Re: histograms
- From: "Charles Schultz" <sacrophyte@xxxxxxxxx>
- To: Joel.Patterson@xxxxxxxxxxx
- Date: Wed, 13 Sep 2006 09:23:04 -0500
I am dealing with something very similar to that myself right now. Can you
provide:
- your Oracle version
- Type of histogram or number of distinct values
- value of cursor_sharing
Your clever developer seems to be on the right track - he would be even more
clever if he did not call himself clever. *grin*
Aside from speculating about what has happened, have you gathered any
evidence? Trace files, information from v$sql_plan, v$sql_shared_cursor?
The Tuning and Performance Guide gives a bunch of rules for sharing:
http://download-east.oracle.com/docs/cd/B19306_01/server.102/b14211/memory.htm#sthref549
A capital letter would fit rule 3, but watch out for rule 5 (your SQL may be
massaged).
On 9/13/06, Joel.Patterson@xxxxxxxxxxx <Joel.Patterson@xxxxxxxxxxx> wrote:
This is from a developer. I'm just trying to help in a timely manner, so
any input or clarifications would be helpful.
JP
"I am strongly under the impression that the default behavior of 10g is to
use the bind variable values provided on the first execution attempt to
determine the execution plan, in the case where those values may matter (
e.g. in the presence of histograms).
So for one query in particular, I did something fairly clever - I actually
know (of the few possible values for the histogrammed column) which are
selective and which are not. I then issue the query with a subtle
difference (capitalization of one letter) depending on whether a selective
value was chosen or not. Thus, if a user runs the non-selective version,
Oracle should give them a different execution plan. By this cleverness, I
should be guaranteed that the plan that is used when a selective value is
chosen, is the plan that was first developed when a selective value was
first sent. But that's what I'm complaining about below - the plan it's
using is appropriate for a non-selective value, so it's as if
a) It didn't use the bound values in determining the plan
b) The histogram wasn't available when it determined the plan I've even
tried "alter system flush shared_pool" to force the regeneration of the
plan, and that doesn't seem to work. That also seems to rule out
b) above, because I can run queries without bind variables that definitely
are sensitive to the histogram."
--
Charles Schultz
Other related posts:
