How does the CBO calculate the selectivity for range predicates on ROWID ?

Question

Hi all,
I'm wondering how the CBO estimate the selectivity for range predicates based on ROWID columns.
For example, for the following query the CBO estimates there's going to be 35004 rows returned instead of 7:
SQL> SELECT count(*)
  FROM intsfi i
WHERE
ROWID>='AAADxyAAWAAHDLIAAB' AND ROWID<='AAADxyAAWAAHDLIAAH';  2    3    4
  COUNT(*)
         7
Elapsed: 00:00:02.31
SQL> select * from table(dbms_xplan.display_cursor(null,null,'iostats last'));
PLAN_TABLE_OUTPUT
SQL_ID  aqbdu2p2t6w0z, child number 1
SELECT count(*)   FROM intsfi i  WHERE  ROWID>='AAADxyAAWAAHDLIAAB' AND
ROWID<='AAADxyAAWAAHDLIAAH'
Plan hash value: 1610739540
| Id  | Operation             | Name    | Starts | E-Rows | A-Rows |   A-Time   | Buffers |
|   0 | SELECT STATEMENT      |         |      1 |        |      1 |00:00:02.31 |   68351 |
|   1 |  SORT AGGREGATE       |         |      1 |      1 |      1 |00:00:02.31 |   68351 |
|*  2 |   INDEX FAST FULL SCAN| INTSFI2 |      1 |  35004 |      7 |00:00:02.31 |   68351 |
Predicate Information (identified by operation id):
   2 - filter((ROWID>='AAADxyAAWAAHDLIAAB' AND ROWID<='AAADxyAAWAAHDLIAAH'))According to Jonathan Lewis' book, for a normal column the selectivity would have been:
(value_column1-value_column2)/(high_value-low_value)+1/num_distinct+1/num_distinct
But here with the ROWID column, how does the CBO make its computation ?
SINGLE TABLE ACCESS PATH
  Single Table Cardinality Estimation for INTSFI[I]
  Table: INTSFI  Alias: I
    Card: Original: 14001681.000000  Rounded: 35004  Computed: 35004.20  Non Adjusted: 35004.20

