How AUTO stats job decides on histograms and sample size?

Question

Hi All,
I am on Oracle 11.2.0.3 .
I have a table with 161M records. In fact it is an IOT. We are running the default stats collection job from Oracle (not even the time is changed !!).
This is the nature of stats on this table
SQL> select column_name, num_distinct, num_nulls, num_buckets, sample_size, histogram
  2  from dba_tab_col_statistics where table_name='TABLE1';
COLUMN_NAME                    NUM_DISTINCT  NUM_NULLS NUM_BUCKETS SAMPLE_SIZE HISTOGRAM
COLUMN1                                   1          0           1   162988917 NONE
COLUMN2                                   0  162988917           0             NONE
COLUMN3                           119808000          0         254        5548 HEIGHT BALANCED
COLUMN4                                3048          0         254        5548 HEIGHT BALANCED
COLUMN5                                   1          0           1   162988917 NONE
COLUMN6                                 173          0          77        5549 FREQUENCY
COLUMN7                            43225088          0           1   162988917 NONE
7 rows selected.I have so many question on this
Why sample size is so small on the columns where histogram exists?
Why sample size is 100% on the columns where there is no histogram? (I am assuming NONE means there is no histogram on this column)
COLUMN6 is a really skewed column. There are actually 173 distinct values in there. How come Oracle found that correctly in such a small sample size?
Thanks in advance

Answer 1

rahulras wrote:
SQL> select column_name, num_distinct, num_nulls, num_buckets, sample_size, histogram
2  from dba_tab_col_statistics where table_name='TABLE1';
COLUMN_NAME                    NUM_DISTINCT  NUM_NULLS NUM_BUCKETS SAMPLE_SIZE HISTOGRAM
COLUMN1                                   1          0           1   162988917 NONE
COLUMN2                                   0  162988917           0             NONE
COLUMN3                           119808000          0         254        5548 HEIGHT BALANCED
COLUMN4                                3048          0         254        5548 HEIGHT BALANCED
COLUMN5                                   1          0           1   162988917 NONE
COLUMN6                                 173          0          77        5549 FREQUENCY
COLUMN7                            43225088          0           1   162988917 NONE
7 rows selected.Why sample size is so small on the columns where histogram exists?
Because the code decided that the histogram it generated at that size was sufficiently accurate - but I can't tell how it came to that conclusion.
Why sample size is 100% on the columns where there is no histogram? (I am assuming NONE means there is no histogram on this column)
COLUMN6 is a really skewed column. There are actually 173 distinct values in there. How come Oracle found that correctly in such a small sample size?
The answer is probably that it didn't. I suspect you have enabled the approximate_ndv mechanism, which applies in the first pass for collecting simple stats, and the num_distinct in every case was therefore very close to perfect with a 100% sample size. The histograms would then have been collected in a second pass. You have only 77 buckets in the frequency histogram on column6, which means Oracle knows about just 77 values for that column despite reporting 173 as the number of distinct values. to my mind, the inconsistency between num_distinct (which should have been pretty accurate) and the number of buckets should have made Oracle collect another, larger sample size of the histogram.
Regards
Jonathan Lewis