Why are there more cursors in 11g for my query containing bind variables?

Oracle introduced a new feature, adaptive cursor sharing, in 11g, to improve the plans that are selected for queries containing bind variables. This feature can result in more cursors for the same query containing bind variables. We'll explain why in this article. Before we get into the details, let's review a little history.

Oracle introduced the bind peeking feature in Oracle 9i. With bind peeking, the Optimizer peeks at the values of user-defined bind variables on the first invocation of a cursor. This allows the optimizer to determine the selectivity of any WHERE clause condition as if literals have been used instead of bind variables, thus improving the quality of the execution plan generated for statements using bind variables.

However, there was a problem with this approach, when the column used in the WHERE clause with the bind contained a data skew. If there is data skew in the column, it is likely that a histogram has been created on this column during statistics gathering. When the optimizer peeks at the value of the user-defined bind variable and chooses a plan, it is not guaranteed that this plan will be good for all possible values for the bind variable. In other words, the plan is optimized for the peeked value of the bind variable, but not for all possible values.

In 11g, the optimizer has been enhanced to allow multiple execution plans to be used for a single statement that uses bind variables. This ensures that the best execution plan will be used depending on the bind value. Let's look at an example to see exactly how this works.

Assume I have simple table emp which has 100,000 rows and has one index called emp_i1 on deptno column.

SQL> desc emp

Name                   Null?    Type
---------------------- -------- ----------------------------------
ENAME                           VARCHAR2(20)
EMPNO                           NUMBER
PHONE                           VARCHAR2(20)
DEPTNO                          NUMBER

There is a data skew in the deptno column, so when I gathered statistics on the emp table, Oracle automatically created a histogram on the deptno column.

SQL> select table_name, column_name, histogram from user_tab_cols;

TABLE_NAME         COLUMN_NAME        HISTOGRAM
------------------ ------------------ ---------------
EMP                DEPTNO             HEIGHT BALANCED
EMP                EMPNO              NONE
EMP                ENAME              NONE
EMP                PHONE              NONE

Now I will execute a simple select on my emp table, which has a single WHERE
clause predicate on the deptno column. The predicate contains a bind variable. We will begin by using the value 9 for this bind variable. The value 9 occurs 10 times in the table, i.e. in 0.0001% of the rows.

SQL> exec :deptno := 9

SQL> select /*ACS_1*/ count(*), max(empno)
2  from emp
3  where deptno = :deptno;


COUNT(*) MAX(EMPNO)
---------- ----------
    10         99

Given how selective the value 9 is, we should expect to get an index range scan for this query. Lets check the execution plan.

SQL> select * from table(dbms_xplan.display_cursor);

PLAN_TABLE_OUTPUT
-------------------------------------------------------------------------
SQL_ID  272gr4hapc9w1, child number 0
------------------------------------------------------------------------
select /*ACS_1*/ count(*), max(empno) from emp where deptno = :deptno

Plan hash value: 3184478295
------------------------------------------------------------------------
| Id | Operation                    | Name  | Rows | Bytes |Cost (%CPU)|
------------------------------------------------------------------------
|  0 | SELECT STATEMENT             |       |      |       |    2 (100)|
|  1 |  SORT AGGREGATE              |       |     1|    16 |           |
|  2 |   TABLE ACCESS BY INDEX ROWID| EMP   |     1|    16 |    2   (0)|
|  3 |    INDEX RANGE SCAN          | EMP_I1|     1|       |    1   (0)|
------------------------------------------------------------------------

So we got the index range scan that we expected. Now let's look at the execution statistics for this statement

SQL> select child_number, executions, buffer_gets,
2    is_bind_sensitive, is_bind_aware
3  from v$sql
4  where sql_text like 'select /*ACS_1%';

CHILD_NUMBER EXECUTIONS BUFFER_GETS IS_BIND_SENSITIVE IS_BIND_AWARE
------------ ---------- ----------- ----------------- -------------
          0          1          53 Y                 N

You can see we have one child cursor that has been executed once and has a small number of buffer gets. We also see that the cursor has been marked bind sensitive. A cursor is marked bind sensitive if the optimizer believes the optimal plan may depend on the value of the bind variable. When a cursor is marked bind sensitive, Oracle monitors the behavior of the cursor using different bind values, to determine if a different plan for different bind values is called for. This cursor was marked bind sensitive because the histogram on the deptno column was used to compute the selectivity of the predicate "where deptno = :deptno". Since the presence of the histogram indicates that the column is skewed, different values of the bind variable may call for different plans.

