Analytics ioug 2011

Analyzing Your Data with Analytic Functions

Carl Dudley
University of Wolverhampton, UK

UKOUG Council
Oracle ACE Director

carl.dudley@wlv.ac.uk

Introduction

Working with Oracle since 1986
Oracle DBA - OCP Oracle7, 8, 9, 10
Oracle DBA of the Year – 2002
Oracle ACE Director
Regular Presenter at Oracle Conferences
Consultant and Trainer
Technical Editor for a number of Oracle texts
UK Oracle User Group Council
Member of IOUC
Day job – University of Wolverhampton, UK

Carl Dudley University of Wolverhampton, UK
2

Analyzing Your Data with Analytic Functions

Overview of Analytic Functions

Ranking Functions

Partitioning

Aggregate Functions

Sliding Windows

Row Comparison Functions

Analytic Function Performance

3

Analytic Functions

 New set of functions introduced in Oracle 8.1.6
— Analytic functions or Window functions
 Intended for OLAP (OnLine Analytic Processing) or data warehouse purposes

 Provide functionality that would require complex conventional SQL
programming or other tools
 Advantages
— Improved performance
• The optimizer “understands” the purpose of the query
— Reduced dependency on report generators and client tools
— Simpler coding

4

Analytic Function Categories

 The analytic functions fall into four categories
Ranking functions
Aggregate functions
Row comparison functions
Statistical functions
 The Oracle documentation describes all of the functions

 Processed as the last step before ORDER BY
— Work on the result set of the query
— Can operate on an intermediate ordering of the rows
— Actions can be based on :
• Partitions of the result set
• A sliding window of rows in the result set

5

Processing Sequence

 There may be several intermediate sort steps if required

Analytic process

WHERE HAVING Intermediate
Rows GROUPING
evaluation evaluation ordering

Analytic
function

Final
ORDER BY Output

6

The Analytic Clause

 Syntax :

<function>(<arguments>) OVER(<analytic clause>)
 The enclosing parentheses are required even if there are no arguments

RANK() OVER (ORDER BY sal DESC)

7

Sequence of Processing

 Being processed just before the final ORDER BY means :
— Analytic functions are not allowed in WHERE and HAVING conditions
• Allowed only in the final ORDER BY clause
 Ordering the final result set
— OVER clause specifies sort order of result set before analytic function is
computed
— Can have multiple analytic functions with different OVER clauses, requiring
multiple intermediate sorts
— Final ordering does not have to match ordering in OVER clause

8

The emp and dept Tables
Analytic Functions
DEPTNO DNAME LOC
emp ------ -------------- --------
10 ACCOUNTING NEW YORK
20 RESEARCH DALLAS
30 SALES Overview of Analytic Functions
CHICAGO
40 OPERATIONS BOSTON
Ranking Functions
EMPNO ENAME JOB MGR HIREDATE SAL COMM DEPTNO
dept ----- ------- --------- ----- ----------- ----- ----- ------
7934 MILLER Partitioning 7782 23-JAN-1982 1300
CLERK 10
7782 CLARK MANAGER 7839 09-JUN-1981 2450 10
7839 KING Aggregate Functions
PRESIDENT 17-NOV-1981 5000 10
7369 SMITH CLERK 7902 17-DEC-1980 800 20
7876 ADAMS
Sliding Windows 12-JAN-1983 1100
CLERK 7788 20
7566 JONES MANAGER 7839 02-APR-1981 2975 20
7902 FORD ANALYST 7566 03-DEC-1981 3000 20
7788 SCOTT Row Comparison Functions
ANALYST 7566 09-DEC-1982 3000 20
7900 JAMES CLERK 7698 03-DEC-1981 950 30
7521 WARD Analytic Function Performance 1250
SALESMAN 7698 22-FEB-1981 500 30
7654 MARTIN SALESMAN 7698 28-SEP-1981 1250 1400 30
7844 TURNER SALESMAN 7698 08-SEP-1981 1500 0 30
7499 ALLEN SALESMAN 7698 20-FEB-1981 1600 300 30
7698 BLAKE MANAGER 7839 01-MAY-1981 2850 30
9

Example of Ranking

 Ranking with ROW_NUMBER
— No handling of ties
• Rows retrieved by the query are intermediately sorted on descending
salary for the analysis
SELECT ROW_NUMBER() OVER( ROWNUMBER SAL ENAME
--------- ---- -----
ORDER BY sal DESC) rownumber 1 5000 KING
,sal 2 3000 SCOTT
,ename 3 3000 FORD
FROM emp 4 2975 JONES
5 2850 BLAKE
ORDER BY sal DESC; 6 2450 CLARK
7 1600 ALLEN
— If the final ORDER BY specifies the same sort 8 1500 TURNER
order as the OVER clause only one sort is required 9 1300 MILLER
— ROW_NUMBER is different from ROWNUM 10 1250 WARD
11 1250 MARTIN
12 1100 ADAMS
13 950 JAMES
14 800 SMITH

10

Different Sort Order in Final ORDER BY

 If the OVER clause sort is different from the final ORDER BY
— An extra sort step is required
SELECT ROW_NUMBER() OVER( ROWNUMBER SAL ENAME
--------- ---- ------
ORDER BY sal DESC) rownumber 12 1100 ADAMS
,sal 7 1600 ALLEN
,ename 5 2850 BLAKE
FROM emp 6 2450 CLARK
ORDER BY ename; 3 3000 FORD
13 950 JAMES
4 2975 JONES
1 5000 KING
11 1250 MARTIN
9 1300 MILLER
2 3000 SCOTT
14 800 SMITH
8 1500 TURNER
10 1250 WARD

11

Multiple Functions With Different Sort Order

 Multiple OVER clauses can be used

SELECT ROW_NUMBER() OVER(ORDER BY sal DESC) sal_n
,sal
,ROW_NUMBER() OVER(ORDER BY comm DESC NULLS LAST) comm_n
,comm
,ename
FROM emp
ORDER BY ename;

12

RANK and DENSE_RANK

 ROW_NUMBER increases even if several rows have identical values
— Does not handle ties
 RANK and DENSE_RANK handle ties
— Rows with the same value are given the same rank
— After the tie value, RANK skips numbers, DENSE_RANK does not
 Ranking using analytic functions has better performance, because the
table is not read repeatedly

13

RANK and DENSE_RANK (continued)

SELECT ROW_NUMBER() OVER(ORDER BY sal DESC) rownumber
,RANK() OVER(ORDER BY sal DESC) rank
,DENSE_RANK() OVER(ORDER BY sal DESC) denserank
,sal
,ename
FROM emp
ORDER BY sal DESC,ename;
ROWNUMBER RANK DENSERANK SAL ENAME
--------- ---- ---------- ----- ------
1 1 1 5000 KING Multiple OVER clauses may be
2 2 2 3000 FORD used specifying different orderings
3 2 2 3000 SCOTT
4 4 3 2975 JONES
5 5 4 2850 BLAKE
6 6 5 2450 CLARK
7 7 6 1600 ALLEN
8 8 7 1500 TURNER
9 9 8 1300 MILLER
10 10 9 1250 MARTIN
11 10 9 1250 WARD
12 12 10 1100 ADAMS
13 13 11 950 JAMES
14 14 12 800 SMITH

14

Analytic Function in ORDER BY

 Analytic functions are computed before the final ordering
— Can be referenced in the final ORDER BY clause
— An alias is used in this case

SELECT RANK() OVER( SAL_RANK SAL ENAME
ORDER BY sal DESC) sal_rank -------- ---- ------
,sal 1 5000 KING
,ename 2 3000 FORD
FROM emp 2 3000 SCOTT
ORDER BY sal_rank 4 2975 JONES
,ename; 5 2850 BLAKE
6 2450 CLARK
7 1600 ALLEN
8 1500 TURNER
9 1300 MILLER
10 1250 MARTIN
10 1250 WARD
12 1100 ADAMS
13 950 JAMES
14 800 SMITH 15

WHERE Conditions

 Analytic (window) functions are computed after the WHERE condition and
hence not available in the WHERE clause

SELECT RANK() OVER(ORDER BY sal DESC) rank
,sal
,ename
FROM emp
WHERE RANK() OVER(ORDER BY sal DESC) <= 5
ORDER BY rank

WHERE RANK() OVER(ORDER BY sal DESC) <= 5
*
ERROR at line 5:
ORA-30483: window functions are not allowed here

16

WHERE Conditions
(continued)

 Use an inline view to force the early processing of the analytic
SELECT *
FROM (SELECT RANK() OVER(ORDER BY sal DESC) rank
,sal
,ename
FROM emp)
WHERE rank <= 5
ORDER BY rank
,ename;

