("EMPNO"= (SELECT MIN("EMPNO")+1 FROM "EMP" "EMP"))
filter("EMPNO"= (SELECT MIN("EMPNO")+1 FROM "EMP" "EMP"))
Note
-----
- dynamic sampling used for this statement (level=2)
Statistics
----------------------------------------------------------
4 recursive calls
0 db block gets
52 consistent gets
32 physical reads
0 redo size
1094 bytes sent via SQL*Net to client
523 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
1 rows processed
SQL>
SQL> set autotrace off
SQL>
Yes, we should have truncated the table, created the new index and re-loaded the data but the results, in 11.2.0.3, would be the same. Notice that this release of Oracle provides an INDEX SKIP SCAN access path; 11.1.0.7 doesn’t provide that index access pathway. As such it’s very likely that the column in the WHERE clause is referencing the second or third column in the index key and if that is the case then Oracle 11.1.0.7 will shift to a full table scan. To circumvent that it will be necessary to create another index, a concatenated index where the leading column is the column of interest as indicated by the WHERE clause. If we recreate the index yet again, swapping the columns so that the leading column is the column referenced in the WHERE clause both 11.1.0.7 and 11.2.0.3 will show the same plan:
SQL> -- Drop the empno index and replace it
SQL> --
SQL> drop index ename_empno_idx;
Index dropped.
SQL> create index empno_ename_idx on emp2(empno, ename);
Index created.
SQL>SQL> --
SQL> -- Try the queries again
SQL> --
SQL> set autotrace on
SQL>
SQL> select *
2 from emp2
3 where empno = (select min(empno)+1 from emp);
EMPNO ENAME JOB MGR HIREDATE SAL COMM DEPTNO UPD
---------- ---------- --------- ---------- --------- ---------- ---------- ---------- ---
7370 SMITH CLERK 7903 17-DEC-80 800 20 YES
Execution Plan
----------------------------------------------------------
Plan hash value: 3006901234
-----------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
-----------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 18 | 738 | 24 (0)| 00:00:01 |
| 1 | TABLE ACCESS BY INDEX ROWID| EMP2 | 18 | 738 | 21 (0)| 00:00:01 |
|* 2 | INDEX RANGE SCAN | EMPNO_ENAME_IDX | 18 | | 3 (0)| 00:00:01 |
| 3 | SORT AGGREGATE | | 1 | 13 | | |
| 4 | TABLE ACCESS FULL | EMP | 18 | 234 | 3 (0)| 00:00:01 |
-----------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("EMPNO"= (SELECT MIN("EMPNO")+1 FROM "EMP" "EMP"))
Note
-----
- dynamic sampling used for this statement (level=2)
Statistics
----------------------------------------------------------
4 recursive calls
0 db block gets
20 consistent gets
2 physical reads
0 redo size
1094 bytes sent via SQL*Net to client
520 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
1 rows processed
SQL>
So, we see that computing statistics on the index in question doesn’t help matters in 11.1.0.7 since, apparently, the problem query is looking for a table column that is not the leading column of that index. And we also see that adding an index, where the leading column is the column of interest, causes that index to be used. It’s good to remember that how the optimizer behaves is version-dependent.
