PL SQL Questions

http://www.sqlsnippets.com/en/home.
htm l PL/SQL Collections

The chart below lists the properties of the three collection types on a set of parameters such as size, ease of modification, persistence, etc. Index By Tables Nested Tables Varrays Bounded i.e. holds a declared number of elements, though this number can be changed at runtime Sequential numbers, starting from one Can be stored in the database using equivalent SQL types, and manipulated through SQL (but with less ease than nested tables) Standard subscripting syntax e.g. color(3) is the 3rd color in varray color
Size Unbounded i.e. the number Unbounded i.e. the of elements it can hold is not number of elements it pre-defined can hold is not predefined Subscript Can be arbitrary numbers or Characteristics strings. Need not be sequential. Database Index by tables can be used Storage in PL/SQL programs only, cannot be stored in the database. Sequential numbers, starting from one Can be stored in the database using equivalent SQL types, and manipulated through SQL.
Referencing Works as key-value pairs. and lookups e.g. Salaries of employees can be stored with unique employee numbers used as subscripts sal(102) := 2000;
Similar to one-column database tables. Oracle stores the nested table data in no particular order. But when you retrieve the nested table into a PL/SQL variable, the rows are given consecutive subscripts starting at 1. Almost like index-by tables, except that
Flexibility to Most flexible. Size can changes increase/ decrease
Not very flexible. You must retrieve and
dynamically. Elements can be added to any position in the list and deleted from any position. Mapping with Hash tables other programming languages
subscript values are not as flexible. Deletions are possible from noncontiguous positions. Sets and bags
update all the elements of the varray at the same time. Arrays
Which Collection Type to Use?

You have all the details about index by tables, nested tables and varrays now. Given a situation, will one should you use for your list data? Here are some guidelines. Use index by tables when:

Your program needs small lookups The collection can be made at runtime in the memory when the package/ procedure is initialized The data volume is unknown beforehand The subscript values are flexible (e.g. strings, negative numbers, non-sequential) You do not need to store the collection in the database
Use nested tables when:

The data needs to be stored in the database The number of elements in the collection is not known in advance The elements of the collection may need to be retrieved out of sequence Updates and deletions affect only some elements, at arbitrary locations Your program does not expect to rely on the subscript remaining stable, as their order may change when nested tables are stored in the database.
Use varrays when:

The data needs to be stored in the database The number of elements of the varray is known in advance The data from the varray is accessed in sequence Updates and deletions happen on the varray as a whole and not on arbitrarily located elements in the varray
Sample Code
Associative Array
Declare
Nested Table
Varray
Declare a collection variable.

aa_0 p.aa_type ;
nt_0 p.nt_type ;
va_0 p.va_type ;
Declare, initialize, and load a collection variable.

aa p.aa_type ; -- cannot load values in -- declaration begin nt p.nt_type := p.nt_type( 'a', 'b' ) ; va p.va_type := p.va_type( 'a', 'b' ) ;
Let's inspect the variables to see what they look like at this point (NULL means the variable is not initialized).
p.print( p.print( p.print( p.print( p.print( 'aa_0 is ' ); aa_0 ); ' ' ); 'aa is ' ); aa ); p.print( p.print( p.print( p.print( p.print( 'nt_0 is ' ); nt_0 ); ' ' ); 'nt is ' ); nt ); p.print( p.print( p.print( p.print( p.print( 'va_0 is ' ); va_0 ); ' ' ); 'va is ' ); va );
aa_0 is NOT NULL .first = .last = .count = .limit = aa is NOT NULL .first = .last = .count = .limit =
and empty NULL NULL 0 NULL and empty NULL NULL 0 NULL
nt_0 is NULL nt is (1) a (2) b .first .last .count .limit
va_0 is NULL va is (1) a (2) b .first .last .count .limit
= = = =
1 2 2 NULL
= = = =
1 2 2 10
Initialize a collection after it has been declared.

-- n/a nt_0 := p.nt_type() ; p.print( nt_0 ); NOT NULL .first = .last = .count = .limit = and empty NULL NULL 0 NULL va_0 := p.va_type(); p.print( va_0 ); NOT NULL .first = .last = .count = .limit = and empty NULL NULL 0 10
Add individual rows to a collection.

-- add 1 row at a time aa(1) := 'a' ; aa(2) := 'b' ; aa(3) := 'c' ; -- add 1 row nt.extend ; nt(3) := 'c' ; -- add 2 rows nt.extend(2) ; nt(4) := 'd' ; nt(5) := 'e' ; -- create two copies -- of row #5 aa(6) := 'e' ; aa(7) := 'e' ; p.print( aa ); nt.extend(2,5) ; p.print( nt ); -- add 1 row va.extend ; va(3) := 'c' ; -- add 2 rows va.extend(2) ; va(4) := 'd' ; va(5) := 'e' ; -- create two copies -- of row #5 va.extend(2,5) ; p.print( va );
aa(4) := 'd' ; aa(5) := 'e' ;
(1) a (2) b (3) c (4) d (5) e (6) e (7) e .first .last .count .limit
= = = =
1 7 7 NULL
(1) a (2) b (3) c (4) d (5) e (6) e (7) e .first .last .count .limit
= = = =
1 7 7 NULL
(1) a (2) b (3) c (4) d (5) e (6) e (7) e .first = 1 .last = 7 .count = 7
.limit = 10
Load a single value from the database into a collection row.

select into from where val aa(2) t val = 'B' ; select into from where val nt(2) t val = 'B' ; select into from where val va(2) t val = 'B' ;
p.print( aa );
p.print( nt );
p.print( va );
(1) a (2) B (3) c (4) d (5) e (6) e (7) e .first .last .count .limit
= = = =
1 7 7 NULL
(1) a (2) B (3) c (4) d (5) e (6) e (7) e .first .last .count .limit
= = = =
1 7 7 NULL
(1) a (2) B (3) c (4) d (5) e (6) e (7) e .first = 1 .last = 7 .count = 7
.limit = 10
Initialize a collection and load it with multiple database values (pre-existing contents will be lost).
select val bulk collect into aa from t ; select val bulk collect into nt from t ; select val bulk collect into va from t ;
p.print( aa );
(1) A (2) B (3) C (4) D (5) E (6) F (7) G .first = 1 .last = 7 .count = 7
p.print( nt );
p.print( va );
.limit = NULL
.limit = NULL
.limit = 10
Test a row's existence by subscript.

p.print ( 'aa.exists(3) is '|| p.tf( aa.exists(3) ) ); p.print ( 'aa.exists(9) is '|| p.tf( aa.exists(9) ) p.print ( 'nt.exists(3) is '|| p.tf( nt.exists(3) ) ); p.print ( 'nt.exists(9) is '|| p.tf( nt.exists(9) ) p.print ( 'va.exists(3) is '|| p.tf( va.exists(3) ) ); p.print ( 'va.exists(9) is '|| p.tf( va.exists(9) )
);
);
);
aa.exists(3) is TRUE
aa.exists(9) is FALSE
nt.exists(3) is TRUE
nt.exists(9) is FALSE
va.exists(3) is TRUE
va.exists(9) is FALSE
Test a row's existence by content. -- use a loop (see below)

p.print ( '''C'' member of nt is '|| p.tf( 'C' member of nt ) ); p.print ( '''X'' member of nt is '|| p.tf( 'X' member of nt )
-- use a loop (see below)
);
'C' member of nt is TRUE
'X' member of nt is FALSE Compare two collections for equality.

-- cannot use "=" with -- two associative arrays nt_0 := nt ;
-- cannot use "=" with
if nt_0 = nt then p.print( 'equal' ); else p.print( 'not equal' );
-- two varrays
end if;
equal
Update a collection row.

aa(1) := 'a' ; aa(3) := 'c' ; nt(1) := 'a' ; nt(3) := 'c' ; va(1) := 'a' ; va(3) := 'c' ;
p.print( aa );
(1) a (2) B (3) c (4) D (5) E (6) F (7) G .first = 1 .last = 7 .count = 7
p.print( nt );
p.print( va );
.limit = NULL
.limit = NULL
.limit = 10
Remove rows from the middle of a collection.
aa.delete(2); aa.delete(3,4);
nt.delete(2); nt.delete(3,4);
-- not possible
p.print( aa );
(1) a (5) E (6) F (7) G .first = 4 .last = 7 .count = 4
p.print( nt );
(1) a (5) E (6) F (7) G .first = 4 .last = 7 .count = 4
p.print( va );
.limit = NULL
.limit = NULL
.limit = 10 Loop through all rows in the collection.

declare i binary_integer ; begin i := aa.first ; while i is not null loop p.print ( i ||'. '|| aa(i) ); i := aa.next(i) ; end loop; declare i binary_integer ; begin i := nt.first ; while i is not null loop p.print ( i ||'. '|| nt(i) ); i := nt.next(i) ; end loop;
for i in nvl(va.first,0) .. nvl(va.last,-1) loop p.print ( i ||'. '|| va(i) );
end loop;
end;
1. a 5. E 6. F
end;
1. a 5. E 6. F
7. G
7. G
1. 2. 3. 4. 5. 6.
7. G Remove row(s) from the end of a collection.

aa.delete(7); aa.delete(5,6); nt.trim; nt.trim(2);
a B c D E F
va.trim; va.trim(2);
p.print( aa );
(4) D .first = 4 .last = 4 .count = 1
p.print( nt );
(4) D .first = 4 .last = 4 .count = 1
p.print( va );
(1) a (2) B (3) c (4) D .first = 1
.limit = NULL
.limit = NULL
.last = 4 .count = 4
.limit = 10
Reuse rows left vacant by earlier delete operations (rows 2,3,4) and trim operations (rows 5,6,7).
aa(2) := 'B' ; aa(3) := 'C' ; aa(4) := 'D' ; -- note we do not need to -- call ".extend" for rows -- 2,3,4 which were -- removed with ".delete" nt(2) := 'B' ; nt(3) := 'C' ; nt(4) := 'D' ; aa(5) := 'E' ; aa(6) := 'F' ; aa(7) := 'G' ; -- we do need to call -- ".extend" for rows 5,6,7 -- which were removed with -- ".trim" nt.extend(3) nt(5) := 'E' nt(6) := 'F' nt(7) := 'G' ; ; ; ; ----we need to call ".extend" first since 5,6,7 were removed with ".trim" ; ; ; ;
va.extend(3) va(5) := 'E' va(6) := 'F' va(7) := 'G'
p.print( va );
p.print( aa );
p.print( nt );
(1) A (2) B (3) C (4) D (5) E (6) F (7) G .first = 1 .last = 7 .count = 7 (1) A (2) B (3) C (4) D (5) E (6) F (7) G .first = 1 .last = 7 .count = 7 (1) a (2) B (3) c (4) D (5) E (6) F (7) G .first = 1 .last = 7 .count = 7
.limit = NULL
.limit = NULL
.limit = 10
Delete all rows in the collection (frees memory too).

aa.delete ; nt.delete ; va.delete ;
p.print( aa );
NOT NULL .first = .last = .count = and empty NULL NULL 0
p.print( nt );
p.print( va );
.limit = NULL
.limit = NULL
.limit = 10
Set a collection to NULL, i.e. uninitialized state. -- not possible

-- "nt := null" will not -- work; use a null -- variable instead declare nt_null p.nt_type ; begin nt := nt_null ; end; -- "va := null" will not -- work; use a null -- variable instead declare va_null p.va_type ; begin va := va_null ; end;
p.print( nt ); NULL
end;
p.print( va ); NULL
The next table presents operational characteristics of each collection type. Associative Characteristic Array The entire collection can be saved in a database column. Rows in the collection retain their order when the entire n/a collection is saved in a database column. Legal subscript datatypes. any -2**31..2**31 Legal subscript value ranges. (for Integers) The collection can be defined to hold a predefined maximum number of rows. There can be gaps between subscripts, e.g. 1,3,8. Y The collection must be initialized before used. The collection can be initialized with multiple rows of data using a single command (i.e. a constructor). The collection must be extended before a new row is added. Two collections can be compared for equality with the "=" operator. The collection can be manipulated in PL/SQL with MULTISET Operators e.g. MULTISET UNION, MULTISET INTERSECT. The collection can be unnested in a query using the TABLE() collection expression.
Nested Table Y Integer 1..2**31
Varray Y Y Integer 1..2**31 Y
Y Y Y Y Y Y Y
Y Y Y
The Difference Between DECODE and CASE
DECODE and CASE statements in Oracle both provide a conditional construct, of this form: if A = n1 then A1 else if A = n2 then A2 else X Databases before Oracle 8.1.6 had only the DECODE function. CASE was introduced in Oracle 8.1.6, as a standard, more meaningful and more powerful function. Everything DECODE can do, CASE can. There is a lot more that you can do with CASE, though, which DECODE cannot which well see in this article.
1. CASE can work with logical operators other than =

DECODE can do an equality check only. CASE is capable of more logical comparisons such as < > etc. To achieve the same effect with DECODE, ranges of data had to be forced into discrete form making unwieldy code. An example of putting employees in grade brackets based on their salaries this can be done elegantly with CASE.
SQL> select ename 2 , case 3 when sal < 1000 4 then 'Grade I' 5 when (sal >=1000 and sal < 2000) 6 then 'Grade II' 7 when (sal >= 2000 and sal < 3000) 8 then 'Grade III' 9 else 'Grade IV' 10 end sal_grade 11 from emp 12 where rownum < 4;
ENAME ---------SMITH ALLEN WARD
SAL_GRADE --------Grade I Grade II Grade II
2. CASE can work with predicates and searchable subqueries

DECODE works with expressions which are scalar values only. CASE can work with predicates and subqueries in searchable form. An example of categorizing employees based on reporting relationship, illustrating these two uses of CASE.
SQL> select e.ename, 2 case 3 -- predicate with "in" 4 -- mark the category based on ename list 5 when e.ename in ('KING','SMITH','WARD') 6 then 'Top Bosses' 7 -- searchable subquery 8 -- identify if this emp has a reportee 9 when exists (select 1 from emp emp1 10 where emp1.mgr = e.empno) 11 then 'Managers' 12 else 13 'General Employees' 14 end emp_category 15 from emp e 16 where rownum < 5; ENAME ---------SMITH ALLEN WARD JONES EMP_CATEGORY ----------------Top Bosses General Employees Top Bosses Managers
3. CASE can work as a PL/SQL construct

DECODE can work as a function inside SQL only. CASE can be a more efficient substitute for IFTHEN-ELSE in PL/SQL.
SQL> declare 2 grade char(1); 3 begin 4 grade := 'b'; 5 case grade 6 when 'a' then dbms_output.put_line('excellent'); 7 when 'b' then dbms_output.put_line('very good'); 8 when 'c' then dbms_output.put_line('good'); 9 when 'd' then dbms_output.put_line('fair'); 10 when 'f' then dbms_output.put_line('poor');
11 12 13 14
else dbms_output.put_line('no such grade'); end case; end; /
PL/SQL procedure successfully completed.
CASE can even work as a parameter to a procedure call, while DECODE cannot.
SQL> var a varchar2(5); SQL> exec :a := 'THREE'; PL/SQL procedure successfully completed. SQL> SQL> 2 3 4 5 6
create or replace procedure proc_test (i number) as begin dbms_output.put_line('output = '||i); end; /
Procedure created. SQL> exec proc_test(decode(:a,'THREE',3,0)); BEGIN proc_test(decode(:a,'THREE',3,0)); END; * ERROR at line 1: ORA-06550: line 1, column 17: PLS-00204: function or pseudo-column 'DECODE' may be used inside a SQL statement only ORA-06550: line 1, column 7: PL/SQL: Statement ignored SQL> exec proc_test(case :a when 'THREE' then 3 else 0 end); output = 3 PL/SQL procedure successfully completed.
4. Careful! CASE handles NULL differently

Check out the different results with DECODE vs NULL.
SQL> select decode(null 2 , null, 'NULL' 3 , 'NOT NULL' 4 ) null_test 5 from dual; NULL ---NULL
SQL> select case null 2 when null 3 then 'NULL' 4 else 'NOT NULL' 5 end null_test 6 from dual; NULL_TES -------NOT NULL
The searched CASE works as does DECODE, though.

SQL> select case 2 when null is null 3 then 'NULL' 4 else 'NOT NULL' 5 end null_test 6* from dual SQL> / NULL_TES -------NULL
5. CASE expects data type consistency, DECODE does not

Compare the two examples DECODE gives you a result, CASE gives a data type mismatch error.
SQL> select decode(2,1,1, 2 '2','2', 3 '3') t 4 from dual; T ---------2 SQL> select case 2 when 1 then '1' 2 when '2' then '2' 3 else '3' 4 end 5 from dual; when '2' then '2' * ERROR at line 2: ORA-00932: inconsistent datatypes: expected NUMBER got CHAR
6. CASE is ANSI SQL-compliant

CASE complies with ANSI SQL. DECODE is proprietary to Oracle.
7. The difference in readability

In very simple situations, DECODE is shorter and easier to understand than CASE, as in:
SQL> -- An example where DECODE and CASE SQL> -- can work equally well, and SQL> -- DECODE is cleaner SQL> select ename 2 , decode (deptno, 10, 'Accounting', 3 20, 'Research', 4 30, 'Sales', 5 'Unknown') as department 6 from emp 7 where rownum < 4; ENAME ---------SMITH ALLEN WARD DEPARTMENT ---------Research Sales Sales
SQL> select ename 2 , case deptno 3 when 10 then 'Accounting' 4 when 20 then 'Research' 5 when 30 then 'Sales' 6 else 'Unknown' 7 end as department 8 from emp 9 where rownum < 4; ENAME ---------SMITH ALLEN WARD DEPARTMENT ---------Research Sales Sales
In complex situations, CASE is shorter and easier to understand. Complicated processing in DECODE, even if technically achievable, is a recipe for messy, unreadable code while the same can be achieved elegantly using CASE.
Grouping Rows with GROUP BY

GROUP BY
Consider a table like this one.
select grp_a, grp_b, val from t order by grp_a, grp_b ; GRP_A ---------a1 a1 a1 a1 a1 a2 a2 a2 GRP_B VAL ---------- ---------b1 10 b1 20 b2 30 b2 40 b2 50 b3 12 b3 22 b3 32
GROUP BY allows us to group rows together so that we can include aggregate functions like COUNT, MAX, and SUM in the result set.
select grp_a, count(*), max( val ), sum( val ) from t GROUP BY GRP_A order by grp_a ; GRP_A COUNT(*) MAX(VAL) SUM(VAL) ---------- ---------- ---------- ---------a1 5 50 150 a2 3 32 66
We can specify multiple columns in the GROUP BY clause to produce a different set of groupings.
select grp_a, grp_b, count(*), max( val ), sum( val ) from t GROUP BY GRP_A, GRP_B order by grp_a, grp_b ; GRP_A ---------a1 a1 a2 GRP_B COUNT(*) MAX(VAL) SUM(VAL) ---------- ---------- ---------- ---------b1 2 20 30 b2 3 50 120 b3 3 32 66
Parentheses may be added around the GROUP BY expression list. Doing so has no effect on the result.
select grp_a, grp_b, count(*), max( val ), sum( val ) from t GROUP BY ( GRP_A, GRP_B ) order by grp_a, grp_b ;
GRP_A ---------a1 a1 a2
GRP_B COUNT(*) MAX(VAL) SUM(VAL) ---------- ---------- ---------- ---------b1 2 20 30 b2 3 50 120 b3 3 32 66
The GROUP BY expression list may be empty. This groups all rows retrieved by the query into a single group. Parentheses are mandatory when specifying an empty set.
select count(*), max( val ), sum( val ) from t GROUP BY () ; COUNT(*) MAX(VAL) SUM(VAL) ---------- ---------- ---------8 50 216
The last example is equivalent to specifying no GROUP BY clause at all, like this.
select count(*), max( val ), sum( val ) from t ; COUNT(*) MAX(VAL) SUM(VAL) ---------- ---------- ---------8 50 216
GROUP BY and DISTINCT

We can use GROUP BY without specifying any aggregate functions in the SELECT list.
select grp_a, grp_b from t GROUP BY GRP_A, GRP_B order by grp_a, grp_b ; GRP_A ---------a1 a1 a2 GRP_B ---------b1 b2 b3
However, the same result is usually produced by specifying DISTINCT instead of using GROUP BY.
select DISTINCT grp_a, grp_b from t order by grp_a, grp_b ; GRP_A ---------a1 a1 a2 GRP_B ---------b1 b2 b3
According to Tom Kyte the two approaches are effectively equivalent (see AskTom "DISTINCT VS, GROUP BY"). Queries that use DISTINCT are typically easier to understand, but the GROUP BY approach can provide an elegant solution to otherwise cumbersome queries when more than one set of groupings is required. For example, to produce a result set that is the union of: distinct values in GRP_A distinct values in GRP_B distinct values in GRP_A + GRP_B the following query would be required if we used DISTINCT
select distinct grp_a, null as grp_b from t union all select distinct null, grp_b from t union all select distinct grp_a, grp_b from t order by 1, 2 ; GRP_A ---------a1 a1 a1 a2 a2 GRP_B ---------b1 b2 b3 b1 b2 b3
but a GROUP BY query could produce the same result with fewer lines of code.
select grp_a, grp_b from t group by cube( grp_a, grp_b ) having grouping_id( grp_a, grp_b ) != 3 order by 1, 2 ; GRP_A ---------a1 a1 a1 a2 a2 GRP_B ---------b1 b2 b3 b1 b2 b3
(We will learn about the CUBE and GROUPING_ID features later in this tutorial.)
GROUP BY and Ordering
All other things being equal, changing the order in which columns appear in the GROUP BY clause has no effect on the way the result set is grouped. For example, this query
select grp_a, grp_b, count(*) from t GROUP BY GRP_A, GRP_B order by grp_a, grp_b ; GRP_A ---------a1 a1 a2 GRP_B COUNT(*) ---------- ---------b1 2 b2 3 b3 3
returns the same results as this one.

select grp_a, grp_b, count(*) from t GROUP BY GRP_B, GRP_A -- columns have been reversed order by grp_a, grp_b ; GRP_A ---------a1 a1 a2 GRP_B COUNT(*) ---------- ---------b1 2 b2 3 b3 3
Gotcha: GROUP BY with no ORDER BY The last two snippets used the same ORDER BY clause in both queries. What happens if we use no ORDER BY clause at all?
select grp_a, grp_b, count(*) from t group by grp_a, grp_b ; GRP_A -----a1 a1 a2 GRP_B COUNT(*) ------ ---------b1 2 b2 3 b3 3
The results are still ordered. Some programmers interpret this as meaning that GROUP BY returns an ordered result set. This is an illusion which is easily proved with the following snippet. Note how the same query now returns rows in a random order given new conditions.
truncate table t; -- this time we insert rows into T using a different order from that -- of the Setup topic insert into t values ( 'a2' , 'b3' , 'c2', 'd2', '32' ) ; insert into t values ( 'a2' , 'b3' , 'c2', 'd2', '22' ) ; insert into t values ( 'a2' , 'b3' , 'c2', 'd2', '12' ) ;
insert insert insert insert insert commit;
into into into into into
t t t t t
values values values values values
( ( ( ( (
'a1' 'a1' 'a1' 'a1' 'a1'
, , , , ,
'b2' 'b2' 'b2' 'b1' 'b1'
, , , , ,
'c2', 'c1', 'c1', 'c1', 'c1',
'd1', 'd1', 'd1', 'd1', 'd1',
'50' '40' '30' '20' '10'
) ) ) ) )
; ; ; ; ;
select grp_a, grp_b, count(*) from t group by grp_a, grp_b ; GRP_A -----a1 a1 a2 GRP_B COUNT(*) ------ ---------b2 3 b1 2 b3 3
-- (your results may vary)
The actual behaviour of GROUP BY without ORDER BY is documented in the SQL Reference Manual as follows. "The GROUP BY clause groups rows but does not guarantee the order of the result set. To order the groupings, use the ORDER BY clause." (See AskTom ; Group by behavior in 10GR2 for another discussion of this issue.)
Duplicate Columns
If a column is used more than once in the SELECT clause it does not need to appear more than once in the GROUP BY clause.
select grp_a, upper(grp_a), count(*) from t group by GRP_A order by grp_a ; GRP_A ---------a1 a2 UPPER(GRP_ COUNT(*) ---------- ---------A1 5 A2 3
If we did include the same column two or more times in the GROUP BY clause it would return the same results as the query above.
select grp_a, upper(grp_a), count(*) from t group by GRP_A, GRP_A order by grp_a ; GRP_A UPPER(GRP_ COUNT(*) ---------- ---------- ---------a1 A1 5
a2
A2
While there is no practical use for the latter syntax in the upcoming topic GROUP_ID we will see how duplicate columns in a GROUPING SETS clause do produce different results than a distinct column list.
SELECT Lists
We may group by table columns that are not in the SELECT list, like GRP_B in the example below.
select grp_a, count(*) from t group by grp_a, GRP_B order by grp_a, grp_b ; GRP_A COUNT(*) ------ ---------a1 2 a1 3 a2 3
However we may not select table columns that are absent from the GROUP BY list, as with GRP_A in this example.
select GRP_A, count(*) from t GROUP BY GRP_B ; select GRP_A, count(*) * ERROR at line 1: ORA-00979: not a GROUP BY expression
Constants
The rules for columns based on constant expressions differ slightly from those for table columns. As with table based columns we can include constant columns in the GROUP BY clause
select 123, 'XYZ', SYSDATE, grp_a, grp_b, count(*) from t group by 123, 'XYZ', SYSDATE, grp_a, grp_b order by grp_a, grp_b ; 123 ---------123 123 123 'XY --XYZ XYZ XYZ SYSDATE ---------2009-06-07 2009-06-07 2009-06-07 GRP_A -----a1 a1 a2 GRP_B COUNT(*) ------ ---------b1 2 b2 3 b3 3
and we can GROUP BY constant columns that are not in the SELECT list.
select grp_a, grp_b, count(*) from t
group by 123, 'XYZ', SYSDATE, grp_a, grp_b order by grp_a, grp_b ; GRP_A -----a1 a1 a2 GRP_B COUNT(*) ------ ---------b1 2 b2 3 b3 3
Unlike table based columns we can select constant columns that are absent from the GROUP BY list.
select 123, 'XYZ', SYSDATE, grp_a, grp_b, count(*) from t GROUP BY GRP_A, GRP_B order by grp_a, grp_b ; 123 ---------123 123 123 'XY --XYZ XYZ XYZ SYSDATE ---------2009-06-07 2009-06-07 2009-06-07 GRP_A -----a1 a1 a2 GRP_B COUNT(*) ------ ---------b1 2 b2 3 b3 3
Note how all three queries returned the same number of rows.
HAVING
When Oracle processes a GROUP BY query the WHERE clause is applied to the result set before the rows are grouped together. This allows us to use WHERE conditions involving columns like GRP_B in the query below, which is not listed in the GROUP BY clause.
select grp_a, count(*) from t WHERE GRP_B in ( 'b2', 'b3' ) group by grp_a order by grp_a ; GRP_A COUNT(*) ---------- ---------a1 3 a2 3
Thia does, however, prevent us from using conditions that involve aggregate values like COUNT(*) that are calculated after the GROUP BY clause is applied. For example, the following will not work.
select grp_a, count(*) from t WHERE COUNT(*) > 4 group by grp_a order by grp_a ; WHERE COUNT(*) > 4 * ERROR at line 3:
ORA-00934: group function is not allowed here
For these types of conditions the HAVING clause can be used.

select grp_a, count(*) from t group by grp_a HAVING COUNT(*) > 4 order by grp_a ; GRP_A COUNT(*) ---------- ---------a1 5
Note that the HAVING clause cannot reference table columns like VAL that are not listed in the GROUP BY clause.
select grp_a, count(*) from t group by grp_a HAVING VAL > 5 order by grp_a ; HAVING VAL > 5 * ERROR at line 4: ORA-00979: not a GROUP BY expression
It can, on the other hand, reference table columns like GRP_A that are in the GROUP BY clause.
select grp_a, count(*) from t group by grp_a HAVING GRP_A = 'a2' order by grp_a ; GRP_A COUNT(*) ---------- ---------a2 3
but doing so yields the same result as using a WHERE clause.

select grp_a, count(*) from t WHERE GRP_A = 'a2' group by grp_a order by grp_a ; GRP_A COUNT(*) ---------- ---------a2 3
Given a choice between the last two snippets I expect using a WHERE clause provides the best performance in most, if not all, cases.
GROUPING SETS
There are times when the results of two or more different groupings are required from a single query. For example, say we wanted to combine the results of these two queries.
select grp_a, count(*) from t group by grp_a order by grp_a ; GRP_A COUNT(*) ---------- ---------a1 5 a2 3 select grp_b, count(*) from t group by grp_b order by grp_b ; GRP_B COUNT(*) ---------- ---------b1 2 b2 3 b3 3
UNION ALL could be used, like this

select grp_a, null, count(*) from t group by grp_a UNION ALL select null, grp_b, count(*) from t group by grp_b order by 1, 2 ; GRP_A NULL COUNT(*) ---------- ---------- ---------a1 5 a2 3 b1 2 b2 3 b3 3
but as of Oracle 9i a more compact syntax is available with the GROUPING SETS extension of the GROUP BY clause. With it the last query can be written as follows.
select grp_a, grp_b, count(*) from t GROUP BY GROUPING SETS ( GRP_A, GRP_B )
order by grp_a, grp_b ; GRP_A GRP_B COUNT(*) ---------- ---------- ---------a1 5 a2 3 b1 2 b2 3 b3 3
It is important to understand how the clause grouping sets( grp_a, grp_b ) used in the last query differs from the clause group by ( grp_a, grp_b ) in the next query.
select grp_a, grp_b, count(*) from t GROUP BY ( GRP_A, GRP_B ) order by grp_a, grp_b ; GRP_A ---------a1 a1 a2 GRP_B COUNT(*) ---------- ---------b1 2 b2 3 b3 3
Note how the last query returned different rows than the GROUPING SETS query did even though both used the term (GRP_A, GRP_B).
GROUPING SETS, Composite Columns, and Empty Sets

Composite Columns You can treat a collection of columns as an individual set by using parentheses in the GROUPING SETS clause. For example, to write a query that returns the equivalent of these two queries
select grp_a, null, count(*) from t GROUP BY GRP_A order by grp_a ; GRP_A N COUNT(*) ---------- - ---------a1 5 a2 3
we could use the following GROUPING SETS clause.

select grp_a, grp_b, count(*) from t GROUP BY GROUPING SETS ( (GRP_A, GRP_B), GRP_A ) order by grp_a, grp_b ; GRP_A ---------a1 a1 a1 a2 a2 GRP_B COUNT(*) ---------- ---------b1 2 b2 3 5 b3 3 3
The term (GRP_A, GRP_B) is called a "composite column" when it appears inside a GROUPING SETS, ROLLUP, or CUBE clause. Empty Sets To add a grand total row to the result set an empty set, specified as (), can be used. In the example below the last row is generated by the empty set grouping.
select grp_a, grp_b, count(*) from t GROUP BY GROUPING SETS ( (GRP_A, GRP_B), () ) order by grp_a, grp_b ; GRP_A ---------a1 a1 a2 GRP_B COUNT(*) ---------- ---------b1 2 b2 3 b3 3 8
Gotcha - Parentheses without GROUPING SETS Outside a GROUPING SETS clause (or ROLLUP or CUBE clauses) a parenthesized expression like (GRP_A, GRP_B) is no different than the same expression without parentheses. For example this query
select grp_a, grp_b, count(*) from t GROUP BY (GRP_A, GRP_B), GRP_A order by grp_a, grp_b ; GRP_A ---------a1 a1 a2 GRP_B COUNT(*) ---------- ---------b1 2 b2 3 b3 3
returns the same results as this query
select grp_a, grp_b, count(*) from t GROUP BY GRP_A, GRP_B, GRP_A order by grp_a, grp_b ; GRP_A ---------a1 a1 a2 GRP_B COUNT(*) ---------- ---------b1 2 b2 3 b3 3
which in turn has the same result set as this one.

