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

limit = 10 Load a single value from the database into a collection row.add 1 row va.create two copies -.count = .first .print( nt_0 ). p. aa(2) := 'b' . -.add 2 rows va.last = .create two copies -.extend .limit = = = = 1 7 7 NULL (1) a (2) b (3) c (4) d (5) e (6) e (7) e . aa(4) := 'd' . va(5) := 'e' .last .count .last = 7 .first . -. -. select into from where val nt(2) t val = 'B' . va(4) := 'd' . nt(5) := 'e' .last . aa(7) := 'e' . nt.of row #5 aa(6) := 'e' .5) . p.va_type(). va(3) := 'c' .count .extend(2. NOT NULL .5) . -.n/a nt_0 := p.first = .first = 1 . . p.print( nt ).first = .Initialize a collection after it has been declared. p.add 2 rows nt.limit = and empty NULL NULL 0 10 Add individual rows to a collection.count = 7 . (1) a (2) b (3) c (4) d (5) e (6) e (7) e . -.limit = = = = 1 7 7 NULL (1) a (2) b (3) c (4) d (5) e (6) e (7) e . NOT NULL . aa(5) := 'e' .add 1 row at a time aa(1) := 'a' . select into from where val va(2) t val = 'B' .add 1 row nt. aa(3) := 'c' .print( va_0 ).extend . -.limit = and empty NULL NULL 0 NULL va_0 := p.print( va ).count = .of row #5 va.nt_type() .print( aa ).extend(2. nt(3) := 'c' .extend(2) .extend(2) . -. nt(4) := 'd' . p. select into from where val aa(2) t val = 'B' .last = . -.

print( aa ).limit = NULL .count = 7 .print ( 'aa. p.exists(3) ) ).exists(9) ) p. select val bulk collect into va from t .p. select val bulk collect into aa from t .print ( 'nt.limit = = = = 1 7 7 NULL (1) a (2) B (3) c (4) d (5) e (6) e (7) e .last = 7 .exists(3) is '|| p.exists(3) is TRUE nt.exists(9) is '|| p.exists(9) is FALSE .exists(9) is '|| p.exists(9) is FALSE nt.limit = 10 Test a row's existence by subscript.print ( 'nt. (1) A (2) B (3) C (4) D (5) E (6) F (7) G .count = 7 p.print( va ).last . ).first = 1 . p. p.exists(9) is '|| p.tf( nt.tf( aa. p.last .print( aa ).last = 7 .print ( 'va.print( nt ).exists(3) ) ).print ( 'va.first = 1 .print ( 'aa. select val bulk collect into nt from t .count = 7 p.last = 7 .exists(3) is TRUE aa. p.limit = = = = 1 7 7 NULL (1) a (2) B (3) c (4) d (5) e (6) e (7) e .last = 7 .tf( va.tf( va. ).exists(9) is FALSE va.exists(3) is '|| p.exists(3) is '|| p.print( nt ). p. aa.first = 1 .exists(3) is TRUE va.count .exists(3) ) ). p.print( va ).first = 1 . (1) A (2) B (3) C (4) D (5) E (6) F (7) G .tf( nt. (1) a (2) B (3) c (4) d (5) e (6) e (7) e .tf( aa.count = 7 .limit = NULL .first .limit = 10 Initialize a collection and load it with multiple database values (pre-existing contents will be lost).first .count . (1) A (2) B (3) C (4) D (5) E (6) F (7) G .exists(9) ) p.exists(9) ) ).

count = 7 . va(3) := 'c' .tf( 'X' member of nt ) -. (1) a (2) B (3) c (4) D (5) E (6) F (7) G .two varrays end if.print ( '''C'' member of nt is '|| p. 'C' member of nt is TRUE 'X' member of nt is FALSE Compare two collections for equality.print ( '''X'' member of nt is '|| p. else p. nt(3) := 'c' . (1) a (2) B (3) c (4) D (5) E (6) F (7) G .last = 7 . -.print( aa ). -.use a loop (see below) p.print( nt ). va(1) := 'a' .limit = NULL . aa(1) := 'a' .count = 7 p.print( 'not equal' ).cannot use "=" with -. (1) a (2) B (3) c (4) D (5) E (6) F (7) G . p. equal Update a collection row.print( va ).last = 7 . nt(1) := 'a' .first = 1 .two associative arrays nt_0 := nt .print( 'equal' ).limit = 10 Remove rows from the middle of a collection.tf( 'C' member of nt ) ). .last = 7 . p.first = 1 .count = 7 p.cannot use "=" with if nt_0 = nt then p.limit = NULL . -.Test a row's existence by content.use a loop (see below) ). -. aa(3) := 'c' .first = 1 .

