Real Time statistics - Part I

Keeping the object statistics upto date is critical to Oracle database performance and stability. Real time statistics introduced in Oracle 19c is one such feature agument the existing statistics. However they are not quite as real-time as the name suggest, to keep their implementation light weight this feature use the table monitoring mechanism and this even limits the amount of information can be collected.
 
And most importanly this feature is available only on engineerred system  and so certainly targeted at the autonomous databases.
 
Oracle 12c introduced online statistics gathering for direct path inserts and CREATE TABLE AS SELECT..
While 19c introduced real-time statistics, which extends online support for conventional DML statements.
 
Lets setup a test case to show how things worked prior to Oracle 19c and then will show how this feature “real-time statistics” will kick-in to provide improvements in Oracle 19c and above.
 
demo@XEPDB1> create table t (
  2     c1 number constraint t_pk primary key ,
  3     c2 number );
 
Table created.
 
demo@XEPDB1> insert /*+ append */ into t(c1,c2)
  2  select rownum, rownum
  3  from dual
  4  connect by level <=10000;
 
10000 rows created.
 
demo@XEPDB1> commit;
 
Commit complete.
 
demo@XEPDB1> select num_rows,blocks,stale_stats
  2  from user_tab_statistics
  3  where table_name ='T';
 
  NUM_ROWS     BLOCKS STALE_S
---------- ---------- -------
     10000         22 NO
 
demo@XEPDB1> select column_name,low_value,high_value,num_distinct,notes
  2  from user_tab_col_statistics
  3  where table_name ='T';
 
COLUMN_NAM LOW_VALUE  HIGH_VALUE NUM_DISTINCT NOTES
---------- ---------- ---------- ------------ -------------------------
C1         C102       C302              10000 STATS_ON_LOAD
C2         C102       C302              10000 STATS_ON_LOAD
 
So created a table and did direct path loads to get the statistics upto date with the latest data, now lets parse few sql’s and validate the estimates
 
demo@XEPDB1> select /*+ gather_plan_statistics */ max(c2), count(*)
  2  from t
  3  where c1 >= 9000;
 
   MAX(C2)   COUNT(*)
---------- ----------
     10000       1001
 
demo@XEPDB1> select * from table( dbms_xplan.display_cursor(format=>'allstats last'));
 
PLAN_TABLE_OUTPUT
----------------------------------------------------------------------------------------------------------------
SQL_ID  6fktndansb5hj, child number 0
-------------------------------------
select /*+ gather_plan_statistics */ max(c2), count(*) from t where c1
>= 9000
 
Plan hash value: 2966233522
 
----------------------------------------------------------------------------------------------
| Id  | Operation          | Name | Starts | E-Rows | A-Rows |   A-Time   | Buffers | Reads  |
----------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT   |      |      1 |        |      1 |00:00:00.01 |      21 |     18 |
|   1 |  SORT AGGREGATE    |      |      1 |      1 |      1 |00:00:00.01 |      21 |     18 |
|*  2 |   TABLE ACCESS FULL| T    |      1 |   1001 |   1001 |00:00:00.01 |      21 |     18 |
----------------------------------------------------------------------------------------------
 
Predicate Information (identified by operation id):
---------------------------------------------------
 
   2 - filter("C1">=9000)
 
With right statistics in place estimates were close to actuals.  Lets do few more data loads and verify the estimates
 
demo@XEPDB1> insert into t(c1,c2)
  2  select rownum+10000, rownum+10000
  3  from dual
  4  connect by level <=1000;
 
1000 rows created.
 
demo@XEPDB1> commit;
 
Commit complete.
 
demo@XEPDB1> select /*+ gather_plan_statistics */ max(c2), count(*)
  2  from t t2
  3  where c1 >= 9000;
 
   MAX(C2)   COUNT(*)
---------- ----------
     11000       2001
 
demo@XEPDB1> select * from table( dbms_xplan.display_cursor(format=>'allstats last'));
 
PLAN_TABLE_OUTPUT
----------------------------------------------------------------------------------------------------------------
SQL_ID  6vsgr79023ggn, child number 0
-------------------------------------
select /*+ gather_plan_statistics */ max(c2), count(*) from t t2 where
c1 >= 9000
 