Gotcha: GROUPING SETS with Constants

When I first started using GROUPING SETS I used constants to produce grand total rows in my result sets, like this.
select grp_a , grp_b , count(*) from t GROUP BY GROUPING SETS ( GRP_A, GRP_B, 0 ) order by grp_a, grp_b ; GRP_A GRP_B COUNT(*) ---------- ---------- ---------a1 5 a2 3 b1 2 b2 3 b3 3 8
The last row in the result set is generated by the "0" grouping. I later learnt that an empty set term, "()", was actually a more appropriate syntactic choice than a constant but I continued to use constants out of habit. After all, both approaches seemed to produce the same results.
select grp_a , grp_b , count(*) from t GROUP BY GROUPING SETS ( GRP_A, GRP_B, () ) order by grp_a, grp_b ; GRP_A GRP_B COUNT(*) ---------- ---------- ---------a1 5 a2 3 b1 2 b2 3 b3 3 8
However, I later ran into a case where the two actually produced different results. Query 1
set null '(null)'
Query 2
set null '(null)'
select select grp_a grp_a , grp_b , grp_b , nvl2( grp_b, 1, 0 ) nvl2_grp_b , nvl2( grp_b, 1, 0 ) nvl2_grp_b , count(*) , count(*) from from t t GROUP BY GROUP BY GROUPING SETS ( GRP_A, GRP_B, () ) GROUPING SETS ( GRP_A, GRP_B, 0 ) order by order by grp_a, grp_b grp_a, grp_b ; ; GRP_A -----a1 a2 (null) (null) (null) (null) GRP_B NVL2_GRP_B COUNT(*) ------ ---------- ---------(null) 0 5 (null) 0 3 b1 1 2 b2 1 3 b3 1 3 (null) 0 8 GRP_A -----a1 a2 (null) (null) (null) (null) GRP_B -----(null) (null) b1 b2 b3 (null) NVL2_GRP_B COUNT(*) ---------- ---------(null) 5 (null) 3 1 2 1 3 1 3 0 8
Note how Query 2 returns "(null)" in the NVL2_GRP_B column and Query 1 does not. This is because "0" appears in both the SELECT list and the GROUP BY clause. Readers who want to understand more about why these two queries differ can reverse engineer the two into their
UNION ALL equivalents using the instructions at Reverse Engineering GROUPING BY Queries. Readers who don't simply need to remember this rule of thumb - always use an empty set term to generate a grand total row, do not use a constant.
ROLLUP
It often happens that a query will have a group A which is a superset of group B which in turn is a superset of group C. When aggregates are required at each level a query like this can be used.
set null '(null)' select grp_a , grp_b , grp_c , count(*) from t group by grouping sets ( ( grp_a, grp_b, grp_c ) , ( grp_a, grp_b ) , ( grp_a ) , () ) order by 1, 2, 3 ; GRP_A ---------a1 a1 a1 a1 a1 a1 a2 a2 a2 (null) GRP_B ---------b1 b1 b2 b2 b2 (null) b3 b3 (null) (null) GRP_C COUNT(*) ---------- ---------c1 2 (null) 2 c1 2 c2 1 (null) 3 (null) 5 c2 3 (null) 3 (null) 3 (null) 8
This arrangement is common enough that SQL actually provides a shortcut for specifying these types of GROUPING SETS clauses. It uses the ROLLUP operator. Here is how the query above looks when implemented with ROLLUP.
select grp_a , grp_b , grp_c , count(*) from t
group by ROLLUP( GRP_A, GRP_B, GRP_C ) order by 1, 2, 3 ; GRP_A ---------a1 a1 a1 a1 a1 a1 a2 a2 a2 (null) GRP_B ---------b1 b1 b2 b2 b2 (null) b3 b3 (null) (null) GRP_C COUNT(*) ---------- ---------c1 2 (null) 2 c1 2 c2 1 (null) 3 (null) 5 c2 3 (null) 3 (null) 3 (null) 8
CUBE
There are times when all combinations of a collection of grouping columns are required, as in this query.
set null '(null)' select grp_a , grp_b , grp_c , count(*) from t group by grouping sets ( ( grp_a, grp_b, grp_c ) , ( grp_a, grp_b ) , ( grp_a, grp_c ) , ( grp_b, grp_c ) , ( grp_a ) , ( grp_b ) , ( grp_c ) , () ) order by 1, 2, 3 ; GRP_A ---------a1 a1 a1 a1 a1 GRP_B ---------b1 b1 b2 b2 b2 GRP_C COUNT(*) ---------- ---------c1 2 (null) 2 c1 2 c2 1 (null) 3
a1 a1 a1 a2 a2 a2 a2 (null) (null) (null) (null) (null) (null) (null) (null) (null) (null)
(null) (null) (null) b3 b3 (null) (null) b1 b1 b2 b2 b2 b3 b3 (null) (null) (null)
c1 c2 (null) c2 (null) c2 (null) c1 (null) c1 c2 (null) c2 (null) c1 c2 (null)
4 1 5 3 3 3 3 2 2 2 1 3 3 3 4 4 8
This arrangement is common enough that SQL provides a shortcut called the CUBE operator to implement it. Here is how the query above looks after re-writing it to use CUBE.
select grp_a , grp_b , grp_c , count(*) from t group by CUBE( GRP_A, GRP_B, GRP_C ) order by 1, 2, 3 ; GRP_A ---------a1 a1 a1 a1 a1 a1 a1 a1 a2 a2 a2 a2 (null) (null) (null) (null) (null) (null) (null) (null) GRP_B ---------b1 b1 b2 b2 b2 (null) (null) (null) b3 b3 (null) (null) b1 b1 b2 b2 b2 b3 b3 (null) GRP_C COUNT(*) ---------- ---------c1 2 (null) 2 c1 2 c2 1 (null) 3 c1 4 c2 1 (null) 5 c2 3 (null) 3 c2 3 (null) 3 c1 2 (null) 2 c1 2 c2 1 (null) 3 c2 3 (null) 3 c1 4
(null) (null)
(null) (null)
c2 (null)
4 8
Concatenated Groupings
The last type of grouping shortcut we will examine is called a Concatenated Grouping. With it one can re-write a query like this one, which effectively performs a cross-product of GRP_A with GRP_B and GRP_C,
select grp_a , grp_b , grp_c , count(*) from t group by grouping sets ( ( grp_a, grp_b ) , ( grp_a, grp_c ) ) order by 1, 2, 3 ; GRP_A ---------a1 a1 a1 a1 a2 a2 GRP_B GRP_C COUNT(*) ---------- ---------- ---------b1 2 b2 3 c1 4 c2 1 b3 3 c2 3
into one like this.

set null '(null)' select grp_a , grp_b , grp_c , count(*) from t group by grp_a , grouping sets( grp_b, grp_c ) order by 1, 2, 3 ; GRP_A GRP_B GRP_C COUNT(*) ---------- ---------- ---------- ----------
a1 a1 a1 a1 a2 a2
b1 b2 (null) (null) b3 (null)
(null) (null) c1 c2 (null) c2
2 3 4 1 3 3
The cross-product effect is more apparent when a query like this one
select grp_a , grp_b , grp_c , count(*) from t group by grouping sets ( ( grp_a, grp_c , ( grp_a, grp_d , ( grp_b, grp_c , ( grp_b, grp_d ) order by 1, 2, 3 ; GRP_A ---------a1 a1 a1 a2 a2 (null) (null) (null) (null) (null) (null) (null)
) ) ) )
GRP_B ---------(null) (null) (null) (null) (null) b1 b1 b2 b2 b2 b3 b3
GRP_C COUNT(*) ---------- ---------c1 4 c2 1 (null) 5 c2 3 (null) 3 c1 2 (null) 2 c1 2 c2 1 (null) 3 c2 3 (null) 3
is re-written into one like this.

select grp_a , grp_b , grp_c , count(*) from t group by grouping sets( grp_a, grp_b ) , grouping sets( grp_c, grp_d ) order by
1, 2, 3 GRP_B ---------(null) (null) (null) (null) (null) b1 b1 b2 b2 b2 b3 b3 GRP_C COUNT(*) ---------- ---------c1 4 c2 1 (null) 5 c2 3 (null) 3 c1 2 (null) 2 c1 2 c2 1 (null) 3 c2 3 (null) 3
GRP_A ---------a1 a1 a1 a2 a2 (null) (null) (null) (null) (null) (null) (null)
Personally I have never found the need to use concatenated groupings. I find that specifically listing the desired groupings in a single GROUPING SETS clause or using a single ROLLUP or CUBE operator makes my queries easier to understand and debug. Concatenated groupings can, however, prove useful in data warehouse queries that deal with hierarchical cubes of data. See Concatenated Groupings for more information.
GROUP_ID
Unlike a regular GROUP BY clause, including the same column more than once in a GROUPING SETS clause produces duplicate rows.
select grp_a, count(*) from t GROUP BY GROUPING SETS ( GRP_A, GRP_A ) order by grp_a ; GRP_A COUNT(*) ---------- ---------a1 5 a1 5 a2 3 a2 3 select grp_a, count(*) from t GROUP BY GROUPING SETS ( GRP_A, GRP_A, GRP_A ) order by grp_a ; GRP_A COUNT(*) ---------- ---------a1 5 a1 5 a1 5 a2 3 a2 3 a2 3
The GROUP_ID function can be used to distinguish duplicates from each other.
select grp_a, count(*), GROUP_ID() from t GROUP BY GROUPING SETS ( GRP_A, GRP_A, GRP_A ) order by grp_a, group_id() ; GRP_A COUNT(*) GROUP_ID() ---------- ---------- ---------a1 5 0 a1 5 1 a1 5 2 a2 3 0 a2 3 1 a2 3 2
In the trivial example above it seems there would be little practical use for GROUP_ID. There are times when more complex GROUP BY clauses can return duplicate rows however. It is in such queries that GROUP_ID proves useful. Note that GROUP_ID will always be 0 in a result set that contains no duplicates.
select grp_a, grp_b, count(*), GROUP_ID() from t GROUP BY GROUPING SETS ( GRP_A, GRP_B ) order by grp_a, grp_b ; GRP_A GRP_B COUNT(*) GROUP_ID() ---------- ---------- ---------- ---------a1 5 0 a2 3 0 b1 2 0 b2 3 0 b3 3 0
Grouping by NULL Values

In the examples used thus far in the tutorial our base table had no null values in it. Let's now look at grouping a table that does contain null values.
set null '(null)' select * from t2 order by grp_a, grp_b ; GRP_A ---------A1 A1 A1 A1 A1 GRP_B VAL ---------- ---------X1 10 X2 40 (null) 20 (null) 30 (null) 50
A2
(null)
60
Now consider the following GROUP BY query.

select grp_a, grp_b, count(*) from t2 group by grp_a, grp_b order by grp_a, grp_b ; GRP_A ---------A1 A1 A1 A2 GRP_B COUNT(*) ---------- ---------X1 1 X2 1 (null) 3 (null) 1
So far so good, but let's use GROUPING SETS next and see what happens.
select grp_a, grp_b, count(*) from t2 GROUP BY GROUPING SETS( (GRP_A, GRP_B), GRP_A ) order by grp_a, grp_b ; GRP_A ---------A1 A1 A1 A1 A2 A2 GRP_B COUNT(*) ---------- ---------X1 1 X2 1 (null) 3 (null) 5 (null) 1 (null) 1
We now have two rows with "(null)" under GRP_B for each GRP_A value, one representing the null values stored in T2.GRP_B and the other representing the set of all values in T2.GRP_B. Gotcha - NVL and NVL2 One might expect that NVL() or NVL2 could be used to distinguish the two nulls, like this
select grp_a , NVL( t2.GRP_B, 'n/a' ) AS GRP_B , nvl2( t2.grp_b, 1, 0 ) as test , count(*) from t2 GROUP BY GROUPING SETS( (GRP_A, GRP_B), GRP_A ) order by grp_a, grp_b ; GRP_A ---------A1 A1 A1 A1 A2 GRP_B TEST COUNT(*) ---------- ---------- ---------X1 1 1 X2 1 1 n/a 0 5 n/a 0 3 n/a 0 1
A2
n/a
but this is not the case because functions in the SELECT list operate on an intermediate form of the result set created after the GROUP BY clause is applied, not before. In the next topic we see how the GROUPING function can help us distinguish the two types of nulls.
GROUPING
The GROUPING function tells us whether or not a null in a result set represents the set of all values produced by a GROUPING SETS, ROLLUP, or CUBE operation. A value of "1" tells us it does, a value of "0" tells us it does not. In the output of the following query two of the four nulls represent the set of all GRP_B values.
set null '(null)' select grp_a , grp_b , count(*) , GROUPING( GRP_A ) GROUPING_GRP_A , GROUPING( GRP_B ) GROUPING_GRP_B from t2 group by grouping sets( (grp_a, grp_b), grp_a ) order by 1 , 2 ; GRP_A ---------A1 A1 A1 A1 A2 A2 GRP_B COUNT(*) GROUPING_GRP_A GROUPING_GRP_B ---------- ---------- -------------- -------------X1 1 0 0 X2 1 0 0 (null) 3 0 0 (null) 5 0 1 (null) 1 0 0 (null) 1 0 1
Of course adding a column with zeros and ones to a report isn't the most user friendly way to distinguish grouped values. However, GROUPING can be used with DECODE to insert labels like "Total" into the result set. Here is one example.
select grp_a as "Group A" , decode ( GROUPING( GRP_B ) , 1, 'Total:' , grp_b ) as "Group B" , count(*) as "Count"
from t2 group by grouping sets( (grp_a, grp_b), grp_a ) order by grp_a , GROUPING( GRP_B ) , grp_b ; Group A ---------A1 A1 A1 A1 A2 A2 Group B Count ---------- ---------X1 1 X2 1 (null) 3 Total: 5 (null) 1 Total: 1
Nulls and Aggregate Functions In this topic we explored working with null values in GROUP BY columns. To learn how aggregate functions like COUNT() and SUM() deal with null values in non-GROUP BY columns see Nulls and Aggregate Functions. Gotcha - ORA-00979 When using ORDER BY we need to be careful with the selection of column aliases. For example, say we attempted this query.
select grp_a , decode( grouping( grp_b ), 1, 'Total:', grp_b ) AS GRP_B , count(*) from t2 group by grouping sets( (grp_a, grp_b), grp_a ) order by grouping( GRP_B ) ; , decode( grouping( grp_b ), 1, 'Total:', grp_b ) AS GRP_B * ERROR at line 3: ORA-00979: not a GROUP BY expression
Note how the table has a column called GRP_B and the SELECT list has a column alias also called GRP_B. In the ORDER BY GROUPING( GRP_B ) clause one might expect the "GRP_B" term to refer to the table column, but Oracle interprets it as referring to the column alias, hence the ORA00979 error. To work around the error we can either prefix the column name with its table name
select
grp_a , decode( grouping( grp_b ), 1, 'Total:', grp_b ) AS GRP_B , count(*) from t2 group by grouping sets( (grp_a, grp_b), grp_a ) order by grouping( T2.GRP_B ) ; GRP_A ---------A1 A1 A1 A2 A1 A2 GRP_B COUNT(*) ---------- ---------(null) 3 X1 1 X2 1 (null) 1 Total: 5 Total: 1
or change the column alias.

select grp_a as "Group A" , decode( grouping( grp_b ), 1, 'Total:', grp_b ) AS "Group B" , count(*) as "Count" from t2 group by grouping sets( (grp_a, grp_b), grp_a ) order by grouping( GRP_B ) ; Group A ---------A1 A1 A1 A2 A1 A2 Group B Count ---------- ---------(null) 3 X1 1 X2 1 (null) 1 Total: 5 Total: 1
GROUPING_ID
In the preceding topic we saw how the GROUPING function could be used to identify null values representing the set of all values produced by a GROUPING SETS, ROLLUP, or CUBE operation. What if we wanted to distinguish entire rows from each other? We could use a number of different GROUPING() calls like this
column bit_vector format a10 select TO_CHAR( GROUPING( GRP_A ) ) || TO_CHAR( GROUPING( GRP_B ) ) AS BIT_VECTOR , DECODE ( TO_CHAR( GROUPING( GRP_A ) ) || TO_CHAR( GROUPING( GRP_B ) ) , '01', 'Group "' || GRP_A || '" Total'
, , , ) ,
'10', 'Group "' || GRP_B || '" Total' '11', 'Grand Total' NULL AS LABEL
count(*) from t2 group by grouping sets ( grp_a, grp_b, () order by GROUPING( GRP_A ) , grp_a , GROUPING( GRP_B ) , grp_b ; BIT_VECTOR ---------01 01 10 10 10 11
LABEL COUNT(*) ------------------------ ---------Group "A1" Total 5 Group "A2" Total 1 Group "X1" Total 1 Group "X2" Total 1 Group "" Total 4 Grand Total 6
but if the number of grouping sets were large concatenating all the required GROUPING() terms together would get cumbersome. Fortunately for us the GROUPING_ID function exists. It yields the decimal value of a bit vector (a string of zeros and ones) formed by concatenating all the GROUPING values for its parameters. The following example shows how it works.
select to_char( grouping( grp_a ) ) || to_char( grouping( grp_b ) ) as bit_vector -- this column is only included for clarity , GROUPING_ID( GRP_A, GRP_B ) , grp_a , grp_b , count(*) from t2 group by grouping sets ( grp_a, grp_b, () ) order by GROUPING_ID( GRP_A, GRP_B ) , grp_a , grp_b ; BIT_VECTOR GROUPING_ID(GRP_A,GRP_B) GRP_A GRP_B COUNT(*) ---------- ------------------------ ---------- ---------- ---------01 1 A1 5 01 1 A2 1 10 2 X1 1 10 2 X2 1 10 2 4
11
Here is how we could use GROUPING_ID to streamline our original query.

select DECODE ( GROUPING_ID( GRP_A, GRP_B ) , 1, 'Group "' || GRP_A || '" Total' , 2, 'Group "' || GRP_B || '" Total' , 3, 'Grand Total' , NULL ) AS LABEL , count(*) from t2 group by grouping sets ( grp_a, grp_b, () ) order by GROUPING_ID( GRP_A, GRP_B ) , grp_a , grp_b ; LABEL COUNT(*) ------------------------ ---------Group "A1" Total 5 Group "A2" Total 1 Group "X1" Total 1 Group "X2" Total 1 Group "" Total 4 Grand Total 6
Composite Columns The following example shows how GROUPING_ID works when a composite column, (GRP_A, GRP_B), is included in the GROUPING SETS clause.
select GROUPING_ID( GRP_A, GRP_B ) , grp_a , grp_b , count(*) from t2 group by grouping sets ( (grp_a, grp_b), grp_a, grp_b, () order by 1 , 2 , 3 ; GROUPING_ID(GRP_A,GRP_B) -----------------------0 0 GRP_A ---------A1 A1
GRP_B COUNT(*) ---------- ---------X1 1 X2 1
0 0 1 1 2 2 2 3
A1 A2 A1 A2 X1 X2
3 1 5 1 1 1 4 6
GROUPING_ID and HAVING GROUPING_ID can also be used in the HAVING clause to filter out unwanted groupings. Say, for example, we started with a query like this one
select grouping_id( grp_a, grp_b ) , grp_a , grp_b , count(*) from t2 group by cube( grp_a, grp_b ) order by 1, 2, 3 ; GROUPING_ID(GRP_A,GRP_B) -----------------------0 0 0 0 1 1 2 2 2 3 GRP_A ---------A1 A1 A1 A2 A1 A2 GRP_B COUNT(*) ---------- ---------X1 1 X2 1 3 1 5 1 X1 1 X2 1 4 6
and then we wanted to exclude the empty set grouping (the one with a GROUPING_ID of "3"). We simply add a HAVING clause as follows.
select grouping_id( grp_a, grp_b ) , grp_a , grp_b , count(*) from t2 group by cube( grp_a, grp_b ) HAVING GROUPING_ID( GRP_A, GRP_B ) != 3 order by
1, 2, 3 GRP_A ---------A1 A1 A1 A2 A1 A2 GRP_B COUNT(*) ---------- ---------X1 1 X2 1 3 1 5 1 X1 1 X2 1 4
GROUPING_ID(GRP_A,GRP_B) -----------------------0 0 0 0 1 1 2 2 2
Reverse Engineering GROUPING BY Queries

At times we are faced with a complex GROUP BY query written by someone else and figuring out the equivalent UNION ALL query can help us better understand its results. This is not as easy as it first may seem. A query like this, for example,
set null (null) select grp_a , grp_b , nvl( grp_b, grp_a ) as nvl_grp_a_b , nvl2( grp_b, 1, 0 ) as nvl2_grp_b , count(*) from t GROUP BY ROLLUP ( GRP_A, GRP_B ) order by grp_a , grp_b ; GRP_A ---------a1 a1 a1 a2 a2 (null) GRP_B ---------b1 b2 (null) b3 (null) (null) NVL_GRP_A_ NVL2_GRP_B COUNT(*) ---------- ---------- ---------b1 1 2 b2 1 3 a1 0 5 b3 1 3 a2 0 3 (null) 0 8
is not simply the result of unioning together three identical subqueries with different GROUP BY clauses.
set null '(null)' select grp_a , grp_b , nvl( grp_b, grp_a ) as nvl_grp_a_b
, nvl2( grp_b, 1, 0 ) as nvl2_grp_b , count(*) from t GROUP BY () UNION ALL select grp_a , grp_b , nvl( grp_b, grp_a ) as nvl_grp_a_b , nvl2( grp_b, 1, 0 ) as nvl2_grp_b , count(*) from t GROUP BY ( GRP_A ) UNION ALL select grp_a , grp_b , nvl( grp_b, grp_a ) as nvl_grp_a_b , nvl2( grp_b, 1, 0 ) as nvl2_grp_b , count(*) from t GROUP BY ( GRP_A, GRP_B ) order by grp_a , grp_b ; grp_a * ERROR at line 2: ORA-00979: not a GROUP BY expression
As you can see, such a query produces an error because the first and second subqueries select columns that are not in the GROUP BY clause. To determine the real equivalent UNION query we can use the following algorithm. Step 1 Replace any ROLLUP or CUBE operators with their equivalent GROUPING SETS operator. In our example the query
select grp_a , grp_b , nvl( grp_b, grp_a ) as nvl_grp_a_b , nvl2( grp_b, 1, 0 ) as nvl2_grp_b , count(*) from t GROUP BY ROLLUP ( GRP_A, GRP_B ) order by
grp_a , grp_b ; GRP_A ---------a1 a1 a1 a2 a2 (null) GRP_B ---------b1 b2 (null) b3 (null) (null) NVL_GRP_A_ NVL2_GRP_B COUNT(*) ---------- ---------- ---------b1 1 2 b2 1 3 a1 0 5 b3 1 3 a2 0 3 (null) 0 8
is replaced with
select grp_a , grp_b , nvl( grp_b, grp_a ) as nvl_grp_a_b , nvl2( grp_b, 1, 0 ) as nvl2_grp_b , count(*) from t GROUP BY GROUPING SETS ( () , ( GRP_A ) , ( GRP_A, GRP_B ) ) order by grp_a , grp_b ; GRP_A ---------a1 a1 a1 a2 a2 (null) GRP_B ---------b1 b2 (null) b3 (null) (null) NVL_GRP_A_ NVL2_GRP_B COUNT(*) ---------- ---------- ---------b1 1 2 b2 1 3 a1 0 5 b3 1 3 a2 0 3 (null) 0 8
Step 2a Next start with a query that groups by only the first term in the GROUPING SETS clause, which is an empty set in our example.
select grp_a , grp_b , nvl( grp_b, grp_a ) as nvl_grp_a_b , nvl2( grp_b, 1, 0 ) as nvl2_grp_b , count(*) from t
GROUP BY () ;
If the SELECT list contains columns that are not in the GROUP BY clause then replace those columns with NULL. In the query above both GRP_A and GRP_B are absent from the GROUP BY clause so we replace all occurrences of these columns in the SELECT list with NULL.
column column column column grp_a grp_b nvl_grp_a_b nvl2_grp_b format format format format a6 a6 a11 999999999
select NULL as grp_a , NULL as grp_b , nvl( NULL, NULL ) as nvl_grp_a_b , nvl2( NULL, 1, 0 ) as nvl2_grp_b , count(*) from t GROUP BY () ; GRP_A GRP_B NVL_GRP_A_B NVL2_GRP_B COUNT(*) ------ ------ ----------- ---------- ---------(null) (null) (null) 0 8
Step 2b Now we repeat the first step using the second term in the GROUPING SETS clause, ( GRP_A ).
select grp_a , grp_b , nvl( grp_b, grp_a ) as nvl_grp_a_b , nvl2( grp_b, 1, 0 ) as nvl2_grp_b , count(*) from t GROUP BY ( GRP_A ) ;
This time GRP_B is in the SELECT list but it is not in the GROUP BY list. We therefore need to replace GRP_B with NULL.
select grp_a , NULL as grp_b , nvl( NULL, grp_a ) as nvl_grp_a_b , nvl2( NULL, 1, 0 ) as nvl2_grp_b , count(*) from t GROUP BY ( GRP_A ) ; GRP_A GRP_B NVL_GRP_A_B NVL2_GRP_B COUNT(*) ------ ------ ----------- ---------- ---------a1 (null) a1 0 5
a2
(null) a2
Step 2c For the last set in the GROUPING SETS clause all selected columns are listed in the GROUP BY clause so no further transformation is needed. We can use the original SELECT list as-is.
select grp_a , grp_b , nvl( grp_b, grp_a ) as nvl_grp_a_b , nvl2( grp_b, 1, 0 ) as nvl2_grp_b , count(*) from t GROUP BY ( GRP_A, GRP_B ) ; GRP_A -----a1 a1 a2 GRP_B -----b1 b2 b3 NVL_GRP_A_B NVL2_GRP_B COUNT(*) ----------- ---------- ---------b1 1 2 b2 1 3 b3 1 3
Step 3 The next step is to combine the three step 2 queries with UNION ALL and add an ORDER BY clause.
select NULL as grp_a , NULL as grp_b , nvl( NULL, NULL ) as nvl_grp_a_b , nvl2( NULL, 1, 0 ) as nvl2_grp_b , count(*) from t group by () UNION ALL select grp_a , NULL as grp_b , nvl( NULL, grp_a ) as nvl_grp_a_b , nvl2( NULL, 1, 0 ) as nvl2_grp_b , count(*) from t group by ( grp_a ) UNION ALL select grp_a , grp_b , nvl( grp_b, grp_a ) as nvl_grp_a_b , nvl2( grp_b, 1, 0 ) as nvl2_grp_b , count(*) from
t group by ( grp_a, grp_b ) ORDER BY GRP_A , GRP_B ; GRP_A -----a1 a1 a1 a2 a2 (null) GRP_B -----b1 b2 (null) b3 (null) (null) NVL_GRP_A_B NVL2_GRP_B COUNT(*) ----------- ---------- ---------b1 1 2 b2 1 3 a1 0 5 b3 1 3 a2 0 3 (null) 0 8
Step 4 (Optional) Lastly we reduce expressions like nvl( NULL, NULL ) and nvl2( NULL , 1, 0 ) to simpler, equivalent terms.
select null as grp_a , null as grp_b , NULL AS NVL_GRP_A_B , 0 AS NVL2_GRP_B , count(*) from t group by () union all select grp_a , null as grp_b , GRP_A AS NVL_GRP_A_B , 0 AS NVL2_GRP_B , count(*) from t group by ( grp_a ) union all select grp_a , grp_b , nvl( grp_b, grp_a ) as nvl_grp_a_b , nvl2( grp_b, 1, 0 ) as nvl2_grp_b , count(*) from t group by ( grp_a, grp_b ) order by grp_a , grp_b ; GRP_A GRP_B NVL_GRP_A_B NVL2_GRP_B COUNT(*)
-----a1 a1 a1 a2 a2 (null)
-----b1 b2 (null) b3 (null) (null)
----------- ---------- ---------b1 1 2 b2 1 3 a1 0 5 b3 1 3 a2 0 3 (null) 0 8
Result The end result of the last step is a query which returns the same rows as the original GROUPING SETS query, which is repeated below for your convenience.
select grp_a , grp_b , nvl( grp_b, grp_a ) as nvl_grp_a_b , nvl2( grp_b, 1, 0 ) as nvl2_grp_b , count(*) from t GROUP BY ROLLUP ( GRP_A, GRP_B ) order by grp_a , grp_b ; GRP_A -----a1 a1 a1 a2 a2 (null) GRP_B -----b1 b2 (null) b3 (null) (null) NVL_GRP_A_B NVL2_GRP_B COUNT(*) ----------- ---------- ---------b1 1 2 b2 1 3 a1 0 5 b3 1 3 a2 0 3 (null) 0 8
Setup
Run the code on this page in SQL*Plus to create the sample tables, data, etc. used by the examples in this section.
create table t ( grp_a varchar2(10) , grp_b varchar2(10) , grp_c varchar2(10) , grp_d varchar2(10) , val number ) ; insert insert insert insert insert insert into into into into into into t t t t t t values values values values values values ( ( ( ( ( ( 'a1' 'a1' 'a1' 'a1' 'a1' 'a2' , , , , , , 'b1' 'b1' 'b2' 'b2' 'b2' 'b3' , , , , , , 'c1', 'c1', 'c1', 'c1', 'c2', 'c2', 'd1', 'd1', 'd1', 'd1', 'd1', 'd2', '10' '20' '30' '40' '50' '12' ) ) ) ) ) ) ; ; ; ; ; ;
insert into t values ( 'a2' , 'b3' , 'c2', 'd2', '22' ) ; insert into t values ( 'a2' , 'b3' , 'c2', 'd2', '32' ) ; commit ; create table t2 ( grp_a varchar2(10) , grp_b varchar2(10) , val number ) ; insert insert insert insert insert insert into into into into into into t2 t2 t2 t2 t2 t2 values values values values values values ( ( ( ( ( ( 'A1' 'A1' 'A1' 'A1' 'A1' 'A2' , , , , , , 'X1' 'X2' null null null null , , , , , , '10' '40' '20' '30' '50' '60' ) ) ) ) ) ) ; ; ; ; ; ;
commit ;
Cleanup
Run the code on this page to drop the sample tables, procedures, etc. created in earlier parts of this section. To clear session state changes (e.g. those made by SET, COLUMN, and VARIABLE commands) exit your SQL*Plus session after running these cleanup commands.
drop table t ; drop table t2 ; exit
Hierarchical Data
This section presents various topics related to hierarchical data (also known as "tree structured" data). An example of hierarchical data is shown below.
KEY ---------nls demo mesg server bin config log ctx admin data delx enlx eslx mig PARENT_KEY ---------(null) nls nls (null) server server config server ctx ctx data data data ctx
It is often useful to order and display such rows using the hierarchical relationship. Doing so yields a result set that looks like this (KEY values are indented to highlight the hierarchy).
KEY_INDENTED --------------nls demo mesg server bin config log ctx admin data delx enlx eslx mig KEY_PATH ------------------------/nls /nls/demo /nls/mesg /server /server/bin /server/config /server/config/log /server/ctx /server/ctx/admin /server/ctx/data /server/ctx/data/delx /server/ctx/data/enlx /server/ctx/data/eslx /server/ctx/mig
In this tutorial we explore various Oracle mechanisms for working with hierarchical data.
Connecting Rows
Say we wanted to take the following directory names from a file system and store them in a database table.
/nls /nls/demo /nls/mesg
/server /server/bin /server/config /server/config/log /server/ctx /server/ctx/admin /server/ctx/data /server/ctx/data/delx /server/ctx/data/enlx /server/ctx/data/eslx /server/ctx/mig
To do this we could use a table with a KEY column, which holds the directory name, and a PARENT_KEY column, which connects the directory to its parent directory. (Directory names like these would not typically be used as primary keys. We are bending the rules here for illustrative purposes.)
select * from t ; KEY ---------nls demo mesg server bin config log ctx admin data delx enlx eslx mig PARENT_KEY ---------(null) nls nls (null) server server config server ctx ctx data data data ctx NAME ---------NLS DATA DEMO SERVER BIN CONFIG LOG CTX ADMIN DATA DELX ENLX ESLX MESG
To connect and order the data in this table using the PARENT_KEY hierarchy we can create a Hierarchical Query using the START WITH and CONNECT BY clauses of the SELECT command. START WITH identifies the topmost rows in the hierarchy. CONNECT BY identifies all subsequent rows in the hierarchy. The following snippet returns rows sorted hierarchically, starting from the root rows (those with no parents) on down through to the leaf rows (those with no children).
select key , level from t START WITH parent_key is null CONNECT BY parent_key = prior key ;
KEY LEVEL ---------- -----nls 1 demo 2 mesg 2 server 1 bin 2 config 2 log 3 ctx 2 admin 3 data 3 delx 4 enlx 4 eslx 4 mig 3
The LEVEL pseudocolumn in the previous result indicates which level in the hierarchy each row is at. The topmost level is assigned a LEVEL of 1. To better illustrate hierarchical relationships the LEVEL column is commonly used to indent selected values, like this.
select lpad( ' ', level-1 ) || key as key_indented , level from t START WITH parent_key is null CONNECT BY parent_key = prior key ; KEY_INDENTED LEVEL --------------- -----nls 1 demo 2 mesg 2 server 1 bin 2 config 2 log 3 ctx 2 admin 3 data 3 delx 4 enlx 4 eslx 4 mig 3
The PRIOR operator in hierarchical queries gives us access to column information from the parent of the current row. It can be used outside the CONNECT BY clause if required.
select lpad( ' ', level-1 ) || key as key_indented ,
PRIOR key PRIOR name from t start with parent_key is null connect by parent_key = prior key ; KEY_INDENTED --------------nls demo mesg server bin config log ctx admin data delx enlx eslx mig PRIOR_KEY ---------(null) nls nls (null) server server config server ctx ctx data data data ctx
as prior_key as prior_name
PRIOR_NAME ---------(null) NLS NLS (null) SERVER SERVER CONFIG SERVER CTX CTX DATA DATA DATA CTX
Changing Direction To traverse the tree in the opposite direction, from leaf to root, simply choose a leaf row as the starting point and apply the PRIOR operator to the PARENT_KEY column instead of the KEY column.
select lpad( ' ', level-1 ) || key as key_indented , level from t START WITH KEY = 'delx' connect by key = PRIOR PARENT_KEY ; KEY_INDENTED LEVEL --------------- -----delx 1 data 2 ctx 3 server 4
Gotchas
CONNECT BY conditions are not applied to rows in level 1 of the hierarchy. In the following snippet note how the KEY <> 'delx' condition did not filter out the row with a KEY value of 'delx'.
select lpad( ' ', level-1 ) || key as key_indented , level from t start with key = 'delx' connect by key = PRIOR PARENT_KEY and KEY <> 'delx' ; KEY_INDENTED LEVEL --------------- -----delx 1 data 2 ctx 3 server 4
Order of Operations
The clauses in hierarchical queries are processed in the following order. 1. join conditions (either in the FROM clause or the WHERE clause) 2. START WITH clause 3. CONNECT BY clause 4. WHERE clause conditions that are not joins. The following two snippets demonstrate how this order of operations affects query results when filter conditions are in the WHERE clause versus when they are in the CONNECT BY clause. Filter Condition in WHERE
select lpad(' ', level-1 ) || key as key_indented , level from t WHERE LEVEL != 3 start with key = 'server' connect by parent_key = prior key -;
Filter Condition in CONNECT BY

select lpad(' ', level-1 ) || key as key_indented , level from t --start with key = 'server' CONNECT BY parent_key = prior key and LEVEL != 3 ;
KEY_INDENTED LEVEL --------------- -----server 1 bin 2 config 2 ctx 2 delx 4 enlx 4 eslx 4
KEY_INDENTED LEVEL --------------- -----server 1 bin 2 config 2 ctx 2
Sorting
Since START WITH and CONNECT BY apply a hierarchical sorting scheme to your data, you should generally not use any features that apply other sorting schemes, such as ORDER BY or GROUP BY, in your hierarchical queries. Doing so would negate the need for START WITH and CONNECT BY in the first place. For example, given data with the following hierarchies
KEY_INDENTED --------------nls demo mesg server bin config log ctx admin data delx enlx eslx mig
the ORDER BY clause in the hierarchical query on the left below destroys the hierarchical order. It yields the same results as if CONNECT BY was not used at all.
Hierarchical Query
select key from t start with parent_key is null connect by parent_key = prior key ORDER BY NAME ;
Regular Query
select key from t ----ORDER BY NAME ;
KEY ---------admin bin config ctx data demo delx mesg enlx eslx log mig nls server
KEY ---------admin bin config ctx data demo delx mesg enlx eslx log mig nls server
ORDER SIBLINGS BY Unlike ORDER BY and GROUP_BY, the ORDER SIBLINGS BY clause will not destroy the hierarchical ordering of queries. It allows you to control the sort order of all rows with the same parent (aka "siblings"). The following examples show how ORDER SIBLINGS BY can be used to sort siblings in ascending and descending order respectively.
Ascending Siblings
select lpad(' ', level-1) || key as key_indented from t start with parent_key is null connect by parent_key = prior key ORDER SIBLINGS BY KEY ASC ; KEY_INDENTED --------------nls demo mesg server bin config log ctx admin data delx enlx eslx
Descending Siblings
select lpad(' ', level-1) || key as key_indented from t start with parent_key is null connect by parent_key = prior key ORDER SIBLINGS BY KEY DESC ; KEY_INDENTED --------------server ctx mig data eslx enlx delx admin config log bin nls mesg
mig
demo
Oracle 8i and Earlier The ORDER SIBLINGS BY clause is only available in Oracle version 9i or greater. For earlier versions a custom, recursive PL/SQL function can be used in place of ORDER SIBLINGS BY.
------------------------------------------------------------- Note: --- This function is only for demonstration purposes. -- In a real application more robust code would be needed -- to guard against things like separator characters -- appearing in KEY values, hierarchical loops in the data, -- etc. -----------------------------------------------------------create or replace function KEY_PATH ( p_key t.key%type , p_separator varchar2 default '/' ) return varchar2 is v_parent_key t.parent_key%type ; v_key t.key%type ; begin select into from where parent_key, key v_parent_key, v_key t key = p_key ;
);
if v_parent_key is null then return ( p_separator || v_key ); else return ( KEY_PATH( v_parent_key, p_separator ) || p_separator || v_key end if;
exception when no_data_found then return( null ); end; / show errors No errors.
Ascending Siblings
Descending Siblings
select select lpad(' ', level-1) || key lpad(' ', level-1) || key as key_indented as key_indented
from t start with parent_key is null connect by parent_key = prior key ORDER BY KEY_PATH( KEY, '/' ) ASC ; KEY_INDENTED --------------nls demo mesg server bin config log ctx admin data delx enlx eslx mig
from t start with parent_key is null connect by parent_key = prior key ORDER BY RPAD( KEY_PATH( KEY, '/' ), 50, '~' ) DESC ; KEY_INDENTED --------------server ctx mig data eslx enlx delx admin config log bin nls mesg demo
Gotchas
KEY_PATH's p_separator character should be a character that
1. does not exist in values under T.KEY 2. sorts lower than all characters that exist in T.KEY
For descending siblings the code RPAD( KEY_PATH( KEY, '/' ), 50, '~' ) should use a length larger than any possible KEY_PATH value ("50" in this example) and it should use a padding character that sorts higher than all characters contained in T.KEY ("~" in this example). Violating these rules can result in incorrectly sorted output.
CONNECT_BY_ISLEAF
The CONNECT_BY_ISLEAF pseudocolumn returns 1 if the current row is a leaf of the tree defined by the CONNECT BY condition, 0 otherwise.
select lpad(' ', level-1 ) || key as key_indented , CONNECT_BY_ISLEAF from t start with key = 'server' connect by parent_key = prior key
; KEY_INDENTED CONNECT_BY_ISLEAF --------------- ----------------server 0 bin 1 config 0 log 1 ctx 0 admin 1 data 0 delx 1 enlx 1 eslx 1 mig 1
It is important to recognize that CONNECT_BY_ISLEAF only considers the tree defined by the CONNECT BY condition, not that of the underlying table data. For example, in table T the rows with a KEY of 'config' and 'ctx' have descendents (children and grandchildren) and are therefore not leaf nodes in that context. In the following query however, those same rows are considered leaf nodes (they have a CONNECT_BY_ISLEAF value of 1) because none of the descendents exist in the tree as defined by the CONNECT BY clause. They are filtered out by the " LEVEL <= 2 " condition.
select lpad(' ', level-1 ) || key as key_indented , connect_by_isleaf from t start with key = 'server' CONNECT BY parent_key = prior key and LEVEL <= 2 -- filters out descendents of "config" and "ctx" ; KEY_INDENTED CONNECT_BY_ISLEAF --------------- ----------------server 0 bin 1 config 1 ctx 1
As we saw happen with the LEVEL column in the preceding tutorial, the order of evaluation of the CONNECT BY and WHERE clauses can also affect the behaviour of the CONNECT_BY_ISLEAF pseudo column. The following example illustrates this. In it, " LEVEL <= 2 " is placed in the WHERE clause, not the CONNECT BY clause as above, causing CONNECT_BY_ISLEAF to be 0 for "config" and "ctx" even though those rows look like leaf nodes in the end result.
select lpad(' ', level-1 ) || key as key_indented , connect_by_isleaf
from t WHERE -- filter out descendents of "config" and "ctx", LEVEL <= 2 -- this time using the WHERE clause start with key = 'server' connect by parent_key = prior key ; KEY_INDENTED CONNECT_BY_ISLEAF --------------- ----------------server 0 bin 1 config 0 ctx 0
CONNECT_BY_ROOT
The CONNECT_BY_ROOT operator returns column information from the root row of the hierarchy.
select lpad(' ', level-1 ) || key as key_indented , level , key , name , CONNECT_BY_ROOT key as root_key , CONNECT_BY_ROOT name as root_name from t start with parent_key is null connect by parent_key = prior key ; KEY_INDENTED LEVEL KEY --------------- ------ ---------nls 1 nls demo 2 demo mesg 2 mesg server 1 server bin 2 bin config 2 config log 3 log ctx 2 ctx admin 3 admin data 3 data delx 4 delx enlx 4 enlx eslx 4 eslx mig 3 mig NAME ---------NLS DATA DEMO SERVER BIN CONFIG LOG CTX ADMIN DATA DELX ENLX ESLX MESG ROOT_KEY ---------nls nls nls server server server server server server server server server server server ROOT_NAME ---------NLS NLS NLS SERVER SERVER SERVER SERVER SERVER SERVER SERVER SERVER SERVER SERVER SERVER
Gotchas
The manual page for CONNECT_BY_ROOT states You cannot specify this operator in the START WITH condition or the CONNECT BY condition. While there would be little use for CONNECT_BY_ROOT in the START WITH condition, which already operates on the root row itself, using CONNECT_BY_ROOT in the CONNECT BY condition can be useful and, in practice, actually works in some cases (as tested in Oracle 10g). In the following example we use CONNECT_BY_ROOT in the CONNECT BY condition to prevent any rows beyond level 3 under only the "server" root row from being included in the results.
select lpad(' ', level-1 ) || key as key_indented , level , CONNECT_BY_ROOT key as root_key from t start with parent_key is null connect by parent_key = prior key and not ( level > 3 and connect_by_root key = 'server' ) ; KEY_INDENTED LEVEL ROOT_KEY --------------- ------ ---------nls 1 nls demo 2 nls mesg 2 nls server 1 server bin 2 server config 2 server log 3 server ctx 2 server admin 3 server data 3 server mig 3 server
The fact that the query above contradicts the documentation yet works without error in 10g suggests a bug in either the documentation or the SQL engine. The Gotchas section of topic CONNECT BY LEVEL Method has an example where using CONNECT_BY_ROOT in CONNECT BY does not work so well.
SYS_CONNECT_BY_PATH
The SYS_CONNECT_BY_PATH function returns a single string containing all the column values encountered in the path from root to node.
select lpad(' ', level-1 ) || key name , as key_indented ,
SYS_CONNECT_BY_PATH( key , '/' ) SYS_CONNECT_BY_PATH( name, '/' ) from t start with parent_key is null connect by parent_key = prior key ; KEY_INDENTED --------------nls demo mesg server bin config log ctx admin data delx enlx eslx mig NAME ---------NLS DATA DEMO SERVER BIN CONFIG LOG CTX ADMIN DATA DELX ENLX ESLX MESG
as key_path as name_path
KEY_PATH ------------------------/nls /nls/demo /nls/mesg /server /server/bin /server/config /server/config/log /server/ctx /server/ctx/admin /server/ctx/data /server/ctx/data/delx /server/ctx/data/enlx /server/ctx/data/eslx /server/ctx/mig
NAME_PATH ------------------------/NLS /NLS/DATA /NLS/DEMO /SERVER /SERVER/BIN /SERVER/CONFIG /SERVER/CONFIG/LOG /SERVER/CTX /SERVER/CTX/ADMIN /SERVER/CTX/DATA /SERVER/CTX/DATA/DELX /SERVER/CTX/DATA/ENLX /SERVER/CTX/DATA/ESLX /SERVER/CTX/MESG
SYS_CONNECT_BY_PATH is only available in Oracle version 9i or greater. For earlier versions a custom, recursive PL/SQL function can be used in place of SYS_CONNECT_BY_PATH.
------------------------------------------------------------- Note: --- This function is only for demonstration purposes. -- In a real application more robust code would be needed -- to guard against things like separator characters -- appearing in KEY values, hierarchical loops in the data, -- etc. -----------------------------------------------------------create or replace function KEY_PATH ( p_key t.key%type , p_separator varchar2 default '/' ) return varchar2 is v_parent_key t.parent_key%type ; v_key t.key%type ; begin select parent_key, key from t into where v_parent_key, v_key key = p_key ;
);
if v_parent_key is null then return ( p_separator || v_key ); else return ( KEY_PATH( v_parent_key, p_separator ) || p_separator || v_key end if;
exception when no_data_found then return( null ); end; / show errors No errors. select lpad(' ', level-1 ) || key as key_indented , name , KEY_PATH( key, '/' ) as KEY_PATH from t start with parent_key is null connect by parent_key = prior key ; KEY_INDENTED --------------nls demo mesg server bin config log ctx admin data delx enlx eslx mig NAME ---------NLS DATA DEMO SERVER BIN CONFIG LOG CTX ADMIN DATA DELX ENLX ESLX MESG KEY_PATH ------------------------/nls /nls/demo /nls/mesg /server /server/bin /server/config /server/config/log /server/ctx /server/ctx/admin /server/ctx/data /server/ctx/data/delx /server/ctx/data/enlx /server/ctx/data/eslx /server/ctx/mig
With this approach an additional function would be needed if the path for another column, like NAME, were required. If no hierarchical information, e.g. LEVEL, other than a path is required then the START WITH and CONNECT BY clauses can be omitted since KEY_PATH already knows how to traverse the hierarchy.
select key , name ,
KEY_PATH( key, '/' ) as KEY_PATH from t order by KEY_PATH ; KEY ---------nls demo mesg server bin config log ctx admin data delx enlx eslx mig NAME ---------NLS DATA DEMO SERVER BIN CONFIG LOG CTX ADMIN DATA DELX ENLX ESLX MESG KEY_PATH ------------------------/nls /nls/demo /nls/mesg /server /server/bin /server/config /server/config/log /server/ctx /server/ctx/admin /server/ctx/data /server/ctx/data/delx /server/ctx/data/enlx /server/ctx/data/eslx /server/ctx/mig
Setup
create table ( key parent_key name ); t varchar2(10) , varchar2(10) , varchar2(10) , null , 'nls' , 'nls' null 'server' 'server' 'config' 'server' 'ctx' 'ctx' 'data' 'data' 'data' 'ctx' , 'NLS' , 'DATA' , 'DEMO' , , , , , , , , , , , 'SERVER' 'BIN' 'CONFIG' 'LOG' 'CTX' 'ADMIN' 'DATA' 'DELX' 'ENLX' 'ESLX' 'MESG' ); ); ); ); ); ); ); ); ); ); ); ); ); );
insert into t values ( 'nls' insert into t values ( 'demo' insert into t values ( 'mesg' insert insert insert insert insert insert insert insert insert insert insert commit; column level format 99999 column key_indented format a15 column root_key format a10 into into into into into into into into into into into t t t t t t t t t t t values values values values values values values values values values values ( ( ( ( ( ( ( ( ( ( (
'server', 'bin' , 'config', 'log' , 'ctx' , 'admin' , 'data' , 'delx' , 'enlx' , 'eslx' , 'mig' ,
column root_name column key_path column name_path set null '(null)'
format a10 format a25 format a25
variable v_target_key varchar2(10)
Cleanup
drop table t ; drop function KEY_PATH ; exit
Materialized Views
A Materialized View is effectively a database table that contains the results of a query. The power of materialized views comes from the fact that, once created, Oracle can automatically synchronize a materialized view's data with its source information as required with little or no programming effort. Materialized views can be used for many purposes, including:

Denormalization Validation Data Warehousing Replication.
This tutorial explores materialized view basics. After completing it you should have enough information to use materialized views effectively in simple applications. For more complex applications links at the end of the tutorial will point to information on advanced features not covered here, e.g. partitioning, refresh groups, updatable materialized views. Terminology With relational views the FROM clause objects a view is based on are called "base tables". With materialized views, on the other hand, these objects are either called "detail tables" (in data warehousing documentation) or "master tables" (in replication documentation and the Oracle Database SQL Reference guide). Since SQL Snippets is concerned mainly with relational uses of materialized views we will avoid the contradictory terms "master" and "detail" all together and instead use the term "base tables", thus remaining consistent with relational view terminology. Materialized Views were originally known as "Snapshots" in early releases of Oracle. This keyword is supported for backward compatibility, but should not be used in new code. You may still see this term in some Oracle 11g materialized view error messages.
Views vs Materialized Views

Like its predecessor the view, materialized views allow you to store the definition of a query in the database. Table View
create view v as select * from t ; select * from T ; KEY ---------1 2 3 4 VAL ----a b c select * from V ; KEY ---------1 2 3 4 VAL ----a b c
Materialized View
create materialized view mv as select * from t ; select * from MV ; KEY ---------1 2 3 4 VAL ----a b c
Unlike views, however, materialized views also store the results of the query in the database. In the following queries note how the rowid's for the table and the view are identical, indicating the view returns the exact same data stored in the table. The rowids of the materialized view, on the other hand, differ from those of the table. This indicates the materialized view is returning a physically separate copy of the table data. Table
select rowid from T order by rowid ; ROWID -----------------AAAgY9AAEAAAAVfAAA AAAgY9AAEAAAAVfAAB AAAgY9AAEAAAAVfAAC AAAgY9AAEAAAAVfAAD
View
select rowid from V order by rowid ; ROWID -----------------AAAgY9AAEAAAAVfAAA AAAgY9AAEAAAAVfAAB AAAgY9AAEAAAAVfAAC AAAgY9AAEAAAAVfAAD
Materialized View
select rowid from MV order by rowid ; ROWID -----------------AAAgZFAAEAAADyEAAA AAAgZFAAEAAADyEAAB AAAgZFAAEAAADyEAAC AAAgZFAAEAAADyEAAD
The difference between views and materialized views becomes even more evident than this when table data is updated. Table
update t set val = upper(val); select * from T ; KEY ---------1 2 3 4 VAL ----A B C select * from V ; KEY ---------1 2 3 4 VAL ----A B C select * from MV ; KEY ---------1 2 3 4 VAL ----a b c
View
Materialized View
Note how, after the update, the view data matches the table data but the materialized view data does not. Data in materialized views must be refreshed to keep it synchronized with its base table. Refreshing can either be done manually, as below, or automatically by Oracle in some cases. Table
select * from T ; KEY ---------1 2 3 4 VAL ----A B C
View
select * from V ; KEY ---------1 2 3 4 VAL ----A B C
Materialized View
select * from MV ; KEY ---------1 2 3 4 VAL ----A B C
execute dbms_mview.refresh( 'MV' );
Now that the materialized view has been refreshed its data matches that of its base table.
Cleanup
drop materialized view mv ; drop view v ; update t set val = lower(val); commit;
REFRESH COMPLETE
There are various ways to refresh the data in a materialized view, the simplest way being a complete refresh. When a complete refresh occurs the materialized view's defining query is executed and the entire result set replaces the data currently residing in the materialized view. The REFRESH COMPLETE clause tells Oracle to perform complete refreshes by default when a materialized view is refreshed.
create materialized view mv REFRESH COMPLETE as select * from t ;
Let's see a complete refresh in action now. We will use the DBMS_MVIEW.REFRESH procedure to initiate it. The "list" parameter accepts a list of materialized views to refresh (in our case we only have one) and the "method" parameter accepts a "C", for Complete refresh.
select key, val, rowid from mv ; KEY ---------1 2 3 4 VAL ----a b c ROWID -----------------AAAWgHAAEAAAAIEAAA AAAWgHAAEAAAAIEAAB AAAWgHAAEAAAAIEAAC AAAWgHAAEAAAAIEAAD
execute DBMS_MVIEW.REFRESH( LIST => 'MV', METHOD => 'C' ); select key, val, rowid from mv ; KEY ---------1 2 3 4 VAL ----a b c ROWID -----------------AAAWgHAAEAAAAIEAAE AAAWgHAAEAAAAIEAAF AAAWgHAAEAAAAIEAAG AAAWgHAAEAAAAIEAAH
Note how the rowids in the second query differ from those of the first, even though the data in table T was unchanged throughout. This is because complete refreshes create a whole new set of data, even when the new result set is identical to the old one. If a materialized view contains many rows and the base table's rows change infrequently refreshing the materialized view completely can be an expensive operation. In such cases it would be better to process only the changed rows. We will explore this type of refresh next.
Cleanup
drop materialized view mv ;
Materialized View Logs

