Thank you for visiting! If you liked the site, please add a bookmark, else add a critical comment! Would you like to visit Greece? Traditional Greek flag This real estate site is only available in Greek! The holy mountain

Sunday, 27 March 2011

A handy way to tune one's SQL

Recently, a french project manager working in a 5 year long European commission project in Brussels, contacted me and complained that their fully working code (in testing  environment), using Oracle Text 10g, failed to meet the customer expectations. In  their RAC production environment the application was unbearably slow. The project deadline is early June this year. Not an unknown situation, isn't it?
One of the most usual issues when moving from the development, or testing  environment to  the production, is setting up properly or updating the database schema and tuning it. In this new post, a fast and handy way for web developers to create, or even drop, indexes, partition tables or materialized views will be demonstrated. No SQL plans or other hard core DBA stuff will be mentioned here, the SQL of the notorious TUHRA ADF model project will be used as an example. So, let's start!
 Consider for instance the following trivial ANSI query for the employees search page based on the email, first and last name criteria entered by the user:


SELECT /* + INDEX_COMBINE( Employees Departments )*/
Employees.EMPLOYEE_ID,
Employees.EMAIL,
Employees.FIRST_NAME,
Employees.LAST_NAME,
Employees.PHONE_NUMBER,
Departments.DEPARTMENT_NAME,
Departments.DEPARTMENT_ID,
Jobs.JOB_TITLE,
Jobs.JOB_ID
FROM EMPLOYEES Employees join DEPARTMENTS Departments on (Employees.DEPARTMENT_ID = Departments.DEPARTMENT_ID)
join JOBS Jobs on (Employees.JOB_ID = Jobs.JOB_ID)
ORDER BY LAST_NAME,FIRST_NAME



When one opens the relevant view object and observes the SQL  query to explain, one sees the image presented below:


The highlighted code below, the  where clause, is actually missing. You need  to enable ADFLogger logging, run the app and search for the actual full query in the output log which is rather time consuming. Besides, when you open sqlplus and enter:

Explain plan for

SELECT * FROM (SELECT /* + INDEX_COMBINE( Employees Departments )*/
Employees.EMPLOYEE_ID,
Employees.EMAIL,
Employees.FIRST_NAME,
Employees.LAST_NAME,
Employees.PHONE_NUMBER,
Departments.DEPARTMENT_NAME,
Departments.DEPARTMENT_ID,
Jobs.JOB_TITLE,
Jobs.JOB_ID
FROM EMPLOYEES Employees join DEPARTMENTS Departments on (Employees.DEPARTMENT_ID = Departments.DEPARTMENT_ID)
join JOBS Jobs on (Employees.JOB_ID = Jobs.JOB_ID)) QRSLT
WHERE ( ( (UPPER(EMAIL) LIKE UPPER('%' || :QueryEmail || '%') ) 
AND (UPPER(FIRST_NAME) LIKE UPPER('%' || :QueryFirstName || '%') )
AND (UPPER(LAST_NAME) LIKE UPPER('%' || :QueryLastName || '%') ) ) )
ORDER BY LAST_NAME,FIRST_NAME

select *
from table(dbms_xplan.display)
/

you get a new plan, which is naturally different. Many people would rightfully argue that creating a function based index, concatenating all the three columns in that order, would solve the tuning problem. But would that really solve it?

To index or not to index? That is the question!

The easiest way to fly is via the automatic pilot. You test fully your app, ideally in isolation and now you  can ask the SQL advisor, thanks to the enterprise manager, following the images below:


 Click on Advisor central to proceed. Next click on sql access advisor in order to create a db job and schedule it to run immediately. Next click on current and recent sql activity, proceed  to select indexes or materialized views  or partition tables, reach the review page where show sql, reveals what is happening under the cover:



DECLARE

taskname
varchar2(30) := 'SQLACCESSTuhra3';
task_desc
varchar2(256) := 'SQL Access Advisor';
task_or_template
varchar2(30) := 'SQLACCESS_EMTASK';
task_id
number := 0;
wkld_name
varchar2(30) := 'SQLACCESSTuhra3_wkld';
saved_rows
number := 0;
failed_rows
number := 0;
num_found
number;
BEGIN
/* Create Task */
dbms_advisor.create_task(DBMS_ADVISOR.SQLACCESS_ADVISOR,task_id,taskname,task_desc,task_or_template);
/* Reset Task */
dbms_advisor.reset_task(taskname);
/* Delete Previous Workload Task Link */
select count(*) into num_found from user_advisor_sqla_wk_map where task_name = taskname and workload_name = wkld_name;IF num_found > 0 THEN
dbms_advisor.delete_sqlwkld_ref(taskname,wkld_name);
END IF;/* Delete Previous Workload */
select count(*) into num_found from user_advisor_sqlw_sum where workload_name = wkld_name;IF num_found > 0 THEN
dbms_advisor.delete_sqlwkld(wkld_name);
END IF;/* Create Workload */
dbms_advisor.create_sqlwkld(wkld_name,
null);/* Link Workload to Task */
dbms_advisor.add_sqlwkld_ref(taskname,wkld_name);
/* Set Workload Parameters */
dbms_advisor.set_sqlwkld_parameter(wkld_name,
'VALID_ACTION_LIST',DBMS_ADVISOR.ADVISOR_UNUSED);
dbms_advisor.set_sqlwkld_parameter(wkld_name,
'VALID_MODULE_LIST',DBMS_ADVISOR.ADVISOR_UNUSED);
dbms_advisor.set_sqlwkld_parameter(wkld_name,
'SQL_LIMIT',DBMS_ADVISOR.ADVISOR_UNLIMITED);
dbms_advisor.set_sqlwkld_parameter(wkld_name,
'VALID_USERNAME_LIST',DBMS_ADVISOR.ADVISOR_UNUSED);
dbms_advisor.set_sqlwkld_parameter(wkld_name,
'VALID_TABLE_LIST',DBMS_ADVISOR.ADVISOR_UNUSED);
dbms_advisor.set_sqlwkld_parameter(wkld_name,
'INVALID_TABLE_LIST',DBMS_ADVISOR.ADVISOR_UNUSED);
dbms_advisor.set_sqlwkld_parameter(wkld_name,
'ORDER_LIST','PRIORITY,OPTIMIZER_COST');
dbms_advisor.set_sqlwkld_parameter(wkld_name,
'INVALID_ACTION_LIST',DBMS_ADVISOR.ADVISOR_UNUSED);
dbms_advisor.set_sqlwkld_parameter(wkld_name,
'INVALID_USERNAME_LIST',DBMS_ADVISOR.ADVISOR_UNUSED);
dbms_advisor.set_sqlwkld_parameter(wkld_name,
'INVALID_MODULE_LIST',DBMS_ADVISOR.ADVISOR_UNUSED);
dbms_advisor.set_sqlwkld_parameter(wkld_name,
'VALID_SQLSTRING_LIST',DBMS_ADVISOR.ADVISOR_UNUSED);
dbms_advisor.set_sqlwkld_parameter(wkld_name,
'INVALID_SQLSTRING_LIST','"@!"');
dbms_advisor.set_sqlwkld_parameter(wkld_name,
'JOURNALING','4');
dbms_advisor.set_sqlwkld_parameter(wkld_name,
'DAYS_TO_EXPIRE','30');
dbms_advisor.import_sqlwkld_sqlcache(wkld_name,
'REPLACE',2,saved_rows,failed_rows);/* Set Task Parameters */
dbms_advisor.set_task_parameter(taskname,
'ANALYSIS_SCOPE','INDEX,TABLE,PARTITION');
dbms_advisor.set_task_parameter(taskname,
'RANKING_MEASURE','PRIORITY,OPTIMIZER_COST');
dbms_advisor.set_task_parameter(taskname,
'DEF_PARTITION_TABLESPACE',DBMS_ADVISOR.ADVISOR_UNUSED);
dbms_advisor.set_task_parameter(taskname,
'TIME_LIMIT',10000);
dbms_advisor.set_task_parameter(taskname,
'MODE','LIMITED');
dbms_advisor.set_task_parameter(taskname,
'STORAGE_CHANGE',DBMS_ADVISOR.ADVISOR_UNLIMITED);
dbms_advisor.set_task_parameter(taskname,
'DML_VOLATILITY','TRUE');
dbms_advisor.set_task_parameter(taskname,
'WORKLOAD_SCOPE','PARTIAL');
dbms_advisor.set_task_parameter(taskname,
'DEF_INDEX_TABLESPACE',DBMS_ADVISOR.ADVISOR_UNUSED);
dbms_advisor.set_task_parameter(taskname,
'DEF_INDEX_OWNER',DBMS_ADVISOR.ADVISOR_UNUSED);
dbms_advisor.set_task_parameter(taskname,
'DEF_MVIEW_TABLESPACE',DBMS_ADVISOR.ADVISOR_UNUSED);
dbms_advisor.set_task_parameter(taskname,
'DEF_MVIEW_OWNER',DBMS_ADVISOR.ADVISOR_UNUSED);
dbms_advisor.set_task_parameter(taskname,
'DEF_MVLOG_TABLESPACE',DBMS_ADVISOR.ADVISOR_UNUSED);
dbms_advisor.set_task_parameter(taskname,
'CREATION_COST','TRUE');
dbms_advisor.set_task_parameter(taskname,
'JOURNALING','4');
dbms_advisor.set_task_parameter(taskname,
'DAYS_TO_EXPIRE','30');/* Execute Task */
dbms_advisor.execute_task(taskname);
END;



