Create a test table:
create table customer (first_name varchar2(20), last_name varchar2(20) not null, address varchar(20));
insert into customer select dbms_random.string('U', 20), dbms_random.string('U', 20), dbms_random.string('U', 20) from dual connect by level <= 100000;
commit;
create index i_ln_fn_0 on customer(last_name, first_name,0); — just to be sure that all rows are indexed
Now the explain plan:
explain plan for
select /*+ FIRST_ROWS(20) */ *
from CUSTOMER
where first_name like 'AB%'
and first_name is not null
order by last_name;
select * from table(dbms_xplan.display);
-------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
-------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 197 | 12411 | 275 (2)| 00:00:04 |
| 1 | SORT ORDER BY | | 197 | 12411 | 275 (2)| 00:00:04 |
|* 2 | TABLE ACCESS FULL| CUSTOMER | 197 | 12411 | 274 (1)| 00:00:04 |
-------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - filter("FIRST_NAME" LIKE 'AB%' AND "FIRST_NAME" IS NOT NULL)
But as I want only the first rows, I'd like to avoid the sort of the whole table. I'd like to have a plan like this:
SELECT STATEMENT
TABLE ACCESS BY ROWID (customer)
INDEX FULL SCAN (i_ln_fn_0)
How to persuade the db to avoid the sort?
The problem is even worse. Even when I use the last_name only, everywhere:
explain plan for
select /*+ FIRST_ROWS(20) */ last_name
from CUSTOMER
where last_name like 'AB%'
and last_name is not null
order by last_name;
select * from table(dbms_xplan.display);
-------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
-------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 17 | 357 | 4 (25)| 00:00:01 |
| 1 | SORT ORDER BY | | 17 | 357 | 4 (25)| 00:00:01 |
|* 2 | INDEX RANGE SCAN| I_LN_FN_0 | 17 | 357 | 3 (0)| 00:00:01 |
-------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("LAST_NAME" LIKE 'AB%')
filter("LAST_NAME" LIKE 'AB%')
Here the sort is really not necessary, but the db still uses it. Why?
Edit: Tested on both
Oracle Database 11g Enterprise Edition Release 11.2.0.1.0 - 64bit Production
PL/SQL Release 11.2.0.1.0 - Production
"CORE 11.2.0.1.0 Production"
TNS for 64-bit Windows: Version 11.2.0.1.0 - Production
NLSRTL Version 11.2.0.1.0 - Production
and
Oracle Database 11g Enterprise Edition Release 11.2.0.2.0 - 64bit Production+
PL/SQL Release 11.2.0.2.0 - Production+
"CORE 11.2.0.2.0Production"+
TNS for IBM/AIX RISC System/6000: Version 11.2.0.2.0 - Production+
NLSRTL Version 11.2.0.2.0 - Production+
with the same plan.
The reason was in wrong value of NLS_SORT param. After changing it to BINARY, the plans began to look like I wanted.
From http://docs.oracle.com/cd/E24693_01/server.11203/e24448/initparams152.htm:
The value of NLS_SORT affects execution plans of queries. Because a standard index cannot be used as a source of values sorted in a linguistic order, an explicit sort operation must usually be performed instead of an index range scan. A functional index on the NLSSORT function may be defined to provide values sorted in a linguistic order and reintroduce the index range scan to the execution plan.
(I've got this answer from Paul Horth in forums.oracle com.)