Question

I had a query where an index was not used when I thought it could be, so I reproduced it out of curiosity:

Create a test_table with 1.000.000 rows (10 distinct values in col, 500 bytes of data in some_data).

CREATE TABLE test_table AS (
  SELECT MOD(ROWNUM,10) col, LPAD('x', 500, 'x') some_data
  FROM dual
  CONNECT BY ROWNUM <= 1000000
);

Create an index and gather table stats:

CREATE INDEX test_index ON test_table ( col );

EXEC dbms_stats.gather_table_stats( 'MY_SCHEMA', 'TEST_TABLE' );

Try to get distinct values of col and the COUNT:

EXPLAIN PLAN FOR
  SELECT col, COUNT(*)
  FROM test_table
  GROUP BY col;

---------------------------------------------------------------------------------
| Id  | Operation          | Name       | Rows  | Bytes | Cost (%CPU)| Time
---------------------------------------------------------------------------------
|   0 | SELECT STATEMENT   |            |    10 |    30 | 15816   (1)| 00:03:10 
|   1 |  HASH GROUP BY     |            |    10 |    30 | 15816   (1)| 00:03:10 
|   2 |   TABLE ACCESS FULL| TEST_TABLE |   994K|  2914K| 15755   (1)| 00:03:10 
--------------------------------------------------------------------------------- 

The index is not used, providing the hint does not change this.

I guess, the index can’t be used in this case, but why?

Answer 1

I ran Peter’s original stuff and reproduced his results. I then applied dcp’s suggestion…

SQL> alter table test_table modify col not null;

Table altered.

SQL> EXEC dbms_stats.gather_table_stats( user, 'TEST_TABLE' , cascade=>true)

PL/SQL procedure successfully completed.

SQL> EXPLAIN PLAN FOR
  2    SELECT col, COUNT(*)
  3    FROM test_table
  4    GROUP BY col;

Explained.

SQL> select * from table(dbms_xplan.display)
  2  /

PLAN_TABLE_OUTPUT
------------------------------------------------------------------------------------
Plan hash value: 2099921975

------------------------------------------------------------------------------------
| Id  | Operation             | Name       | Rows  | Bytes | Cost (%CPU)| Time     |
------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT      |            |    10 |    30 |   574   (9)| 00:00:07 |
|   1 |  HASH GROUP BY        |            |    10 |    30 |   574   (9)| 00:00:07 |
|   2 |   INDEX FAST FULL SCAN| TEST_INDEX |  1000K|  2929K|   532   (2)| 00:00:07 |
------------------------------------------------------------------------------------

9 rows selected.

SQL>

The reason this matters, is because NULL values are not included in a normal B-TREE index, but the GROUP BY has to include NULL as a grouping “value” in your query. By telling the optimizer that there are no NULLs in col it is free to use the much more efficient index (I was getting an elapsed time of almost 3.55 seconds with the FTS). This is a classic example of how metadata can influence the optimizer.

Incidentally, this is obviously a 10g or 11g database, because it uses the HASH GROUP BY algorithm, instead of the older SORT (GROUP BY) algorithm.