Answer 1

Hi Jonathan,
Some Clarifications
=============
DELETE /*+ ROWID(I) */ FROM INTSFI I WHERE
(I.DAVAL<=TO_DATE('12032008','DDMMYYYY') AND (EXISTS(SELECT 1 FROM
INTSFI S WHERE S.COINT=I.COINT AND S.NUCPT=I.NUCPT AND S.CTSIT=I.CTSIT
AND NVL(S.RGCID,-1)=NVL(I.RGCID,-1) AND S.CODEV=I.CODEV AND
S.COMAR=I.COMAR AND S.DAVAL>I.DAVAL) AND I.COMAR IN (SELECT P.COMAR
FROM PURMAR P WHERE P.NUPUR=1))) AND ROWID>='AAADxyAAWAAHDLIAAB' AND
ROWID<='AAADxyAAWAAHDLIAAH'
Plan hash value: 1677274993
| Id  | Operation                      | Name    | Starts | E-Rows | A-Rows |   A-Time   | Buffers |  OMem |  1Mem | Used-Mem |
|   0 | DELETE STATEMENT               |         |      1 |        |      0 |00:00:05.94 |   53247 |    |          |          |
|   1 |  DELETE                        | INTSFI  |      1 |        |      0 |00:00:05.94 |   53247 |    |          |          |
|*  2 |   HASH JOIN SEMI               |         |      1 |   9226 |      7 |00:00:05.94 |   53180 |   783K|   783K|  471K (0)|
|   3 |    NESTED LOOPS                |         |      1 |   9226 |      7 |00:00:00.01 |      10 |    |          |          |
|*  4 |     TABLE ACCESS BY ROWID RANGE| INTSFI  |      1 |   9226 |      7 |00:00:00.01 |       6 |    |          |          |
|*  5 |     INDEX UNIQUE SCAN          | PURMAR1 |      7 |      1 |      7 |00:00:00.01 |       4 |    |          |          |
|   6 |    INDEX FAST FULL SCAN        | INTSFI1 |      1 |     14M|   7543K|00:00:01.73 |   53170 |    |          |          |
Predicate Information (identified by operation id):
   2 - access("S"."COINT"="I"."COINT" AND "S"."NUCPT"="I"."NUCPT" AND "S"."CTSIT"="I"."CTSIT" AND
              NVL("S"."RGCID",(-1))=NVL("I"."RGCID",(-1)) AND "S"."CODEV"="I"."CODEV" AND "S"."COMAR"="I"."COMAR")
       filter("S"."DAVAL">"I"."DAVAL")
   4 - access(ROWID>='AAADxyAAWAAHDLIAAB' AND ROWID<='AAADxyAAWAAHDLIAAH')
       filter("I"."DAVAL"<=TO_DATE(' 2008-03-12 00:00:00', 'syyyy-mm-dd hh24:mi:ss'))
   5 - access("P"."NUPUR"=1 AND "I"."COMAR"="P"."COMAR")
When I force the NESTED LOOP SEMI JOIN the query runs faster:
DELETE /*+ ROWID(I) */ FROM INTSFI I WHERE
(I.DAVAL<=TO_DATE('12032008','DDMMYYYY') AND (EXISTS(SELECT /*+ NL_SJ
*/ 1 FROM INTSFI S WHERE S.COINT=I.COINT AND S.NUCPT=I.NUCPT AND
S.CTSIT=I.CTSIT AND NVL(S.RGCID,-1)=NVL(I.RGCID,-1) AND S.CODEV=I.CODEV
AND S.COMAR=I.COMAR AND S.DAVAL>I.DAVAL) AND I.COMAR IN (SELECT P.COMAR
FROM PURMAR P WHERE P.NUPUR=1))) AND ROWID>='AAADxyAAWAAHDLIAAB' AND
ROWID<='AAADxyAAWAAHDLIAAH'
Plan hash value: 2031485112
| Id  | Operation                      | Name    | Starts | E-Rows | A-Rows |   A-Time   | Buffers |
|   0 | DELETE STATEMENT               |         |      1 |        |      0 |00:00:00.01 |      94 |
|   1 |  DELETE                        | INTSFI  |      1 |        |      0 |00:00:00.01 |      94 |
|   2 |   NESTED LOOPS SEMI            |         |      1 |   9226 |      7 |00:00:00.01 |      27 |
|   3 |    NESTED LOOPS                |         |      1 |   9226 |      7 |00:00:00.01 |       9 |
|*  4 |     TABLE ACCESS BY ROWID RANGE| INTSFI  |      1 |   9226 |      7 |00:00:00.01 |       5 |
|*  5 |     INDEX UNIQUE SCAN          | PURMAR1 |      7 |      1 |      7 |00:00:00.01 |       4 |
|*  6 |    INDEX RANGE SCAN            | INTSFI1 |      7 |     14M|      7 |00:00:00.01 |      18 |
Predicate Information (identified by operation id):
   4 - access(ROWID>='AAADxyAAWAAHDLIAAB' AND ROWID<='AAADxyAAWAAHDLIAAH')
       filter("I"."DAVAL"<=TO_DATE(' 2008-03-12 00:00:00', 'syyyy-mm-dd hh24:mi:ss'))
   5 - access("P"."NUPUR"=1 AND "I"."COMAR"="P"."COMAR")
   6 - access("S"."COINT"="I"."COINT" AND "S"."NUCPT"="I"."NUCPT" AND
              "S"."CTSIT"="I"."CTSIT" AND "S"."CODEV"="I"."CODEV" AND "S"."COMAR"="I"."COMAR" AND
              "S"."DAVAL">"I"."DAVAL")
       filter(NVL("S"."RGCID",(-1))=NVL("I"."RGCID",(-1)))the above post is from Ahmed AANGOUR
Case 1 - . If you check Plan hash value: 16772749938
=====
TABLE ACCESS BY ROWID RANGE| INTSFI  For every row access from INTSFI - it fetches a record from INDEX UNIQUE SCAN | PURMAR1
If we check A-rows = 9226
9226 * 7 = 64582 request across the table - perhaps with hint of rowid it fetches exact rows from PURMAR1
in this case i think going for hash join with ordered hints (jonathan as you suggest go for leading hint's instead of ordered) - from INTSFI - PURMAR1 - instead of going for IN clause would get the rows that satifies the ("P"."NUPUR"=1 AND "I"."COMAR"="P"."COMAR")
|*  2 |   HASH JOIN SEMI               |         |      1 |   9226 |      7 |00:00:05.94 |   53180 |   783K|   783K|  471K (0)|
|   3 |    NESTED LOOPS                |         |      1 |   9226 |      7 |00:00:00.01 |      10 |    |          |          |
|*  4 |     TABLE ACCESS BY ROWID RANGE| INTSFI  |      1 |   9226 |      7 |00:00:00.01 |       6 |    |          |          |
|*  5 |     INDEX UNIQUE SCAN          | PURMAR1 |      7 |      1 |      7 |00:00:00.01 |       4 |    |          |          |My understanding with above plan would change to
HASH JOIN
TABLE ACCESS BY ROWID RANGE| INTSFI
INDEX UNIQUE SCAN | PURMAR1
HASH JOIN
INDEX FAST FULL SCAN | INTSFI1
Which migt be feasible.
2 .
DELETE /*+ ROWID(I) */ FROM INTSFI I WHERE
(I.DAVAL<=TO_DATE('12032008','DDMMYYYY') AND (EXISTS(SELECT /*+ NL_SJ
*/ 1 FROM INTSFI S WHERE S.COINT=I.COINT AND S.NUCPT=I.NUCPT AND
S.CTSIT=I.CTSIT AND NVL(S.RGCID,-1)=NVL(I.RGCID,-1) AND S.CODEV=I.CODEV
AND S.COMAR=I.COMAR AND S.DAVAL>I.DAVAL) AND I.COMAR IN (SELECT P.COMAR
FROM PURMAR P WHERE P.NUPUR=1))) AND ROWID>='AAADxyAAWAAHDLIAAB' AND
ROWID<='AAADxyAAWAAHDLIAAH'Ahmed AANGOUR, modified the query by /*+ NL_SJ */ hint, Instead of that in to remove the most of the rows as we join the tables using subquery, I still doubt it
to go push_predicate hints - still doubt it.
Jonathan your comments are most valuable in the above two cases..
Looking forward to calrify my understanding with concepts of indexes for above test cases
- Pavan Kumar N