Plan hash value: 2966233522
 
-------------------------------------------------------------------------------------
| Id  | Operation          | Name | Starts | E-Rows | A-Rows |   A-Time   | Buffers |
-------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT   |      |      1 |        |      1 |00:00:00.01 |      30 |
|   1 |  SORT AGGREGATE    |      |      1 |      1 |      1 |00:00:00.01 |      30 |
|*  2 |   TABLE ACCESS FULL| T    |      1 |   1001 |   2001 |00:00:00.01 |      30 |
-------------------------------------------------------------------------------------
 
Predicate Information (identified by operation id):
---------------------------------------------------
 
   2 - filter("C1">=9000)
 
 
20 rows selected.
 
demo@XEPDB1>
 
with few more 1000 rows got added to the table, but those changes were not yet available on statistics and due to that estimates were misleading.
 
But when we get into 19c database, the optimizer will keep track off all the dml changes in the data dictionary, And these statistics were feed back into the optimizer while optimising the future sql’s.
Here is an example of this performed in 19c.
 
demo@PDB19> insert into t(c1,c2)
  2  select rownum+10000, rownum+10000
  3  from dual
  4  connect by level <=1000;
 
1000 rows created.
 
demo@PDB19> commit;
 
Commit complete.
 
demo@PDB19> select num_rows,blocks,notes,stale_stats
  2  from user_tab_statistics
  3  where table_name ='T';
 
  NUM_ROWS     BLOCKS NOTES                     STALE_S
---------- ---------- ------------------------- -------
     10000         22                           NO
 
Here is where the 19c enhancements comes into play, when we query the data dictionary, we now have additional statistics and this new statistics we have is derived from normal DML activities that have occurred in the database and this is not a full GATHER Statistics option. The real time statistics can keep trace of simple column level statistics like whether low/high value of columns have changed.
 
demo@PDB19> select column_name,low_value,high_value,num_distinct,notes
  2  from user_tab_col_statistics
  3  where table_name ='T';
 
COLUMN_NAM LOW_VALUE  HIGH_VALUE NUM_DISTINCT NOTES
---------- ---------- ---------- ------------ -------------------------
C1         C102       C302              10000 STATS_ON_LOAD
C2         C102       C302              10000 STATS_ON_LOAD
C1         C102       C3020737                STATS_ON_CONVENTIONAL_DML
C2         C102       C3020737                STATS_ON_CONVENTIONAL_DML
 
demo@PDB19> select /*+ gather_plan_statistics */ max(c2), count(*)
  2  from t t1
  3  where c1 >= 9000;
 
   MAX(C2)   COUNT(*)
---------- ----------
     11000       2001
 
demo@PDB19> select * from table( dbms_xplan.display_cursor(format=>'allstats last'));
 
PLAN_TABLE_OUTPUT
----------------------------------------------------------------------------------------------
SQL_ID  4q9b758gr91m8, child number 0
-------------------------------------
select /*+ gather_plan_statistics */ max(c2), count(*) from t t1 where
c1 >= 9000
 
Plan hash value: 2966233522
 
---------------------------------------------------------------------------------------------
| Id  | Operation                  | Name | Starts | E-Rows | A-Rows |   A-Time   | Buffers |
---------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT           |      |      1 |        |      1 |00:00:00.01 |      30 |
|   1 |  SORT AGGREGATE            |      |      1 |      1 |      1 |00:00:00.01 |      30 |
|*  2 |   TABLE ACCESS STORAGE FULL| T    |      1 |   1709 |   2001 |00:00:00.01 |      30 |
---------------------------------------------------------------------------------------------
 
Predicate Information (identified by operation id):
---------------------------------------------------
 
   2 - storage("C1">=9000)
       filter("C1">=9000)
 
Note
-----
   - dynamic statistics used: statistics for conventional DML
 
Now the estimates is much close to actuals, we can expect a little variation from there from the true values, but the real-time statistics is not about performing a full gather statistics operation on a table.
So some elements of the statistics (like NDV , histograms etc) are not feasible to keep upto date everytime DML is performed. But having improved rowcounts yeild benefit to execution plan.
 
More to come, stay Tuned.