Now let's change the value of the bind variable to 10, which is the most popular value for the deptno column. It occurs 99900 times in the table, i.e in 99.9% of the rows.

SQL>  exec :deptno := 10

SQL> select /*ACS_1*/ count(*), max(empno)
2  from emp
3  where deptno = :deptno;

COUNT(*) MAX(EMPNO)
---------- ----------
 99900     100000

We expect to get the same plan as before for this execution because Oracle initially assumes it can be shared. Let's check:

SQL> select * from table(dbms_xplan.display_cursor);

PLAN_TABLE_OUTPUT
------------------------------------------------------------------------
SQL_ID  272gr4hapc9w1, child number 0
------------------------------------------------------------------------
select /*ACS_1*/ count(*), max(empno) from emp where deptno = :deptno

Plan hash value: 3184478295
------------------------------------------------------------------------
| Id | Operation                    | Name  | Rows | Bytes |Cost (%CPU)|
------------------------------------------------------------------------
|  0 | SELECT STATEMENT             |       |      |       |    2 (100)|
|  1 |  SORT AGGREGATE              |       |     1|    16 |           |
|  2 |   TABLE ACCESS BY INDEX ROWID| EMP   |     1|    16 |    2   (0)|
|  3 |    INDEX RANGE SCAN          | EMP_I1|     1|       |    1   (0)|
------------------------------------------------------------------------

The plan is still an index range scan as before, but if we look at the execution statistics, we should see two executions and a big jump in the number of buffer gets from what we saw before.

SQL> select child_number, executions, buffer_gets,
 2    is_bind_sensitive, is_bind_aware
 3  from v$sql
 4  where sql_text like 'select /*ACS_1%';

CHILD_NUMBER EXECUTIONS BUFFER_GETS IS_BIND_SENSITIVE IS_BIND_AWARE
------------ ---------- ----------- ----------------- -------------
          0          2        1007 Y                 N

You should also note that the cursor is still only marked bind sensitive and not bind aware at this point. So let's re-execute the statement using the same popular value, 10.

SQL> exec :deptno := 10

SQL> select /*ACS_1*/ count(*), max(empno)
2  from emp
3  where deptno = :deptno;

COUNT(*)  MAX(EMPNO) 
---------- -----------
99900      100000

Behind the scenes during the first two executions, Oracle was monitoring the behavior of the queries, and determined that the different bind values caused the data volumes manipulated by the query to be significantly different. Based on this difference, Oracle "adapts" its behavior so that the same plan is not always shared for this query. Hence a new plan is generated based on the current bind value, 10.

Let's check what the new plan is.

SQL> select * from table(dbms_xplan.display_cursor);

PLAN_TABLE_OUTPUT
--------------------------------------------------------------------
SQL_ID  272gr4hapc9w1, child number 1
--------------------------------------------------------------------
select /*ACS_1*/ count(*), max(empno) from emp where deptno = :deptno

Plan hash value: 2083865914
--------------------------------------------------------------------                                                                       
| Id  | Operation            | Name  | Rows   | Bytes | Cost (%CPU)|
--------------------------------------------------------------------
|   0 | SELECT STATEMENT     |       |        |       |   240 (100)|
|   1 |  SORT AGGREGATE      |       |     1  |   16  |            |
|*  2 |   TABLE ACCESS FULL  | EMP   | 95000  | 1484K |   240   (1)|
--------------------------------------------------------------------                                                                        

Given how unselective the value 10 is in the table, it's not surprising that the new plan is a full table scan. Now if we display the execution statistics we should see an additional child cursor (#1) has been created. Cursor #1 should show a number of buffers gets lower than cursor #0 and it is marked both bind sensitive and bind aware. A bind aware cursor may use different plans for different bind values, depending on how selective the predicates containing the bind variable are.

