Re: Very Strange Query Access Plan
- From: "Greg Rahn" <greg@xxxxxxxxxxxxxxxxxx>
- To: Thomas.Mercadante@xxxxxxxxxxxxxxxxx
- Date: Thu, 4 Oct 2007 11:12:28 -0700
Since you are on 9.2.0.7 the default METHOD_OPT did not collect
histograms. It wasn't until 10g that the default changed to collect
them.
I think your data is a poster child for a problem that I have seen
several times in recent years. This problem is with the sample of
data that DBMS_STATS is gathering with. You don't mentioned what your
ESTIMATE_PERCENT is (if you are using DBMS_STATS.AUTO_SAMPLE_SIZE or a
fixed %). If the data has a significant skew, generally a high number
of NDV with a very few values dominating the set of values,
AUTO_SAMPLE_SIZE can have challenges getting a representative sample.
This is due to the way AUTO_SAMPLE_SIZE works: that being an algorithm
based on convergence of the values observed. With skewed data, often
times the convergence happens before an representative NDV is
calculated. In the case of a fixed %, the chosen value is too low to
get a representative NDV count. I went into a real-world example last
year at Oracle OpenWorld 2006 which is available here:
http://structureddata.org/presentations/
In your case you have >27% (5M of 18M) of the data having 2 values.
That is pretty significant skew. In this case I would recommend
finding a sample size that yields representative NDV counts or if you
feel comfortable because you know your data well, setting the NDV
manually.
The AUTO_SAMPLE_SIZE in 11g has been enhanced to eliminate the issue
around NDV and skewed data. More on that here:
http://structureddata.org/2007/09/17/oracle-11g-enhancements-to-dbms_stats/
Hope this helps.
On 10/4/07, Mercadante, Thomas F (LABOR)
<Thomas.Mercadante@xxxxxxxxxxxxxxxxx> wrote:
> John,
>
> We were gathering stats with no "METHOD_OPT" option. And according to
> an Oracle SR, the calculation for density is *not* 1/NDV, but:
> DENSITY = SUM(1..NDV)(nocc^2)/(T^2)
> where T is the number of elements sampled, adjusted like nocc
> (i.e. values that span histogram buckets are removed).
> basically, for each distinct value (i.e. NDV) we count the number of
> occurences of that value (the nocc value) tossing any value that spans
> a histogram bucket."
>
> The SR Tech said that the simpler calculation "is a rough approximation
> of the formula above."
>
> Funny thing - I tested several scenarios and the first calculation seems
> to hold.
>
> He suggested trying histograms with a various number of buckets and
> testing the result, taking a 10046 trace to see what is happening. If I
> am not satisfied with my results, to submit a (possible) bug report.
>
> The skew of the data in this table is the real problem.
>
> 18,000,000 rows.
> Ssn column:
>
> 1,289,561 rows with a value of "undefined"
> 3,656,617 rows with a value of null
> 625,018 distinct values.
>
> So 4.8 million rows of bad data.
>
> Now, try and find a time to test this without killing my users!
>
> Tom
--
Regards,
Greg Rahn
http://structureddata.org
--
//www.freelists.org/webpage/oracle-l
Other related posts:
