Function Based Index And Selectivity

Question

Hi All,
I have some doubts w.r.t FBI.I am on 10gR2 (10.2.0.4) with Solaris 5.9
I was under impression that FBI does not provide guaranteed index access and CBO choose access pattern purely on basis of available stats and query selectivity.
However, many a times i found that CBO is going for FBI even in case when FTS provides better query elapsed time.
I created following test case:
create table fbi_test (id number,flag varchar2(1));
begin
for i in 1..1000000
loop
insert into fbi_test values(i,'Y');
end loop;
end;
commit;
begin
for i in 1..10
loop
insert into fbi_test values(i,'N');
end loop;
end;
commit;
ANALYZE TABLE FBI_TEST COMPUTE STATISTICS;
CREATE INDEX fbi_test_FBI
ON fbi_test (CASE WHEN flag = 'Y' THEN 1 ELSE NULL END);
Autotrace for FBI ACCESS
SQL> select *from fbi_test where (CASE WHEN flag = 'Y' THEN 1 ELSE NULL END)=1;
1000000 rows selected.
Elapsed: 00:00:18.43
Execution Plan
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)|
| 0 | SELECT STATEMENT | | 10000 | 50000 | 342 (1)|
| 1 | TABLE ACCESS BY INDEX ROWID| FBI_TEST | 10000 | 50000 | 342 (1)|
|* 2 | INDEX RANGE SCAN | FBI_TEST_FBI | 4000 | | 1958 (1)|
Statistics
0 recursive calls
0 db block gets
136812 consistent gets
0 physical reads
0 redo size
22180292 bytes sent via SQL*Net to client
733814 bytes received via SQL*Net from client
66668 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
1000000 rows processed
Autotrace for FTS
SQL> select *from fbi_test where flag = 'Y';
1000000 rows selected.
Elapsed: 00:00:16.56
Execution Plan
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)|
| 0 | SELECT STATEMENT | | 500K| 2441K| 371 (9)|
|* 1 | TABLE ACCESS FULL| FBI_TEST | 500K| 2441K| 371 (9)|
Statistics
0 recursive calls
0 db block gets
68372 consistent gets
0 physical reads
0 redo size
22180292 bytes sent via SQL*Net to client
733814 bytes received via SQL*Net from client
66668 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
1000000 rows processed
FYI...
SQL> show parameter opt
NAME TYPE VALUE
filesystemio_options string asynch
object_cache_optimal_size integer 102400
optimizer_dynamic_sampling integer 2
optimizer_features_enable string 10.2.0.3
optimizer_index_caching integer 0
optimizer_index_cost_adj integer 100
optimizer_mode string ALL_ROWS
optimizer_secure_view_merging boolean TRUE
plsql_optimize_level integer 2
My questions are,
1.why oracle optimizer is going for FBI with high cardinality of 100000 rows?
2.Why cost of FTS is high (371) as compare to FBI (342) eventhough FTS is having fewer IO (68372) + Less Elapsed Time?
3.Why Optimizer is considering ELAPSED TIME during plan generation?
Any inpute would be highly appreciated.

Answer 1

user635930 wrote:
Hi All,
I have some doubts w.r.t FBI.I am on 10gR2 (10.2.0.4) with Solaris 5.9
Execution Plan
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)|
| 0 | SELECT STATEMENT | | 10000 | 50000 | 342 (1)|
| 1 | TABLE ACCESS BY INDEX ROWID| FBI_TEST | 10000 | 50000 | 342 (1)|
|* 2 | INDEX RANGE SCAN | FBI_TEST_FBI | 4000 | | 1958 (1)|
---------------------------------------------------------------------------------You're seeing three different effects here.
First - for the table access by index, Oracle has "lost" the cost of the index range scan - notice that the total cost of the query is 342, but the cost of the index access is 1958. The total cost of the query should be 2,300 and Oracle should have chosen the full tablescan automatically.
Second, the stats on the index show just one distinct value for "distinct_keys", and the optimizer has decided (for no reason I can think of - it may be a bug) to assume a 0.4% selectivity on the index.
Third, the estimated cardinality of the table and index lines differs. Whatever Oracle has done in the index line has been forgotten, and the cardinality of the table line has been based on the predicate given by your case statement and, as a "complex function", that predicate has been given a selectivity of 1% - hence the 10,000 rows estimate.
The combination of unsuitable statistics, an extreme case, and a couple of quirks in the optimizer mean that the chosen path is clearly unsuitable.
[Addendum]: It just occurred to me that part of the problem is that you collected stats on the table before you created the index. Given you're running 10g, the 'create index' would automatically generate index stats at the same time - but since it's a function-based index, there's a "virtual column" created for the table as well, and that column won't have any statistics on it - which is why you get the "fixed percentage" selectivities.
Regards
Jonathan Lewis
http://jonathanlewis.wordpress.com
http://www.jlcomp.demon.co.uk
"Science is more than a body of knowledge; it is a way of thinking" Carl Sagan
Edited by: Jonathan Lewis on Nov 29, 2008 1:15 PM