As mentioned earlier, complete refreshes of materialized views can be expensive operations. Fortunately there is a way to refresh only the changed rows in a materialized view's base table. This is called fast refreshing. Before a materialized view can perform a fast refresh however it needs a mechanism to capture any changes made to its base table. This mechanism is called a Materialized View Log. We can create a materialized view log on our test table, T, like this.
describe T Name Null? Type -------------------------------------------- ------------------------------------KEY NOT NULL NUMBER VAL VARCHAR2(5) create materialized view log on t ;
Note how the materialized view log is not given a name. This is because a table can only ever have one materialized view log related to it at a time, so a name is not required. To see what a materialized view log looks like we can examine the table used to implement it. In practice developers other than Dizwell never actually need to reference this table, but showing it here helps illustrate materialized view log behaviour.
describe MLOG$_T Name Null? -------------------------------------------- -------KEY SNAPTIME$$ DMLTYPE$$ OLD_NEW$$ CHANGE_VECTOR$$ Type ---------------------NUMBER DATE VARCHAR2(1) VARCHAR2(1) RAW(255)
The MLOG$_T.KEY column mirrors the base table's primary key column T.KEY. The other MLOG$ columns are system generated.
select * from MLOG$_T ; no rows selected
The query above shows that a materialized view log is initially empty upon creation. Rows are automatically added to MLOG$_T when base table T is changed.
UPDATE t set val = upper( val ) where KEY = 1 ;
INSERT into t ( KEY, val ) values ( 5, 'e' ); column dmltype$$ format a10 select key, dmltype$$ from MLOG$_T ; KEY ---------1 5 DMLTYPE$$ ---------U I
If the changes affecting T are rolled back, so are the changes to MLOG$_T.
rollback ; Rollback complete. select key, dmltype$$ from MLOG$_T ; no rows selected
WITH PRIMARY KEY To include the base table's primary key column in a materialized view log the WITH PRIMARY KEY clause can be specified.
drop materialized view log on t ; create materialized view log on t WITH PRIMARY KEY ; desc mlog$_t Name Null? -------------------------------------------- -------KEY SNAPTIME$$ DMLTYPE$$ OLD_NEW$$ CHANGE_VECTOR$$ Type --------------------NUMBER DATE VARCHAR2(1) VARCHAR2(1) RAW(255)
Note how MLOG$_T contains T's primary key column, T.KEY. This materialized view log is equivalent to the one created earlier in this topic, which did not have a WITH clause, because WITH PRIMARY KEY is the default option when no WITH clause is specified. WITH ROWID To include rowids instead of primary keys WITH ROWID can be specified.
drop materialized view log on t ; create materialized view log on t WITH ROWID ;
desc mlog$_t Name Null? -------------------------------------------- -------M_ROW$$ SNAPTIME$$ DMLTYPE$$ OLD_NEW$$ CHANGE_VECTOR$$
Type --------------------VARCHAR2(255) DATE VARCHAR2(1) VARCHAR2(1) RAW(255)
Note how the KEY column was replaced by the M_ROW$$ column, which contains rowids from table T. A materialized view log can also be created with both a rowid and a primary key column.
drop materialized view log on t ; create materialized view log on t WITH ROWID, PRIMARY KEY ; desc mlog$_t Name Null? -------------------------------------------- ------------------------------------KEY M_ROW$$ SNAPTIME$$ DMLTYPE$$ OLD_NEW$$ CHANGE_VECTOR$$ Type NUMBER VARCHAR2(255) DATE VARCHAR2(1) VARCHAR2(1) RAW(255)
In this case both KEY and M_ROW$$ appear in the log table. WITH SEQUENCE A special SEQUENCE column can be include in the materialized view log to help Oracle apply updates to materialized view logs in the correct order when a mix of Data Manipulation (DML) commands, e.g. insert, update and delete, are performed on multiple base tables in a single transaction.
drop materialized view log on t ; create materialized view log on t WITH SEQUENCE ; create materialized view log on t2 WITH SEQUENCE ; INSERT into T values ( 5, 'e' ); INSERT into T2 values ( 60, 3, 300 ); UPDATE T set val = upper(val) where key = 5 ; UPDATE T2 set amt = 333 where key = 60 ; commit; select SEQUENCE$$, key, dmltype$$ from mlog$_T ; SEQUENCE$$ KEY DMLTYPE$$
---------- ---------- ---------60081 5 I 60083 5 U select SEQUENCE$$, key, dmltype$$ from mlog$_T2 ; SEQUENCE$$ KEY DMLTYPE$$ ---------- ---------- ---------60082 60 I 60084 60 U
Since mixed DML is a common occurrence SEQUENCE will be specified in most materialized view logs. In fact, Oracle recommends it. "Oracle recommends that the keyword SEQUENCE be included in your materialized view log statement unless you are sure that you will never perform a mixed DML operation (a combination of INSERT, UPDATE, or DELETE operations on multiple tables)." -- from Creating Materialized Views: Materialized View Logs" WITH Column List The WITH clause can also contain a list of specific base table columns. In the next snippet we include the VAL column.
drop materialized view log on t ; create materialized view log on t WITH ( VAL ); desc mlog$_t Name Null? -------------------------------------------- ------------------------------------KEY VAL SNAPTIME$$ DMLTYPE$$ OLD_NEW$$ CHANGE_VECTOR$$ select * from t ; KEY ---------1 2 3 4 5 VAL ----a b c E Type NUMBER VARCHAR2(5) DATE VARCHAR2(1) VARCHAR2(1) RAW(255)
UPDATE t set val = 'f' where key = 5 ; column old_new$$ format a10
select key, val, old_new$$ from mlog$_t ; KEY VAL OLD_NEW$$ ---------- ----- ---------5 E O
INCLUDING NEW VALUES Clause In the last snippet we see that the VAL column contains values as they existed before the update operation, aka the "old" value. There is no need to store the new value for an update because it can be derived by applying the change vector (a RAW value stored in CHANGE_VECTOR$$, which Oracle uses internally during refreshes) to the old value. In some situations, which we will identify in later topics, it helps to have both the old value and the new value explicitly saved in the materialized view log. We can do that using the INCLUDING NEW VALUES clause, like this.
drop materialized view log on T ; create materialized view log on t with sequence ( VAL ) INCLUDING NEW VALUES ; update t set val = 'g' where key = 5 ; column old_new$$ format a9 select sequence$$, key, val, old_new$$ from mlog$_t order by sequence$$ ; SEQUENCE$$ KEY VAL ---------- ---------- ----60085 5 f 60086 5 g OLD_NEW$$ --------O N
Note how both the old and the new values are stored in the same column, VAL. The OLD_NEW$$ column identifies the value as either an old or a new value. Gotcha - Commas The syntax diagrams for the CREATE MATERIALIZED VIEW LOG command indicate a comma is required between each component of the WITH clause. However this does not appear to be the case when the component is a column list, e.g. "( VAL )".
drop materialized view log on t ; create materialized view log on t with sequence, ( VAL ), primary key ; create materialized view log on t with sequence, ( VAL ), primary key * ERROR at line 1: ORA-00922: missing or invalid option
Omitting the comma before the column list works better.
create materialized view log on t with sequence ( VAL ), primary key; Materialized view log created.
Gotcha - DBMS_REDEFINITION The DBMS_REDEFINITION package has certain restrictions related to materialized view logs. In Oracle 10g these restrictions are: Tables with materialized view logs defined on them cannot be redefined online. For materialized view logs and queue tables, online redefinition is restricted to changes in physical properties. No horizontal or vertical subsetting is permitted, nor are any column transformations. The only valid value for the column mapping string is NULL. -- from Oracle Database Administrator's Guide 10g Release 2 (10.2) - Restrictions for Online Redefinition of Tables In Oracle 11g they are: After redefining a table that has a materialized view log, the subsequent refresh of any dependent materialized view must be a complete refresh. For materialized view logs and queue tables, online redefinition is restricted to changes in physical properties. No horizontal or vertical subsetting is permitted, nor are any column transformations. The only valid value for the column mapping string is NULL. -- from Oracle Database Administrator's Guide 11g Release 1 (11.1) - Restrictions for Online Redefinition of Tables Cleanup
delete t2 ; delete t ; insert into t select * from t_backup ; insert into t2 select * from t2_backup ; commit; drop materialized view log on t ; drop materialized view log on t2 ;
REFRESH FAST
Now that we know how materialized view logs track changes to base tables we can use them to perform fast materialized view refreshes, i.e. refreshes where only the individual materialized view rows affected by base table changes are updated. This is also called "incremental" refreshing. Earlier in this tutorial we saw how the rowids for each row in a materialized view changed after a complete refresh. Now let's see what happens to a materialized view's rowids after a fast refresh. First we use the REFRESH FAST clause to specify that the default refresh method should be fast.
create materialized view log on t with sequence ; create materialized view mv REFRESH FAST as select * from t ; select key, val, rowid from mv ; KEY ---------1 2 3 4 VAL ----a b c ROWID -----------------AAAWm+AAEAAAAaMAAA AAAWm+AAEAAAAaMAAB AAAWm+AAEAAAAaMAAC AAAWm+AAEAAAAaMAAD
Now we refresh the materialized view. The "F" value for the "method" parameter ensures the refresh will be a Fast one.
execute dbms_mview.refresh( list => 'MV', method => 'F' ); select key, val, rowid from mv ; KEY ---------1 2 3 4 VAL ----a b c ROWID -----------------AAAWm+AAEAAAAaMAAA AAAWm+AAEAAAAaMAAB AAAWm+AAEAAAAaMAAC AAAWm+AAEAAAAaMAAD
The rowids did not change. Thus, with a fast refresh the materialized view data is not touched when no changes have been made to the base table, unlike a complete refresh where each row would have been created anew. Now let's update a row in the base table.
update t set val = 'XX' where key = 3 ; commit; execute dbms_mview.refresh( list => 'MV', method => 'F' ); select key, val, rowid from mv ; KEY ---------1 2 3 4 VAL ----a b XX ROWID -----------------AAAWm+AAEAAAAaMAAA AAAWm+AAEAAAAaMAAB AAAWm+AAEAAAAaMAAC AAAWm+AAEAAAAaMAAD
Still no change in the rowids. In row 3 we can see that VAL changed from "c" to "XX" though, telling us that row 3 was updated during the refresh. Defaults The REFRESH FAST clause of the CREATE MATERIALIZED VIEW command tells Oracle what type of refresh to perform when no refresh option is specified. A materialized view created with REFRESH FAST can still be refreshed completely if required though. In the following example note how, even though MV was created above with the REFRESH FAST clause, all its rowids change after the refresh. This indicates that a complete refresh was performed.
execute dbms_mview.refresh( list => 'MV', method => 'C' ); select key, val, rowid from mv ; KEY ---------1 2 3 4 VAL ----a b XX ROWID -----------------AAAWm+AAEAAAAaMAAE AAAWm+AAEAAAAaMAAF AAAWm+AAEAAAAaMAAG AAAWm+AAEAAAAaMAAH
Similarly a materialized view created with REFRESH COMPLETE can be fast refreshed (assuming the materialized view is capable of being fast refreshed, we'll learn more about this later).
drop materialized view mv ; create materialized view mv REFRESH COMPLETE as select * from t ; select key, val, rowid from mv ; KEY ---------1 2 3 4 VAL ----a b XX ROWID -----------------AAAWnBAAEAAAAaMAAA AAAWnBAAEAAAAaMAAB AAAWnBAAEAAAAaMAAC AAAWnBAAEAAAAaMAAD
execute dbms_mview.refresh( list => 'MV', method => 'F' ); select key, val, rowid from mv ; KEY ---------1 2 VAL ----a b ROWID -----------------AAAWnBAAEAAAAaMAAA AAAWnBAAEAAAAaMAAB
3 XX 4
AAAWnBAAEAAAAaMAAC AAAWnBAAEAAAAaMAAD
Note how none of the rowids in MV changed, indicating a fast refresh. Cleanup
drop materialized view mv ; drop materialized view log on t ; update t set val = 'c' where key = 3 ; commit ;
Purging Materialized View Logs

Oracle automatically purges rows in the materialized view log when they are no longer needed. In the example below note how the log table is empty after the refresh.
create materialized view log on t ; create materialized view mv refresh fast as select * from t ; select count(*) from mlog$_t ; COUNT(*) ---------0 insert into t values ( 5, 'e' ) ; commit; select count(*) from mlog$_t ; COUNT(*) ---------1 execute dbms_mview.refresh( list => 'MV', method => 'F' ); select count(*) from mlog$_t ; COUNT(*) ---------0
DBMS_MVEW.PURGE_LOG If a materialized view log needs to be purged manually for some reason a procedure called DBMS_MVEW.PURGE_LOG can be used.
select count(*) from mlog$_t ; COUNT(*) ---------0 update t set val = 'X' where key = 5 ; commit; select count(*) from mlog$_t ; COUNT(*) ---------1 execute DBMS_MVIEW.PURGE_LOG( master => 'T', num => 9999, flag => 'delete' ) ; select count(*) from mlog$_t ; COUNT(*) ---------0
The "num" and "flag" parameters can be used to partially purge the log. See the PURGE_LOG manual page for further details. Once a materialized view log has been purged any materialized views dependent on the deleted rows cannot be fast refreshed. Attempting a fast refresh will raise an error.
execute dbms_mview.refresh( list => 'MV', method => 'F' ); BEGIN dbms_mview.refresh( list => 'MV', method => 'F' ); END; * ERROR at line 1: ORA-12034: materialized view log on "SCOTT"."T" younger than last refresh ORA-06512: at "SYS.DBMS_SNAPSHOT", line 2537 ORA-06512: at "SYS.DBMS_SNAPSHOT", line 2743 ORA-06512: at "SYS.DBMS_SNAPSHOT", line 2712 ORA-06512: at line 1
Such materialized views will need to be refreshed completely.

select * from mv ; KEY ---------1 2 3 4 5 VAL ----a b c e
execute dbms_mview.refresh( list => 'MV', method => 'C' ); select * from mv ;
KEY ---------1 2 3 4 5
VAL ----a b c X
Cleanup
delete from t where key = 5 ; commit; drop materialized view mv ; drop materialized view log on t ;
REFRESH FAST Categories

There are three ways to categorize a materialized view's ability to be fast refreshed. 1. It can never be fast refreshed. 2. It can always be fast refreshed. 3. It can be fast refreshed after certain kinds of changes to the base table but not others. For the first case Oracle will raise an error if you try to create such a materialized view with its refresh method defaulted to REFRESH FAST. In the example below table T does not have a materialized view log on it. Materialized views based on T cannot therefore be fast refreshed. If we attempt to create such a materialized view we get an error.
create materialized view MV REFRESH FAST as select * from t2 ; as select * from t2 * ERROR at line 3: ORA-23413: table "SCOTT"."T2" does not have a materialized view log
For the second case materialized views are created without error, obviously, and will always be fast refreshed unless a complete refresh is explicitly requested. The third case is a little trickier. The next example demonstrates why.
select * from t2 ; KEY T_KEY AMT ---------- ---------- ---------10 1 100 20 1 300 30 1 200 40 2 250
50
150
create materialized view log on t2 with primary key, rowid, sequence ( t_key, amt ) including new values ; create materialized view mv REFRESH FAST as select t_key, max( amt ) amt_max from t2 group by t_key ; select rowid, t_key, amt_max from mv ; ROWID T_KEY AMT_MAX ------------------ ---------- ---------AAAhMzAAEAAAEG8AAA 1 300 AAAhMzAAEAAAEG8AAB 2 250
So far everything works as expected. We created a materialized view log and created a materialized view with fast refresh as its default refresh method. Let's try inserting a row into the base table.
insert into t2 values ( 5, 2, 500 ); commit; execute dbms_mview.refresh( list => 'MV', method => 'F' ); select rowid, t_key, amt_max from mv ; ROWID T_KEY AMT_MAX ------------------ ---------- ---------AAAhMzAAEAAAEG8AAA 1 300 AAAhMzAAEAAAEG8AAB 2 500
Again, it worked as expected. The view was fast refreshed (the rowid's did not change after the DBMS_MVIEW.REFRESH command) and the materialized view correctly shows 500 as the maximum value for rows with T_KEY = 2. Now let's try deleting a row from the base table.
delete from t2 where key = 5 ; commit; execute dbms_mview.refresh( list => 'MV', method => 'F' ); BEGIN dbms_mview.refresh( list => 'MV', method => 'F' ); END; * ERROR at line 1: ORA-32314: REFRESH FAST of "SCOTT"."MV" ORA-06512: at "SYS.DBMS_SNAPSHOT", line ORA-06512: at "SYS.DBMS_SNAPSHOT", line ORA-06512: at "SYS.DBMS_SNAPSHOT", line ORA-06512: at line 1
unsupported after deletes/updates 2255 2461 2430
This time we received an error when we attempted a fast refresh. The reason is because this type of materialized view is an "insert-only" materialized view, i.e. it is only fast refreshable for inserts and direct loads, not updates or deletes. (We will see why it is an insert-only view in the next topic, DBMS_MVIEW.EXPLAIN_MVIEW.) To synchronize an insert-only materialized view after a delete we need to do a complete refresh.
execute dbms_mview.refresh( list => 'MV', method => 'C' ); select rowid, t_key, amt_max from mv ; ROWID T_KEY AMT_MAX ------------------ ---------- ---------AAAhMzAAEAAAEG8AAC 1 300 AAAhMzAAEAAAEG8AAD 2 250
Restrictions on Fast Refresh So how do we know whether a materialized view can be fast refreshed each time, sometimes, or never? One way would be to learn all the documented restrictions for fast refreshable materialized views. Here are some of them. In general materialized views cannot be fast refreshed if the base tables do not have materialized view logs or the defining query: contains an analytic function contains non-repeating expressions like SYSDATE or ROWNUM contains RAW or LONG RAW data types contains a subquery in the SELECT clause contains a MODEL clause contains a HAVING clause contains nested queries with ANY, ALL, or NOT EXISTS contains a CONNECT BY clause references remote tables in different databases references remote tables in a single database and defaults to the ON COMMIT refresh mode references other materialized views which are not join or aggregate materialized views. There are even more restrictions for materialized views containing joins, aggregates, UNION ALL, subqueries, etc. They are documented in various sections of a few different manuals and are too numerous and complex to repeat here. The following links can help you find them if required though.

CREATE MATERIALIZED VIEW - FAST Clause General Restrictions on Fast Refresh Restrictions on Fast Refresh on Materialized Views with Joins Only Restrictions on Fast Refresh on Materialized Views with Aggregates Restrictions on Fast Refresh on Materialized Views with UNION ALL Restrictions for Materialized Views with Subqueries
Restrictions for Materialized Views with Unions Containing Subqueries Restrictions for Using Multitier Materialized Views Restrictions for Materialized Views with Collection Columns
Fortunately there is a second, simpler alternative for determining whether a materialized view is fast refreshable or not. It uses the DBMS_MVIEW.EXPLAIN_MVIEW utility which we will explore next. Cleanup
drop materialized view mv ; drop materialized view log on t2 ;
DBMS_MVIEW.EXPLAIN_MVIEW
As we saw in the preceding topic, predicting whether or not a materialized view is fast refreshable can be complicated. The DBMS_MVIEW.EXPLAIN_MVIEW utility can simplify this task however. Full details on how the utility works are available at the preceding link. The material below will help you use the utility effectively. MV_CAPABILITIES_TABLE There are two ways to get the output from DBMS_MVIEW.EXPLAIN_MVIEW, via a table or via a varray. To use the table method the current schema must contain a table called MV_CAPABILITIES_TABLE. The full, documented CREATE TABLE command for MV_CAPABILITIES_TABLE can be found on UNIX systems at $ORACLE_HOME/rdbms/admin/utlxmv.sql. It is also available in Oracle's documentation at Oracle Database Data Warehousing Guide - Basic Materialized Views - Using MV_CAPABILITIES_TABLE (see Gotcha for a related bug). Here is an abridged version.
create table MV_CAPABILITIES_TABLE ( statement_id varchar(30) , mvowner varchar(30) , mvname varchar(30) , capability_name varchar(30) , possible character(1) , related_text varchar(2000) , related_num number , msgno integer , msgtxt varchar(2000) , seq number ) ;
VARRAY Output Using DBMS_MVIEW.EXPLAIN_MVIEW with the table output method typically involves 1. deleting old rows from MV_CAPABILITIES_TABLE 2. running DBMS_MVIEW.EXPLAIN_MVIEW
3. selecting new rows from MV_CAPABILITIES_TABLE. To save time in this tutorial we will use DBMS_MVIEW.EXPLAIN_MVIEW's varray output option instead and supplement it with a custom function called MY_MV_CAPABILITIES.
create or replace function my_mv_capabilities ( p_mv in varchar2 , p_capability_name_filter in varchar2 default '%' , p_include_pct_capabilities in varchar2 default 'N' , p_linesize in number default 80 ) return clob as -------------------------------------------------------------------------------- From http://www.sqlsnippets.com/en/topic-12884.html --- Parameters: --p_mv -o this value is passed to DBMS_MVIEW.EXPLAIN_MVIEW's "mv" parameter -o it can contain either a query, CREATE MATERIALIZED VIEW command text, -or a materialized view name --p_capability_name_filter -o use either REFRESH, REWRITE, PCT, or the default --p_include_pct_capabilities -Y - capabilities like REFRESH_FAST_PCT are included in the report -N - capabilities like REFRESH_FAST_PCT are not included in the report --p_linesize -o the maximum size allowed for any line in the report output -o data that is longer than this value will be word wrapped --- Typical Usage: --set long 5000 -select my_mv_capabilities( 'MV_NAME' ) as mv_report from dual ; --o the value 5000 is arbitraty; any value big enough to contain the -report output will do -------------------------------------------------------------------------------pragma autonomous_transaction ; v_nl constant char(1) := unistr( '\000A' ); -- new line v_previous_possible char(1) := 'X' ; v_capabilities sys.ExplainMVArrayType ;
v_output clob ; begin dbms_mview.explain_mview( mv => p_mv, msg_array => v_capabilities ) ; for v_capability in ( select capability_name , possible , related_text , msgtxt from table( v_capabilities ) where capability_name like '%' || upper( p_capability_name_filter ) || '%' and not ( capability_name like '%PCT%' and upper(p_include_pct_capabilities) = 'N' ) order by mvowner , mvname , possible desc , seq ) loop ------------------------------------------------------------- print section heading -----------------------------------------------------------if v_capability.possible <> v_previous_possible then v_output := v_output || v_nl || case v_capability.possible when 'T' then 'Capable of: when 'Y' then 'Capable of: when 'F' then 'Not Capable when 'N' then 'Not Capable else v_capability.possible end || v_nl ; end if; v_previous_possible := v_capability.possible ; ------------------------------------------------------------- print section body -----------------------------------------------------------declare
' ' of: ' of: ' || ':'
v_indented_line_size varchar2(3) := to_char( p_linesize - 5 ); begin -- print capability name indented 2 spaces v_output := v_output || v_nl || ' ' || v_capability.capability_name || v_nl ; -- print related text indented 4 spaces and word wrapped if v_capability.related_text is not null then v_output := v_output || regexp_replace ( v_capability.related_text || ' ' , '(.{1,' || v_indented_line_size || '} |.{1,' || v_indented_line_size || '})' , ' \1' || v_nl ) ; end if; -- print message text indented 4 spaces and word wrapped if v_capability.msgtxt is not null then v_output := v_output || regexp_replace ( v_capability.msgtxt || ' ' , '(.{1,' || v_indented_line_size || '} |.{1,' || v_indented_line_size || '})' , ' \1' || v_nl ) ; end if; end; end loop; commit ; return( v_output ); end;
/ show errors No errors.
This completes our preparations. Now let's see DBMS_MVIEW.EXPLAIN_VIEW in action. DBMS_MVIEW.EXPLAIN_MVIEW With a Query DBMS_MVIEW.EXPLAIN_MVIEW can analyze three different types of materialized view code: 1. a defining query 2. a CREATE MATERIALIZED VIEW command 3. an existing materialized view. Here is an example that explains a simple query which could appear as the defining query in a CREATE MATERIALIZED VIEW command.
set long 5000 select my_mv_capabilities( 'SELECT * FROM T', 'REFRESH' ) as mv_report from dual ; MV_REPORT ------------------------------------------------------------------------------Capable of: REFRESH_COMPLETE Not Capable of: REFRESH_FAST REFRESH_FAST_AFTER_INSERT SCOTT.T the detail table does not have a materialized view log REFRESH_FAST_AFTER_ONETAB_DML see the reason why REFRESH_FAST_AFTER_INSERT is disabled REFRESH_FAST_AFTER_ANY_DML see the reason why REFRESH_FAST_AFTER_ONETAB_DML is disabled
(Descriptions of each capability name are available at Table 8-7 CAPABILITY_NAME Column Details. A list of messages and related text is available at Table 8-8 MV_CAPABILITIES_TABLE Column Details.) The EXPLAIN_MVIEW output above shows that fast refresh is not possible in this case because T has no materialized view log.
Note that DBMS_MVIEW.EXPLAIN_MVIEW can report on a materialized view's refresh, rewrite, and partition change tracking (PCT) capabilities. For now we will only examine refresh capabilities. Rewrite capabilities will be covered in Query Rewrite Restrictions and Capabilities. DBMS_MVIEW.EXPLAIN_MVIEW With CREATE MATERIALIZED VIEW Now let's create a materialized view log on T and then use EXPLAIN_MVIEW to explain the capabilities of an entire CREATE MATERIALIZED VIEW command.
create materialized view log on t ; select my_mv_capabilities ( 'CREATE MATERIALIZED VIEW MV REFRESH FAST AS SELECT * FROM T' , 'REFRESH' ) as mv_report from dual ; MV_REPORT ------------------------------------------------------------------------------Capable of: REFRESH_COMPLETE REFRESH_FAST REFRESH_FAST_AFTER_INSERT REFRESH_FAST_AFTER_ONETAB_DML REFRESH_FAST_AFTER_ANY_DML
This time we see that a materialized view using our simple query could be fast refreshable in all cases. DBMS_MVIEW.EXPLAIN_MVIEW With Existing Materialized View For our last example we will explain an existing materialized view, the insert-only one we saw in the preceding topic REFRESH FAST Categories.
create materialized view log on t2 with primary key, rowid, sequence ( t_key, amt ) including new values ; create materialized view mv refresh fast as select t_key, max( amt ) amt_max from t2 group by t_key
; select my_mv_capabilities( 'MV', 'REFRESH' ) as mv_report from dual ; MV_REPORT ------------------------------------------------------------------------------Capable of: REFRESH_COMPLETE REFRESH_FAST REFRESH_FAST_AFTER_INSERT Not Capable of: REFRESH_FAST_AFTER_ONETAB_DML mv uses the MIN or MAX aggregate functions REFRESH_FAST_AFTER_ONETAB_DML COUNT(*) is not present in the select list REFRESH_FAST_AFTER_ANY_DML see the reason why REFRESH_FAST_AFTER_ONETAB_DML is disabled
Here we see that fast refresh is available after inserts, but not other types of DML. Note also that the "REFRESH_FAST" capability will appear whenever at least one of the other REFRESH_FAST_% capabilities is available. It does not mean the materialized view is fast refreshable in all cases. Gotcha Both the $ORACLE_HOME/rdbms/admin/utlxmv.sql file and the CREATE TABLE command at Oracle Database Data Warehousing Guide - Basic Materialized Views - Using MV_CAPABILITIES_TABLE state the values in MV_CAPABILITIES_TABLE.POSSIBLE will either be "T" or "F".
CREATE TABLE MV_CAPABILITIES_TABLE ... POSSIBLE CHARACTER(1), -- T = capability is possible -- F = capability is not possible ...
In actual use we can see the values are really "Y" and "N".
delete from mv_capabilities_table ; execute dbms_mview.explain_mview( 'select * from t' ); commit; column possible format a8
select distinct POSSIBLE from mv_capabilities_table ; POSSIBLE -------Y N
The values "T" and "F" are, however, used when DBMS_MVIEW.EXPLAIN_MVIEW output is saved to a varray. Cleanup
set long 80 drop materialized view mv ; drop materialized view log on t ; drop materialized view log on t2 ;
REFRESH FORCE
In REFRESH FAST Categories and DBMS_MVIEW.EXPLAIN_MVIEW we saw an insert-only materialized view which could be fast refreshed after inserts into the base table but needed a complete refresh after other types of DML. With these types of materialized views it is often most convenient to let Oracle decide which refresh method is best. The REFRESH FORCE method does just that. It performs a FAST refresh if possible, otherwise it performs a COMPLETE refresh.
create materialized view log on t2 with primary key, rowid, sequence ( t_key, amt ) including new values ; create materialized view mv REFRESH FORCE as select t_key, max( amt ) amt_max from t2 group by t_key ; select rowid, t_key, amt_max from mv ; ROWID T_KEY AMT_MAX ------------------ ---------- ---------AAAWpLAAEAAAAaMAAA 1 300 AAAWpLAAEAAAAaMAAB 2 250
First let's try an insert and a refresh.

insert into t2 values ( 5, 2, 500 ); commit; execute dbms_mview.refresh( list => 'MV' );
select rowid, t_key, amt_max from mv ; ROWID T_KEY AMT_MAX ------------------ ---------- ---------AAAWpLAAEAAAAaMAAA 1 300 AAAWpLAAEAAAAaMAAB 2 500
Since the rowids did not change but the AMT_MAX values did we can tell that a FAST refresh was performed. Now let's try a delete followed by a refresh.
delete from t2 where key = 5 ; commit; execute dbms_mview.refresh( list => 'MV' ); select rowid, t_key, amt_max from mv ; ROWID T_KEY AMT_MAX ------------------ ---------- ---------AAAWpLAAEAAAAaMAAC 1 300 AAAWpLAAEAAAAaMAAD 2 250
In the REFRESH FAST Categories topic we received an "ORA-32314: REFRESH FAST of "SCOTT"."MV" unsupported after deletes/updates" error at this point. This time with REFRESH FORCE we did not. Instead Oracle performed a COMPLETE refresh (note how the rowids for each row changed). Cleanup
NEVER REFRESH
If for some reason we need to prevent refresh operations of any sort, FAST or COMPLETE, on our materialized views we can use the NEVER REFRESH method.
create materialized view mv NEVER REFRESH as select * from t ; select * from mv ; KEY ---------1 2 3 4 VAL ----a b c
Let's see what happens when we update the base table and then attempt a refresh.
update t set val = upper(val) ; commit ; execute dbms_mview.refresh( 'MV' ); BEGIN dbms_mview.refresh( 'MV' ); END; * ERROR at line 1: ORA-23538: cannot explicitly refresh a NEVER REFRESH materialized view ("MV") ORA-06512: at "SYS.DBMS_SNAPSHOT", line 2537 ORA-06512: at "SYS.DBMS_SNAPSHOT", line 2743 ORA-06512: at "SYS.DBMS_SNAPSHOT", line 2712 ORA-06512: at line 1
Oracle prevented the refresh by raising an error. I cannot see a practical reason for having a materialized view with NEVER REFRESH set at all times. (If you know of any please let me know using the Comments link below.) NEVER REFRESH can come in handy though when refresh operations on a materialized view need to be prevented temporarily during maintenance or debugging operations. In this case the materialized view's refresh mode can be changed to NEVER REFRESH using the ALTER MATERIALIZED VIEW command. Cleanup
drop materialized view mv ; update t set val = lower(val) ; commit ;
ON DEMAND
Up to this point in the tutorial we have always refreshed our materialized views manually with the DBMS_MVIEW.REFRESH command. This is know as ON DEMAND refreshing and it is the default refresh mode when none is specified in the CREATE MATERIALIZED VIEW command. In other words this
create materialized view mv as select * from t ;
is equivalent to this.
drop materialized view mv ; create materialized view mv REFRESH ON DEMAND as select * from t ;
To refresh ON DEMAND materialized views we explicitly call one of the following procedures.
DBMS_MVIEW.REFRESH DBMS_MVIEW.REFRESH_ALL_MVIEWS DBMS_MVIEW.REFRESH_DEPENDENT
Here is an example that uses DBMS_MVIEW.REFRESH.

insert into t values ( 5, 'e' ); commit; select * from mv where key = 5 ; no rows selected execute DBMS_MVIEW.REFRESH( 'MV' ); select * from mv where key = 5 ; KEY VAL ---------- ----5 e
Cleanup
drop materialized view mv ; delete from t where key = 5 ; commit;
ON COMMIT
In some situations it would be convenient to have Oracle refresh a materialized view automatically whenever changes to the base table are committed. This is possible using the ON COMMIT refresh mode. Here is an example.
create materialized view log on t ; create materialized view mv REFRESH FAST ON COMMIT as select * from t ; select rowid, key, val from mv ; ROWID KEY VAL ------------------ ---------- ----AAAXNGAAEAAAAasAAA 1 a AAAXNGAAEAAAAasAAB 2 b AAAXNGAAEAAAAasAAC 3 c AAAXNGAAEAAAAasAAD 4
Let's see what happens to the view in the course of an insert operation.
insert into t values ( 5, 'e' );
select rowid, key, val from mv ; ROWID KEY VAL ------------------ ---------- ----AAAXNGAAEAAAAasAAA 1 a AAAXNGAAEAAAAasAAB 2 b AAAXNGAAEAAAAasAAC 3 c AAAXNGAAEAAAAasAAD 4
Nothing happend yet. Let's issue a COMMIT.

