Re: Differences in Range Scan Performance between Binary and NLS Indexes
Date: Sat, 10 May 2008 10:49:52 +0100
Message-ID: <8dOdnQMd__BQ8rjVnZ2dnUVZ8umdnZ2d@bt.com>
"Pat" <pat.casey_at_service-now.com> wrote in message
news:b0cf7d61-3547-48e0-a6e4-f2cdbfcde83a_at_27g2000hsf.googlegroups.com...
I have a table set with about 340k rows in it. User's execute queries
against this table for arbitrary search strings which are resolved
against the name table in the column. The table contains mixed case,
but the users expect to match their search regardless of case. So, for
example, both of these should match a search term of "badg":
ARabidBadger
somebadglowercase
A typical search, for all those entries whose name contains "badg" would look like this:
SELECT "sys_id" FROM (SELECT /*+ FIRST_ROWS(250) */ a.*, ROWNUM rnum FROM (SELECT cmdb_ci0."sys_id" FROM cmdb_ci cmdb_ci0 WHERE lower(cmdb_ci0."name") LIKE '%badge%' ORDER BY lower(cmdb_ci0."name")) a where ROWNUM <= 250) WHERE rnum > 0
To make this query efficient, I added a covering index
create index cmdb_ci_lower_name_sys_id on cmdb_ci (lower("name"), "sys_id");
And our query plan looks like this and completes in about 120 ms. This was worst case since only 1 row was returned. Short search strings like "b" will hit their stop key quickly and exit after finding 250 matches without having to exhaust the index.
Still, worst case, this query does a full index scan in about 120ms;
SQL> SELECT "sys_id" FROM (SELECT /*+ FIRST_ROWS(250) */ a.*, ROWNUM rnum FROM (SELECT cmdb_ci0."sys_id" FROM cmdb_ci cmdb_ci0 WHERE lower(cmdb_ci0."name") LIKE '%badge%' ORDER BY lower(cmdb_ci0."name")) a where ROWNUM <= 250) WHERE rnum > 0;
Elapsed: 00:00:00.12
Execution Plan
Plan hash value: 424525705
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)|
Time |
| 0 | SELECT STATEMENT | | 250 | 11750 | 46 (0)|
00:00:01 |
|* 1 | VIEW | | 250 | 11750 | 46 (0)| 00:00:01 |
|* 2 | COUNT STOPKEY | | | | | |
| 3 | VIEW | | 251 | 8534 | 46 (0)| 00:00:01
|
|* 4 | INDEX FULL SCAN| CMDB_CI_LOWER_NAME_SYS_ID | 16997 |
813K| 46 (0)| 00:00:01 |
Predicate Information (identified by operation id):
1 - filter("RNUM">0)
2 - filter(ROWNUM<=250)
4 - filter(LOWER("name") LIKE '%badge%')
Statistics
141 recursive calls
0 db block gets
2964 consistent gets
0 physical reads
0 redo size
439 bytes sent via SQL*Net to client
400 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
1 rows processed
Recently though, we started storing NLS characters in this column (european customers), so we'll have values like:
badger
bädger <-- note the accented a
The users want both of the above to collate next to each other.
So we, not unnaturally, put the database in linguisitc mode:
alter session set nls_sort=german;
alter session set nls_comp=linguistic;
Then we issue the same query. The query plan here involves a sort, but the query returns a grand total of 1 record, so the sort isn't material to the response time.
Note that the same query, running the same index full scan against the same index, now takes 1.23 second, roughly 10X as long as it took when running in binary mode.