RANK SAL ENAME
---------- ---------- ----------
1 5000 KING
2 3000 FORD
2 3000 SCOTT
4 2975 JONES
5 2850 BLAKE
— Inline view is processed before the WHERE clause

17

Grouping, Aggregate Functions and Analytics

 Rank the departments by number of employees
SELECT deptno
,COUNT(*) employees
,RANK() OVER(ORDER BY COUNT(*) DESC) rank
FROM emp
GROUP BY deptno
ORDER BY employees
,deptno;

DEPTNO EMPLOYEES RANK
------ ---------- ---------
10 3 3
20 5 2
30 6 1
 Analytic functions are illegal in the HAVING clause
— The workaround is the same; use an inline view
— Ordering subclause may not reference a column alias

18

Analytic Functions


Ranking Functions

Partitioning

Aggregate Functions

Sliding Windows



19

Partitioning

 Analytic functions can be applied to logical groups within the result set
rather than the full result set
— Partitions
... OVER(PARTITION BY mgr ORDER BY sal DESC)

— PARTITION BY specifies the grouping
— ORDER BY specifies the ordering within each group
— Not connected with database table partitioning
 If partitioning is not specified, the full result set behaves as one partition
 NULL values are grouped together in one partition, as in GROUP BY

 Can have multiple analytic functions with different partitioning subclauses

20

Partitioning Example
 Rank employees by salary within their manager
SELECT ename
,mgr
,sal
,RANK() OVER(PARTITION BY mgr ORDER BY sal DESC)
m_rank
FROM emp
ORDER BY mgr
,m_rank;
ENAME MGR SAL M_RANK
---------- ---------- ---------- ----------
SCOTT 7566 3000 1
FORD 7566 3000 1
ALLEN 7698 1600 1
TURNER 7698 1500 2
WARD 7698 1250 3
MARTIN 7698 1250 3
JAMES 7698 950 5
MILLER 7782 1300 1
ADAMS 7788 1100 1
JONES 7839 2975 1
BLAKE 7839 2850 2
CLARK 7839 2450 3
SMITH 7902 800 1
KING 5000 1
21

Result Sets With Different Partitioning

 Rank the employees by salary within their manager, within the year they
were hired, as well as overall

SELECT ename
,sal
,manager
,RANK()
OVER(PARTITION BY mgr
ORDER BY sal DESC) m_rank
,TRUNC(TO_NUMBER(TO_CHAR(date_hired,'YYYY'))) year_hired
,RANK()
OVER(PARTITION BY TRUNC(TO_NUMBER(TO_CHAR(date_hired,'YYYY'))
ORDER BY sal DESC) d_rank
,RANK() OVER(ORDER BY sal DESC) rank
FROM emp
ORDER BY rank
,ename;

22

Result Sets With Different Partitioning (continued)

ENAME SAL MGR M_RANK YEAR_HIRED D_RANK RANK
------- ---- ---- ---------- ---------- ---------- ----------
KING 5000 1 1981 1 1
FORD 3000 7566 1 1981 2 2
SCOTT 3000 7566 1 1987 1 2
JONES 2975 7839 1 1981 3 4
BLAKE 2850 7839 2 1981 4 5
CLARK 2450 7839 3 1981 5 6
ALLEN 1600 7698 1 1981 6 7
TURNER 1500 7698 2 1981 7 8
MILLER 1300 7782 1 1982 1 9
MARTIN 1250 7698 3 1981 8 10
WARD 1250 7698 3 1981 8 10
ADAMS 1100 7788 1 1987 2 12
JAMES 950 7698 5 1981 10 13
SMITH 800 7902 1 1980 1 14

23

Hypothetical Rank

 Rank a specified hypothetical value (2999) in a group ('what-if' query)
SELECT RANK(2999) WITHIN GROUP (ORDER BY sal DESC) H_S_rank
,PERCENT_RANK(2999) WITHIN GROUP (ORDER BY sal DESC) PR
,CUME_DIST(2999) WITHIN GROUP (ORDER BY sal DESC) CD
FROM emp;
H_S_RANK PR CD
-------- ---------- ----------
4 .214285714 .266666667
3/14 4/15

SELECT deptno
,RANK(20,'CLERK') WITHIN GROUP
(ORDER BY deptno DESC,job ASC) H_D_J_rank
FROM emp
GROUP BY deptno; A clerk in 20 would be higher than anyone in 10
DEPTNO H_D_J_RANK A clerk would be third in ascending job
------ ----------
10 1 order in department 20 (below analysts)
20 3
A clerk in 20 would be lower than anyone in 30 (6 employees)
30 7
24

Frequent Itemsets (dbms_frequent_itemset)

 Typical question
— When a customer buys product x, how likely are they to also buy product y?
SELECT CAST(itemset AS fi_char) itemset
,support
,length
,total_tranx
FROM TABLE(DBMS_FREQUENT_ITEMSET.FI_TRANSACTIONAL(
CURSOR(SELECT TO_CHAR(sales.cust_id)
Minimum fraction of different ,TO_CHAR(sales.prod_id)
FROM sh.sales
'Documentation' customers ,sh.products
having this combination WHERE products.prod_id = sales.prod_id
AND products.prod_subcategory = 'Documentation'),
0.5,
include items 2, mimimum items in set
3, Number of
NULL, maximum items in set Different customers
exclude items
NULL));

ITEMSET SUPPORT LENGTH TOTAL_TRANX
-------------------------------------- --------- ---------- -----------
FI_CHAR('40', '41') 3692 2 6077
FI_CHAR('40', '42') 2 or 3 items per set 3900 2 6077
FI_CHAR('40', '45') 3482 2 6077
FI_CHAR('41', '42') Number of instances 3163 2 6077
FI_CHAR('40', '41', '42') 3141 3 6077

25

Frequent Itemsets (continued)

 Need to create type to accommodate the set
— Ranking functions can AS TABLE OF itemset
CREATE TYPE fi_char be applied to theVARCHAR2(100);
 The total transactions (TOTAL_TRANX) is the number of different customers
involved with any product within the set of products under examination
SELECT COUNT(DISTINCT cust_id)
FROM sales prod_ids for
WHERE prod_id BETWEEN 40 AND 45; 'Documentation'
COUNT(DISTINCTCUST_ID)
----------------------
6077
— Ranking functions can be applied to the itemset
 Itemsets containing certain items can be included/excluded
,CURSOR(SELECT * FROM table(fi_char(40,45)))
Include any sets
,CURSOR(SELECT * FROM table(fi_char(42))) involving 40 or 45

Exclude any sets
involving 42

26

Plan of Itemset Query

 Only one full table scan of sales
--------------------------------------------------------------------------------
|Id | Operation | Name |Rows |
--------------------------------------------------------------------------------
| 0| SELECT STATEMENT | | 8|
| 1| FIC RECURSIVE ITERATION | | |
| 2| FIC LOAD ITEMSETS | | |
| 3| FREQUENT ITEMSET COUNTING | | 8|
| 4| SORT GROUP BY NOSORT | | |
| 5| BITMAP CONVERSION COUNT | | |
| 6| FIC LOAD BITMAPS | | |
| 7| SORT CREATE INDEX | | 500|
| 8| BITMAP CONSTRUCTION | | |
| 9| FIC ENUMERATE FEED | | |
| 10| SORT ORDER BY | |43755|
|*11| HASH JOIN | |43755|
| 12| TABLE ACCESS BY INDEX ROWID| PRODUCTS | 3 |
|*13| INDEX RANGE SCAN | PRODUCTS_PROD_SUBCAT_IX | 3 |
| 14| PARTITION RANGE ALL | | 918K|
| 15| TABLE ACCESS FULL | SALES | 918K|
| 16| TABLE ACCESS FULL | SYS_TEMP_0FD9D6605_153B1EE| |
--------------------------------------------------------------------------------

27

Applying Analytics to Frequent Itemsets

SELECT itemset, support, length, total_tranx, rnk
FROM (SELECT itemset, support, length, total_tranx
,RANK() OVER (PARTITION BY length ORDER BY support DESC) rnk
FROM (SELECT CAST(ITEMSET AS fi_char) itemset
,support
,length
,total_tranx
FROM TABLE(dbms_frequent_itemset.fi_transactional
(CURSOR(SELECT TO_CHAR(sales.cust_id)
,TO_CHAR(sales.prod_id)
FROM sh.sales
,sh.products
WHERE products.prod_id = sales.prod_id
AND products.prod_subcategory = 'Documentation')
,0.5
,2
,3
,NULL
,NULL))))
WHERE rnk < 4;

ITEMSET SUPPORT LENGTH TOTAL_TRANX RNK
-------------------------------- ---------- ---------- ----------- ----------
FI_CHAR('40', '42') 3900 2 6077 1
FI_CHAR('40', '41') 3692 2 6077 2
FI_CHAR('40', '45') 3482 2 6077 3
FI_CHAR('40', '41', '42') 3141 3 6077 1

28

Analytic Functions


Ranking Functions

Partitioning

Aggregate Functions

Sliding Windows



29

Expanding Windows

Partition (first) or entire result set OVER (ORDER BY col_name)
ROWS BETWEEN UNBOUNDED
Window PRECEDING AND CURRENT ROW

Default value for window setting -
produces an expanding window

Partition (second)

Sliding Windows

Partition (first) or entire result set
OVER (ORDER BY col_name)
ROWS BETWEEN 2 PRECEDING
3 ROWS Window
5 ROWS AND 2 FOLLOWING

Produces a sliding window

Partition (second)

Aggregate Functions

 Aggregate functions can be used as analytic functions
— Must be embedded in the OVER clause
 Analytic aggregate values can be easily included within row-level reports
— Analytic functions are applied after computation of result set
— Optimizer often produces a better execution plan
 Aggregate level is determined by the partitioning subclause
— Similar effect to GROUP BY clause
— If no partitioning subclause, aggregate is across the complete result set

32

Aggregate Functions – the OVER Clause

SELECT deptno SELECT deptno
,AVG(sal) ,AVG(sal) OVER (PARTITION BY deptno) avg_dept
FROM emp ,AVG(sal) OVER () avg_all
GROUP BY deptno; FROM emp;

DEPTNO AVG(SAL) DEPTNO AVG_DEPT AVG_ALL No subclause
---------- ---------- ---------- ---------- ----------
30 1566.66667 10 2916.66667 2073.21429
20 2175 10 2916.66667 2073.21429
10 2916.66667 10 2916.66667 2073.21429
20 2175 2073.21429
20 2175 2073.21429
20 2175 2073.21429
20 2175 2073.21429 Analytic aggregates
20 2175 2073.21429
30 1566.66667 2073.21429 cause no reduction
30 1566.66667 2073.21429 in rows
30 1566.66667 2073.21429
30 1566.66667 2073.21429
30 1566.66667 2073.21429
30 1566.66667 2073.21429

 Could easily include row-level data
— e.g. ename and sal

33

Analytic versus Conventional SQL Performance

Average sal
 The requirement per department
— Data at different levels of grouping
ENAME SAL DEPTNO AVG_DEPT AVG_ALL Overall
------ ---- ------ ---------- ---------- average sal
CLARK 2450 10 2916.66667 2073.21429
KING 5000 10 2916.66667 2073.21429
MILLER 1300 10 2916.66667 2073.21429
JONES 2975 20 2175 2073.21429
FORD 3000 20 2175 2073.21429
ADAMS 1100 20 2175 2073.21429
SMITH 800 20 2175 2073.21429
SCOTT 3000 20 2175 2073.21429
WARD 1250 30 1566.66667 2073.21429
TURNER 1500 30 1566.66667 2073.21429
ALLEN 1600 30 1566.66667 2073.21429
JAMES 950 30 1566.66667 2073.21429
BLAKE 2850 30 1566.66667 2073.21429
MARTIN 1250 30 1566.66667 2073.21429

34

Conventional SQL Performance

SELECT r.ename,r.sal,g.deptno,g.ave_dept,a.ave_all
FROM emp r
,(SELECT deptno,AVG(sal) ave_dept
FROM emp GROUP BY deptno) g
,(SELECT AVG(sal) ave_all
FROM emp) a
WHERE g.deptno = r.deptno
ORDER BY r.deptno;
-----------------------------------------------
| Id | Operation | Name | Rows |
-----------------------------------------------
| 0 | SELECT STATEMENT | | 15 |
| 1 | MERGE JOIN | | 15 | 1M row emp table :
| 2 | SORT JOIN | | 3 |
| 3 | NESTED LOOPS | | 3 | 48.35 seconds
| 4 | VIEW | | 1 |
| 5 | SORT AGGREGATE | | 1 | 230790 consistent gets
| 6 | TABLE ACCESS FULL| EMP | 14 |
| 7 | VIEW | | 3 |
| 8 | SORT GROUP BY | | 3 |
|* 10 | SORT JOIN | | 14 |
| 11 | TABLE ACCESS FULL | EMP | 14 |
-----------------------------------------------

35


SELECT ename,sal,deptno
,AVG(sal) OVER (PARTITION BY deptno) ave_dept
,AVG(sal) OVER () ave_all
FROM emp;

-------------------------------------------
------------------------------------------- 1M row emp table :
| 0 | SELECT STATEMENT | | 14 | 21.20 seconds
| 1 | WINDOW SORT | | 14 |
| 2 | TABLE ACCESS FULL| EMP | 14 | 76930 consistent gets
-------------------------------------------

36

Aggregating Over an Ordered Set of Rows –
Running Totals

 The ORDER BY clause creates an expanding window (running total) of rows
SELECT empno
,ename
,sal
,SUM(sal) OVER(ORDER BY empno) run_total
FROM emp5
ORDER BY empno;
EMPNO ENAME SAL RUN_TOTAL
----- ------ ---- ---------
7369 SMITH 800 800
7499 ALLEN 1600 2400
7521 WARD 1250 3650 -------------------------------
7566 JONES 2975 6625 |Id| Operation | Name|
-------------------------------
7654 MARTIN 1250 7875 | 0| SELECT STATEMENT | |
7698 BLAKE 2850 10725 | 1| WINDOW SORT | |
7782 CLARK 2450 13175 | 2| TABLE ACCESS FULL| EMP5|
7788 SCOTT 3000 16175 -------------------------------
7839 KING 5000 21175
7844 TURNER 1500 22675
7876 ADAMS 1100 23775 emp table of 5000 rows
7900 JAMES 950 24725 0.07 seconds
7902 FORD 3000 27725 33 consistent gets
7934 MILLER 1300 29025
: : : : No index necessary
37

Running Total With Conventional SQL (1)

 Self-join solution
SELECT e1.empno
,e1.sal
,SUM(e2.sal) 13.37 seconds
FROM emp5 e1, emp5 e2
WHERE e2.empno <= e1.empno 66 consistent gets
GROUP BY e1.empno, e1.sal
ORDER BY e1.empno;
-------------------------------------------------
| Id | Operation | Name |
-------------------------------------------------
| 0 | SELECT STATEMENT | |
| 1 | SORT GROUP BY | |
| 2 | MERGE JOIN | |
| 3 | SORT JOIN | |
| 4 | TABLE ACCESS BY INDEX ROWID| EMP5 |
| 5 | INDEX FULL SCAN | PK_EMP5|
|* 6 | SORT JOIN | |
| 7 | TABLE ACCESS FULL | EMP5 |
-------------------------------------------------

38

Running Total With Conventional SQL (2)

 Subquery in SELECT list solution – column expression
SELECT empno
,ename
,sal 4.62 seconds
,(SELECT SUM(sal) sumsal
FROM emp5 97948 consistent gets
WHERE empno <= b.empno) a
FROM emp5 b
ORDER BY empno;

-----------------------------------------------
| Id | Operation | Name |
-----------------------------------------------
| 0 | SELECT STATEMENT | |
| 1 | SORT AGGREGATE | |
| 2 | TABLE ACCESS BY INDEX ROWID| EMP5 |
|* 3 | INDEX RANGE SCAN | PK_EMP5|
| 4 | TABLE ACCESS BY INDEX ROWID | EMP5 |
| 5 | INDEX FULL SCAN | PK_EMP5|
-----------------------------------------------

39

Aggregate Functions With Partitioning

 Find average salary of employees within each manager
— Use PARTITION BY to specify the grouping
SELECT ename, mgr, sal
,ROUND(AVG(sal) OVER(PARTITION BY mgr)) avgsal
,sal - ROUND(AVG(sal) OVER(PARTITION BY mgr)) diff
FROM emp;

ENAME MGR SAL AVGSAL DIFF
---------- ------- ---------- ---------- ----------
SCOTT 7566 3000 3000 0
FORD 7566 3000 3000 0
ALLEN 7698 1600 1310 290
WARD 7698 1250 1310 -60
JAMES 7698 950 1310 -360
TURNER 7698 1500 1310 190
MARTIN 7698 1250 1310 -60
MILLER 7782 1300 1300 0
ADAMS 7788 1100 1100 0
JONES 7839 2975 2758 217
CLARK 7839 2450 2758 -308
BLAKE 7839 2850 2758 92
SMITH 7902 800 800 0
KING 5000 5000 0