commit; select rowid, key, val from mv ; ROWID KEY VAL ------------------ ---------- ----AAAXNGAAEAAAAasAAA 1 a AAAXNGAAEAAAAasAAB 2 b AAAXNGAAEAAAAasAAC 3 c AAAXNGAAEAAAAasAAD 4 AAAXNGAAEAAAAatAAA 5 e
Note how the materialized view was automatically fast refreshed after the COMMIT command. No call to DBMS_MVIEW.REFRESH was required. Restrictions Materialized views can only refresh ON COMMIT in certain situations. 1. The materialized view cannot contain object types or Oracle-supplied types. 2. The base tables will never have any distributed transactions applied to them. The first case produces an error during the CREATE MATERIALIZED VIEW command.
-- this materialized view is not fast refreshable -- because the materialized view contains an Oracle-supplied type create materialized view mv2 REFRESH FAST ON COMMIT as select key, val, sys_xmlgen( val ) as val_xml from t ; as select key, val, sys_xmlgen( val ) as val_xml from t * ERROR at line 3: ORA-12054: cannot set the ON COMMIT refresh attribute for the materialized view
The second case generates an error when a distributed transaction is attempted on the base table. In the following example materialized view MV (created at the top of this page) was created with REFRESH FAST. Attempting a distributed transaction on its base table, T, will therefore raise an error.
insert into t select key+10, val from T@REMOTE ; commit; commit * ERROR at line 1: ORA-02050: transaction 5.21.5632 rolled back, some remote DBs may be in-doubt ORA-02051: another session in same transaction failed
(REMOTE is a database link which loops back to the current account.) ON DEMAND materialized views have no such restriction, as the following snippet demonstrates.
alter materialized view mv refresh ON DEMAND ; insert into t select key+10, val from T@REMOTE ; commit; select * from t ; KEY ---------1 2 3 4 5 11 12 13 14 15 VAL ----a b c e a b c e
-- cleanup test data in preparation for next section delete from t where key >= 5 ; commit ;
Gotcha The SQL Language Reference manual says this about the ON COMMIT clause. "Specify ON COMMIT to indicate that a fast refresh is to occur whenever the database commits a transaction that operates on a master table of the materialized view." -- Oracle Database SQL Language Reference: CREATE MATERIALIZED VIEW When I first read this I assumed it meant that "REFRESH COMPLETE ON COMMIT" is not allowed. I also assumed that specifying "REFRESH ON COMMIT" is equivalent to specifying "REFRESH FAST ON COMMIT". The following examples prove neither is correct however.
create materialized view mv2 REFRESH COMPLETE ON COMMIT as select key, val from t ;
As we can see the CREATE MATERIALZIED view command succeeded even though COMPLETE, not FAST, was specified with ON COMMIT. The next example examines the behavior of "REFRESH ON COMMIT" without a specified refresh method.
drop materialized view log on t ; -- fast refreshable materialized views on T can no longer be created on T -- because it has no materialized view log drop materialized view mv2 ; create materialized view mv2 REFRESH ON COMMIT as select key, val from t ; select rowid, key, val from mv2 ; ROWID KEY VAL ------------------ ---------- ----AAAXNMAAEAAAAakAAA 1 a AAAXNMAAEAAAAakAAB 2 b AAAXNMAAEAAAAakAAC 3 c AAAXNMAAEAAAAakAAD 4 insert into t values ( 5, 'e' ); commit ; select rowid, key, val from mv2 ; ROWID KEY VAL ------------------ ---------- ----AAAXNMAAEAAAAakAAE 1 a AAAXNMAAEAAAAakAAF 2 b AAAXNMAAEAAAAakAAG 3 c AAAXNMAAEAAAAakAAH 4 AAAXNMAAEAAAAakAAI 5 e
The fact that all the rowid's in MV2 changed after the INSERT transaction committed confirms that a complete refresh took place during the commit. "REFRESH ON COMMIT" is not therefore equivalent to "REFRESH FAST ON COMMIT". In fact, when no REFRESH method is specified the default behaviour is "REFRESH FORCE" regardless of whether ON COMMIT is used or not. Given these observations I can only conclude the documentation is either in error or misleading when it says "specify ON COMMIT to indicate that a fast refresh is to occur". Cleanup
drop materialized view mv ; drop materialized view mv2 ; delete from t where key >= 5 ; commit ;
Constraints
System Generated Constraints When a materialized view is created Oracle may add system generated constraints to its underlying table (i.e. the table containing the results of the query, not to be confused with a base table). In the following example note how Oracle automatically adds a primary key constraint to the table called "MV", which is part of the materialized view also called "MV".
create materialized view mv as select key, val from t ; column constraint_name format a20 column constraint_type format a15 column index_name format a15 select constraint_name, constraint_type, index_name from user_constraints where TABLE_NAME = 'MV' ; CONSTRAINT_NAME CONSTRAINT_TYPE INDEX_NAME -------------------- --------------- --------------SYS_C0019948 P SYS_C0019948
In the next example Oracle automatically adds a check constraint.

drop materialized view mv ; describe t2 Name ------------------------------------------------------------------------KEY T_KEY AMT Null? Type -------NOT NULL NUMBER NOT NULL NUMBER NOT NULL NUMBER
create materialized view log on t2 with primary key, rowid, sequence ( t_key, amt ) including new values ; create materialized view mv refresh fast on commit as select t_key, count(*) row_count from t2 group by t_key ; column search_condition format a30 select constraint_name, constraint_type, search_condition
from where
user_constraints table_name = 'MV' ;
CONSTRAINT_NAME CONSTRAINT_TYPE SEARCH_CONDITION -------------------- --------------- -----------------------------SYS_C0019949 C "T_KEY" IS NOT NULL
Adding Your Own Constraints If necessary we can create our own constraints on materialized view tables in addition to the ones Oracle may add. When the materialized view is in ON COMMIT mode these constraints effectively constrain the materialized view's base tables. Let's see this in action by creating a check constraint on MV.
select * from t2 ; KEY T_KEY AMT ---------- ---------- ---------10 1 100 20 1 300 30 1 200 40 2 250 50 2 150 alter table mv -- note we used "alter table" here add CONSTRAINT MY_CONSTRAINT CHECK ( ROW_COUNT <= 3 ) DEFERRABLE ; select constraint_name, constraint_type, search_condition from user_constraints where table_name = 'MV' ; CONSTRAINT_NAME -------------------SYS_C0019949 MY_CONSTRAINT CONSTRAINT_TYPE SEARCH_CONDITION --------------- -----------------------------C "T_KEY" IS NOT NULL C ROW_COUNT <= 3
Now any attempt to create more than 3 rows per group in table T2 will generate an error at commit time.
insert into T2 values ( 5, 1, 500 ); commit; commit * ERROR at line 1: ORA-12008: error in materialized view refresh path ORA-02290: check constraint (SCOTT.MY_CONSTRAINT) violated
Implementing multirow validation rules such as this one properly is not possible using check constraints on regular tables. Implementing them using triggers can be difficult if not impossible. With materialized views they are declared using a few lines of code and are virtually bullet proof when applied correctly. We will learn more about this powerful multirow validation approach in a
future SQL Snippets tutorial so stay tuned! In the mean time Ask Tom "Declarative Integrity" has some good information on the subject. Gotcha When we created MY_CONSTRAINT above we use an ALTER TABLE command. Curiously enough an ALTER MATERIALIZED VIEW command would have worked too.
ALTER MATERIALIZED VIEW mv add constraint my_second_constraint check ( row_count < 4 ) deferrable ; select constraint_name, constraint_type, search_condition from user_constraints where table_name = 'MV' ; CONSTRAINT_NAME -------------------SYS_C0019949 MY_CONSTRAINT MY_SECOND_CONSTRAINT CONSTRAINT_TYPE SEARCH_CONDITION --------------- -----------------------------C "T_KEY" IS NOT NULL C ROW_COUNT <= 3 C row_count < 4
The Oracle manual page for ALTER MATERIALIZED VIEW however does not indicate that constraints can be added this way. Until the documentation says this is legal it is best to use ALTER TABLE. Cleanup
Indexes
When a materialized view is created Oracle may add system generated indexes to its underlying table (i.e. the table containing the results of the query, not to be confused with a base table). In the following example note how Oracle automatically adds an index to implement the system generated primary key we saw in the preceding topic, Constraints.
create materialized view mv as select key, val from t ; column index_name format a15 column column_name format a15 select index_name , i.uniqueness , ic.column_name from user_indexes i inner join user_ind_columns ic
using ( index_name ) where i.table_name = 'MV' ; INDEX_NAME UNIQUENES COLUMN_NAME --------------- --------- --------------SYS_C0019959 UNIQUE KEY
In the next example Oracle automatically generates a function based index.

drop materialized view mv ; create materialized view log on t2 with primary key, rowid, sequence ( t_key, amt ) including new values ; create materialized view mv refresh fast on commit as select t_key, COUNT(*) ROW_COUNT from t2 group by t_key ; column column_expression format a35 select index_name , i.uniqueness , ic.column_name , ie.column_expression from user_indexes i inner join user_ind_columns ic left outer join user_ind_expressions ie using ( index_name ) using ( index_name ) where ic.table_name = 'MV' ; INDEX_NAME UNIQUENES COLUMN_NAME COLUMN_EXPRESSION --------------- --------- --------------- ----------------------------------I_SNAP$_MV UNIQUE SYS_NC00003$ SYS_OP_MAP_NONNULL("T_KEY")
(Note that SYS_OP_MAP_NONNULL is an undocumented Oracle function. Do not attempt to use it in your own code. See Nulls and Equality: SQL Only for additional info.) Adding Your Own Indexes We can add out own indexes to MV just as we would a regular table. In the following example we will add an index on the T_KEY column.
create index MY_INDEX on mv ( T_KEY ) ; select index_name , i.uniqueness , ic.column_name from user_indexes i inner join user_ind_columns ic using ( index_name ) where i.table_name = 'MV' ; INDEX_NAME --------------I_SNAP$_MV MY_INDEX UNIQUENES --------UNIQUE NONUNIQUE COLUMN_NAME --------------SYS_NC00003$ T_KEY
To confirm that Oracle uses our index in queries let's turn SQL*Plus's Autotrace feature on and execute a query.
set autotrace on explain set linesize 95 select * from mv where t_key = 2 ; T_KEY ROW_COUNT ---------- ---------2 2 Execution Plan ---------------------------------------------------------Plan hash value: 2793437614 -----------------------------------------------------------------------------------------| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time | -----------------------------------------------------------------------------------------| 0 | SELECT STATEMENT | | 1 | 26 | 2 (0)| 00:00:01 | | 1 | MAT_VIEW ACCESS BY INDEX ROWID| MV | 1 | 26 | 2 (0)| 00:00:01 | |* 2 | INDEX RANGE SCAN | MY_INDEX | 1 | | 1 (0)| 00:00:01 | -----------------------------------------------------------------------------------------Predicate Information (identified by operation id): --------------------------------------------------2 - access("T_KEY"=2)
Note ----- dynamic sampling used for this statement
Note how the optimizer chose an INDEX RANGE SCAN from MY_INDEX in step 2. Cleanup
ENABLE QUERY REWRITE

Materialized views can be useful for pre-calculating and storing derived values such as AMT_MAX in the following snippet.
create materialized view log on t2 with primary key, rowid, sequence ( t_key, amt ) including new values ; create materialized view mv refresh fast on commit as select t_key, MAX( AMT ) AMT_MAX from t2 group by t_key ;
Such materialized views make queries like this

select t_key, amt_max FROM MV order by t_key ; T_KEY AMT_MAX ---------- ---------1 300 2 250
faster than its equivalent query.
select t_key, max( amt ) as amt_max FROM T2 group by t_key order by t_key ; T_KEY AMT_MAX ---------- ---------1 300 2 250
Wouldn't it be nice if Oracle could use the information in MV to resolve this last query too? If your database has a feature called Query Rewrite available and enabled this happens automatically. To see it in action we first need to make the materialized view available to Query Rewrite like this.
alter materialized view mv ENABLE QUERY REWRITE ;
(See Gotcha - ORA-00439 below if you encounter an ORA-00439 error at this step.) Note that materialized views which do not include the ENABLE QUERY REWRITE clause will have Query Rewrite disabled by default. Next we collect statistics on the materialized view to help Oracle optimize the query rewrite process.
execute dbms_stats.gather_table_stats( user, 'MV' ) ;
Finally we can confirm Oracle will use the materialized view in queries by turning SQL*Plus's Autotrace feature on.
set autotrace on explain set linesize 95 select t_key, max( amt ) as amt_max FROM T2 group by t_key order by t_key ; T_KEY AMT_MAX ---------- ---------1 300 2 250 Execution Plan ---------------------------------------------------------Plan hash value: 446852971 ------------------------------------------------------------------------------------| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time | ------------------------------------------------------------------------------------| 0 | SELECT STATEMENT | | 2 | 14 | 4 (25)| 00:00:01 | | 1 | SORT ORDER BY | | 2 | 14 | 4 (25)| 00:00:01 | | 2 | MAT_VIEW REWRITE ACCESS FULL| MV | 2 | 14 | 3 (0)| 00:00:01 | -------------------------------------------------------------------------------------
Note how the optimizer chose to access MV for its pre-calculated MAX(AMT) values in line 2 even though the query itself made no mention of MV. Without the Query Rewrite feature the execution plan would look like this.
alter session set QUERY_REWRITE_ENABLED = FALSE ; select t_key, max( amt ) as amt_max FROM T2 group by t_key order by t_key ; T_KEY AMT_MAX ---------- ---------1 300 2 250 Execution Plan ---------------------------------------------------------Plan hash value: 50962384 --------------------------------------------------------------------------| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time | --------------------------------------------------------------------------| 0 | SELECT STATEMENT | | 5 | 130 | 4 (25)| 00:00:01 | | 1 | SORT GROUP BY | | 5 | 130 | 4 (25)| 00:00:01 | | 2 | TABLE ACCESS FULL| T2 | 5 | 130 | 3 (0)| 00:00:01 | --------------------------------------------------------------------------Note ----- dynamic sampling used for this statement
Note how the optimizer chose to access T2 this time. Each time this query is executed it has to recalculate MAX(VAL) from the information in T2 for each group, a more expensive approach than simply selecting pre-calculated column values from MV is. Gotcha - ORA-00439 The materialized view query rewrite feature is not available in Oracle XE and some other Oracle configurations. If you attempt to use ENABLE QUERY REWITE in an Oracle database where the feature is not enabled you will receive an ORA-00439 error.
create materialized view mv2 refresh fast on commit ENABLE QUERY REWRITE as select t_key , count(*) as row_count , count(amt) as amt_count from t2 group by t_key ;
from t2 * ERROR at line 9: ORA-00439: feature not enabled: Materialized view rewrite
Cleanup
alter session set query_rewrite_enabled = true ; set autotrace off drop materialized view mv ; drop materialized view log on t2 ;
Query Rewrite Restrictions and Capabilities

Restrictions Materialized views with the following characteristics cannot have query rewrite enabled: the defining query references functions which are not DETERMINISTIC an expression in the defining query is not repeatable; e.g. an expression containing the USER pseudo column or the SYSTIMESTAMP function. Attempting to violate these restrictions results in an error.
create materialized view mv ENABLE QUERY REWRITE as select key, val, USER from t ; as select key, val, USER from t * ERROR at line 3: ORA-30353: expression not supported for query rewrite
Capabilities A few different materialized view query rewrite capabilities exist. In EXPLAIN_MVIEW we used a utility called MY_MV_CAPABILITIES to explore a materialized view's refresh capabilities. In the snippets below we will use this same utility to explore rewrite capabilities. First lets look at a simple, single table materialized view with query rewrite disabled.
create materialized view mv DISABLE QUERY REWRITE as select key, val from t ; set long 5000 select my_mv_capabilities( 'MV', 'REWRITE' ) as mv_report from dual ; MV_REPORT
------------------------------------------------------------------------------Not Capable of: REWRITE REWRITE_FULL_TEXT_MATCH query rewrite is disabled on the materialized view REWRITE_PARTIAL_TEXT_MATCH query rewrite is disabled on the materialized view REWRITE_GENERAL query rewrite is disabled on the materialized view
This materialized view obviously has no rewrite capabilities available to it. (Descriptions of each capability name are available at Table 8-7 CAPABILITY_NAME Column Details.) Enabling query rewrite on the materialized view changes this.
alter materialized view mv ENABLE QUERY REWRITE ; select my_mv_capabilities( 'MV', 'REWRITE' ) as mv_report from dual ; MV_REPORT ------------------------------------------------------------------------------Capable of: REWRITE REWRITE_FULL_TEXT_MATCH REWRITE_PARTIAL_TEXT_MATCH REWRITE_GENERAL
Now all rewrite capabilities are available. If the materialized view happened to referenced a remote table then some rewrite capabilities would be available, but not others.
drop materialized view mv ; create materialized view mv enable query rewrite as select key, val from T@REMOTE ; select my_mv_capabilities( 'MV', 'REWRITE' ) as mv_report from dual ; MV_REPORT -------------------------------------------------------------------------------
Capable of: REWRITE REWRITE_PARTIAL_TEXT_MATCH REWRITE_GENERAL Not Capable of: REWRITE_FULL_TEXT_MATCH T mv references a remote table or view in the FROM list
Cleanup
drop materialized view mv ;
Join Queries
So far in this tutorial we have only seen materialized views based on a single table. Materialized views can also be created on multi-table queries to store the pre-calculated results of expensive join operations. Here is a simple example.
select * from t ; KEY ---------1 2 3 4 VAL ----a b c
select * from t2 ; KEY T_KEY AMT ---------- ---------- ---------10 1 100 20 1 300 30 1 200 40 2 250 50 2 150 create materialized view mv as select t.key t_key , t.val t_val , t2.key t2_key , t2.amt t2_amt from t, t2 where t.key = t2.t_key ;
select t_key, t_val, t2_key, t2_amt from mv ; T_KEY ---------1 1 1 2 2 T_VAL T2_KEY T2_AMT ----- ---------- ---------a 10 100 a 20 300 a 30 200 b 40 250 b 50 150
REFRESH FAST For a materialized view with only joins (no aggregates, unions, subqueries, etc.) to be fast refreshable certain restrictions beyond the General Restrictions on Fast Refresh must be met. These additional restrictions are: materialized view logs with rowids must exist for all of the defining query's base tables the SELECT clause cannot contain object type columns the defining query cannot have a GROUP BY clause or aggregates rowid columns for each table instance in the FROM clause must appear in the SELECT clause. -- from Restrictions on Fast Refresh on Materialized Views with Joins Only In addition to these restrictions there are some recommended practices for using join queries. They are as follows. "If a materialized view contains joins but no aggregates, then having an index on each of the join column rowids in the detail table will enhance refresh performance greatly, because this type of materialized view tends to be much larger than materialized views containing aggregates." -- from Refreshing Materialized Views: Tips for Refreshing Materialized Views Without Aggregates "After you create the materialized view, you must collect statistics on it using the DBMS_STATS package. Oracle Database needs the statistics generated by this package to optimize query rewrite." -- from CREATE MATERIALIZED VIEW. The Prototype Applying these restrictions and recommendations to our test case above yields the following prototypical materialized view with joins. Whenever I need to create this type of materialized view in an application I use the code below as a starting point to remind me of the requirements.
drop materialized view mv ; create materialized view log on t with rowid, sequence ; create materialized view log on t2 with rowid, sequence ; create materialized view mv
refresh fast on commit enable query rewrite as select t.key t_key t.val t_val t2.key t2_key t2.amt t2_amt t.rowid t_row_id t2.rowid t2_row_id from t, t2 where t.key = t2.t_key ;
, , , , ,
create index mv_i1 on mv ( t_row_id ) ; create index mv_i2 on mv ( t2_row_id ) ; execute dbms_stats.gather_table_stats( user, 'MV' ) ;
Whenever we create a fast refreshable view we should use our EXPLAIN_MVIEW utility, MY_MV_CAPABILITIES, to confirm it can be refreshed in all required situations.
set long 5000 select my_mv_capabilities( 'MV', 'REFRESH' ) as mv_report from dual ; MV_REPORT ------------------------------------------------------------------------------Capable of: REFRESH_COMPLETE REFRESH_FAST REFRESH_FAST_AFTER_INSERT REFRESH_FAST_AFTER_ONETAB_DML REFRESH_FAST_AFTER_ANY_DML
Now let's test drive our new MV. First, here are MV's initial contents.
select t_key, t_val, t2_key, t2_amt from mv ; T_KEY ---------1 1 1 2 T_VAL T2_KEY T2_AMT ----- ---------- ---------a 10 100 a 20 300 a 30 200 b 40 250
2 b
50
150
Now let's do some DML on both base tables and see the effect on MV.
insert into t2 values ( 60, 3, 300 ) ; update t set val = upper(val) ; commit; select t_key, t_val, t2_key, t2_amt from mv ; T_KEY ---------1 1 1 2 2 3 T_VAL T2_KEY T2_AMT ----- ---------- ---------A 10 100 A 20 300 A 30 200 B 40 250 B 50 150 C 60 300
Both changes are reflected in MV, as expected. Query Rewrite Materialized views containing joins can be used by the query rewrite facility (see ENABLE QUERY REWRITE).
select my_mv_capabilities( 'MV', 'REWRITE' ) as mv_report from dual ; MV_REPORT ------------------------------------------------------------------------------Capable of: REWRITE REWRITE_FULL_TEXT_MATCH REWRITE_PARTIAL_TEXT_MATCH REWRITE_GENERAL
Gotcha - ANSI Join Syntax When we attempt to create a materialized view with the ANSI join syntax equivalent of the defining query used above we are surprisingly rewarded with an ORA error.
create materialized view mv2 refresh fast as
select t.key t_key t.val t_val t2.key t2_key t2.amt t2_amt t.rowid t_row_id t2.rowid t2_row_id from T INNER JOIN T2 ON
, , , , , ( T.KEY = T2.T_KEY )
T INNER JOIN T2 ON ( T.KEY = T2.T_KEY ) * ERROR at line 12: ORA-12015: cannot create a fast refresh materialized view from a complex query
While this behaviour appears to be a bug at first glance, Metalink note 420856.1 explains that it is really an undocumented limitation of fast refresh materialized views. An examination of the EXPLAIN_MVIEW results for this case points to some behind-the-scenes transformations with ANSI syntax which may be causing the limitation.
select my_mv_capabilities ( 'create materialized view mv2 refresh fast as select t.key t_key , t.val t_val , t2.key t2_key , t2.amt t2_amt , t.rowid t_row_id , t2.rowid t2_row_id from T INNER JOIN T2 ON ( T.KEY = T2.T_KEY )' , 'REFRESH_FAST_AFTER_INSERT' ) as mv_report from dual ; MV_REPORT ------------------------------------------------------------------------------Not Capable of: REFRESH_FAST_AFTER_INSERT inline view or subquery in FROM list not supported for this type MV REFRESH_FAST_AFTER_INSERT inline view or subquery in FROM list not supported for this type MV REFRESH_FAST_AFTER_INSERT view or subquery in from list
Cleanup
drop materialized view mv ; drop materialized view log on t ; drop materialized view log on t2 ; delete t2 ; delete t ; insert into t select * from t_backup ; insert into t2 select * from t2_backup ; commit;
Aggregate Queries
In addition to materialized views based on join queries, materialized views containing aggregate functions are also possible. Here is a simple example.
select * from t2 order by key ; KEY T_KEY AMT ---------- ---------- ---------10 1 100 20 1 300 30 1 200 40 2 250 50 2 150 create materialized view mv as select t_key t_key , SUM(AMT) AMT_SUM from t2 group by t_key ; select * from mv order by t_key ; T_KEY AMT_SUM ---------- ---------1 600 2 400
REFRESH FAST For a materialized view with only aggregates (no joins, unions, subqueries, etc.) to be fast refreshable certain restrictions beyond the General Restrictions on Fast Refresh must be met. These
additional restrictions are fully documented at Restrictions on Fast Refresh on Materialized Views with Aggregates. For our current test case the most significant restrictions are these. all base tables must have materialized view logs that: o "Contain all columns from the table referenced in the materialized view." o "Specify with ROWID and INCLUDING NEW VALUES." o "Specify the SEQUENCE clause if the table is expected to have a mix of inserts/direct-loads, deletes, and updates." aggregates in the defining query must be either SUM, COUNT, AVG, STDDEV, VARIANCE, MIN, or MAX the defining query's SELECT clause must contain all the columns listed in the GROUP BY clause In addition to these restrictions some additional columns may be required in the defining query to allow it to be fast refreshable in all cases. The table below summarized these requirements. Materialized Views with Aggregates: Requirements for Refresh Fast After Any DML Aggregate COUNT(expr) MIN(expr) MAX(expr) SUM(expr) SUM(col) AVG(expr) Additional Aggregates Required COUNT(*) COUNT(*) COUNT(*) COUNT(*) COUNT(expr) COUNT(*) SUM(expr) SUM(expr*expr) SUM(expr*expr) "col" must have a NOT NULL constraint Optional Aggregates Note
defining query must have no WHERE clause defining query must have no WHERE clause
COUNT(*) COUNT(expr) COUNT(*) STDDEV(expr) COUNT(expr) SUM(expr) COUNT(*) VARIANCE(expr) COUNT(expr) SUM(expr)
(For insert-only materialized views see Table 8-2 Requirements for Materialized Views with Aggregates.) Oracle recommends including the Optional Aggregates expressions to obtain the most efficient and accurate fast refresh of the materialized view. Recommendations
The recommendation about gathering statistics that we saw in the Join Queries topic also applies to materialized views with aggregates. "After you create the materialized view, you must collect statistics on it using the DBMS_STATS package. Oracle Database needs the statistics generated by this package to optimize query rewrite." -- from CREATE MATERIALIZED VIEW. Additionally we also expect that our GROUP BY columns will often be specified in WHERE or JOIN clauses. To improve the performance of such queries we will therefore add indexes to our materialized view's GROUP BY columns. The Prototype Applying these restrictions and recommendations to our test case above yields the following prototypical materialized view with aggregates. Whenever I need to create this type of materialized view in an application I use the code below as a starting point to remind me of the requirements.
drop materialized view mv ; create materialized view log on t2 with rowid, sequence ( t_key, amt ) including new values ; create materialized view mv refresh fast on commit enable query rewrite as select t_key , sum(amt) as amt_sum , count(*) as row_count , count(amt) as amt_count from t2 group by t_key ; create index mv_i1 on mv ( t_key ) ; execute dbms_stats.gather_table_stats( user, 'MV' ) ;
Whenever we create a fast refreshable view we should use our EXPLAIN_MVIEW utility, MY_MV_CAPABILITIES, to confirm it can be refreshed in all required situations.
set long 5000 select my_mv_capabilities( 'MV', 'REFRESH' ) as mv_report from dual ; MV_REPORT ------------------------------------------------------------------------------Capable of:
REFRESH_COMPLETE REFRESH_FAST REFRESH_FAST_AFTER_INSERT REFRESH_FAST_AFTER_ONETAB_DML REFRESH_FAST_AFTER_ANY_DML
Now let's test drive our new MV. First, here are MV's initial contents.
select * from mv order by t_key ; T_KEY AMT_SUM ROW_COUNT AMT_COUNT ---------- ---------- ---------- ---------1 600 3 3 2 400 2 2
Now let's do some DML on the base table and see the effect on MV.
insert into t2 values ( 60, 3, 300 ) ; update t2 set amt = 0 where t_key = 2 ; commit; select * from mv order by t_key ; T_KEY AMT_SUM ROW_COUNT AMT_COUNT ---------- ---------- ---------- ---------1 600 3 3 2 0 2 2 3 300 1 1
Both changes are reflected in MV, as expected. Query Rewrite Materialized views containing aggregates can be used by the query rewrite facility (see ENABLE QUERY REWRITE).
alter materialized view mv enable query rewrite ; select my_mv_capabilities( 'MV', 'REWRITE' ) as mv_report from dual ; MV_REPORT -------------------------------------------------------------------------------
Capable of: REWRITE REWRITE_FULL_TEXT_MATCH REWRITE_PARTIAL_TEXT_MATCH REWRITE_GENERAL
Gotcha - Insert-Only Materialized Views On Commit We know that COUNT(*), and sometimes COUNT(expr), must be included in our materialized views for them to be fast refreshable in all cases, but what happens if we do not include these columns? Let's find out.
create materialized view mv2 refresh fast on commit enable query rewrite as select t_key , sum(amt) as amt_sum -count(*) as row_count , -count(amt) as amt_count from t2 group by t_key ; select * from mv2 order by t_key ; T_KEY AMT_SUM ---------- ---------1 600 2 0 3 300
Let's try an INSERT.

insert into t2 values ( 70, 3, 900 ) ; commit ; select * from mv2 order by t_key ; T_KEY AMT_SUM ---------- ---------1 600 2 0 3 1200
Looks good. The view was fast refreshed after the transaction committed. In topic REFRESH FAST Categories we saw how an insert-only ON DEMAND materialized view similar to this one raised an error when we attempted to fast refresh it manually after a DELETE transaction. Let's see how our ON COMMIT version behaves after a DELETE.
delete from t2 where t_key = 1 ; 3 rows deleted. commit; Commit complete. select * from mv2 order by t_key ; T_KEY AMT_SUM ---------- ---------1 600 2 0 3 1200
Oops, all the rows for T_KEY = 1 were deleted from T2 but the group still appears in MV2. The materialized view did not refresh on commit and no errors were generated. Let's try synchronizing MV2 manually using DBMS_MVIEW.REFRESH.
execute DBMS_MVIEW.REFRESH( 'MV2', 'c' ) select * from mv2 order by t_key ; T_KEY AMT_SUM ---------- ---------2 0 3 1200
That's a little better. So we've confirmed we have another insert-only materialized view, except this time we won't get any warnings or errors if a commit fails to trigger a fast refresh. When I first learned materialized views I stumbled across this behaviour by accident and found it puzzling. After all, when one creates a materialized view specifying that it should REFRESH FAST ON COMMIT it seems reasonable to assume it will always refresh fast on commit. The manual page for CREATE MATERIALIZED VIEW did not mention insert-only materialized views so I had no clue to their existence, until I re-read the page a third time and followed up on this seemingly inconsequential little comment.
"(The REFRESH clause) only sets the default refresh options. For instructions on actually implementing the refresh, refer to Oracle Database Advanced Replication and Oracle Database Data Warehousing Guide." -- from CREATE MATERIALIZED VIEW This eventually lead me to learn about insert-only refreshing and how indispensable the DBMS_MVIEW.EXPLAIN_MVIEW utility is when working with fast refreshable materialized views. Let's see what DBMS_MVIEW.EXPLAIN_MVIEW has to say about MV2.
select my_mv_capabilities( 'MV2', 'REFRESH' ) as mv_report from dual ; MV_REPORT ------------------------------------------------------------------------------Capable of: REFRESH_COMPLETE REFRESH_FAST REFRESH_FAST_AFTER_INSERT Not Capable of: REFRESH_FAST_AFTER_ONETAB_DML AMT_SUM SUM(expr) without COUNT(expr) REFRESH_FAST_AFTER_ONETAB_DML COUNT(*) is not present in the select list REFRESH_FAST_AFTER_ANY_DML see the reason why REFRESH_FAST_AFTER_ONETAB_DML is disabled
The report tells us MV2 is fast refreshable after insert, as we saw earlier, but not after other types of DML. This is how to recognize an insert-only materialized view. So the lesson here is do not assume materialized views created with REFRESH FAST ON COMMIT will always refresh fast on commit. Always check it with DBMS_MVEW.EXPLAIN_MVIEW to see whether or not it is an "insert-only" materialized view. Cleanup
drop materialized view mv ; drop materialized view mv2 ; drop materialized view log on t2 ; delete t2 ; insert into t2 select * from t2_backup ; commit;
Nested Materialized Views

