Is my DB2 optimizer a tad drunk?

Posted by David Yahalom under DB2 LUW

Today I have noticed a strange phenomena in one of our DB2 databases. When doing a low zero-selectivity query (that is, query without a WHERE clause), the DB2 optimizer preferred to use an index scan rather than do a table scan.

That seemed a bit strange so I tried running the same query twice. Once without any indexes on the underlying table (to force a table scan) and then again with (to allow the optimizer use an index).

The results really surprised me.

According to the optimizer, the table scan query will perform slower than the index scan query.

I know that with Oracle systems, the rule of the thumb is that using a full table scan would perform faster than an index scan when the total number of rows retrieved from a table is between 1% and 15% of table data.

It is very logical. If the database has to read the entire table wouldn’t it be faster just to read the blocks directly from the table container rather than doing an INDEX READ + ROWID lookup for the entire table data?

The way I see it:

SELECT * FROM TABLE

Performing a TABLE SCAN to get the data: X I/O blocks read from the table.
Performing an INDEX SCAN to get the data: X I/O blocks read from the table + Y I/O blocks read from the index.

Or am I wrong?

Any of you can shed some light on why the DB2 optimizer reports the table scan as slower than an index scan in this case?

Here is the real world example of what I mean:

SELECT COUNT *
FROM TABLE1;

Using an INDEX:

Access Plan:
-----------
Total Cost: 72284.4
Query Degree: 1
Rows
RETURN
( 1)
Cost
I/O
|
95047
FETCH
( 2)
72284.4
26250.8
/---+---
95047 95047
RIDSCN TABLE: USER
( 3) TABLE1
556.668
275.886
|
95047
SORT
( 4)
556.667
275.886
|
95047
IXSCAN
( 5)
427.788
275.886
|
95047
INDEX: USER
TABLE1_IDX1

Without an INDEX:

Access Plan:
-----------
Total Cost: 73526
Query Degree: 1
Rows
RETURN
( 1)
Cost
I/O
|
95047
TBSCAN
( 2)
73526
27725
|
95047
TABLE: USER
TABLE1