6.trim. 1. G 1.limit = NULL .count = 1 p. while i is not null loop p.print ( i ||'.4).print( nt ). -.limit = 10 Loop through all rows in the collection. nt. nt. a 5. end loop. aa. 3.print ( i ||'.delete(2). 7. declare i binary_integer . p. 5.last = 7 .last = 7 . end. end loop.trim(2).first = 4 .count = 4 p.trim(2).first .delete(7). (1) a (2) B (3) c (4) D .4). (1) a (5) E (6) F (7) G . G Remove row(s) from the end of a collection. 1.6). begin i := aa. nt.trim.0) .-1) loop p.aa.not possible p.print( aa ).limit = NULL . E 6. a B c D E F va. while i is not null loop p. a 5.count = 7 .first .limit = NULL . (1) a (2) B (3) c (4) D (5) E (6) F (7) G .. va. F end. (4) D .first = 4 . 4.first. begin i := nt.delete(2). E 6. nt. F 7.limit = NULL . 2.count = 4 p. '|| va(i) ).delete(5.last. declare i binary_integer .first = 4 .last = 4 . aa. i := aa. (4) D . '|| aa(i) ).print( va ). aa.last = 7 . for i in nvl(va.print ( i ||'. i := nt.next(i) .first = 4 . '|| nt(i) ).delete(3.next(i) . G 7.first = 1 .last = 4 . nvl(va.count = 1 p.first = 1 . (1) a (5) E (6) F (7) G .print( nt ).delete(3. end loop.print( va ).print( aa ).

va. . NOT NULL .limit = NULL .count = 7 (1) a (2) B (3) c (4) D (5) E (6) F (7) G .print( nt ).count = 7 (1) A (2) B (3) C (4) D (5) E (6) F (7) G .call ".first = 1 . . NOT NULL . aa(6) := 'F' . nt(4) := 'D' .print( aa ).4) and trim operations (rows 5.last = 7 . -.2.last = 7 . ----we need to call ".first = . aa(4) := 'D' .delete" nt(2) := 'B' .7).delete .last = 4 . aa(5) := 'E' .print( va ).count = 4 . .extend" for rows 5.extend" first since 5.last = .last = .delete . aa(3) := 'C' . (1) A (2) B (3) C (4) D (5) E (6) F (7) G .print( va ).limit = 10 Delete all rows in the collection (frees memory too).".note we do not need to -.first = . aa(2) := 'B' . NOT NULL .extend(3) va(5) := 'E' va(6) := 'F' va(7) := 'G' p.".we do need to call -. -. . p. va.7 were removed with ".3.limit = NULL .count = and empty NULL NULL 0 p.trim" . aa(7) := 'G' . .first = .3.6.count = and empty NULL NULL 0 .trim" nt.first = 1 .6. aa.last = 7 .last = .delete . .first = 1 .extend" for rows -.count = and empty NULL NULL 0 p. nt(3) := 'C' .6..print( aa ).print( nt ).7 -.4 which were -.extend(3) nt(5) := 'E' nt(6) := 'F' nt(7) := 'G' .limit = 10 Reuse rows left vacant by earlier delete operations (rows 2.count = 7 . p. p.which were removed with -.removed with ". nt.

. Two collections can be compared for equality with the "=" operator.8. NULL / The next table presents operational characteristics of each collection type.limit = 10 Set a collection to NULL.print( nt ). 1. e. NULL end. MULTISET INTERSECT.2**31 Y Y Y Y Y Y Y Y Y Y Y Y . begin nt := nt_null .print( va ). end."nt := null" will not -. use a null -. The collection can be manipulated in PL/SQL with MULTISET Operators e. Y The collection must be initialized before used.g.va_type .limit = NULL .limit = NULL . Nested Table Y Integer 1. Associative Characteristic Array The entire collection can be saved in a database column. uninitialized state. begin va := va_null . use a null -.. a constructor).g.work. -.2**31 Varray Y Y Integer 1.not possible -."va := null" will not -. p. any -2**31. end. i.variable instead declare nt_null p.. Rows in the collection retain their order when the entire n/a collection is saved in a database column.3..nt_type . -. There can be gaps between subscripts. (for Integers) The collection can be defined to hold a predefined maximum number of rows. The collection can be unnested in a query using the TABLE() collection expression. p. The collection must be extended before a new row is added.work.e.2**31 Legal subscript value ranges.e.variable instead declare va_null p. Legal subscript datatypes. MULTISET UNION. The collection can be initialized with multiple rows of data using a single command (i.

of this form: if A = n1 then A1 else if A = n2 then A2 else X Databases before Oracle 8. 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. 1. Everything DECODE can do. as a standard. CASE can work with logical operators other than ‘=’ DECODE can do an equality check only. . To achieve the same effect with DECODE. though. ranges of data had to be forced into discrete form making unwieldy code. more meaningful and more powerful function. which DECODE cannot – which we’ll see in this article.1.6 had only the DECODE function. SQL> select ename 2 .1. An example of putting employees in grade brackets based on their salaries – this can be done elegantly with CASE. CASE can. CASE was introduced in Oracle 8. There is a lot more that you can do with CASE.6. CASE is capable of more logical comparisons such as < > etc.The Difference Between DECODE and CASE DECODE and CASE statements in Oracle both provide a conditional construct.