Sometimes a single materialized view will not meet our requirements. For example, given this base table
select t_key, amt, key from t2 order by t_key, amt, key ; T_KEY AMT KEY ---------- ---------- ---------1 100 10 1 200 30 1 300 20 2 150 50 2 250 40
say we wanted a fast refreshable materialized view defined with the following query.
select t_key t_key , max(amt) amt_max , max(key) keep ( dense_rank last order by amt ) as t2_key_of_amt_max from t2 group by t_key ; T_KEY AMT_MAX T2_KEY_OF_AMT_MAX ---------- ---------- ----------------1 300 20 2 250 40
(T2_KEY_OF_AMT_MAX identifies the KEY value associated with the highest AMT value in each group.) As always the first step is to create a materialized view log on T2.
create materialized view log on t2 with rowid , sequence ( key, t_key, amt ) including new values ;
Now let's see what the MY_MV_CAPABILITIES utility (created in topic DBMS_MVIEW.EXPLAIN_MVIEW) tells us about our query.
set long 5000 select
my_mv_capabilities( 'select t_key t_key , max(amt) amt_max , max(key) keep ( dense_rank LAST order by amt ) as t2_key_of_amt_max from t2 group by t_key' , 'REFRESH' ) as mv_report from dual ; MV_REPORT ------------------------------------------------------------------------------Capable of: REFRESH_COMPLETE Not Capable of: REFRESH_FAST REFRESH_FAST_AFTER_INSERT aggregate function nested within an expression REFRESH_FAST_AFTER_ONETAB_DML see the reason why REFRESH_FAST_AFTER_INSERT is disabled REFRESH_FAST_AFTER_ONETAB_DML mv uses the MIN or MAX aggregate functions REFRESH_FAST_AFTER_ANY_DML see the reason why REFRESH_FAST_AFTER_ONETAB_DML is disabled
Though not entirely obvious from the report, it turns out our query is not fast refreshable because the LAST aggregate function which we used to implement T2_KEY_OF_AMT_MAX is not one of the fast refreshable aggregates SUM, COUNT, AVG, STDDEV, VARIANCE, MIN and MAX (see Restrictions on Fast Refresh on Materialized Views with Aggregates). Let's try writing the query using a subquery instead of LAST.
select my_mv_capabilities( 'select t_key t_key , max(amt) amt_max , max(key) as t2_key_of_amt_max from t2 where ( t_key, t2.amt ) in ( select t_key, max(amt) from t2 group by t_key
) group by t_key' , 'REFRESH' ) as mv_report from dual ; MV_REPORT ------------------------------------------------------------------------------Capable of: REFRESH_COMPLETE Not Capable of: REFRESH_FAST REFRESH_FAST_AFTER_INSERT subquery in mv REFRESH_FAST_AFTER_ONETAB_DML see the reason why REFRESH_FAST_AFTER_INSERT is disabled REFRESH_FAST_AFTER_ONETAB_DML mv uses the MIN or MAX aggregate functions REFRESH_FAST_AFTER_ANY_DML see the reason why REFRESH_FAST_AFTER_ONETAB_DML is disabled
It looks like a subquery will not work either. Perhaps an analytic approach will work?
select my_mv_capabilities( 'select distinct t_key t_key , max( amt ) over ( partition by t_key ) as amt_max , last_value( key ) over ( partition by t_key order by amt range between unbounded preceding and unbounded following ) as t2_key_of_amt_max from t2' , 'REFRESH' ) as mv_report from dual ; MV_REPORT ------------------------------------------------------------------------------Capable of: REFRESH_COMPLETE
Not Capable of: REFRESH_FAST REFRESH_FAST_AFTER_INSERT DISTINCT clause in select list in mv REFRESH_FAST_AFTER_INSERT DISTINCT clause in select list in mv REFRESH_FAST_AFTER_INSERT window function in mv REFRESH_FAST_AFTER_ONETAB_DML see the reason why REFRESH_FAST_AFTER_INSERT is disabled REFRESH_FAST_AFTER_ANY_DML see the reason why REFRESH_FAST_AFTER_ONETAB_DML is disabled
This last approach did not work either, which is a bit of a relief actually since the technique is rather crass. We need to rethink our approach. Since deriving the desired result set is conceptually a three step process 1. find the highest AMT value 2. find the highest KEY value per AMT 3. join the results of steps (1) and (2) together on the AMT column perhaps three separate materialized views would work? The materialized views for steps 1 and 2, which we will call MV1 and MV2, can be based on table T2 and can be refreshed independently of each other. However the materialized view for step 3, which we will call MV3, will need to be based on MV1 and MV2 and will need to refresh after they do. Fortunately Oracle allows for a materialized view like MV3 and automatically manages the refresh order when all three views are refreshable on commit. Materialized views like MV3 are called "Nested Materialized Views". Note the term "Nested Materialized View" does not refer to MV1 and MV2, even though they could be thought of as being "nested" within MV3. Restrictions and Recommendations As always, before creating a type of materialized view we have not tried before we must be aware of its restrictions. For nested materialized views they are these. The base materialized views must contain joins or aggregates. The defining query must contain joins or aggregates. All base objects, whether they are tables or materialized views, must each have materialized view logs. If REFRESH FAST is specified then all materialized views in any chain related to the materialized view must also specify REFRESH FAST. Note that all base objects in a nested materialized view, regardless of whether they are tables or materialized views, are treated as tables.
We are now ready to craft our three step solution.

create materialized view MV1 refresh fast on commit as select t_key , max(amt) amt_max , count(amt) amt_count , count(*) row_count from t2 group by t_key ; create materialized view log on mv1 with rowid , sequence ( t_key, amt_max, amt_count, row_count ) including new values ; select my_mv_capabilities( 'MV1', 'REFRESH' ) as mv_report from dual ; MV_REPORT ------------------------------------------------------------------------------Capable of: REFRESH_COMPLETE REFRESH_FAST REFRESH_FAST_AFTER_INSERT REFRESH_FAST_AFTER_ANY_DML create materialized view MV2 refresh fast on commit as select t_key , amt , max(key) max_key_per_amt , count(*) row_count from t2 group by t_key, amt ; create materialized view log on mv2 with rowid , sequence
( t_key, max_key_per_amt, row_count ) including new values ; select my_mv_capabilities( 'MV2', 'REFRESH' ) as mv_report from dual ; MV_REPORT ------------------------------------------------------------------------------Capable of: REFRESH_COMPLETE REFRESH_FAST REFRESH_FAST_AFTER_INSERT REFRESH_FAST_AFTER_ANY_DML create materialized view MV3 refresh fast on commit as select mv1.t_key , mv1.amt_max , mv2.max_key_per_amt as t2_key_of_amt_max , mv1.rowid mv1_rowid , mv2.rowid mv2_rowid from mv1, mv2 where mv1.t_key = mv2.t_key and mv1.amt_max = mv2.amt ; select my_mv_capabilities( 'MV3', 'REFRESH' ) as mv_report from dual ; MV_REPORT ------------------------------------------------------------------------------Capable of: REFRESH_COMPLETE REFRESH_FAST REFRESH_FAST_AFTER_INSERT REFRESH_FAST_AFTER_ONETAB_DML REFRESH_FAST_AFTER_ANY_DML
We finally have a fast refreshable materialized view solution. Let's confirm that MV3, the nested one, contains the correct results.
select t_key, amt_max, t2_key_of_amt_max from mv3 order by t_key ; T_KEY AMT_MAX T2_KEY_OF_AMT_MAX ---------- ---------- ----------------1 300 20 2 250 40
Good. It matches the results returned by the first query we tried which used the LAST function. Now let's put all three materialized views through their paces. First we perform a few mixed DML transactions.
insert into t2 values ( 60, 3, 450 ); insert into t2 values ( 70, 3, 550 ); update t2 set amt = 300 where key = 30 ; commit; delete from t2 where key = 70 ; update t2 set amt = 650 where key = 60 ; commit; select t_key, amt, key from t2 order by t_key, amt, key ; T_KEY AMT KEY ---------- ---------- ---------1 100 10 1 300 20 1 300 30 2 150 50 2 250 40 3 650 60
Now we check MV3 to see if it contains the correct info.

select t_key, amt_max, t2_key_of_amt_max from mv3 order by t_key ; T_KEY AMT_MAX T2_KEY_OF_AMT_MAX ---------- ---------- ----------------1 300 30 2 250 40 3 650 60
It does. Mission accomplished. Cleanup
drop materialized view mv1 ; drop materialized view mv2 ; drop materialized view mv3 ; drop materialized view log on t2 ; delete t2 ; insert into t2 select * from t2_backup ; commit;
Setup
create table t ( key number , val varchar2(5) ) ; insert insert insert insert commit; create table t_backup as select * from t ; create table t2 ( key number , t_key number , amt number ) ; insert into insert into insert into insert into insert into commit; t2 t2 t2 t2 t2 into into into into t t t t values values values values
primary key
( ( ( (
1, 2, 3, 4,
'a' 'b' 'c' null
); ); ); );
primary key not null references t not null values values values values values ( ( ( ( ( 10, 20, 30, 40, 50, 1, 1, 1, 2, 2, 100 300 200 250 150 ) ) ) ) ) ; ; ; ; ;
create table t2_backup as select * from t2 ;
Cleanup
drop table t2 ; drop table t ; drop table t_backup ; drop table t2_backup ; drop function my_mv_capabilities ; -------------------------------------------------------------------------------- WARNING!! --- MV_CAPABILITIES_TABLE is an Oracle table; do not drop it from your schema -- unless you specifically created it for this tutorial and no longer wish to -- use it ------------------------------------------------------------------------------drop table mv_capabilities_table ; exit
MODEL Clause
This section presents tutorials on the MODEL clause of the SELECT command. Introduced in Oracle 10g, the MODEL clause is a powerful feature that gives you the ability to change any cell in the query's result set using data from any other cell (similar to the way a spreadsheet works). It also adds procedural features to SQL previously available only through PL/SQL calls. For example, with MODEL you can take a simple table like this
KEY -----1 2 3 4 5 6 7 8 9 10 GROUP_1 ---------A A A B B B C C GROUP_2 ---------a1 a2 a3 a1 a2 a3 a1 a2 a1 a2 DATE_VAL NUM_VAL ---------- ------2005-01-01 100 2005-06-12 200 300 2006-02-01 2006-06-12 300 2005-01-01 100 2006-06-12 100 2005-02-01 2005-02-01 200 800
and, with a single command, create a report containing ad-hoc totals like this.
set null "" select case when key like 'Total%' then key else null end as total , group_1 , group_2 , num_val from t model dimension by ( cast(key as varchar2(20)) as key , nvl( group_1, 'n/a' ) as group_1 , nvl( group_2, 'n/a' ) as group_2 ) measures( num_val ) rules ( num_val[ 'Total 1 - A + C', null, null ] = sum(num_val)[any,group_1 in ('A','C'),any] , num_val[ 'Total 2 - A + a2', null, null ] = sum(num_val)[any,'A',any] + sum(num_val)[any,group_1 <> 'A','a2'] , num_val[ 'Total 3 - n/a', null, null ] = sum(num_val)[any,'n/a',any] , num_val[ 'Total 4 - a1 + a3', null, null ] =
sum(num_val)[any,any,group_2 in ('a1','a3')] ) order by group_1 , group_2 , total nulls first ; TOTAL GROUP_1 -------------------- ---------A A A B B B C C n/a n/a Total 1 - A + C Total 2 - A + a2 Total 3 - n/a Total 4 - a1 + a3 GROUP_2 NUM_VAL ---------- ------a1 100 a2 200 a3 300 a1 a2 300 a3 100 a1 100 a2 a1 200 a2 800 700 1700 1000 800
You can also use MODEL'S procedural features to produce results that are difficult, inefficient, or impossible to do with a non-MODEL SELECT command. Here is an example.
set null "(null)" column string format a40 select group_1, substr( string, 2 ) as string from t where num_val is not null model return updated rows partition by ( group_1 ) dimension by ( row_number() over (partition by group_1 order by num_val) as position ) measures ( cast( num_val as varchar2(65) ) as string ) -- Note 1 rules upsert iterate( 6 ) until ( presentv(string[iteration_number+2],1,0) = 0 ) ( string[0] = string[0] || ',' || string[iteration_number+1] ) order by group_1 ; GROUP_1 ---------A B C (null) STRING ---------------------------------------100,200,300 100,300 100 200,800
(This last technique is explained fully in another section of SQL Snippets at Rows to String: MODEL Method 1.)
Though powerful, the MODEL clause is also somewhat complex and this can be intimidating when you read about it for the first time. The tutorials to follow will therefore present very simple MODEL examples to help you quickly become comfortable with its many features. Before continuing it is important to know that everything in the MODEL clause is evaluated after all other clauses in the query, except for SELECT DISTINCT and ORDER BY. Knowing this will help you better understand the examples in this section's tutorials.
DIMENSION BY
In this tutorial we learn about the DIMENSION BY component of the MODEL clause. DIMENSION BY specifies which columns in a SELECT statement are dimension columns, which for our purposes can be thought of as any column that serves to identify each row in the result of a SELECT statement. By default, the dimension columns in a MODEL clause must produce a unique key for the result set. See the Oracle Database Data Warehousing Guide 10g Release 2 (10.2) - Glossary for a formal definition. Before we begin please note that, on its own, DIMENSION BY has little visible effect on the output of the SELECT statement. Most of the examples below would produce the same result as one with no MODEL clause at all. This is because we are not trying to manipulate the results just yet. We are simply seeing how to specify our dimension columns, which is a precursor for learning to manipulate results in subsequent pages. Consider the following table.
select key, key_2, group_1, group_2, num_val from t order by key ; KEY -----1 2 3 4 5 6 7 8 9 10 KEY_2 ----T-1 T-2 T-3 T-4 T-5 T-6 T-7 T-8 T-9 T-10 GROUP_1 ---------A A A B B B C C GROUP_2 NUM_VAL ---------- ------a1 100 a2 200 a3 300 a1 a2 300 a3 100 a1 100 a2 a1 200 a2 800
We see that KEY, KEY_2, and (GROUP_1, GROUP_2) all uniquely identify each row in the table. They are therefore dimension column candidates. To let Oracle know which column(s) we plan to use as dimensions we compose a MODEL clause like this. (Ignore the MEASURES and RULES clauses for now. We will explore those later.)
select key, num_val from t model DIMENSION BY ( KEY ) measures ( num_val ) rules () order by key ; KEY NUM_VAL ------ ------1 100 2 200 3 300 4 5 300 6 100 7 100 8 9 200 10 800
Multiple Dimensions If needed, you can define more than one dimension column, as this example shows.
select group_1, group_2, num_val from t model DIMENSION BY ( GROUP_1, GROUP_2 ) measures ( num_val ) rules () order by group_1, group_2 ; GROUP_1 ---------A A A B B B C C GROUP_2 NUM_VAL ---------- ------a1 100 a2 200 a3 300 a1 a2 300 a3 100 a1 100 a2 a1 200 a2 800
You can even include columns in the DIMENSION BY clause which are not required to uniquely identify each result row.
select key, date_val, num_val from t model DIMENSION BY ( KEY, DATE_VAL ) -- date_val not required to uniquely identify row measures ( num_val ) rules () order by key ; KEY -----1 2 3 4 5 6 7 8 9 10 DATE_VAL NUM_VAL ---------- ------2005-01-01 100 2005-06-12 200 300 2006-02-01 2006-06-12 300 2005-01-01 100 2006-06-12 100 2005-02-01 2005-02-01 200 800
Aliasing You cannot use SELECT clause aliases in DIMENSION BY. Here are some examples of aliases that will cause errors.
select KEY AS KEY_3, num_val from t model dimension by ( KEY_3 ) measures ( num_val ) rules () ; dimension by ( KEY_3 ) * ERROR at line 7: ORA-00904: "KEY_3": invalid identifier
select KEY * 10 AS KEY_3, num_val from t model
dimension by ( KEY_3 ) measures ( num_val ) rules ()
dimension by ( KEY_3 ) * ERROR at line 7: ORA-00904: "KEY_3": invalid identifier
select ROWNUM AS KEY_3, num_val from t model dimension by ( KEY_3 ) measures ( num_val ) rules () ; dimension by ( KEY_3 ) * ERROR at line 7: ORA-00904: "KEY_3": invalid identifier
You can however alias such expressions directly in DIMENSION BY.

select KEY_3, num_val from t model DIMENSION BY ( KEY AS KEY_3 ) measures ( num_val ) rules () order by key_3 ; KEY_3 NUM_VAL ---------- ------1 100 2 200 3 300 4 5 300 6 100 7 100 8 9 200 10 800 select KEY_3, num_val
from t model DIMENSION BY ( KEY * 10 AS KEY_3 ) measures ( num_val ) rules () order by key_3 ; KEY_3 NUM_VAL ---------- ------10 100 20 200 30 300 40 50 300 60 100 70 100 80 90 200 100 800 select KEY_3, num_val from t model DIMENSION BY ( ROWNUM AS KEY_3 ) measures ( num_val ) rules () order by key_3 ; KEY_3 NUM_VAL ---------- ------1 100 2 200 3 300 4 5 300 6 100 7 100 8 9 200 10 800
Uniqueness By default, if your DIMENSION BY columns do not give you a unique key for your result set you will get an error.
select group_2, num_val from
t model DIMENSION BY ( GROUP_2 ) -- group_2 is not unique measures ( num_val ) rules () order by group_2 ; t * ERROR at line 5: ORA-32638: Non unique addressing in MODEL dimensions
This rule can be relaxed somewhat by specifying UNIQUE SINGLE REFERENCE.

select group_2, num_val from t model UNIQUE SINGLE REFERENCE dimension by ( group_2 ) measures ( num_val ) rules () order by group_2 ; GROUP_2 NUM_VAL ---------- ------a1 100 a1 100 a1 200 a1 a2 800 a2 200 a2 300 a2 a3 300 a3 100
-- group_2 is not unique
Note that UNIQUE SINGLE REFERENCE affects the types of RULES you can define. This is explained further in Expressions and Cell References.
MEASURES
In this tutorial we learn about the MEASURES component of the MODEL clause. MEASURES specifies which columns in a SELECT are measure columns, which for our purposes can be thought of as any column containing a measurable quantity like a price or a length. See the Oracle Database Data Warehousing Guide 10g Release 2 (10.2) - Glossary for a formal definition.
Before we begin please note that, on its own, MEASURES has little visible effect on the output of the SELECT statement. Most of the examples below would produce the same result as one with no MODEL clause at all. This is because we are not trying to manipulate the results just yet. We are simply seeing how to specify our measure columns, which is a precursor to manipulating the results. We will see how to actually manipulate our output when we explore the RULES clause in subsequent tutorials. Before we see MEASURES in action first consider the following table.
select key, group_1, group_2, date_val, num_val from t order by key ; KEY -----1 2 3 4 5 6 7 8 9 10 GROUP_1 ---------A A A B B B C C GROUP_2 ---------a1 a2 a3 a1 a2 a3 a1 a2 a1 a2 DATE_VAL NUM_VAL ---------- ------2005-01-01 100 2005-06-12 200 300 2006-02-01 2006-06-12 300 2005-01-01 100 2006-06-12 100 2005-02-01 2005-02-01 200 800
If we decide to use KEY as our sole dimension column, then all other columns are available for use as measure columns. To let Oracle know we want to use the NUM_VAL column as our measure we can compose a MODEL clause like this.
select key, num_val from t model dimension by ( key ) MEASURES ( NUM_VAL ) rules () order by key ; KEY NUM_VAL ------ ------1 100 2 200 3 300 4 5 300 6 100
7 8 9 10
100 200 800
If we want to include more measure columns we do it like this.

select key, date_val, num_val from t model dimension by ( key ) MEASURES ( DATE_VAL, NUM_VAL ) rules () order by key ; KEY -----1 2 3 4 5 6 7 8 9 10 DATE_VAL NUM_VAL ---------- ------2005-01-01 100 2005-06-12 200 300 2006-02-01 2006-06-12 300 2005-01-01 100 2006-06-12 100 2005-02-01 2005-02-01 200 800
You can define measures using constants and expressions instead of simple column names, like this.
select key, num_val, num_val_2, date_val_2, note from t model dimension by ( key ) MEASURES ( num_val , NUM_VAL * 10 AS NUM_VAL_2 , SYSDATE AS DATE_VAL_2 , 'A BRIEF NOTE' AS NOTE ) rules( ) order by key ; KEY NUM_VAL NUM_VAL_2 DATE_VAL_2 NOTE ------ ------- ---------- ---------- ------------
1 2 3 4 5 6 7 8 9 10
100 200 300 300 100 100 200 800
1000 2007-02-28 A BRIEF 2000 2007-02-28 A BRIEF 3000 2007-02-28 A BRIEF 2007-02-28 A BRIEF 3000 2007-02-28 A BRIEF 1000 2007-02-28 A BRIEF 1000 2007-02-28 A BRIEF 2007-02-28 A BRIEF 2000 2007-02-28 A BRIEF 8000 2007-02-28 A BRIEF
NOTE NOTE NOTE NOTE NOTE NOTE NOTE NOTE NOTE NOTE
Nulls
Nulls and Aggregate Functions
This tutorial demonstrates how aggregate functions deal with null values. Techniques for generating results that ignore nulls and results that include nulls are highlighted. Ignoring Nulls According to the SQL Reference Manual section on Aggregate Functions: All aggregate functions except COUNT(*) and GROUPING ignore nulls. You can use the NVL function in the argument to an aggregate function to substitute a value for a null. COUNT never returns null, but returns either a number or zero. For all the remaining aggregate functions, if the data set contains no rows, or contains only rows with nulls as arguments to the aggregate function, then the function returns null. This means that, given a table with values like this
GROUP_KEY ---------Group-1 Group-1 Group-2 Group-2 Group-2 Group-2 Group-2 Group-3 Group-3 Group-3 VAL ---------(null) (null) a a z z (null) A A Z
aggregate functions like MAX , MIN , and COUNT will return values that for the most part ignore nulls, like these.
select group_key , MAX( VAL ) MIN( VAL ) COUNT( * ) COUNT( VAL ) COUNT( DISTINCT VAL ) from t1 group by group_key order by group_key ;
max_val , min_val , count_all_rows , count_val , count_distinct_val
GROUP_KEY MAX_VAL MIN_VAL COUNT_ALL_ROWS COUNT_VAL COUNT_DISTINCT_VAL ---------- ---------- ---------- -------------- ---------- -----------------Group-1 (null) (null) 2 0 0
Group-2 Group-3
z Z
a A
5 3
4 3
2 2
Note how MAX_VAL contains the same results for Group-2 and Group-3, even though Group-2 contains null VAL values and Group-3 does not. Note also that only COUNT_ALL_ROWS returned a count that included null values. The other two versions of COUNT() ignored null values. Including Nulls For mathematical aggregate functions like AVG, MEDIAN, and SUM including nulls in the calculation is of little practical use. For aggregate functions like MAX, MIN, and COUNT(DISTINCT ...) however, we sometimes need results that take nulls into account. For example, using the same test data as above
GROUP_KEY ---------Group-1 Group-1 Group-2 Group-2 Group-2 Group-2 Group-2 Group-3 Group-3 Group-3 VAL ---------(null) (null) a a z z (null) A A Z
we may wish to produce results like these.

GROUP_KEY ---------Group-1 Group-2 Group-3 MAX_VAL ---------(null) (null) Z MIN_VAL COUNT_DISTINCT_VAL_1 ---------- -------------------(null) 1 a 3 A 2
For the MAX and MIN cases it helps to take the statement "all aggregate functions except COUNT(*) and GROUPING ignore nulls" with a grain of salt. Fortunately for us there are aggregate functions in addition to COUNT(*) and GROUPING which do not ignore nulls. Two of them are the FIRST and LAST functions, which we can use as follows to give us MAX and MIN results that include nulls.
select group_key, MAX( VAL ) KEEP ( DENSE_RANK LAST ORDER BY VAL ) max_val , MIN( VAL ) KEEP ( DENSE_RANK FIRST ORDER BY VAL ) min_val from t1 group by group_key order by group_key ; GROUP_KEY MAX_VAL MIN_VAL
---------Group-1 Group-2 Group-3
---------(null) (null) Z
---------(null) a A
For the COUNT( DISTINCT VAL ) case two possible approaches for including nulls are demonstrated below.
select group_key, COUNT( DISTINCT DUMP(VAL) ) count_distinct_val_1 , COUNT( DISTINCT VAL ) + MAX( NVL2(VAL,0,1) ) count_distinct_val_2 from t1 group by group_key order by group_key ; GROUP_KEY COUNT_DISTINCT_VAL_1 COUNT_DISTINCT_VAL_2 ---------- -------------------- -------------------Group-1 1 1 Group-2 3 3 Group-3 2 2
Be careful with the DUMP approach though since DUMP's output is truncated at 4000 characters. If the VAL column contained values whose DUMP output is truncated then the results can be incorrect. DENSE_RANK and RANK Two more aggregate functions where including nulls in the calculation may be necessary are the DENSE_RANK and RANK functions. Fortunately, as with FIRST and LAST, DENSE_RANK and RANK include nulls by default. For example, given test data like this (analytic value rankings are included for clarity)
GROUP_KEY ---------Group-1 Group-1 Group-2 Group-2 Group-2 Group-2 Group-2 Group-3 Group-3 Group-3 VAL VAL_DENSE_RANK VAL_RANK ---------- -------------- ---------(null) 1 1 (null) 1 1 a a z z (null) A A Z 1 1 2 2 3 1 1 2 1 1 3 3 5 1 1 3
The following results show how the aggregate versions of DENSE_RANK and RANK do not ignore nulls.
select group_key ,
DENSE_RANK( NULL ) WITHIN GROUP ( ORDER BY VAL ) null_dense_rank_within_group , RANK( NULL ) WITHIN GROUP ( ORDER BY VAL ) null_rank_within_group from t1 group by group_key order by group_key ; GROUP_KEY NULL_DENSE_RANK_WITHIN_GROUP NULL_RANK_WITHIN_GROUP ---------- ---------------------------- ---------------------Group-1 1 1 Group-2 3 5 Group-3 3 4
Gotchas Some people reading these two sentences from the manual All aggregate functions except COUNT(*) and GROUPING ignore nulls. You can use the NVL function in the argument to an aggregate function to substitute a value for a null. may infer that aggregate functions can be made to treat null values the same way they treat nonnull values by simply using NVL to substitute nulls with some non-null value. A simple application of this logic can lead to trouble however. For example, say we choose to substitute all null values with a 'z', like this.
select group_key , max( NVL( VAL, 'z' ) ) max_val , min( NVL( VAL, 'z' ) ) min_val , count( distinct NVL( VAL, 'z' ) ) count_distinct_val from t1 group by group_key order by group_key ; GROUP_KEY ---------Group-1 Group-2 Group-3 MAX_VAL ---------z z Z MIN_VAL COUNT_DISTINCT_VAL ---------- -----------------z 1 a 2 A 2
Note how none of the columns above contain the desired results which, as you will recall, are these.
GROUP_KEY ---------Group-1 Group-2 Group-3 MAX_VAL ---------(null) (null) Z MIN_VAL COUNT_DISTINCT_VAL_1 ---------- -------------------(null) 1 a 3 A 2
A simple application of NVL clearly will not do then. Taking the NVL idea a little further programmers sometimes employ more complex solutions such as this one.
select group_key , DECODE( MAX( NVL( VAL, '~' ) ), '~', NULL, MAX( VAL ) ) max_val ,
DECODE( MIN( NVL( VAL, '~' ) ), '~', NULL, MIN( VAL ) ) min_val , COUNT( distinct NVL( VAL, '~' ) ) count_distinct_val from t1 group by group_key order by group_key ; GROUP_KEY ---------Group-1 Group-2 Group-3 MAX_VAL ---------(null) (null) Z MIN_VAL COUNT_DISTINCT_VAL ---------- -----------------(null) 1 a 3 A 2
Without some mechanism to ensure '~' and strings that sort higher than '~' never appear in VAL however, these solutions will fail if such values are ever inserted into the table. For example, given this data
insert into t1 values ( 10, 'Group-4', null insert into t1 values ( 10, 'Group-4', '~' ); );
insert into t1 values ( 10, 'Group-5', null ); insert into t1 values ( 10, 'Group-5', '~~~' );
the results should be

GROUP_KEY ---------Group-4 Group-5 MAX_VAL ---------(null) (null) MIN_VAL COUNT_DISTINCT_VAL_1 ---------- -------------------~ 2 ~~~ 2
but using the NVL approach gives us these, incorrect results.