You end up as shown:


 where one can see the whole set of the sql statements issued. If you click on schedule implementation the recommended indexes are created, while clicking show sql gives you the script to run when you wish, on your own:




Rem SQL Access Advisor: Version 11.2.0.1.0 - Production
Rem

Rem
Username: SYSTEM
Rem
Task: SQLACCESSTuhra3
Rem
Execution date:
Rem

/* RETAIN INDEX "HR"."IMAGE_USAGES_PK" */
/* RETAIN INDEX "HR"."DEPT_ID_PK" */
/* RETAIN INDEX "HR"."JOB_ID_PK" */
/* RETAIN INDEX "HR"."IMAGE_USAGES_ASSOCIATED_IDX" */
/* RETAIN INDEX "HR"."JOBS_TITLE_IDX" */
/* RETAIN INDEX "HR"."EMP_DEPARTMENT_IX" */
/* RETAIN INDEX "HR"."EMP_EMAIL_UK" */
/* RETAIN INDEX "HR"."IMAGES_PK" */
/* RETAIN INDEX "APEX_030200"."WWV_FLOW_PLATFORM_PREFS_FKIDX" */
/* RETAIN INDEX "HR"."EMP_EMP_ID_PK" */
CREATE INDEX "HR"."EMPLOYEES_NAME_IDX"ON "HR"."EMPLOYEES"
("LAST_NAME","FIRST_NAME")
COMPUTE STATISTICS;





Thus, the aforementioned concatenated index is not being used, and is just decelerating inserts and taking up  valuable disk space.We can safely drop it for now. This does not mean that we are done! Since the production db will have millions of rows, other statistics,etc, the db job results, i.e.  about indexing, may frequently change. The morale of this story is that the earliest one starts to test against the full production data,  the higher the possibility of project success. Performance tuning is not an afterthought, but a gradually ongoing process, which must ideally begin with the first application release. Nonetheless, not many developers follow this rule. Cynicals say that experience is just a set of past, bitter failures. Germans say: Durch Schaden wird mann klug. The man on the street tends to forget more easily, whatever has been unpleasant in the past.  It seems that IT staff does not have that privilege!












No comments:

Post a Comment