Looking at the execution statistics:

SQL> select child_number, executions, buffer_gets,
 2    is_bind_sensitive, is_bind_aware
 3  from v$sql
 4  where sql_text like 'select /*ACS_1%';

CHILD_NUMBER EXECUTIONS BUFFER_GETS IS_BIND_SENSITIVE IS_BIND_AWARE
------------ ---------- ----------- ----------------- -------------
          0          2        1007 Y                 N
          1          1         821 Y                 Y  

we see that there is a new cursor, which represents the plan which uses a table scan. But if we execute the query again with a more selective bind value, we should use the index plan:

SQL> exec :deptno := 9

SQL> select /*ACS_1*/ count(*), max(empno)
2  from emp
3  where deptno = :deptno;

COUNT(*) MAX(EMPNO)
---------- ----------
    10         99

SQL> select * from table(dbms_xplan.display_cursor);

PLAN_TABLE_OUTPUT
------------------------------------------------------------------------
SQL_ID  272gr4hapc9w1, child number 2
------------------------------------------------------------------------
select /*ACS_1*/ count(*), max(empno) from emp where deptno = :deptno

Plan hash value: 3184478295
------------------------------------------------------------------------
| Id | Operation                    | Name  | Rows | Bytes |Cost (%CPU)|
------------------------------------------------------------------------
|  0 | SELECT STATEMENT             |       |      |       |    2 (100)|
|  1 |  SORT AGGREGATE              |       |     1|    16 |           |
|  2 |   TABLE ACCESS BY INDEX ROWID| EMP   |     1|    16 |    2   (0)|
|  3 |    INDEX RANGE SCAN          | EMP_I1|     1|       |    1   (0)|
------------------------------------------------------------------------

The proper plan was chosen, based on the selectivity produced by the current bind value.

There is one last interesting thing to note about this. If we look at the execution statistics again, there are three cursors now:

SQL> select child_number, executions, buffer_gets,
 2    is_bind_sensitive, is_bind_aware, is_shareable
 3  from v$sql
 4  where sql_text like 'select /*ACS_1%';

CHILD_NUMBER EXECUTIONS BUFFER_GETS IS_B_SENS IS_B_AWAR IS_SHAR
------------ ---------- ----------- --------- --------- ----------
          0          2         957 Y         N         N
          1          1         765 Y         Y         Y
          2          2           6 Y         Y         Y

The original cursor was discarded when the cursor switched to bind aware mode. This is a one-time overhead. Note that the cursor is marked as not shareable (is_shareable is "N"), which means that this cursor will be among the first to be aged out of the cursor cache, and that it will no longer be used. In other words, it is just waiting to be garbage collected.

There is one other reason that you may notice additional cursors for such a query in 11g. When a new bind value is used, the optimizer tries to find a cursor that it thinks will be a good fit, based on similarity in the bind value's selectivity. If it cannot find such a cursor, it will create a new one (like above, when one (#1) was created for unselective "10" and one (#2) was created for highly-selective "9"). If the plan for the new cursor is the same as one of the existing cursors, the two cursors will be merged, to save space in the cursor cache. This will result in one being left behind that is in a not shareable state. This cursor will be aged out first if there is crowding in the cursor cache, and will not be used for future executions.

Q & A
Instead of answering the questions in your comments one by one, I am going to summarize the questions and provide my answers here.

Q: Is this behavior managed by 11g optimizer automatically and we don't need cursor_sharing anymore?
A: We have not changed the behavior of the cursor_sharing parameter yet, for backwards compatibility purposes. So if you set it to similar, adaptive cursor sharing will only kick in for queries where the literals are replace with binds. We hope that in the future, this feature will persuade people to set cursor_sharing to force.

Q: Would it have any impact like holding library cache latches for longer time to search for appropriate child cursor.
A: Any additional overhead in matching a cursor is always a concern, and we strive to minimize the impact. There is of course some increase in the code path to match a bind-aware cursor, since it requires more intelligent checks. This feature should not, however, impact cursors which are not yet marked bind-aware.

Q: What triggers a cursor to be marked "bind sensitive"?
A: Our goal is to consider many types of predicates where the selectivity can change when the bind value changes. In this first version of the feature, we only handle equality predicates where a histogram exists on the column and range predicates (with or without histogram). We do not currently consider LIKE predicates, but it is on the top of our list for future work.

Q: Also it sounds like the optimizer is using the number of rows returned to decided that it's time for a new plan...
A: I am not going to go into the details of the "special sauce" for how we decide to mark a cursor bind-aware. The number of rows processed is one input.

Q: Are you planning a hint to mark statements as bind-aware ?
A: Yes, we plan to add this in the future. This will allow users to bypass the startup cost of automatically determining that a query is a good candidate for bind-aware cursor sharing.