40

Analytics on Aggregates

 Analytics are processed last

SELECT deptno
,SUM(sal)
,SUM(SUM(sal)) OVER () Totsal
,SUM(SUM(sal)) OVER (ORDER BY deptno) Runtot_deptno
,SUM(SUM(sal)) OVER (ORDER BY SUM(sal)) Runtot_sumsal
FROM emp
GROUP BY deptno
ORDER BY deptno;

DEPTNO SUM(SAL) TOTSAL RUNTOT_DEPTNO RUNTOT_SUMSAL
------ -------- ------ ------------- -------------
10 8750 29025 8750 8750
+ sum(20)
20 10875 29025 19625 29025 + sum(30)
+ sum(20)
+ sum(30)
30 9400 29025 29025 18150

41

Aggregate Functions and the WHERE clause

 Analytic functions are applied after production of the complete result set
— Rows excluded by the WHERE clause are not included in the aggregate value
 Include only employees whose name starts with a ‘S’ or ‘M’
— The average is now only for those rows starting with 'S' or 'M'
SELECT ename
,sal
,ROUND(AVG(sal) OVER()) avgsal
,sal - ROUND(AVG(sal) OVER()) diff
FROM emp
WHERE ename LIKE 'S%'
OR ename LIKE 'M%';

ENAME SAL AGSAL DIFF
------ ---- ----- -----
SMITH 800 1588 -788
MARTIN 1250 1588 338
SCOTT 3000 1588 1412
MILLER 1300 1588 -288

42

RATIO_TO_REPORT

 Each row’s fraction of total salary can easily be found when the total salary
value is available
— Example: sal/SUM(sal) OVER()
— The function RATIO_TO_REPORT performs this calculation

SELECT ename
,sal
,SUM(sal) OVER() sumsal
,sal/SUM(sal) OVER() ratio
,RATIO_TO_REPORT(sal) OVER() ratio_rep
FROM emp;

43

RATIO_TO_REPORT (continued)

 The query on the previous slide gives this result
ENAME SAL SUMSAL RATIO RATIO_REP
---------- ------- ---------- ---------- ----------
SMITH 800 29025 .027562446 .027562446
ALLEN 1600 29025 .055124892 .055124892
WARD 1250 29025 .043066322 .043066322
JONES 2975 29025 .102497847 .102497847
MARTIN 1250 29025 .043066322 .043066322
BLAKE 2850 29025 .098191214 .098191214
CLARK 2450 29025 .084409991 .084409991
SCOTT 3000 29025 .103359173 .103359173
KING 5000 29025 .172265289 .172265289
TURNER 1500 29025 .051679587 .051679587
ADAMS 1100 29025 .037898363 .037898363
JAMES 950 29025 .032730405 .032730405
FORD 3000 29025 .103359173 .103359173
MILLER 1300 29025 .044788975 .044788975

44

Analytic Functions


Ranking Functions

Partitioning

Aggregate Functions

Sliding Windows



45

Sliding Windows

 The OVER clause can have a sliding window subclause
— Not permitted without ORDER BY subclause
— Specifies size of window (set of rows) to be processed by the analytic
function
— Window defined relative to current row
• Slides through result set as different rows become current
 Size of window is governed by ROWS or RANGE
— ROWS
• physical offset, a number of rows relative to the current row
— RANGE
• logical offset, a value interval relative to value in current row
 Syntax for sliding window :
— BETWEEN <starting point> AND <ending point>

46

Sliding Windows Example

 For each employee, show the sum of the salaries of the preceding, current,
and following employee (row)
— Window includes current row as well as the preceding and following ones
— Must have order subclause for “preceding” and “following” to be meaningful
— First row has no preceding row and last row has no following row

SELECT ename
,sal
,SUM(sal) OVER(ORDER BY sal DESC
AND 1 FOLLOWING) sal_window
FROM emp
ORDER BY sal DESC
,ename;

47

Sliding Windows Example (continued)

ENAME SAL SAL_WINDOW Calculation:
---------- ---------- ----------
KING 5000 8000 =5000+3000
FORD 3000 11000 =5000+3000+3000
SCOTT 3000 8975 =3000+3000+2975
JONES 2975 8825 =3000+2975+2850
BLAKE 2850 8275 =2975+2850+2450
CLARK 2450 6900 =2850+2450+1600
ALLEN 1600 5550 =2450+1600+1500
TURNER 1500 4400 =1600+1500+1300
MILLER 1300 4050 =1500+1300+1250
MARTIN 1250 3800 =1300+1250+1250
WARD 1250 3600 =1250+1250+1100
ADAMS 1100 3300 =1250+1100+950
JAMES 950 2850 =1100+950+800
SMITH 800 1750 =950+800

48

Partitioned Sliding Windows

 Partitioning can be used with sliding windows
— A sliding window does not span partitions

SELECT ename
,job
,sal
,SUM(sal) OVER(PARTITION BY job
ORDER BY sal DESC
FROM emp
ORDER BY job
,sal DESC
,ename;

49

Partitioned Sliding Windows (continued)

ENAME JOB SAL
Calculation
SAL_WINDOW
---------- --------- ---------- =3000+3000
---------- =3000+3000
FORD ANALYST 3000
6000 =1300+1100
SCOTT ANALYST 3000 =1300+1100+950
6000 =1100+950+800
=950+800
MILLER CLERK 1300
2400 =2975+2850
ADAMS CLERK 1100 =2975+2850+2450
3350 =2850+2450
JAMES CLERK 950
=5000
2850
SMITH CLERK 800 =1600+1500
1750 =1600+1500+1250
=1500+1250+1250
JONES MANAGER 2975
=1250+1250
5825
BLAKE MANAGER 2850
8275 Carl Dudley University of Wolverhampton, UK
50

Sliding Window With Logical (RANGE) Offset

 Physical offset
— Specified number of rows

 Logical offset
— A RANGE of values
• Numeric or date
— Values in the ordering column indirectly determine number of rows in
window

SELECT ename
,sal
RANGE BETWEEN 150 PRECEDING
FROM emp
ORDER BY sal DESC
,ename;

51

Sliding Window With Logical (RANGE) Offset
(continued)

ENAME SAL SAL_WINDOW
---------- ---------- ----------
KING 5000 5000
FORD 3000 8975
SCOTT 3000 8975
JONES 2975 8975
Range for this row is
BLAKE 2850 11825
3000 to 2775
CLARK 2450 2450
ALLEN 1600 1600
TURNER 1500 3100
MILLER 1300 3800
MARTIN 1250 3800
WARD 1250 3800
ADAMS 1100 3600
JAMES 950 2050
SMITH 800 1750

52

UNBOUNDED and CURRENT ROW

 Sliding windows have starting and ending points
— BETWEEN <starting point> AND <ending point>
 Ways for specifying starting and ending points
— UNBOUNDED PRECEDING specifies the first row as starting point
— UNBOUNDED FOLLOWING specifies the last row as ending point
— CURRENT ROW specifies the current row
 Create a window that grows with each row in ename order
— The RANGE clause is not necessary if a running total is required (default)
SELECT ename
,sal
,SUM(sal) OVER(ORDER BY ename
RANGE BETWEEN UNBOUNDED PRECEDING
AND CURRENT ROW) run_total
FROM emp
ORDER BY ename;

53

Keywords UNBOUNDED and CURRENT ROW
(continued)

 Running Total
— Produced by default 'expanding' window when window not specified
ENAME SAL RUN_TOTAL Explanation:
---------- ---------- ----------
ADAMS 1100 1100 =1100
ALLEN 1600 2700 =1600+1100
BLAKE 2850 5550 =2700+2850
CLARK 2450 8000 =5550+2450
FORD 3000 11000 =8000+3000
JAMES 950 11950 =11000+950
JONES 2975 14925 =11950+2975
KING 5000 19925 =14925+5000
MARTIN 1250 21175 =19925+1250
MILLER 1300 22475 =21175+1300
SCOTT 3000 25475 =22475+3000
SMITH 800 26275 =25475+800
TURNER 1500 27775 =26275+1500
WARD 1250 29025 =27775+1250

54

Keywords UNBOUNDED and CURRENT ROW
(continued)

 Be aware of the subtle difference between RANGE and ROWS in this context
— Apparent only when adjacent rows have equal values

SELECT ename
,sal
ROWS BETWEEN UNBOUNDED PRECEDING
AND CURRENT ROW) row_tot
RANGE BETWEEN UNBOUNDED PRECEDING
AND CURRENT ROW)
range_tot
,SUM(sal) OVER(ORDER BY sal DESC) default_tot
FROM EMP
ORDER BY sal DESC
,ename;

55

Difference between ROWS and RANGE

 Ford and Scott fall within the same range - also applies to Martin and Ward
— For example Scott is included in range when the value for Ford is calculated

ENAME SAL ROW_TOT RANGE_TOT DEFAULT_TOT
---------- ---------- ---------- --------- -----------
KING 5000 5000 5000 5000
FORD 3000 8000 11000 11000
SCOTT 3000 11000 11000 11000
JONES 2975 13975 13975 13975
BLAKE 2850 16825 16825 16825
CLARK 2450 19275 19275 19275
ALLEN 1600 20875 20875 20875
TURNER 1500 22375 22375 22375
MILLER 1300 23675 23675 23675
MARTIN 1250 24925 26175 26175
WARD 1250 26175 26175 26175
ADAMS 1100 27275 27275 27275
JAMES 950 28225 28225 28225
SMITH 800 29025 29025 29025

56

Time Intervals

 Sliding windows are often based on time intervals

 Example: Compare the salary of each employee to the maximum and
minimum salaries of hirings made within three months of their own hiring
date

SELECT ename
,hiredate
,sal
,MIN(sal) OVER(ORDER BY hiredate
RANGE BETWEEN INTERVAL '3' MONTH PRECEDING
AND INTERVAL '3' MONTH FOLLOWING) min
,MAX(sal) OVER(ORDER BY hiredate
RANGE BETWEEN INTERVAL '3' MONTH PRECEDING
AND INTERVAL '3' MONTH FOLLOWING) max
FROM emp;

57

Time Intervals
(continued)

 Sliding time window

ENAME HIREDATE SAL MIN MAX
---------- --------- ---------- ---------- ----------
SMITH 17-DEC-80 800 800 1600
ALLEN 20-FEB-81 1600 800 2975
WARD 22-FEB-81 1250 800 2975
JONES 02-APR-81 2975 1250 2975
BLAKE 01-MAY-81 2850 1250 2975
CLARK 09-JUN-81 2450 1500 2975
TURNER 08-SEP-81 1500 950 5000
MARTIN 28-SEP-81 1250 950 5000
KING 17-NOV-81 5000 950 5000
JAMES 03-DEC-81 950 950 5000
FORD 03-DEC-81 3000 950 5000
MILLER 23-JAN-82 1300 950 5000
SCOTT 09-DEC-82 3000 1100 3000
ADAMS 12-JAN-83 1100 1100 3000

58

Analytic Functions


Ranking Functions

Partitioning

Aggregate Functions

Sliding Windows



59

LAG and LEAD Functions

 Useful for comparing values across rows
— Need to specify count of rows which separate target row from current row
• No need for self-join
— LAG provides access to a row at a given offset prior to the current position
— LEAD provides access to a row at a given offset after the current position
{LAG | LEAD} ( value_expr [, offset] [, default] )
OVER ( [query_partition_clause] order_by_clause )
— offset is an optional parameter and defaults to 1
— default is an optional parameter and is the value returned if offset falls
outside the bounds of the table or partition
• In this case, NULL will be returned if no default is specified

60

LAG/LEAD Simple Example

SELECT hiredate
,sal AS salary
,LAG(sal,1) OVER (ORDER BY hiredate) AS LAG1
,LEAD(sal,1) OVER (ORDER BY hiredate) AS LEAD1
FROM emp;
HIREDATE SALARY LAG1 LEAD1
--------- ---------- ---------- ----------
17-DEC-80 800 1600
20-FEB-81 1600 800 1250
22-FEB-81 1250 1600 2975
Comparison of salaries
02-APR-81 2975 1250 2850
01-MAY-81 2850 2975 2450 with those for nearest
09-JUN-81 2450 2850 1500 recruits in terms of
08-SEP-81 1500 2450 1250 proximity of hiredates
28-SEP-81 1250 1500 5000
17-NOV-81 5000 1250 950
03-DEC-81 950 5000 3000
03-DEC-81 3000 950 1300
23-JAN-82 1300 3000 3000
09-DEC-82 3000 1300 1100
12-JAN-83 1100 3000

61

FIRST_VALUE and LAST_VALUE

 Hold first or last value in a partition (based on ordering) as a start point
SELECT empno, deptno, hiredate
,FIRST_VALUE(hiredate)
OVER (PARTITION BY deptno ORDER BY hiredate) firstdate
,hiredate - FIRST_VALUE(hiredate)
OVER (PARTITION BY deptno ORDER BY hiredate) Day_Gap
FROM emp
EMPNO DEPTNOdeptno, Day_Gap; DAY_GAP
ORDER BY HIREDATE FIRSTDATE
----- ------ --------- --------- -------
7782 10 09-JUN-81 09-JUN-81 0 Days after hiring of first
7839 10 17-NOV-81 09-JUN-81 161 employee in this department
7934 10 23-JAN-82 09-JUN-81 228
7369 20 17-DEC-80 17-DEC-80 0
7566 20 02-APR-81 17-DEC-80 106
7902 20 03-DEC-81 17-DEC-80 351
7788 20 09-DEC-82 17-DEC-80 722
7876 20 12-JAN-83 17-DEC-80 756
Works with partitioning and
7499 30 20-FEB-81 20-FEB-81 0
7521 30 22-FEB-81 20-FEB-81 2 windowing subclauses
7698 30 01-MAY-81 20-FEB-81 70
7844 30 08-SEP-81 20-FEB-81 200
7654 30 28-SEP-81 20-FEB-81 220
7900 30 03-DEC-81 20-FEB-81 286

62

Influence of Window on LAST_VALUE

SELECT deptno,ename,sal
,LAST_VALUE(ename) OVER (PARTITION BY deptno
ORDER BY sal
ROWS BETWEEN UNBOUNDED PRECEDING
AND UNBOUNDED FOLLOWING) AS hsal1
,LAST_VALUE(ename) OVER (PARTITION BY deptno
ORDER BY sal) AS hsal2
FROM emp
ORDER BY deptno,sal;

DEPTNO ENAME SAL HSAL1 HSAL2
------ ------ ---- ---------- ---------- Last value in
10 MILLER 1300 KING MILLER expanding window
10 CLARK 2450 KING CLARK (based on range)
10 KING 5000 KING KING
20 SMITH 800 SCOTT SMITH
20 ADAMS 1100 SCOTT ADAMS
20 JONES 2975 SCOTT JONES
20 FORD 3000 SCOTT SCOTT
20 SCOTT 3000 SCOTT SCOTT
30 JAMES 950 BLAKE JAMES
30 MARTIN 1250 BLAKE WARD 63

Ignoring Nulls in First and Last Values

SELECT ename
,FIRST_VALUE (ename) OVER (PARTITION BY deptno
ORDER BY ename) fv
,LAST_VALUE (ename) OVER (PARTITION BY deptno
ORDER BY ename) lv
Highest value (1400) is
,comm 'kept' for null values
,FIRST_VALUE (comm) OVER (PARTITION BY deptno
ORDER BY comm) fv_comm
,LAST_VALUE (comm) OVER (PARTITION BY deptno
ORDER BY comm) lv_comm
,LAST_VALUE (comm IGNORE NULLS) OVER (PARTITION BY deptno
ORDER BY comm) lv_ignore
FROM emp
WHERE deptno = 30;

ENAME FV LV COMM FV_COMM LV_COMM LV_IGNORE
---------- ---------- ---------- ---------- ---------- ---------- ----------
ALLEN ALLEN ALLEN 300 0 300 300
BLAKE ALLEN BLAKE 0 1400
JAMES ALLEN JAMES 0 1400
MARTIN ALLEN MARTIN 1400 0 1400 1400
TURNER ALLEN TURNER 0 0 0 0
WARD ALLEN WARD 500 0 500 500