select group_key , decode( max( nvl(val, '~' ) ), '~', null, max( val ) ) max_val , decode( min( nvl(val, '~' ) ), '~', null, min( val ) ) min_val , count( distinct nvl( val, '~' ) ) count_distinct_val from t1 where group_key in ( 'Group-4', 'Group-5' ) group by group_key order by group_key ; GROUP_KEY ---------Group-4 Group-5 MAX_VAL ---------(null) ~~~ MIN_VAL COUNT_DISTINCT_VAL ---------- -----------------(null) 1 (null) 2
To avoid these gotchas simply use the non-NVL alternatives presented under "Including Nulls" above.
Nulls and Equality

In SQL you should always consider the effect of null values when comparing two values for equality (or any type of comparison for that matter). Consider a table where two of its columns can contain null values.
select * from t; C1 --1 2 3 4 C2 ---------A A (null) (null) C3 ---------A B A (null)
If we attempt a SELECT statement like the following we will only get row 1.
select * from t where c2 = c3 ; C1 C2 C3 --- ---------- ---------1 A A
Row 4 is not returned because, in SQL, a null is not considered to be equal to or unequal to any value (including another null). If this is the behavior you need, then read no further. However, if you need a query that returns row 1 and row 4 then try one of the solutions in the subtopics to follow.
SQL + PL/SQL
These techniques work in both SQL and PL/SQL. OR with IS NULL While a bit cumbersome, this basic solution is the easiest to understand and implement.
select * from t where ( C2 = C3 OR ( C2 IS NULL AND C3 IS NULL ) ); C1 --1 4 C2 ---------A (null) C3 ---------A (null)
begin for r in ( select * from t ) loop
if ( R.C2 = R.C3 OR ( R.C2 IS NULL AND R.C3 IS NULL ) ) then dbms_output.put_line( 'Row ' || r.c1 || ' contains matching values.' ); end if; end loop; end; . / Row 1 contains matching values. Row 4 contains matching values.
NVL The following NVL approach is a popular one.

select * from t where nvl( c2, 'x' ) = nvl( c3, 'x' ) ; C1 --1 4 C2 ---------A (null) C3 ---------A (null)
One problem with this solution is that the replacement value "x", or whatever value you choose to use, might be inserted into columns C2 or C3 some day. This would cause a SELECT statement that has been working properly until that day to all of a sudden start returning the wrong answer, like this.
insert into t values( 5, 'x', null ); commit; select * from t where nvl( c2, 'x' ) = nvl( c3, 'x' ) ; C1 --1 4 5 C2 ---------A (null) x C3 ---------A (null) (null)
The trick to making this solution bullet proof is to choose a replacement value that can never appear in either of the columns being compared. If we look at the table definition for T
desc t Name Null? Type ---------------------------------------------- -------------------------------------C1 NUMBER C2 VARCHAR2(10)
C3
VARCHAR2(10)
we see that values in C2 and C3 can be at most 10 characters long. Any replacement value larger than 10 characters is therefore guaranteed to never appear in either column (assuming the sizes of C2 or C3 are never expanded).
select * from t where nvl( c2, '12345678901' ) = nvl( c3, '12345678901' ) ; C1 --1 4 C2 ---------A (null) C3 ---------A (null)
Custom Function If you do these comparisons frequently you may wish to create a custom database function like this one.
create function SAME( p_1 in varchar2, p_2 in varchar2 ) return varchar2 is begin return ( case when p_1 is null and p_2 is null then 'Y' when p_1 = p_2 then 'Y' else 'N' end ); end; / select * from t where SAME( C2, C3 ) = 'Y' ; C1 --1 4 C2 ---------A (null) C3 ---------A (null)
begin for r in ( select * from t ) loop if SAME( R.C2, R.C3 ) = 'Y' then dbms_output.put_line( 'Row ' || r.c1 || ' contains matching values.' ); end if; end loop; end; .
/ Row 1 contains matching values. Row 4 contains matching values.
With this approach however one function is required for comparing NUMBER values, one for VARCHAR2 values, one for DATE values, etc.
SQL Only
The following techniques, while more compact than the solutions presented in SQL + PL/SQL , unfortunately only work in SQL commands, not PL/SQL commands. DECODE One approach uses the DECODE function. Unlike the "=" operator, DECODE treats two nulls as equivalent. To return rows where two columns contain the same value we can therefore use a command like the following.
select * from t where DECODE( C2, C3, 'Y', 'N' ) = 'Y' ; C1 --1 4 C2 ---------A (null) C3 ---------A (null)
DUMP Another approach uses the DUMP function.

select * from t where DUMP(C2) = DUMP(C3) ; C1 --1 4 C2 ---------A (null) C3 ---------A (null)
This approach has a couple of limitations however. One, the output of the DUMP function is truncated at 4000 characters. If the values being compared produce truncated DUMP output then the comparison can produce false positives. Here is an example.
select 'Oops! This row should not be returned.' as result from dual where DUMP( LPAD( 'A', 4000) ) = DUMP( LPAD( 'B', 4000 ) ) ; RESULT -------------------------------------Oops! This row should not be returned.
Two, C2 and C3 must be the exact same datatype for the comparison to work. Comparing compatible datatypes such as VARCHAR2 and CHAR will fail to match any rows.
select 'Oops! This row should be returned, but it is not.' as result from t where dump( c2 ) = dump( 'A' ) ; no rows selected
Examining the output of DUMP shows why this occurs. The "typ=" part of the DUMP output for both terms differs because column C2 is datatype 1, VARCHAR2, and the literal 'A' is datatype 96, CHAR.
column varchar2_val format a30 column char_val format a30 fold_before select dump( c2 ) as varchar2_val , dump( 'A' ) as char_val from t where c1 = 1 ; VARCHAR2_VAL -----------------------------CHAR_VAL -----------------------------Typ=1 Len=1: 65 Typ=96 Len=1: 65
SYS_OP_MAP_NONNULL Another approach that some have proposed uses the undocumented function SYS_OP_MAP_NONNULL.
select * from t where SYS_OP_MAP_NONNULL( C2 ) = SYS_OP_MAP_NONNULL( C3 ) ; C1 --1 4 C2 ---------A (null) C3 ---------A (null)
As with the other solutions on this page, SYS_OP_MAP_NONNULL does not work in PL/SQL.
begin for r in ( select * from t ) loop
if ( SYS_OP_MAP_NONNULL( R.C2 ) = SYS_OP_MAP_NONNULL( R.C3 ) ) then dbms_output.put_line( 'Row ' || r.c1 || ' contains matching values.' ); end if; end loop; end; . / if ( SYS_OP_MAP_NONNULL( R.C2 ) = SYS_OP_MAP_NONNULL( R.C3 ) ) * ERROR at line 6: ORA-06550: line 6, column 10: PLS-00201: identifier 'SYS_OP_MAP_NONNULL' must be declared ORA-06550: line 6, column 5: PL/SQL: Statement ignored
It also has a length limitation.

select * from dual where SYS_OP_MAP_NONNULL( LPAD( 'A', 4000 ) ) = SYS_OP_MAP_NONNULL( LPAD( 'B', 4000 ) ) ; select * * ERROR at line 1: ORA-01706: user function result value was too large
While undocumented features such as this one are compelling, their behavior or availability can change at any time making them a risky thing to include in your code. They also make support and maintenance harder for others who need to work with your code and are not familiar with the feature.
Setup
Run the code on this page in SQL*Plus to create the sample tables, data, etc. used by the examples in this section. Be sure to read Using SQL Snippets before executing any of these setup steps.
create table t ( c1 number , c2 varchar2(10) , c3 varchar2(10) ); insert insert insert insert into into into into t t t t values( values( values( values( 1, 2, 3, 4, 'A' 'A' null null , , , , 'A' 'B' 'A' null ); ); ); );
commit; set null '(null)' set numformat 99 set serveroutput on
Cleanup
Run the code on this page to drop the sample tables, procedures, etc. created in earlier parts of this section. To clear session state changes (e.g. those made by SET, COLUMN, and VARIABLE commands) exit your SQL*Plus session after running these cleanup commands. Be sure to read Using SQL Snippets before executing any of these setup steps.
drop table t ; drop function same ; exit
Integer Series Generators

Sometimes, having a way to create a series of integers greatly simplifies certain queries. For example, if your data looks like this:
select * from t ; DAY_OF_WEEK VAL ----------- ---------1 100 3 300 4 400 5 500
and you want a report that looks like this

DAY_OF_WEEK VAL ----------- ---------0 1 100 2 3 300 4 400 5 500 6
It would be useful to have a table with the numbers 0 to 6 in it so you could write an outer join query like this.
select day_of_week, t.val from days_of_the_week d left outer join t using ( day_of_week ) order by day_of_week ; DAY_OF_WEEK VAL ----------- ---------0 1 100 2 3 300 4 400 5 500 6
If you expect to write lots of queries that use the same series of integers and they are based on real world phenomena then creating a table like DAYS_OF_THE_WEEK can be the best solution.
Occassionally however, you may need a different set of integers just for one specific query, for adhoc reports, or for a system you do not have CREATE TABLE privileges on. In these cases it may be impractical or impossible to create a dedicated table that meets your needs. Fortunately there are flexible, generic techniques for generating integers. The tutorials in this section demonstrate a few of them. The feature graph below will help you decide which method is best for you.
Feature Intege MOD r Table EL Y N ROWNUM Type CONNECT CUB Pipelined + a Big Construct BY LEVEL E Function Table or Y Y Y Y Y Y N Y N Y
Pure SQL solution; no custom objects N required Works in versions Y prior to 10g
Performance comparison charts for all these methods are available at the end of the section on the Performance Comparison - Small Numbers and Performance Comparison - Large Numbers pages.
Integer Table Method

This tutorial demonstrates how to generate a series of integers using a generic integer table. Other techniques are discussed in the topics listed in the menu to the left. One of the most straightforward ways to generate a series of integers is by adding a generic integer table to your application. You can create such a table like this.
create table integers ( integer_value integer primary key ) organization index ; Table created.
Since this table only has a single indexed column we specified organization index to make this an Index-Organized table and save storage space. To load the table a simple loop like the following will do the trick.
begin for i in -5 .. 10 loop insert into integers values ( i ); end loop; commit; end; / PL/SQL procedure successfully completed.
We used -5 and 10 as the limits of our series in this example. In practice you would choose limits that anticipate the smallest and largest integers you will ever need.
select * from integers ; INTEGER_VALUE -------------5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10
Later, when you need a specific series of integers you can use the INTEGERS table like this.
select i.integer_value as day_of_week , t.val from integers i left outer join t on ( i.integer_value = t.day_of_week ) where i.integer_value between 0 and 6 order by i.integer_value ; DAY_OF_WEEK VAL ----------- ---------0 1 100 2 3 300 4 400 5 500 6
MODEL Method
This tutorial demonstrates how to generate a series of integers using the MODEL clause of the SELECT command. (You can learn more about MODEL at SQL Features Tutorials: MODEL
Clause.) This technique only works with Oracle versions starting at 10g. Other techniques are discussed in the tutorials listed in the menu to the left. With this technique you can generate a series of integers starting at "1" using a query like this.
select integer_value from dual where 1=2 model dimension by ( 0 as key ) measures ( 0 as integer_value ) rules upsert ( integer_value[ for key from 1 to 10 increment 1 ] = cv(key) ) ; INTEGER_VALUE ------------1 2 3 4 5 6 7 8 9 10
Chaning the INCREMENT value lets us control the difference between successive values in the series.
select integer_value from dual where 1=2 model dimension by ( 0 as key ) measures ( 0 as integer_value ) rules upsert ( integer_value[ for key from 2 to 10 INCREMENT 2 ] = cv(key) ) ; INTEGER_VALUE ------------2 4 6 8 10
We can use bind variables to make the solution more generic.

variable v_first_key variable v_last_key variable v_increment execute :V_FIRST_KEY execute :V_LAST_KEY execute :V_INCREMENT number number number := 1 := 5 := 2
select key, integer_value from dual where 1=2 model dimension by ( 0 as key ) measures ( 0 as integer_value ) rules upsert ( integer_value[ for key from :V_FIRST_KEY to :V_LAST_KEY increment 1 ] = nvl2( integer_value[cv()-1], integer_value[cv()-1] + :V_INCREMENT, cv(key) ) ) ; KEY INTEGER_VALUE ---------- ------------1 1 2 3 3 5 4 7 5 9
When v_last_key is NULL or less than v_first_key no rows are returned.

execute :v_first_key := 1 PL/SQL procedure successfully completed. execute :v_last_key := null
PL/SQL procedure successfully completed. / no rows selected execute :v_last_key := 0 PL/SQL procedure successfully completed. / no rows selected execute :v_last_key := -5 PL/SQL procedure successfully completed. / no rows selected
Day of the Week Case Study
We can apply this technique to the day of the week scenario presented at the start of this chapter as follows.
select day_of_week , t.val from ( select day_of_week from dual where 1=2 model dimension by ( 0 as key ) measures ( 0 as day_of_week ) rules upsert ( day_of_week[ for key from 0 to 6 increment 1 ] = cv(key) ) ) i left outer join t using ( day_of_week ) order by day_of_week ; DAY_OF_WEEK VAL ----------- ---------0 1 100 2 3 300 4 400 5 500 6
Gotchas
Descending Series
If you need a descending series of integers this attempt will not work.
select integer_value from dual where 1=2 model dimension by ( 0 as key ) measures ( 0 as integer_value ) rules upsert ( integer_value[ for key FROM 3 TO 1 increment 1 ] = cv(key) ) ; no rows selected
Instead, do it this way

select integer_value from dual
where 1=2 model dimension by ( 0 as key ) measures ( 0 as integer_value ) rules upsert ( integer_value[ for key from 3 to 1 DECREMENT 1 ] = cv(key) ) ORDER BY INTEGER_VALUE DESC ; INTEGER_VALUE ------------3 2 1
or this way.
select integer_value from dual where 1=2 model dimension by ( 0 as key ) measures ( 0 as integer_value ) rules upsert ( integer_value[ for key from 1 TO 3 INCREMENT 1 ] = cv(key) ) ORDER BY INTEGER_VALUE DESC ; INTEGER_VALUE ------------3 2 1
WHERE 1=2
It is important to note that everything in the MODEL clause is evaluated after all other clauses in the query, except for SELECT DISTINCT and ORDER BY. Using the WHERE 1=2 clause ensures the query starts with an empty result set when MODEL rules are first applied to the rows returned by the SELECT ... FROM ... WHERE portion of the query. While it would be possible to omit the WHERE 1=2 clause using an approach like this
select integer_value from dual model dimension by ( 1 as key ) measures ( 1 as integer_value ) rules upsert ( integer_value[ for key from 1 to 10 increment 1 ] = cv(key) ) ; INTEGER_VALUE ------------1 2 3
4 5 6 7 8 9 10
this query causes the result set to always contain at least one row both before and after the MODEL rules are applied. This is not a problem for queries that always return one or more rows like this one,
select key, integer_value from dual model dimension by ( 4 as key ) measures ( 4 as integer_value ) rules upsert ( integer_value[ for key from 4 to 8 increment 1 ] = cv(key) ) ; KEY INTEGER_VALUE ---------- ------------4 4 5 5 6 6 7 7 8 8
but if the code is later parameterized and the TO bound is ever null or less than the FROM bound then the query will incorrectly return 1 row instead of the required zero rows for these cases.
variable v_first_key variable v_last_key number number
execute :v_first_key := 3 execute :v_last_key := 0 select key, integer_value from dual model dimension by ( :v_first_key as key ) measures ( :v_first_key as integer_value ) rules upsert ( integer_value[ for key from :v_first_key to :v_last_key increment 1 ] = cv(key) ) ; KEY INTEGER_VALUE ---------- ------------3 3
RETURN UPDATED ROWS
An alternative to using WHERE 1=2 would be to instead include a RETURN UPDATED ROWS clause, like this
select integer_value from dual model RETURN UPDATED ROWS dimension by ( 1 as key ) measures ( 1 as integer_value ) rules upsert ( integer_value[ for key from 1 TO 3 increment 1 ] = cv(key) ) ; INTEGER_VALUE ------------1 2 3 select integer_value from dual model RETURN UPDATED ROWS dimension by ( 1 as key ) measures ( 1 as integer_value ) rules upsert ( integer_value[ for key from 3 TO 0 increment 1 ] = cv(key) ) ; no rows selected
but using WHERE 1=2 to ensure the query always starts with an empty set seems like a cleaner way to work than starting with one row and then relying on RETURN UPDATED ROWS to return that row in some cases but not others.
INCREMENT and Bind Variables

Unlike the FROM and TO bounds, we cannot use a variable in the INCREMENT value (as tested in 10g).
variable v_first_key variable v_last_key variable v_increment execute :v_first_key execute :v_last_key execute :v_increment number number number := 1 := 9 := 2
select key, integer_value from dual where 1=2 model dimension by ( 0 as key ) measures ( 0 as integer_value ) rules upsert
( integer_value[ for key from :v_first_key to :v_last_key INCREMENT :v_increment ] = cv(key) ) ; ( integer_value[ for key from :v_first_key to :v_last_key INCREMENT :v_increment ] ERROR at line 8: ORA-32626: illegal bounds or increment in MODEL FOR loop
ROWNUM + a Big Table Method

This tutorial demonstrates how to generate a series of integers using the ROWNUM pseudocolumn and any available table with as many rows in it as the number of integers required. Other techniques are discussed in the tutorials listed in the menu to the left. Prerequisites Before using this solution you need to find a table with at least as many rows in it as the number of integers you need to generate. I.e. if you need a series of 10 integers then you need to find a table or view that will always have at least 10 rows in it. The data dictionary view ALL_OBJECTS is a popular choice for this method. The Solution Once you have identified a table with a sufficient number of rows simply select ROWNUM from it to generate the required integer series, like this.
select rownum from all_objects where rownum <= 10 ; ROWNUM ---------1 2 3 4 5 6 7 8 9 10
We can apply this technique to our day of the week scenario as follows.
select day_of_week , t.val from ( select rownum - 1 as day_of_week
from all_objects where rownum <= 7 ) i left outer join t using ( day_of_week ) order by day_of_week ; DAY_OF_WEEK VAL ----------- ---------0 1 100 2 3 300 4 400 5 500 6
CONNECT BY LEVEL Method

This tutorial demonstrates how to generate a series of integers using a novel application of the CONNECT BY clause first posted by Mikito Harakiri at Ask Tom "how to display selective record twice in the query?". Other techniques are discussed in the tutorials listed in the menu to the left. With this technique you can generate a series of integers starting at "1" using a query like this.
select level from dual connect by level <= 10 ; LEVEL ---------1 2 3 4 5 6 7 8 9 10
Queries Without PRIOR The query above is a special case of a more general type of query, those that do not use the PRIOR operator. Applying the technique to a table with two rows, "a" and "b", yields some insight into how such queries work.
break on level duplicates skip 1 column path format a10
select from connect by order by LEVEL ---------1 1 2 2 2 2 3 3 3 3 3 3 3 3
level, sys_connect_by_path( key, '/' ) as path, key t4 level <= 3 level, path ; PATH ---------/a /b /a/a /a/b /b/a /b/b /a/a/a /a/a/b /a/b/a /a/b/b /b/a/a /b/a/b /b/b/a /b/b/b KEY --a b a b a b a b a b a b a b
Without a CONNECT BY condition that uses PRIOR it appears Oracle returns all possible hierarchy permutations. This effect may be useful where an exponentially increasing number of output rows is required. The effect also proves useful in situations where more than one integer series is required from a single query. See Multiple Integer Series: CONNECT BY LEVEL Method for more details and an example. Note there is some debate about whether queries without PRIOR in the CONNECT BY clause are legal or not. This is discussed further in the "Gotchas" section below. Variables The original syntax for this technique works fine when the number of rows is hardcoded to a value greater than or equal to 1. If the number of rows is set with a bind variable whose value can be 0, negative, or null however, the technique may not work as expected. It always generates at least one row in these cases.
clear breaks variable v_total_rows number execute :v_total_rows := 0 select level from dual connect by level <= :v_total_rows ;
LEVEL ---------1 execute :v_total_rows := -5 PL/SQL procedure successfully completed. / LEVEL ---------1 1 row selected. execute :v_total_rows := null PL/SQL procedure successfully completed. / LEVEL ---------1 1 row selected.
A simple WHERE clause fixes this behaviour.

execute :v_total_rows := 0 PL/SQL procedure successfully completed. select from WHERE connect level dual :V_TOTAL_ROWS >= 1 by level <= :v_total_rows ;
no rows selected execute :v_total_rows := -5 PL/SQL procedure successfully completed. / no rows selected execute :v_total_rows := null
PL/SQL procedure successfully completed. / no rows selected execute :v_total_rows := 3 PL/SQL procedure successfully completed. / LEVEL ---------1 2 3 3 rows selected.
Day of the Week Case Study In the next snippet we apply the technique to the day of the week scenario we examined in prior tutorials.
select day_of_week , t.val from ( select level - 1 as day_of_week from dual connect by level <= 7 ) i left outer join t using( day_of_week ) order by day_of_week ; DAY_OF_WEEK VAL ----------- ---------0 1 100 2 3 300 4 400 5 500 6
Gotchas
To Use PRIOR or Not to Use PRIOR, That is the Question
Laurent Schneider argues in his blog post Bible of Oracle that a clause like CONNECT BY LEVEL <= 10 is an illegal construct since it has no expressions qualified with the PRIOR operator, as dictated by this statement in the SQL Reference Manual "in a hierarchical query, one expression in condition must be qualified with the PRIOR operator to refer to the parent row." -- Oracle Database SQL Reference 10g Release 2 (10.2) Others argue that this statement is a documentation bug. The fact the CONNECT BY clause works without error in Oracle 10g and some 9i versions somewhat supports this view. Note that the following queries, seemingly equivalent to the CONNECT BY LEVEL <= 10 solution, do not produce the desired 10 rows of output (as tested in Oracle 10g).
select level from dual connect by level <= 10 AND PRIOR DUMMY = DUMMY ; ERROR: ORA-01436: CONNECT BY loop in user data
select level from dual connect by level <= 10 AND PRIOR 1 = 1 ; ERROR: ORA-01436: CONNECT BY loop in user data
The following variation may be more legal than the original solution since it includes a PRIOR condition and does not produce a CONNECT BY loop, but the PL/SQL call it contains makes it perform worse (from Re: Creating N Copies of a Row using "CONNECT BY CONNECT_BY_ROOT" - Volder).
select level from dual connect by level <= 10 and PRIOR DBMS_RANDOM.VALUE IS NOT NULL ; LEVEL ---------1 2 3 4 5 6
7 8 9 10
The jury is still out on using CONNECT BY LEVEL to generate integers. Until there is a definitive answer, be aware there is a risk the technique may not work in future versions.
Issues
Issues with this technique, or variations of it, have been reported in Oracle versions earlier than 10.2. I have not tested these myself but here are some posts that describe problems.

Ask Tom "Can there be an infinite DUAL?" - Weird results Ask Tom "Can there be an infinite DUAL?" - Weird Results (bug?) Ask Tom "how to display selective record twice in the query?"- minor simplification CONNECT BY Generator Rules | Ask Mr. Ed
In Oracle 9i, if you try the CONNECT BY LEVEL technique and get a single row when expecting muliple rows, like this
select level from dual connect by level < 10 ; LEVEL ---------1
putting the query in an inline view, as in this snippet, may help (I have not tested this).
select * from (select level from dual connect by level < 10) ; LEVEL ---------1 2 3 4 5 6 7 8 9 9 rows selected.
Acknowledgements Mikito Harakiri, Tom Kyte, Laurent Schneider and other posters at these threads.

Ask Tom "how to display selective record twice in the query?" Ask Tom "Can there be an infinite DUAL?"
CUBE Method
This tutorial demonstrates how to generate a series of integers using the CUBE clause of the SELECT statement. Other techniques are discussed in the tutorials listed in the menu to the left. Here are some examples that generate a series of integers using CUBE.
To return 4 rows (2^2):
select rownum from ( select 1 from dual group by cube( 1, 2 ) ) ; ROWNUM ---------1 2 3 4

select rownum from ( select 1 from dual group by cube( 1, 2, 3 ) ) ; ROWNUM ---------1 2 3 4 5 6 7 8

ROWNUM ---------1 2 3 4 select rownum 5 from 6 ( 7 select 1 8 from dual 9 group by cube( 1, 2, 3, 4 ) 10 ) ; 11 12 13 14 15 16
To return 9 rows:
select rownum from ( select 1 ROWNUM ---------1 2
from dual group by cube( 1, 2, 3, 4 ) ) where rownum <= 9 ;
3 4 5 6 7 8 9
We can apply this technique to our day of the week scenario with this query.
with days_of_the_week as ( select rownum - 1 as day_of_week from ( select 1 from dual group by cube( 1, 2, 3 ) ) where rownum <= 7 ) select day_of_week , t.val from days_of_the_week left outer join t using ( day_of_week ) order by day_of_week ; DAY_OF_WEEK VAL ----------- ---------0 1 100 2 3 300 4 400 5 500 6
For more details about how this method works see CUBE - Explained. Gotchas
Number of Arguments in CUBE

Ensure the numeric literal in the WHERE clause is less than or equal to 2^(number of CUBE arguments), otherwise you will not get the correct number of rows, as in this example which attempts to generate 7 integers but only succeeds in generating 4.
select rownum as integer_value from ( select 1 from t2 group by cube ( 1, 2 ) -- will only generate 2^2 rows
) where rownum <= 7 -- can only be <= 4, 3, 2, or 1 in this query, not 5, 6, 7, ... ; INTEGER_VALUE ------------1 2 3 4
Inline View
Attempting to use rownum without the inline view will cause errors or incorrect results.
select rownum from t2 group by cube( 1, 2 ) ; select rownum * ERROR at line 1: ORA-00979: not a GROUP BY expression
select rownum from t2 group by rownum, cube( 1, 2 ) ; ROWNUM ---------1 1 1 1
CUBE Method - Explained

To understand how the integer series generator described in the CUBE Method tutorial works we will start with a simple query, transform it into a query that uses CUBE the traditional way, and then turn it into an integer series generator for the values 1 to 4. (Read the Oracle manual page on the CUBE grouping operation first if you are not already familiar with this feature.)
set null "(null)" select * from t2 ; select c1, c2, sum( c3 ) as sum_c3 from t2 GROUP BY C1, C2 ; C1 C2 C3 ------ ------ ---------x y 42 C1 C2 SUM_C3 ------ ------ ---------x y 42
select c1, c2, sum( c3 ) as sum_c3 from t2 group by CUBE( c1, c2 );
C1 -----(null) (null) x x C1 -----(null) (null) x x
C2 SUM_C3 ------ ---------(null) 42 y 42 (null) 42 y 42 C2 ANY_LITERAL ------ ----------(null) 1 y 1 (null) 1 y 1
select c1, c2, 1 AS ANY_LITERAL from t2 group by cube( c1, c2 );
SELECT 1 from t2 group by cube ( c1, c2 ) ;
1 ---------1 1 1 1
1 select 1 ---------from t2 1 group by CUBE ( 1, 2 ) -- see Note 1 1 ; 1 1 select ROWNUM AS INTEGER_VALUE from ( select 1 from t2 group by cube ( 1, 2 ) ); select rownum as integer_value from ( select 1 from t2 group by cube ( 1, 2 ) ) where ROWNUM <= 3 ; INTEGER_VALUE ------------1 2 3 4
INTEGER_VALUE ------------1 2 3
Note 1: In this technique it does not matter what literals you use in the arguments to CUBE. You could use arguments like 1, 1 or 'a','b' and still get the same number of rows. The important part is how many literals you include. Two literals will give you four rows (2^2), three literals will give you eight rows (2^3), four literals will give you sixteen rows (2^4), etc. I like to use arguments like 1,2,3,4,5,6,7 because it is
easier to tell there are 7 arguments (which produce 2^7 rows) with this approach than with an argument list like 1,1,1,1,1,1,1.
Type Constructor Expression Method

This tutorial demonstrates how to generate a series or set of integers using Type Constructor Expressions for collection types. Other techniques are discussed in the tutorials listed in the menu to the left. Prerequisites This solution requires a nested table type or varry type. We will use one called INTEGER_TABLE_TYPE created in the Setup topic for this section. If you do not have privileges to create a type like this see Setup - Note 1.
desc integer_table_type integer_table_type TABLE OF NUMBER(38)
The Solution If you need a manageable number of integers, like 10 or 20, you can use a simple query like this one.
select column_value from table( integer_table_type( 1,2,3,4,5,6,7,8,9,10 ) ) ; COLUMN_VALUE -----------1 2 3 4 5 6 7 8 9 10
This method is unique from the others in this section in that it lends itself well to creating sets of non-sequential integers as well as sequential series.
select column_value from table( integer_table_type( 1,1,4,4,4,8,10 ) ) ; COLUMN_VALUE -----------1 1 4 4
4 8 10
Applying the technique to our day of the week scenario yields this query.
select i.column_value as day_of_week , t.val from table( integer_table_type( 0,1,2,3,4,5,6 ) ) i left outer join t on ( i.column_value = t.day_of_week ) order by i.column_value ; DAY_OF_WEEK VAL ----------- ---------0 1 100 2 3 300 4 400 5 500 6
If you require more integers than you care to list in a type constructor expression see the Type Constructor + Cartesian Product tutorial for a variation of this technique. Gotchas If we specify more than 999 arguments in a type constructor it will generate a ORA-00939: too many arguments for function error (as tested in Oracle 10g).
Type Constructor + Cartesian Product Method

This tutorial demonstrates how to generate a series of integers using Type Constructor Expressions for collection types and Cartesian Products. Other techniques are discussed in the tutorials listed in the menu to the left. Prerequisites This solution requires a nested table type or varry type. We will use one called INTEGER_TABLE_TYPE created in the Setup topic for this section. If you do not have privileges to create a type like this see Setup - Note 1.
The Solution
If you require a large number of integers then listing them all in a type constructor expression, like the solutions in the Type Constructor Expression Method tutorial, may be difficult or impossible. In this case you can use a Cartesian product with your type constructor expressions to generate a large number of rows with a small amount of code. Here are some examples. This query returns 9 rows (3x3).
ROWNUM ---------1 2 3 4 5 6 7 8 9
select rownum from table( integer_table_type( 1,2,3 ) ) i1, table( integer_table_type( 1,2,3 ) ) i2 ;
This query returns 12 rows (3x4).

