Comments on: Cursor_sharing : a picture is worth a 1000 words

By: Shared Pool | My Blog

Shared Pool | My Blog — Thu, 02 Oct 2014 21:09:48 +0000

[…] http://dboptimizer.com/2010/06/08/cursor_sharing-a-picture-is-worth-a-1000-words/ […]

By: The CURSOR_SHARING Parameter is a Silver Bullet – What is Wrong with this Quote? « Charles Hooper's Oracle Notes

Sun, 03 Jul 2011 18:03:52 +0000

[...] Kyle Hailey and Randolf Geist discuss the CURSOR_SHARING parameter in a blog article titled “Cursor_sharing: a picture is worth a 1000 words“ [...]

By: Kyle Hailey

Kyle Hailey — Sun, 18 Jul 2010 19:31:37 +0000

By: Randolf

Randolf — Thu, 10 Jun 2010 08:29:56 +0000

Kyle,

I think the main issue with the CURSOR_SHARING is that it creates the child cursors unconditionally if the bind is marked as unsafe - so for each unique literal value you end up with a separate child cursor, but this is independent from the execution plan generated - so you might end up with literally thousand(s) of child cursors all having the same execution plan.

This is were Adaptive Cursor Sharing is different, since it tries to "group" the input values to minimize the number of child cursors.

However the potential drawback is that it will need a few "bad" executions to have Adaptive Cursor Sharing kick in which might not be acceptable in all cases since these "bad" executions might bring your system already to a halt.

Randolf

By: Kyle Hailey

Kyle Hailey — Thu, 10 Jun 2010 03:24:01 +0000

Interesting to note that the same problem shows up for equality
SELECT count(*) FROM t1 where c1=val1 and c2=val2
As Dion Cho pointed out in one of your links, maybe it has something as well to do with table stats missing, as I didn't create stats on the table.
Which ever the case, it was definitely considered a bug back when I was at Oracle in 2005. Surprised to see similar is still an issue in 11gR2
At one point if I recall correctly, a cursor was only allowed to have around 1000 children and after that it got errors, of the genre ORA-600. On 11gR2 I saw up to 2500 child cursors in the above example
Thanks for the note and links

By: Randolf

Randolf — Wed, 09 Jun 2010 19:25:53 +0000

Kyle,

I don't think that this is a "bug" per se, but intended behaviour of CURSOR_SHARING=SIMILAR.

There are more reasons why the child cursor won't be shared than histograms. Basically anything that could lead to a different execution plan might lead to a bind variable being marked as "unsafe" which can be seen from some flag in the 10046 trace file if I remember correctly.

As far as I know this is the list of reasons that lead to "unsafe" binds:

- predicates on columns with histograms
- range based predicates
- predicates on partition keys
- Dynamic Sampling

For more details, see e.g.

Oracle Forums thread 1

Oracle Forums thread 2

So your example above seems to fall into the category "range-based predicates".

Randolf