64

NTH_VALUE
SELECT deptno
SELECT deptno
,ename
,ename
,sal
,sal
,FIRST_VALUE(sal) OVER (PARTITION BY deptno
,FIRST_VALUE(sal) OVER (PARTITION sal deptno
ORDER BY BY DESC)
- NTH_VALUE(sal,2) FROMORDER BY sal (PARTITION BY deptno
FIRST OVER DESC)
- NTH_VALUE(sal,3) FROM FIRST OVER (PARTITION sal deptno t2_diff
ORDER BY BY DESC)
FROM emp; ORDER BY sal DESC) t2_diff
FROM emp;
DEPTNO ENAME SAL T2_DIFF
---------- ---------- ---- ------- Could use
10 KING
DEPTNO ENAME 5000SAL T2_DIFF
10 CLARK 2450 2550 FROM LAST
------ ---------- ---------- ----------
10 MILLER 1300 2550
10 KING SCOTT
20 5000
3000 0
0??
10 CLARK
20 FORD 2450
3000 0
10 MILLER
20 JONES 1300
2975 0 3700
20 ADAMS
20 SCOTT 1100
3000 0 Reports difference between first and
20 SMITH
20 FORD 800
3000 0 second member of each partition
30 BLAKE
20 JONES 2850
2975 25
30 ALLEN 1600 1250
20 ADAMS
30 TURNER 1100 1250
1500 25
20 SMITH
30 MARTIN 1250800 1250 25
30 BLAKE
30 WARD 2850 1250
1250
30 JAMES
30 ALLEN 1250
1600 1250
30 TURNER 1500 1350
30 MARTIN 1250 1350
65

LISTAGG Function

 Example - show columns in indexes in an ordered list

SELECT table_name
,index_name
,LISTAGG(column_name,’;’) WITHIN GROUP (
ORDER BY column_position) “Column List”
FROM user_ind_columns
GROUP BY table_name
,index_name;

TABLE_NAME INDEX_NAME Column List
------------ ------------------ -----------------------------
EMP EMP_PK EMPNO
PROJ_ASST SYS_C0011223 PROJNO;EMPNO;START_DATE
DEPT DEPT$DIVNO_DEPTNO DIVNO;DEPTNO

66

FIRST and LAST

SELECT empno
 Compare each employee's ,deptno
,TO_CHAR(hiredate,'YYYY') Hire_Yr
salary with the average ,sal
salary of the first year of ,TRUNC(AVG(sal) KEEP (DENSE_RANK FIRST
ORDER BY TO_CHAR(hiredate,'YYYY') )
hirings of their department OVER (PARTITION BY deptno)) Avg_Sal_Yr1_Hire
FROM emp
— Must use KEEP ORDER BY deptno, empno, Hire_Yr;
— Must use DENSE_RANK EMPNO DEPTNO HIRE_YR SAL AVG_SAL_YR1_HIRE
----- ---------- ------- ------- ----------------
7782 10 1981 2450 3725
7839 10 1981 5000 3725
7934 10 1982 1300 3725

7369 20 1980 800 800
7566 20 1981 2975 800
7788 20 1982 3000 800
7876 20 1983 1100 800
7902 20 1981 3000 800

7499 30 1981 1600 1566
7521 30 1981 1250 1566
7654 30 1981 1250 1566
7698 30 1981 2850 1566
7844 30 1981 1500 1566
7900 30 1981 950 1566

67

FIRST and LAST (Continued)

 Compare salaries to the average SELECT empno
of the 'LAST' department ,deptno
,TO_CHAR(hiredate,'YYYY') Hire_Yr
— Note no ORDER BY inside the ,sal
,TRUNC(AVG(sal) KEEP (DENSE_RANK LAST
OVER clause ORDER BY deptno )
— No support for any OVER () ) AVG_SAL_LAST_DEPT
FROM emp
<window> clause ORDER BY deptno, empno, Hire_Yr;

EMPNO DEPTNO Hire_Yr SAL AVG_SAL_LAST_DEPT
----- ------ ------- ---- -----------------
7782 10 1981 2450 1566
7839 10 1981 5000 1566
7934 10 1982 1300 1566
7369 20 1980 800 1566
7566 20 1981 2975 1566
7788 20 1982 3000 1566
7876 20 1983 1100 1566
7902 20 1981 3000 1566
7499 30 1981 1600 1566
7521 30 1981 1250 1566
7654 30 1981 1250 1566
7698 30 1981 2850 1566
7844 30 1981 1500 1566
7900 30 1981 950 1566

68

Bus Times
SELECT route,stop,bus,TO_CHAR(bustime,'DD-MON-YYYY HH24.MI.SS') bustime
FROM bustimes ORDER BY route,stop,bustime;
ROUTE STOP BUS BUSTIME
---------- -------- -------- --------------------
1 1 10 01-MAR-2011 12.17.33
1 1 30 01-MAR-2011 12.58.10
1 1 20 01-MAR-2011 13.58.41 Times for 5 buses stopping
1 1 40 01-MAR-2011 14.06.13 at 5 stops on route 1
1 1 50 01-MAR-2011 14.11.45
1 2 10 01-MAR-2011 12.56.19
1 2 30 01-MAR-2011 13.00.09
1 2 40 01-MAR-2011 14.20.45
1 2 50 01-MAR-2011 14.24.01
1 2 20 01-MAR-2011 14.31.04
1 3 10 01-MAR-2011 13.58.53
1 3 40 01-MAR-2011 14.35.58
1 3 20 01-MAR-2011 14.58.41
1 3 50 01-MAR-2011 15.18.09
1 3 30 01-MAR-2011 15.28.33
1 4 10 01-MAR-2011 14.17.33
1 4 40 01-MAR-2011 15.11.26
1 4 30 01-MAR-2011 15.30.30
1 4 20 01-MAR-2011 15.42.25
1 4 50 01-MAR-2011 15.55.54
1 5 40 01-MAR-2011 15.51.14
1 5 50 01-MAR-2011 16.02.19
1 5 20 01-MAR-2011 16.18.09
1 5 10 01-MAR-2011 16.30.21
1 5 30 01-MAR-2011 16.47.58

69

Journey Times of Buses Between Stops

SELECT route
,stop
,bus
,TO_CHAR(bustime,'dd/mm/yy hh24:mi:ss') bus_stop_time
,TO_CHAR(LAG(bustime,1)
OVER (PARTITION BY bus
ORDER BY route,stop,bustime)
,'dd/mm/yy hh24:mi:ss') prev_bus_stop_time
,SUBSTR(NUMTODSINTERVAL(bustime - LAG(bustime,1)
ORDER BY route,stop,bustime),'DAY'),12,8) time_between_stops
,SUBSTR(NUMTODSINTERVAL(bustime - FIRST_VALUE(bustime)
ORDER BY route,stop,bustime),'DAY'),12,8) jrny_time
FROM bustimes;

70

Journey Times of Buses Between Stops (cont'd)
ROUTE STOP BUS BUS_STOP_TIME PREV_BUS_STOP_TIM TIME_BET JRNY_TIM
----- ---- --- ----------------- ----------------- -------- --------
1 1 10 01/03/11 12:17:33 00:00:00
1 2 10 01/03/11 12:56:19 01/03/11 12:17:33 00:38:46 00:38:46
1 3 10 01/03/11 13:58:53 01/03/11 12:56:19 01:02:34 01:41:20
1 4 10 01/03/11 14:17:33 01/03/11 13:58:53 00:18:40 02:00:00
1 5 10 01/03/11 16:30:21 01/03/11 14:17:33 02:12:48 04:12:48
1 1 20 01/03/11 13:58:41 00:00:00
1 2 20 01/03/11 14:31:04 01/03/11 13:58:41 00:32:23 00:32:23
1 3 20 01/03/11 14:58:41 01/03/11 14:31:04 00:27:37 01:00:00
1 4 20 01/03/11 15:42:25 01/03/11 14:58:41 00:43:44 01:43:44
1 5 20 01/03/11 16:18:09 01/03/11 15:42:25 00:35:44 02:19:28
1 1 30 01/03/11 12:58:10 00:00:00
1 2 30 01/03/11 13:00:09 01/03/11 12:58:10 00:01:59 00:01:59
1 3 30 01/03/11 15:28:33 01/03/11 13:00:09 02:28:24 02:30:23
1 4 30 01/03/11 15:30:30 01/03/11 15:28:33 00:01:57 02:32:20
1 5 30 01/03/11 16:47:58 01/03/11 15:30:30 01:17:28 03:49:48
1 1 40 01/03/11 14:06:13 00:00:00
1 2 40 01/03/11 14:20:45 01/03/11 14:06:13 00:14:32 00:14:32
1 3 40 01/03/11 14:35:58 01/03/11 14:20:45 00:15:13 00:29:45
1 4 40 01/03/11 15:11:26 01/03/11 14:35:58 00:35:28 01:05:13
1 5 40 01/03/11 15:51:14 01/03/11 15:11:26 00:39:48 01:45:01
1 1 50 01/03/11 14:11:45 00:00:00
1 2 50 01/03/11 14:24:01 01/03/11 14:11:45 00:12:16 00:12:16
1 3 50 01/03/11 15:18:09 01/03/11 14:24:01 00:54:08 01:06:24
1 4 50 01/03/11 15:55:54 01/03/11 15:18:09 UK
Carl Dudley University of Wolverhampton, 00:37:45 01:44:09
1 5 50 01/03/11 16:02:19 01/03/11 15:55:54 00:06:25 71
01:50:34