ROWNUM ---------1 2 3 4 5 6 7 8 9 10 11 12
select rownum from table( integer_table_type( 1,2,3 ) ) i1, table( integer_table_type( 1,2,3,4 ) ) i2 ;
A query like this can return up to 10,000 rows (10^4), though we won't prove this by diplaying them all here. Listing 15 of them should suffice.
ROWNUM ---------1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
with i as ( select * from table ( integer_table_type( 1,2,3,4,5,6,7,8,9,10 ) ) ) select rownum from i,i,i,i where rownum <= 15 ;
How Cartesian Products Work When you do not specify a join between tables Oracle combines each row of one table with each row in the other to produce every possible row combination. This produces a result set with
(number of rows in Table 1) x (number of rows in Table 2)
rows in it. The following query illustrates this.

select rownum , i1.column_value i1_column_value, i2.column_value i2_column_value from table( integer_table_type( 1,2 ) ) i1, table( integer_table_type( 10,20,30 ) ) i2 ; ROWNUM I1_COLUMN_VALUE I2_COLUMN_VALUE ---------- --------------- --------------1 1 10 2 1 20 3 1 30 4 2 10 5 2 20 6 2 30
Pipelined Function Method

This tutorial demonstrates how to generate a series of integers using a Pipelined Function. Other techniques are discussed in the tutorials listed in the menu to the left. Prerequisites This solution requires a nested table type or varry type. We will use one called INTEGER_TABLE_TYPE created in the Setup topic for this section. If you do not have privileges to create a type like this see Setup - Note 1.
You will also need the following custom database function.

create function integer_series ( p_lower_bound in number, p_upper_bound in number ) return integer_table_type pipelined as begin
for i in p_lower_bound .. p_upper_bound
loop pipe row(i); end loop; return; end; /
The Solution Now that we have our prerequisites in place, here is how you use the INTEGER_SERIES function.
select * from table( integer_series(-5,7) ) ; COLUMN_VALUE ------------5 -4 -3 -2 -1 0 1 2 3 4 5 6 7
We can apply this technique to our day of the week scenario with this query.
select i.column_value as day_of_week , t.val from table( integer_series(0,6) ) i left outer join t on ( i.column_value = t.day_of_week ) order by i.column_value ; DAY_OF_WEEK VAL ----------- ---------0 1 100 2 3 300 4 400 5 500 6
Performance Comparison - Small Numbers

The following tables show performance metrics for one run each of the eight integer series generation techniques described in the preceeding tutorials. Integer Table Method MODEL Method ROWNUM + a Big Table Method CONNECT BY LEVEL Method CUBE Method Type Constructor Expression Method Type Constructor + Cartesian Product Method Pipelined Function Method Each run generated a series of integers from 1 to 100. See the log file from these tests for more details. Statistics The following table shows database statistics where values for one method differ by more than 100 from another method.
Integer ROWNUM CONNECT BY Type Type Constructor Pipelined METRIC_NAME Table MODEL + Big Table LEVEL CUBE Constructor + Cartesian Product Function ---------------------------------------- ------------ ----------------------- ------------ ------------ ------------ -----------------------------Elapsed Time (1/100 sec) 3 3 3 2 5 262 3 4 session pga memory max 262,144 262,144 262,144 262,144 262,144 262,144 262,144 327,680 session pga memory 196,608 65,536 131,072 131,072 196,608 65,536 131,072 327,680 redo size 2,744 2,640 2,684 2,684 2,684 2,684 2,684 2,684 sorts (rows) 2,071 2,071 2,071 2,072 2,199 2,071 2,081 2,076 session uga memory 0 0 0 0 0 65,464 65,464 65,464
See Statistics Descriptions for a description of each metric. Latch Gets The following table shows total latch gets for each method.
Integer ROWNUM CONNECT BY Type Type Constructor Pipelined METRIC_NAME Table MODEL + Big Table LEVEL CUBE Constructor + Cartesian Product Function ---------------------------------------- ------------ ----------------------- ------------ ------------ ------------ -----------------------------cache buffers chains 206 163 221 163 163 1,208 181 249 row cache objects 136 96 129 87 84 135 114 171 library cache 92 77 86 80 77 113 89 179 shared pool 70 25 26 24 25 85 29 45 session idle bit 56 55 56 55 55 56 55 57 library cache pin 50 43 48 43 43 57 45 89 library cache lock 26 20 28 20 20 40 25 72 enqueues 23 16 17 16 16 26 16 20 enqueue hash chains 22 16 16 16 16 26 16 18 shared pool simulator 12 9 7 9 9 16 10 17 object queue header operation 8 12 12 12 12 305 12 15 redo allocation 8 8 8 8 8 18 8 8 cache buffers lru chain 7 6 6 6 6 396 6 7 SQL memory manager workarea list latch 6 10 6 6 6 73 6 6 session allocation 6 2 4 3 2 2 2 6 sort extent pool 4 4 4 4 4 4 4 4 session switching 4 4 4 4 4 4 4 4 kks stats 4 2 2 2 2 4 2 2 simulator hash latch 4 0 10 0 0 134 0 1 simulator lru latch 4 0 10 0 0 130 0 1 PL/SQL warning settings 3 3 3 3 3 3 3 3 compile environment latch 2 1 2 1 1 1 1 3 object stats modification 1 1 2 1 1 1 1 2 library cache lock allocation 1 1 1 1 1 1 1 2
dml lock allocation 1 1 1 1 FOB s.o list latch 0 1 0 0 OS process 0 3 0 0 messages 0 1 0 40 channel operations parent latch 0 1 0 18 channel handle pool latch 0 1 0 0 OS process allocation 0 1 0 0 process allocation 0 1 0 0 process group creation 0 1 0 0 checkpoint queue latch 0 0 0 269 redo writing 0 0 0 13 active checkpoint queue latch 0 0 0 13 loader state object freelist 0 0 0 12 virtual circuit buffers 0 0 0 9 virtual circuit queues 0 0 0 7 parallel query alloc buffer 0 0 0 4 user lock 0 0 0 4 session timer 0 0 0 3 library cache load lock 0 0 0 2 virtual circuits 0 0 0 2 active service list 0 0 0 2 library cache pin allocation 0 0 0 1 resmgr:actses active list 0 0 0 1 XDB unused session pool 0 0 0 1 KMG MMAN ready and startup request latch 0 0 0 1 resmgr:free threads list 0 0 0 1
1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0
1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 0 0 1 0 0 0
0 0 ------------ ----------------------- ------------ ------------ ------------ -----------------------------sum 757 575 709 575 559 3,242 632 986
Techniques that use a small number of latches scale better than techniques that use a large number of latches. Warning: Results on your own systems with your own data will differ from these results. Results will even differ from one set of test runs to the next on the same machine. Run your own tests and average the results from multiple runs before making performance decisions.
Performance Comparison - Large Numbers

The following tables show performance metrics for one run each of six integer series generation techniques described in the preceeding tutorials. Integer Table Method MODEL Method ROWNUM + a Big Table Method CONNECT BY LEVEL Method Type Constructor + Cartesian Product Method Pipelined Function Method Each run generated a series of integers from 1 to 100,000. Note that the Type Constructor Expression Method technique was excluded from this comparison because it can only be used to generate up to 999 different values. The CUBE Method technique was excluded from this test because it failed to complete in under 10 minutes. See the log file from these tests for more details. Statistics The following table shows database statistics where values for one method differ by more than 100 from another method.
Integer ROWNUM CONNECT BY Type Constructor Pipelined METRIC_NAME Table MODEL Table LEVEL + Cartesian Product Function ---------------------------------------- ------------ ----------------------- ------------ ------------------- -----------Elapsed Time (1/100 sec) 59 349 68 67 57 544 session pga memory max 262,144 4,784,128 262,144 2,031,616 262,144 327,680 session uga memory max 261,964 4,533,984 261,964 2,016,252 261,964 261,964 session pga memory 196,608 65,536 131,072 0 262,144 327,680 session logical reads 6,879 45 6,927 45 78 111 consistent gets 6,840 6 6,888 6 39 72
+ Big
consistent gets from cache 6,888 6 39 no work - consistent read gets 6,860 0 0 buffer is not pinned count 0 0 22 DB time 27 37 22 CPU used when call started 27 35 22 CPU used by this session 27 35 21 session uga memory 0 0 65,464
37 34 34 34 65,464
6,840 72 6,822 5 6,658 29 29 29 0
6 0 0 311 309 309 0
See Statistics Descriptions for a description of each metric. Latch Gets The following table shows total latch gets for each method.
Integer ROWNUM CONNECT BY Type Constructor Pipelined METRIC_NAME Table MODEL Table LEVEL + Cartesian Product Function ---------------------------------------- ------------ ----------------------- ------------ ------------------- -----------cache buffers chains 13,946 836 13,927 163 196 258 session idle bit 13,376 13,375 13,376 13,375 13,375 13,377 simulator lru latch 424 172 461 0 0 1 simulator hash latch 424 172 461 0 0 1 row cache objects 139 96 129 87 132 171 cache buffers lru chain 106 522 6 6 6 7 library cache 92 83 85 79 119 20,172 object queue header operation 79 365 12 12 15 18 checkpoint queue latch 55 237 0 0 11 43 library cache pin 50 49 48 43 54 13,417 shared pool 47 20 18 17 45 49 library cache lock 30 24 30 22 34 79 enqueues 21 96 17 16 16 100
+ Big
enqueue hash chains 16 16 16 messages 4 0 8 shared pool simulator 6 9 16 redo allocation 8 8 8 SQL memory manager workarea list latch 6 6 6 channel operations parent latch 0 0 6 session allocation 4 2 2 session switching 4 4 4 sort extent pool 4 4 4 kks stats 2 2 2 PL/SQL warning settings 3 3 3 redo writing 0 0 1 active checkpoint queue latch 0 0 1 compile environment latch 2 1 1 object stats modification 1 2 1 library cache lock allocation 1 1 1 dml lock allocation 1 1 1 session timer 0 0 1 KMG MMAN ready and startup request latch 0 0 1 object queue header heap 0 0 0 JS queue state obj latch 0 0 0 active service list 0 0 0 qmn task queue latch 0 0 0 In memory undo latch 0 0 0 OS process allocation 1 0 0 resmgr:actses active list 0 0 0 resmgr:schema config 0 0 0 kwqbsn:qsga 0 0 0
99 48 17 12 140 29 6 4 4 2 3 8 1 3 1 2 1 2 1 0 36 10 0 2 2 1 1 0
20 10 10 9 6 6 6 4 4 4 3 2 2 2 1 1 1 1 1 1 0 0 0 0 0 0 0 0
97 28 9 12 144 22 2 4 4 2 3 7 3 1 1 1 1 1 1 0 36 10 4 2 1 1 1 1
Shared B-Tree 0 0 library cache load lock 0 0 library cache pin allocation 0 0 mostly latch-free SCN 0 0 undo global data 0 0 lgwr LWN SCN 0 0 Consistent RBA 0 0 FOB s.o list latch 0 0
0 2 1 0 0 0 0 0
0 2 1 1 1 1 1
0 0 0 0 0 0 0
1 0 0 0 0 0 0
0 0 1 ------------ ----------------------- ------------ ------------------- -----------sum 28,883 16,447 28,633 13,879 14,089 48,136
Techniques that use a small number of latches scale better than techniques that use a large number of latches. Warning: Results on your own systems with your own data will differ from these results. Results will even differ from one set of test runs to the next on the same machine. Run your own tests and average the results from multiple runs before making performance decisions.
Multiple Integer Series

The preceding topics showed how to generate a single series of integers. Sometimes many different integer series are required in the same query. For example, given a data table like this
KEY QTY --- ---------a (null) b 0 c 1 d 2 e 3
we sometimes need queries that generate results like these

KEY QTY INTEGER_VALUE --- ---------- ------------a (null) (null) b c d d e e e 0 (null) 1 2 2 3 3 3 1 1 2 1 2 3
where a row like the one where KEY='d' needs the integer series "1,2" but the row where KEY='e' needs the series "1,2,3". The subtopics in this section demonstrate various ways to accomplish this.
Join Method
Many of the single series solutions presented earlier can be easily adapted to generate multiple series with a simple join. Here is an example of how this is done using the Integer Table Method technique.
create table integers ( integer_value integer primary key ) organization index ; insert into insert into insert into insert into insert into commit; integers integers integers integers integers values values values values values ( ( ( ( ( 1 2 3 4 5 ); ); ); ); );
set null "(null)" break on key duplicates skip 1 select key, qty, integer_value from t3 left outer join integers on ( integers.integer_value <= t3.qty ) order by key, integer_value ; KEY QTY INTEGER_VALUE --- ---------- ------------a (null) (null) b c d d e e e 0 (null) 1 2 2 3 3 3 1 1 2 1 2 3
With the Type Constructor Expression Method technique it would look like this.
select key, qty, column_value as integer_value from t3 left outer join table( integer_table_type(1,2,3,4,5) ) integers on ( integers.column_value <= t3.qty ) order by key, integer_value ; KEY QTY INTEGER_VALUE --- ---------- ------------a (null) (null) b c d d e e e 0 (null) 1 2 2 3 3 3 1 1 2 1 2 3
In both queries note that we first generate more integers than required and then filter out the excess values via a join condition.
MODEL Method
With MODEL queries there is no need to use the join technique described in Join Method. We can generate multiple series by applying a FOR loop to each row in the base table with the aid of a PARTITION BY clause.
set null "(null)" break on key duplicates skip 1 select key, key2, qty, integer_value from t3 model PARTITION BY ( KEY ) dimension by ( 1 as key2 ) measures ( qty, cast( null as integer ) as integer_value ) rules ( integer_value[ FOR KEY2 FROM 1 TO QTY[1] INCREMENT 1 ] = cv(key2) ) order by key, integer_value ; KEY KEY2 QTY INTEGER_VALUE --- ---------- ---------- ------------a 1 (null) (null) b c d d e e e 1 1 1 2 (null) 1 2 (null) 3 (null) 0 (null) 1 2 3 1 1 2 1 2 3
CONNECT BY LEVEL Method

With the CONNECT BY LEVEL approach there is also no need to use the Join Method. Multiple integer series can be created using a query like this one (the PATH column is included to illustrate how the query works)
set null "(null)" break on key duplicates skip 1 column path format a10
select key, qty, level as integer_value, sys_connect_by_path( key, '/' ) as path from t3 where qty >= 1 connect by KEY = PRIOR KEY and prior dbms_random.value is not null and level <= t3.qty order by key, integer_value ; KEY QTY INTEGER_VALUE PATH --- ---------- ------------- ---------c 1 1 /c d d e e e 2 2 3 3 3 1 /d 2 /d/d 1 /e 2 /e/e 3 /e/e/e
or this approach (only possible on version 10g or greater)

select key, qty, level as integer_value, sys_connect_by_path( key, '/' ) as path from t3 where qty >= 1 connect by KEY = CONNECT_BY_ROOT KEY and level <= t3.qty order by key, integer_value ; KEY QTY INTEGER_VALUE PATH --- ---------- ------------- ---------c 1 1 /c d d e e e 2 2 3 3 3 1 /d 2 /d/d 1 /e 2 /e/e 3 /e/e/e
Note these approaches will not work for rows where no integer series is required, like the rows with KEY in ( 'a', 'b' ). Gotchas
The CONNECT_BY_ROOT technique may not work without error in all cases and it may not work in Oracle versions beyond 10g. This is because it violates two restrictions documented at Hierarchical Query Operators: 1. "In a hierarchical query, one expression in the CONNECT BY condition must be qualified by the PRIOR operator." 2. "You cannot specify (CONNECT_BY_ROOT) in the START WITH condition or the CONNECT BY condition." The fact that the query above contradicts the documentation yet works without error in 10g suggests a bug in either the documentation or the SQL engine. On my system the following variation of the CONNECT_BY_ROOT query raised some rather severe ORA errors casting further doubt on the technique's reliability (do not run this query on your own systems).
select key, qty, level as integer_value from t3 start with qty >= 1 connect by KEY = CONNECT_BY_ROOT KEY and level <= t3.qty order by key, integer_value ; ERROR at line 1: ORA-03113: end-of-file on communication channel ERROR: ORA-03114: not connected to ORACLE From the .trc file: ORA-07445: exception encountered: core dump [ACCESS_VIOLATION] [__VInfreq__msqopnws+2740] [PC:0x30E2580] [ADDR:0x2C] [UNABLE_TO_READ] []
Bulk Collect
Executing sql statements in plsql programs causes a context switch between the plsql engine and the sql engine. Too many context switches may degrade performance dramatically. In order to reduce the number of these context switches we can use a feature named bulk binding. Bulk binding lets us to transfer rows between the sql engine and the plsql engine as collections. Bulk binding is available for select, insert, delete and update statements. Bulk collect is the bulk binding syntax for select statements. One of the things i usuallly come accross is that developers usually tend to use cursor for loops to process data. They declare a cursor, open it, fetch from it row by row in a loop and process the row they fetch.
Declare Cursor c1 is select column_list from table_name>; Rec1 c1%rowtype; Begin Open c1; Loop; Fetch c1 into r1; Exit when c1%notfound; --process rows... End loop; End;
Here is a simple test case to compare the performance of fetching row by row and using bulk collect to fetch all rows into a collection.
SQL> create table t_all_objects as select * from all_objects; Table created. SQL> insert into t_all_objects select * from t_all_objects; 3332 rows created. SQL> r 1* insert into t_all_objects select * from t_all_objects 6664 rows created. ---replicated a couple of times SQL> select count(*) from t_all_objects; COUNT(*) ---------213248 SQL> declare cursor c1 is select object_name from t_all_objects; 2 3 rec1 c1%rowtype; 4 begin 5 open c1; 6 loop
7 8 9 10 11 12 13 14
fetch c1 into rec1; exit when c1%notfound; null; end loop; end; /
PL/SQL procedure successfully completed. Elapsed: 00:00:44.75 SQL> declare 2 cursor c1 is select object_name from t_all_objects; 3 type c1_type is table of c1%rowtype; 4 rec1 c1_type; 5 begin 6 open c1; 7 8 fetch c1 bulk collect into rec1; 9 10 11 end; 12 / PL/SQL procedure successfully completed. Elapsed: 00:00:05.32
As can be clearly seen, bulk collecting the rows shows a huge performance improvement over fetching row by row. The above method (which fetched all the rows) may not be applicable to all cases. When there are many rows to process, we can limit the number of rows to bulk collect, process those rows and fetch again. Otherwise process memory gets bigger and bigger as you fetch the rows.
SQL> declare 2 cursor c1 is select object_name from t_all_objects; 3 type c1_type is table of c1%rowtype; 4 rec1 c1_type; 5 begin 6 open c1; 7 loop 8 fetch c1 bulk collect into rec1 limit 200; 9 for i in 1..rec1.count loop 10 null; 11 end loop; 12 exit when c1%notfound; 13 end loop; 14 15 16 end; 17 / PL/SQL procedure successfully completed. Elapsed: 00:00:04.07
Deciding When to Use Bulk Binds

PL/SQL code that uses bulk binds will be slightly more complicated and somewhat more prone to programmer bugs than code without bulk binds, so you need to ask yourself if the improved runtime performance will justify the expense. No universal rule exists to dictate when bulk binds are worthwhile and when they are not. However, the cost of adding a few lines of code is so slight that I would lean toward using bulk binds when in doubt. A PL/SQL program that reads a dozen rows from a cursor will probably see no noticeable benefit from bulk binds. The same goes for a program that issues five or six UPDATE statements. However, a program that reads 1,000 rows from a cursor or performs that many similar UPDATE statements will most likely benefit from bulk binds. If you have the luxury of time, you can test your code both with and without bulk binds. Running both versions of the code through SQL trace and TKPROF will yield reports from which you may derive a wealth of information.
A Simple Program With and Without Bulk Binds

In this section we will look at a simple program written both with and without bulk binds. We'll look at TKPROF reports that demonstrate the impact bulk binds can have. The discussion of the TKPROF reports will help you see how to interpret TKPROF output in order to assess the impact of bulk binds on your application. Consider the following excerpts from a TKPROF report:
************************************************************************ DECLARE CURSOR c_orders IS SELECT order_id, currency_code, amount_local /* no bulk bind */ FROM open_orders; v_amount_usd NUMBER; BEGIN FOR r IN c_orders LOOP v_amount_usd := currency_convert (r.amount_local, r.currency_code); UPDATE open_orders /* no bulk bind */ SET amount_usd = v_amount_usd WHERE order_id = r.order_id; END LOOP; COMMIT; END; call count ------- -----Parse 1 Execute 1 Fetch 0 ------- -----total 2 cpu elapsed disk query current -------- ---------- -------- -------- -------0.05 0.04 0 0 1 10.55 11.40 0 0 0 0.00 0.00 0 0 0 -------- ---------- -------- -------- -------10.60 11.45 0 0 1 rows -------0 1 0 -------1
************************************************************************ SELECT order_id, currency_code, amount_local /* no bulk bind */ FROM open_orders call count ------- -----Parse 1 Execute 1 Fetch 30287 ------- -----total 30289 cpu elapsed disk query current -------- ---------- -------- -------- -------0.00 0.00 0 0 0 0.00 0.00 0 0 0 1.08 1.10 0 30393 0 -------- ---------- -------- -------- -------1.08 1.10 0 30393 0 rows -------0 0 30286 -------30286
************************************************************************ UPDATE open_orders /* no bulk bind */ SET amount_usd = :b2 WHERE order_id = :b1 call count ------- -----Parse 1 Execute 30286 Fetch 0 ------- -----total 30287 cpu elapsed disk query current -------- ---------- -------- -------- -------0.00 0.00 0 0 0 7.19 7.32 1 60576 31022 0.00 0.00 0 0 0 -------- ---------- -------- -------- -------7.19 7.33 1 60576 31022 rows -------0 30286 0 -------30286
As you can see, this is a very simple program that does not use bulk binds. (The code borders on being silly; please recognize it is for illustrative purposes only.) The PL/SQL engine used 10.55 CPU seconds to run this code (this figure does not include CPU time used by the SQL engine). There were 30,287 fetch calls against the cursor, requiring 30,393 logical reads and 1.08 CPU seconds. The UPDATE statement was executed 30,286 times, using 7.19 CPU seconds. Now consider the following excerpts from another TKPROF report:
************************************************************************ DECLARE CURSOR c_orders IS SELECT order_id, currency_code, amount_local /* bulk bind */ FROM open_orders; TYPE t_num_array IS TABLE OF NUMBER INDEX BY BINARY_INTEGER; TYPE t_char_array IS TABLE OF VARCHAR2(10) INDEX BY BINARY_INTEGER; v_order_ids t_num_array; v_currency_codes t_char_array; v_amounts_local t_num_array; v_amounts_usd t_num_array; v_row_count NUMBER := 0; BEGIN OPEN c_orders; LOOP FETCH c_orders BULK COLLECT INTO v_order_ids, v_currency_codes, v_amounts_local LIMIT 100; EXIT WHEN v_row_count = c_orders%ROWCOUNT;
v_row_count := c_orders%ROWCOUNT; FOR i IN 1..v_order_ids.count LOOP v_amounts_usd(i) := currency_convert (v_amounts_local(i), v_currency_codes(i)); END LOOP; FORALL i IN 1..v_order_ids.count UPDATE open_orders /* bulk bind */ SET amount_usd = v_amounts_usd(i) WHERE order_id = v_order_ids(i); END LOOP; CLOSE c_orders; COMMIT; END; call count ------- -----Parse 1 Execute 1 Fetch 0 ------- -----total 2 cpu elapsed disk query current -------- ---------- -------- -------- -------0.03 0.03 0 0 0 0.60 0.62 0 0 0 0.00 0.00 0 0 0 -------- ---------- -------- -------- -------0.63 0.66 0 0 0 rows -------0 1 0 -------1
************************************************************************ SELECT order_id, currency_code, amount_local /* bulk bind */ FROM open_orders call count ------- -----Parse 1 Execute 1 Fetch 303 ------- -----total 305 cpu elapsed disk query current -------- ---------- -------- -------- -------0.00 0.00 0 0 0 0.00 0.00 0 0 0 0.48 0.59 0 4815 0 -------- ---------- -------- -------- -------0.48 0.59 0 4815 0 rows -------0 0 30286 -------30286
************************************************************************ UPDATE open_orders /* bulk bind */ SET amount_usd = :b1 WHERE order_id = :b2 call count ------- -----Parse 1 Execute 303 Fetch 0 ------- -----total 304 cpu elapsed disk query current -------- ---------- -------- -------- -------0.00 0.00 0 0 0 3.75 8.38 0 30895 31021 0.00 0.00 0 0 0 -------- ---------- -------- -------- -------3.75 8.38 0 30895 31021 rows -------0 30286 0 -------30286
This code uses bulk binds to do the same thing as the first code sample, but works with data 100 rows at a time instead of one row at a time. We can see that CPU time used by the PL/SQL engine has reduced from 10.55 seconds to 0.60 seconds. There were only 303 fetch calls against the cursor instead of 30,287, bringing logical reads down from 30,393 to 4,815 and CPU time down from 1.08 seconds to 0.48 seconds. The UPDATE statement was executed only 303 times instead of 30,286, reducing CPU time from 7.19 seconds to 3.75 seconds.
In this example it would appear that bulk binds were definitely worthwhile CPU time was reduced by about 75%, elapsed time by 50%, and logical reads by 50%. Although bulk binds are indeed beneficial here, the benefit is not truly as rosy as it appears in these TKPROF reports. The SQL trace facility imparts an overhead that is proportional to the number of parse, execute, and fetch calls to the SQL engine. Since bulk binds reduce the number of SQL calls, SQL trace adds much less overhead to code that uses bulk binds. While these TKPROF reports suggest that in this example bulk binds shaved about 50% off of the elapsed time, the savings were about 25% when SQL trace was not enabled. This is still a significant savings. Thus using bulk binds in your PL/SQL programs can certainly be worth the effort. Just remember that SQL trace can inflate the perceived benefit.
Conclusion
Bulk binds allow PL/SQL programs to interact more efficiently with the SQL engine built into Oracle, enabling your PL/SQL programs to use less CPU time and run faster. The Oracle Call Interface has supported array processing for 15 years or more, and the increased efficiency it brings is well understood. It is nice to see this benefit available to PL/SQL programmers as well. PL/SQL bulk binds are not hard to implement, and can offer significant performance improvements for certain types of programs.

PL SQL Questions

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

PL SQL Questions

Uploaded by

Copyright:

Available Formats

http://www.sqlsnippets.com/en/home.

htm l PL/SQL Collections

Flexibility to Most flexible. Size can changes increase/ decrease

Not very flexible. You must retrieve and

Which Collection Type to Use?

Use nested tables when:

Use varrays when:

Declare a collection variable.

Declare, initialize, and load a collection variable.

nt_0 is NULL nt is (1) a (2) b .first .last .count .limit

va_0 is NULL va is (1) a (2) b .first .last .count .limit

Initialize a collection after it has been declared.

Add individual rows to a collection.

aa(4) := 'd' ; aa(5) := 'e' ;

Load a single value from the database into a collection row.

Test a row's existence by subscript.

Test a row's existence by content. -- use a loop (see below)

-- use a loop (see below)

'C' member of nt is TRUE

'X' member of nt is FALSE Compare two collections for equality.

-- cannot use "=" with

if nt_0 = nt then p.print( 'equal' ); else p.print( 'not equal' );

Update a collection row.

Remove rows from the middle of a collection.

.limit = 10 Loop through all rows in the collection.

for i in nvl(va.first,0) .. nvl(va.last,-1) loop p.print ( i ||'. '|| va(i) );

7. G Remove row(s) from the end of a collection.

va.extend(3) va(5) := 'E' va(6) := 'F' va(7) := 'G'

Delete all rows in the collection (frees memory too).

Set a collection to NULL, i.e. uninitialized state. -- not possible

Nested Table Y Integer 1..2**31

Varray Y Y Integer 1..2**31 Y

The Difference Between DECODE and CASE

1. CASE can work with logical operators other than =

ENAME ---------SMITH ALLEN WARD

SAL_GRADE --------Grade I Grade II Grade II

2. CASE can work with predicates and searchable subqueries

3. CASE can work as a PL/SQL construct

else dbms_output.put_line('no such grade'); end case; end; /

PL/SQL procedure successfully completed.

create or replace procedure proc_test (i number) as begin dbms_output.put_line('output = '||i); end; /

4. Careful! CASE handles NULL differently

The searched CASE works as does DECODE, though.

5. CASE expects data type consistency, DECODE does not

6. CASE is ANSI SQL-compliant

7. The difference in readability

Grouping Rows with GROUP BY

GRP_B COUNT(*) MAX(VAL) SUM(VAL) ---------- ---------- ---------- ---------b1 2 20 30 b2 3 50 120 b3 3 32 66

GROUP BY and DISTINCT

GROUP BY and Ordering

returns the same results as this one.

insert insert insert insert insert commit;

into into into into into

values values values values values

'a1' 'a1' 'a1' 'a1' 'a1'

'b2' 'b2' 'b2' 'b1' 'b1'

'c2', 'c1', 'c1', 'c1', 'c1',

'd1', 'd1', 'd1', 'd1', 'd1',

'50' '40' '30' '20' '10'

-- (your results may vary)

ORA-00934: group function is not allowed here

For these types of conditions the HAVING clause can be used.

but doing so yields the same result as using a WHERE clause.