By AndyBaker on Apr 18, 2013
On a client site and had an issue with a 3rd party product to do with an inbuilt query executing from an application. A specific query was taking forever to run and the execution plan was not ideal. Nothing obvious had shown anything useful such as hints or indexes etc.
Obviously as the code is submitted from a product itself therefore it could not be changed! Table and Index statistics were up to date and the query had always been pretty poor in performance when examining the AWR history.
However during an initial analysis of the code a much better performing execution plan was located, the difference between the 2 pieces of code were simply that one was using a bind variable and the second was a hard coded value given as a test case. There are documented examples of how to simulate code in PLSQL with bind variables so that this issue does not actually occur, however the interesting thing was that without the bind variable the plan was just what the customer wanted.
The real answer to the above question would be to look at dynamic sampling / statistics gathering etc to ensure nothing is being influenced due to the bind variables (such as bind variable peeking or histograms etc), however a really simple solution was how to get the explain plan of SQL A over to SQL B with no actual code change.
There is a blog entry already showing how to introduce hints and also parallelism when a code change is not possible: https://blogs.oracle.com/Bakers_Byte/entry/applying_a_hint_to_sql
However this is a test case where we simply want to link the execution plan of SQL A to that of SQL B.
Please Note these scripts were written only for demonstration purposes. They are not optimized and they have almost no error checking, so be careful!
- The first thing required is to know the SQL_ID for the SQL that you wish to change the explain plan. (poor performer)
- The Second thing required to know is the SQL_ID and Plan_Hash_Value for the SQL that has the desired execution plan.
Then it is simply a matter of running the following PL/SQL block:
The above once executed created a baseline entry in dba_sql_baselines. By executing the following query this could be seen :
SQLPLUS > select fixed,accepted,enabled,created,sql_text,sql_handle from dba_sql_plan_baselines;
Other useful baseline queries are:
SQLPLUS> select * from table(dbms_xplan.display_cursor(‘&1’,&2));
NOTE: above needs &1 set to the sql_id and &2 the child number. At the end of the explain plan you can see if the baseline has been used or not.
SQLPLUS> select * from table(dbms_xplan.display_sql_plan_baseline(sql_handle=>’xxxxxx’, format=>’basic’));
NOTE: xxxxxx is the sql_handle from dba_sql_plan_baselines.
As it could be seen the baseline was taken and being used so the final step was also to fix the baseline:
If the baseline is not applicable anymore or did not work then it is easy to tidy up and remove it as below:
Conclusion: Although the long term solution is still to review the code and database to see really why the bind variable execution was giving a considerably poorer execution plan to that of the hard coded SQL. The issue was at least resolved with a nice work around due to being able to use a baseline and influence the explain plan as desired.