Average Wait Times for a Bus

SELECT v.route
,v.stop
,v.bus
,v.bustime
,v.prev_bus_time
,SUBSTR(NUMTODSINTERVAL(v.numgap,'DAY'),12,8) wait_for_next_bus
,CASE WHEN bustime = FIRST_VALUE(bustime)
OVER (PARTITION BY stop
THEN SUBSTR(NUMTODSINTERVAL(AVG(v.numgap)
OVER (PARTITION BY stop),'DAY'),12,8)
ELSE NULL END ave_wait
FROM (SELECT route
,stop
,bus
,TO_CHAR(bustime,'dd/mm/yy hh24:mi:ss') bustime
,TO_CHAR(LAG(bustime,1)
,'dd/mm/yy hh24:mi:ss') prev_bus_time
,bustime - LAG(bustime,1)
ORDER BY route,stop,bustime) numgap
FROM bustimes) v;

72

Average Waiting Times for a Bus (continued)

ROUTE STOP BUS BUSTIME PREV_BUS_TIME WAIT_FOR AVE_WAIT
----- ---- --- ------------------ ----------------- -------- --------
1 1 10 01/03/11 12:17:33 00:28:33
1 1 30 01/03/11 12:58:10 01/03/11 12:17:33 00:40:37
1 1 20 01/03/11 13:58:41 01/03/11 12:58:10 01:00:31
1 1 40 01/03/11 14:06:13 01/03/11 13:58:41 00:07:32
1 1 50 01/03/11 14:11:45 01/03/11 14:06:13 00:05:32
1 2 10 01/03/11 12:56:19 00:23:41
1 2 30 01/03/11 13:00:09 01/03/11 12:56:19 00:03:50
1 2 40 01/03/11 14:20:45 01/03/11 13:00:09 01:20:36
1 2 50 01/03/11 14:24:01 01/03/11 14:20:45 00:03:16
1 2 20 01/03/11 14:31:04 01/03/11 14:24:01 00:07:03
1 3 10 01/03/11 13:58:53 00:22:25
1 3 40 01/03/11 14:35:58 01/03/11 13:58:53 00:37:05
1 3 20 01/03/11 14:58:41 01/03/11 14:35:58 00:22:43
1 3 50 01/03/11 15:18:09 01/03/11 14:58:41 00:19:28
1 3 30 01/03/11 15:28:33 01/03/11 15:18:09 00:10:24
1 4 10 01/03/11 14:17:33 00:24:35
1 4 40 01/03/11 15:11:26 01/03/11 14:17:33 00:53:53
1 4 30 01/03/11 15:30:30 01/03/11 15:11:26 00:19:04
1 4 20 01/03/11 15:42:25 01/03/11 15:30:30 00:11:55
1 4 50 01/03/11 15:55:54 01/03/11 15:42:25 00:13:29
1 5 40 01/03/11 15:51:14 00:14:11
1 5 50 01/03/11 16:02:19 01/03/11 15:51:14 00:11:05
1 5 20 01/03/11 16:18:09 01/03/11 16:02:19 00:15:50
1 5 10 01/03/11 16:30:21 01/03/11 16:18:09 00:12:12
1 5 30 01/03/11 16:47:58 01/03/11 16:30:21 00:17:37

73

Analytic Functions


Ranking Functions

Partitioning

Aggregate Functions

Sliding Windows



74

Finding Holes in 'Sequences'
SELECT DISTINCT prod_id
FROM sales
ORDER BY prod_id;  Sales table has 918843 rows
— Gap in prod_ids from 48 to 113
PROD_ID
-------
:
46
47
48
113
114
115
SELECT:prod_id
,next_prod_id
FROM ( SELECT prod_id
,LEAD(prod_id) OVER(ORDER BY prod_id) next_prod_id
FROM sales)
WHERE next_prod_id - prod_id > 1;

PROD_ID NEXT_PROD_ID Elapsed time : 3.17 secs
---------- ------------
48 113
75

Eliminating Duplicate rows

 dup_emp table has 3670016 rows with unique empno values and no primary key
INSERT INTO dup_emp SELECT * FROM dup_emp WHERE empno = 1;
— dup_emp now has one extra duplicate row
 Use conventional SQL to eliminate the duplicate row
DELETE FROM dup_emp y WHERE ROWID <>(SELECT MAX(ROWID)
FROM dup_emp WHERE y.empno = empno);
1 row deleted.
Elapsed: 00:01:38.76
 -------------------------------------------------
 | Id | Operation | Name | Rows |
 -------------------------------------------------
 | 0 | DELETE STATEMENT | | 3670K|
 | 1 | DELETE | DUP_EMP | |
 |* 2 | HASH JOIN | | 3670K|
 | 3 | VIEW | VW_SQ_1 | 3670K|
 | 4 | SORT GROUP BY | | 3670K|
 | 5 | TABLE ACCESS FULL| DUP_EMP | 3670K|
 | 6 | TABLE ACCESS FULL | DUP_EMP | 3670K|
 -------------------------------------------------

76

Eliminating Duplicate rows (continued)

 Use the ranking function to efficiently eliminate the same duplicate row
— ORDER BY clause is necessary so NULL is used as a dummy
DELETE FROM dup_emp WHERE ROWID IN
(SELECT rid
FROM (SELECT ROWID rid
,ROW_NUMBER() OVER (PARTITION BY empno
ORDER BY NULL) rnk
FROM dup_emp)
WHERE rnk > 1);
1 row deleted.
Elapsed: 00:00:19.61
---------------------------------------------------------
--------------------------------------------------------- Similar story with
| 0 | DELETE STATEMENT | | 1 | index on empno
| 1 | DELETE | DUP_EMP | |
| 2 | NESTED LOOPS | | 1 |
| 3 | VIEW | VW_NSO_1 | 3670K|
| 4 | SORT UNIQUE | | 1 |
|* 5 | VIEW | | 3670K|
| 6 | WINDOW SORT | | 3670K|
| 7 | TABLE ACCESS FULL | DUP_EMP | 3670K|
| 8 | TABLE ACCESS BY USER ROWID| DUP_EMP | 1 |

77


 Example based on sales table in sh schema
— 918843 rows, 72 different prod_ids
PROD_ID CUST_ID TIME_ID CHANNEL_ID PROMO_ID QUANTITY_SOLD AMOUNT_SOLD
------- ---------- --------- ---------- ---------- ------------- -----------
46 11702 15-FEB-98 3 999 1 24.92
125 942 27-MAR-98 3 999 1 16.86
46 6406 17-JUL-98 2 999 1 24.83
127 4080 11-SEP-98 3 999 1 38.14
14 19810 20-JUL-98 3 999 1 1257.35
123 3076 24-OCT-98 3 999 1 64.38
48 11403 28-OCT-98 2 999 1 12.95
148 6453 27-MAR-99 2 999 1 20.25
119 609 27-NOV-99 4 999 1 6.54
30 4836 13-DEC-99 2 999 1 10.15
31 1698 17-FEB-00 3 999 1 9.47
119 22354 09-FEB-00 2 999 1 7.75
114 6609 01-JUN-00 3 999 1 21.06
21 8539 28-AUG-00 3 999 1 1097.9
143 11073 14-JAN-01 3 999 1 21.59
119 2234 18-FEB-01 3 999 1 7.51
43 488 25-JUN-01 3 999 1 47.63
27 1577 17-SEP-01 4 999 1 46.16
: : : : : : :

78

Analytic Function Performance - Scenario

 Number of times products are on order

PROD_ID COUNT(*)
SELECT prod_id ------- ----------
,COUNT(*) 22 3441
FROM sh.sales 25 19557
GROUP BY prod_id; 30 29282
34 13043
42 12116
43 8340
123 139
129 7557
138 5541
13 6002
28 16796
116 17389
120 19403
: :

79

nth Best Product – "Conventional" SQL Solution

 Find nth ranked product in terms of numbers of orders for each product

SELECT prod_id
,ycnt
FROM (SELECT prod_id
,COUNT(*) ycnt
FROM sh.sales y
GROUP BY prod_id) v
WHERE &position - 1 = (SELECT COUNT(*)
FROM (SELECT COUNT(*) zcnt
FROM sh.sales z
5 GROUP BY prod_id) w
WHERE w.zcnt > v.ycnt);