SQL> SELECT "sys_id" FROM (SELECT /*+ FIRST_ROWS(250) */ a.*, ROWNUM rnum FROM (SELECT cmdb_ci0."sys_id" FROM cmdb_ci cmdb_ci0 WHERE lower(cmdb_ci0."name") LIKE '%badge%' ORDER BY lower(cmdb_ci0."name")) a where ROWNUM <= 250) WHERE rnum > 0;
Elapsed: 00:00:01.23
Execution Plan
Plan hash value: 3626452865
| Id | Operation | Name | Rows | Bytes |TempSpc| Cost
(%CPU)| Time |
| 0 | SELECT STATEMENT | | 250 | 11750 | | 870 (3)|
00:00:11 |
|* 1 | VIEW | | 250 | 11750 | | 870 (3)| 00:00:11
|
|* 2 | COUNT STOPKEY | | | | | | |
| 3 | VIEW | | 16997 | 564K| | 870 (3)|
00:00:11 |
|* 4 | SORT ORDER BY STOPKEY| | 16997 | 813K| 2008K|
870 (3)| 00:00:11 |
|* 5 | INDEX FAST FULL SCAN| CMDB_CI_LOWER_NAME_SYS_ID | 16997
| 813K| | 659 (3)| 00:00:08 |
Predicate Information (identified by operation id):
1 - filter("RNUM">0)
2 - filter(ROWNUM<=250)
4 - filter(ROWNUM<=250)
5 - filter(LOWER("name") LIKE '%badge%')
Statistics
134 recursive calls
0 db block gets
2978 consistent gets
2929 physical reads
0 redo size
439 bytes sent via SQL*Net to client
400 bytes received via SQL*Net from client 2 SQL*Net roundtrips to/from client
1 sorts (memory)
0 sorts (disk)
1 rows processed
If I take off the order by clause so we use exactly the same plan as our original query, we still get:
SQL> SELECT "sys_id" FROM (SELECT /*+ FIRST_ROWS(250) */ a.*, ROWNUM rnum FROM (SELECT cmdb_ci0."sys_id" FROM cmdb_ci cmdb_ci0 WHERE lower(cmdb_ci0."name") LIKE '%badge%' ) a where ROWNUM <= 250) WHERE rnum > 0;
Elapsed: 00:00:01.20
Execution Plan
Plan hash value: 1941712263
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)|
Time |
| 0 | SELECT STATEMENT | | 250 | 11750 | 11 (0)|
00:00:01 |
|* 1 | VIEW | | 250 | 11750 | 11 (0)| 00:00:01 |
|* 2 | COUNT STOPKEY | | | | | |
|* 3 | INDEX FAST FULL SCAN| CMDB_CI_LOWER_NAME_SYS_ID | 251 |
12299 | 11 (0)| 00:00:01 |
Predicate Information (identified by operation id):
1 - filter("RNUM">0)
2 - filter(ROWNUM<=250)
3 - filter(LOWER("name") LIKE '%badge%')
Statistics
134 recursive calls
0 db block gets
2979 consistent gets
0 physical reads
0 redo size
439 bytes sent via SQL*Net to client
400 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
1 rows processed
So my questions are is:
- Am I doing something wrong here? Why does putting the system into NLS_SORT and NLS_COMP make this big a difference on performance?
- Is there an index I should add here that would allow some form of efficient querying across this set? I know I'm going to have to do a full scan here (contains queries being what they are), but why is my full scan 10X slower in linguistic mode?
- Is there some other recommended approach to getting linguistic collation working efficiently? Shadow columns aren't really practical here (and even then I'm not sure what I'd put in the shadow unless I wrote my own german to english unaccenter).
Any help would be much appreciated.
A couple of points.
In starting case, you are doing an "INDEX FULL SCAN". Because the "order by" clause matches an index that can be used for the "like" - Oracle knows that if it walks the index in order then it can stop as soon as it has found enough data without sorting.
In the second case, Oracle has to do the "table scan" - but in fact, since all the necessary data is available in an index, it can do an "index FAST full scan", which is the tablescan mechanism (multiblock read) applied to an index segment. Because of this, it has to read the whole data set, and then sort it to get the first 250. This gives you three possible reasons for the extra time:
a) time for reading all the data b) time for sorting c) physical reads - a 'full scan' will buffer the index blocks,a 'fast full scan' may not, unless the index is a small one. This means you have to check your test conditions very carefully to see that you are comparing like with like, and also creating a test that will reflect the production behaviour closely.
In the last case (without the order by), Oracle could stop the fast full scan early (after finding 250) rows, and doesn't have to sort. It also happens that you don't do any physical reads - so the index blocks were buffered for this test, when they weren't for the previous test.
Note - your third test wasn' the same as your first - the third one is still doing the FAST full scan of the second test.
-- Regards Jonathan Lewis http://jonathanlewis.wordpress.com Author: Cost Based Oracle: Fundamentals http://www.jlcomp.demon.co.uk/cbo_book/ind_book.html The Co-operative Oracle Users' FAQ http://www.jlcomp.demon.co.uk/faq/ind_faq.htmlReceived on Sat May 10 2008 - 04:49:52 CDT