Slow Response Time Joining Multiple Domain Indexes

Question

Hi All,
I am working with a schema that has several full text searchable columns along with several other convential columns. I noticed that when I join on 2 Domain Indexes, the response time for query completion nearly doubles. However, If I were to combine my 2 clob columns into 1 clob, the extra cost of finding the intersection of 2 rows sets can be saved..
In my query, I am taking 2 sets of random high probability words (the top 1000 sorted by token_count DESC).
NOTE: All of my query execution times are taken with words not previously used to avoid caching by the engine..
HERE IS THE SLOW VERSION OF THE QUERY WHICH REFERENCES THE BODY CLOB TWICE:
SELECT count(NSS_ID) FROM jk_test_2 WHERE
CONTAINS (body, '( STANDARDS and HELP ) ' ) > 0
AND
CONTAINS (body, '( WORKING and LIMITED ) ' ) > 0 ;
THE EXPLAIN PLAN shows the intersection being calculated:
| Id | Operation               | Name | Rows | Bytes | Cost (%CPU)| Time |
| 0 | SELECT STATEMENT          |     | 1 | 99 | 3836     (0)| 00:00:47 |
| 1 | SORT AGGREGATE           |     | 1 | 99 |     |     |
| 2 | BITMAP CONVERSION TO ROWIDS     |     | | |     |     |
| 3 | BITMAP AND               |     | | |     |     |
| 4 | BITMAP CONVERSION FROM ROWIDS|     | | |     |     |
| 5 | SORT ORDER BY          |     | | |     |     |
|* 6 | DOMAIN INDEX          | JK_BODY_NDX | | | 1284     (0)| 00:00:16 |
| 7 | BITMAP CONVERSION FROM ROWIDS|     | | |     |     |
| 8 | SORT ORDER BY          |     | | |     |     |
|* 9 | DOMAIN INDEX          | JK_BODY_NDX | | | 2547     (0)| 00:00:31 |
Predicate Information (identified by operation id):
6 - access("CTXSYS"."CONTAINS"("BODY",'( PURCHASE and POSSIBLE)')>0 AND
     "CTXSYS"."CONTAINS"("BODY",'(NATIONAL and OFFICIAL)')>0)
9 - access("CTXSYS"."CONTAINS"("BODY",'(NATIONAL and OFFICIAL)')>0)
I RAN 3 QUERIES and got these times:
Elapsed: 00:00:00.25
Elapsed: 00:00:00.21
Elapsed: 00:00:00.27
HERE IS THE QUERY RE-WRITTEN INTO A DIFFERENT FORMAT WHICH COMBINES THE 2 PARTS INTO 1:
SELECT count(NSS_ID) FROM jk_test_2 WHERE
CONTAINS (body, '( ( STANDARDS and HELP ) AND ( WORKING and LIMITED ) ) ' ) > 0;
The Plan is now:
| Id | Operation     | Name | Rows | Bytes | Cost (%CPU)| Time |
| 0 | SELECT STATEMENT |     | 1 | 99 | 3207     (0)| 00:00:39 |
| 1 | SORT AGGREGATE |     | 1 | 99 |     |     |
|* 2 | DOMAIN INDEX     | JK_BODY_NDX | 5 | 495 | 3207     (0)| 00:00:39 |
Predicate Information (identified by operation id):
2 - access("CTXSYS"."CONTAINS"("BODY",'( ( FORM and GIVE ) AND (WEB and AGREED) ) ')>0)
I RAN 3 QUERIES using new words and got these times:
Elapsed: 00:00:00.12
Elapsed: 00:00:00.11
Elapsed: 00:00:00.13
Although the cost is only %15 lower, it executes twice as fast. Also, the same improvement is gained if OR's are used instead of ands:
With --> CONTAINS (BODY,'( ( ( FORM OR GIVE ) and (WEB OR AGREED) ) ') >0
My 2 timings are .25 and .50 with the OR'ed clause above getting the better response time..
BASED ON THIS, On my project I am tempted to merged 2 totally seperate clob columns into 1, to get the better response time. The 2 columns are 1) Body Text and 2) Codes. They LOGICALLY DO NOT BELONG WITH ONE ANOTHER. This would be very ackward for the project as it would cause a lot of fudging of the data in many places throughout the system. I have done this testing taking averages of 500 unique queries, my indexes are up to date with full statistics computed on my tables and indexes. Response time is HIGHLY CRITICAL for this project.
Does anyone have any advice that would allow me to obtain the good response time avoiding the ackwardness of globbing all the data into one clob ?
Joek

Answer 1

You might try adding sections and querying using within.