PROD_ID YCNT
------- ----------
33 22768

Elapsed: 00:00:24.09

80

"Conventional" SQL Solution - Trace

----------------------------------------------------------------------------
| Id | Operation | Name | Rows | Cost |
----------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 72 | 134|
|* 1 | FILTER | | | |
| 2 | VIEW | | 72 | 67|
| 3 | HASH GROUP BY | | 72 | 67|
| 4 | PARTITION RANGE ALL | | 918K| 29|
| 5 | BITMAP CONVERSION COUNT | | 918K| 29|
| 6 | BITMAP INDEX FAST FULL SCAN | SALES_PROD_BIX | | |
| 7 | SORT AGGREGATE | | 1 | |
| 8 | VIEW | | 4 | 67|
|* 9 | FILTER | | | |
| 10 | SORT GROUP BY | | 4 | 67|
| 12 | BITMAP CONVERSION TO ROWIDS | | 918K| 29|
| 13 | BITMAP INDEX FAST FULL SCAN| SALES_PROD_BIX | | |
----------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
1 - filter( (SELECT COUNT(*) FROM (SELECT COUNT(*) "ZCNT" FROM
"SH"."SALES" "Z" GROUP BY "PROD_ID" HAVING COUNT(*)>:B1) "W")=4)
9 - filter(COUNT(*)>:B1)
Statistics
----------------------------------------------------------
29 consistent gets
72 sorts (memory)

81

nth Best Product – "Failed" SQL Solution


SELECT prod_id
,ycnt
,COUNT(*) ycnt
FROM sh.sales y
GROUP BY prod_id) v
FROM (SELECT ycnt FROM v) w
WHERE w.ycnt > v.ycnt);

*
ERROR at line 8:
ORA-04044: procedure, function, package,
or type is not allowed here

82

nth Best Product – Factored Subquery Solution


WITH v AS
(SELECT prod_id
,COUNT(*) ycnt
FROM sh.sales y
GROUP BY prod_id)
5
SELECT prod_id
,ycnt
FROM v
FROM (SELECT ycnt
FROM v) w
WHERE w.ycnt > v.ycnt);

PROD_ID YCNT
------- ----------
33 22768

Elapsed: 00:00:00.07

83

Factored Subquery Solution - Trace

---------------------------------------------------------------------------------------
---------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | 71 |
| 1 | TEMP TABLE TRANSFORMATION | | | |
| 2 | LOAD AS SELECT | | | |
| 3 | HASH GROUP BY | | 72 | 67 |
| 4 | PARTITION RANGE ALL | | 918K| 29 |
| 5 | BITMAP CONVERSION COUNT | | 918K| 29 |
|* 7 | FILTER | | | |
| 8 | VIEW | | 72 | 2 |
| 9 | TABLE ACCESS FULL | SYS_TEMP_0FD9D661A_14D8441 | 72 | 2 |
| 10 | SORT AGGREGATE | | 1 | |
|* 11 | VIEW | | 72 | 2 |
| 12 | TABLE ACCESS FULL | SYS_TEMP_0FD9D661A_14D8441 | 72 | 2 |
---------------------------------------------------------------------------------------

---------------------------------------------------

7 - filter( (SELECT COUNT(*) FROM (SELECT /*+ CACHE_TEMP_TABLE ("T1") */ "C0"
"PROD_ID","C1" "YCNT"
"SYS"."SYS_TEMP_0FD9D661A_14D8441" "T1") "V" WHERE "YCNT">:B1)=4)
11 - filter("YCNT">:B1)

Statistics
----------------------------------------------------------
355 consistent gets
0 sorts (memory)

84

nth Best Product – Analytic Function Solution


SELECT prod_id
,vcnt
,vcnt
,RANK() OVER (ORDER BY vcnt DESC) rnk
,COUNT(*) vcnt
FROM sh.sales z
GROUP BY z.prod_id)) qry 5
WHERE qry.rnk = &position;

PROD_ID YCNT
------- ----------
33 22768

Elapsed: 00:00:00.01

85

Analytic Function Solution - Trace

--------------------------------------------------------------------------
--------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 72 | 105|
|* 1 | VIEW | | 72 | 105|
|* 2 | WINDOW SORT PUSHED RANK | | 72 | 105|
| 3 | HASH GROUP BY | | 72 | 105|
| 5 | BITMAP CONVERSION COUNT | | 918K| 29|
--------------------------------------------------------------------------

---------------------------------------------------

1 - filter("QRY"."RNK"=5)
2 - filter(RANK() OVER ( ORDER BY COUNT(*) DESC )<=5)

Statistics
----------------------------------------------------------
116 consistent gets
1 sorts (memory)

86


 Defining the PARTITION BY and ORDER BY clauses on indexed columns
will provide optimum performance
— For example, a composite index on (deptno, hiredate) columns will
prove effective
 Analytic functions still provide acceptable performance in absence of
indexes but need to do sorting for computing based on partition and order
by columns
— If the query contains multiple analytic functions, sorting and partitioning on
two different columns should be avoided if they are both not indexed

87

Performance

 Hiding analytics in views can prevent the use of indexes
— SUM(sal) has to be computed across all rows before the analysis
CREATE OR REPLACE VIEW vv AS
SELECT *, SUM(sal) OVER (PARTITION BY deptno) Deptno_Sum_Sal
FROM emp;
SELECT * FROM vv WHERE empno = 7900;
EMPNO ENAME JOB MGR HIREDATE SAL COMM DEPTNO DEPTNO_SUM_SAL
----- ----- ----- ---- --------- ---- ---- ------ --------------
7900 JAMES CLERK 7698 03-DEC-81 950 30 9400
--------------------------------------------
--------------------------------------------
|* 1 | VIEW | VV | 14 |
| 2 | WINDOW SORT | | 14 |
--------------------------------------------
SELECT * FROM emp WHERE empno = 7900;
------------------------------------------------------------
------------------------------------------------------------
| 1 | TABLE ACCESS BY INDEX ROWID| EMP | 1 |
|* 2 | INDEX UNIQUE SCAN | SYS_C0017750 | 1 |
------------------------------------------------------------

88

Steamy Windows

SELECT empno, ename, sal, deptno
,SUM(sal) OVER (PARTITION BY deptno ORDER BY sal) sumsal

FROM emp ORDER BY deptno, sal;

EMPNO ENAME SAL DEPTNO SUMSAL
---------- ---------- ---------- ---------- ----------
7934 MILLER 1300 10 1300
7782 CLARK 2450 10 3750
7839 KING 5000 10 8750
7369 SMITH 800 20 800
7876 ADAMS 1100 20 1900
7566 JONES 2975 20 4875
7788 SCOTT 3000 20 10875
7902 FORD 3000 20 10875
7900 JAMES 950 30 950
7654 MARTIN 1250 30 3450
7521 WARD 1250 30 3450
7844 TURNER 1500 30 4950
7499 ALLEN 1600 30 6550
7698 BLAKE 2850 30 9400
Default window is RANGE BETWEEN UNBOUNDED PRECEDING AND CURRENT
ROW

89

Steamy Windows (continued)

SELECT empno, ename, sal, deptno
FROM emp
WHERE ename LIKE '%M%'
ORDER BY deptno ,sal

---------- ---------- ---------- ---------- ----------
7934 MILLER 1300 10 1300
7369 SMITH 800 20 800
7876 ADAMS 1100 20 1900
7900 JAMES 950 30 950
7654 MARTIN 1250 30 2200

SELECT * Includes WARD who is in department 30
FROM (SELECT empno, ename, sal, deptno and has a salary of 1250. which is within
the RANGE with MARTIN
FROM emp )
WHERE ename LIKE '%M%'
ORDER BY deptno ,sal;

---------- ---------- ---------- ---------- ----------
7934 MILLER 1300 10 1300
7369 SMITH 800 20 800
7876 ADAMS 1100 20 1900
7900 JAMES 950 30 950 90

In the Final Analysis

So we have discussed
 The ranking of data using analytic functions
 Partitioning datasets from queries
 Using aggregate functions in analytic scenarios
 How to apply sliding windows to query results
 Comparing values across rows
 Performance characteristics

91

Analytic Functions

Carl Dudley
University of Wolverhampton, UK

UKOUG Council
Oracle ACE Director

carl.dudley@wlv.ac.uk

92

Analytics ioug 2011

More Related Content

What's hot (8)

Viewers also liked (9)

Similar to Analytics ioug 2011 (20)

Recently uploaded (20)

Analytics ioug 2011

Editor's Notes