CASE can be a more efficient substitute for IFTHEN-ELSE in PL/SQL. SQL> declare 2 grade char(1).put_line('good'). 2 case 3 -. 7 when 'b' then dbms_output.ENAME ---------SMITH ALLEN WARD SAL_GRADE --------Grade I Grade II Grade II 2.'WARD') 6 then 'Top Bosses' 7 -. 8 when 'c' then dbms_output.ename. CASE can work as a PL/SQL construct DECODE can work as a function inside SQL only. SQL> select e.'SMITH'. CASE can work with predicates and subqueries in searchable form. An example of categorizing employees based on reporting relationship. . 5 case grade 6 when 'a' then dbms_output.mgr = e.ename in ('KING'.put_line('excellent').put_line('poor').mark the category based on ename list 5 when e.put_line('very good').searchable subquery 8 -. 3 begin 4 grade := 'b'. 9 when 'd' then dbms_output. ENAME ---------SMITH ALLEN WARD JONES EMP_CATEGORY ----------------Top Bosses General Employees Top Bosses Managers 3. 10 when 'f' then dbms_output.put_line('fair').empno) 11 then 'Managers' 12 else 13 'General Employees' 14 end emp_category 15 from emp e 16 where rownum < 5. CASE can work with predicates and searchable subqueries DECODE works with expressions which are scalar values only.predicate with "in" 4 -.identify if this emp has a reportee 9 when exists (select 1 from emp emp1 10 where emp1. illustrating these two uses of CASE.

output = 3 PL/SQL procedure successfully completed. BEGIN proc_test(decode(:a. 'NULL' 3 .put_line('output = '||i). / PL/SQL procedure successfully completed. SQL> var a varchar2(5). column 17: PLS-00204: function or pseudo-column 'DECODE' may be used inside a SQL statement only ORA-06550: line 1. PL/SQL procedure successfully completed. NULL ---NULL .'THREE'. END. * ERROR at line 1: ORA-06550: line 1. end case.11 12 13 14 else dbms_output. CASE can even work as a parameter to a procedure call.3. while DECODE cannot. column 7: PL/SQL: Statement ignored SQL> exec proc_test(case :a when 'THREE' then 3 else 0 end). SQL> SQL> 2 3 4 5 6 create or replace procedure proc_test (i number) as begin dbms_output.0)). 4. end.0)). SQL> exec :a := 'THREE'. SQL> select decode(null 2 . end.'THREE'.3. SQL> exec proc_test(decode(:a. 'NOT NULL' 4 ) null_test 5 from dual. / Procedure created. null.put_line('no such grade'). Careful! CASE handles NULL differently Check out the different results with DECODE vs NULL.

NULL_TES -------NOT NULL The “searched CASE” works as does DECODE.1. 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. SQL> select decode(2. DECODE does not Compare the two examples – DECODE gives you a result. CASE is ANSI SQL-compliant CASE complies with ANSI SQL.SQL> select case null 2 when null 3 then 'NULL' 4 else 'NOT NULL' 5 end null_test 6 from dual.'2'. DECODE is proprietary to Oracle. 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. . 3 '3') t 4 from dual. CASE expects data type consistency. though. 2 '2'.1. CASE gives a data type mismatch error.

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

GRP_A ---------a1 a1 a2 GRP_B COUNT(*) MAX(VAL) SUM(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. sum( val ) from t GROUP BY GRP_A order by grp_a . GRP_B ) order by grp_a.---------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. grp_b .---------. grp_b.Grouping Rows with GROUP BY GROUP BY Consider a table like this one. GRP_A ---------a1 a1 a1 a1 a1 a2 a2 a2 GRP_B VAL ---------. and SUM in the result set. MAX.---------b1 2 20 30 b2 3 50 120 b3 3 32 66 Parentheses may be added around the GROUP BY expression list. grp_b. sum( val ) from t GROUP BY ( GRP_A. Doing so has no effect on the result. select grp_a. max( val ).---------. max( val ). grp_b . select grp_a. sum( val ) from t GROUP BY GRP_A. grp_b . GRP_A COUNT(*) MAX(VAL) SUM(VAL) ---------. count(*). count(*). grp_b. val from t order by grp_a. max( val ). GRP_B order by grp_a.---------. count(*). select grp_a. .---------. select grp_a.

the same result is usually produced by specifying DISTINCT instead of using GROUP BY. like this.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. select count(*). sum( val ) from t GROUP BY () .---------8 50 216 GROUP BY and DISTINCT We can use GROUP BY without specifying any aggregate functions in the SELECT list. grp_b . COUNT(*) MAX(VAL) SUM(VAL) ---------.---------. This groups all rows retrieved by the query into a single group. select DISTINCT grp_a. GRP_A ---------a1 a1 a2 GRP_B ---------b1 b2 b3 . grp_b from t GROUP BY GRP_A. select count(*). grp_b from t order by grp_a. COUNT(*) MAX(VAL) SUM(VAL) ---------.---------8 50 216 The last example is equivalent to specifying no GROUP BY clause at all.---------. grp_b . GRP_B order by grp_a. select grp_a. sum( val ) from t .---------. max( val ). GRP_A ---------a1 a1 a2 GRP_B ---------b1 b2 b3 However. Parentheses are mandatory when specifying an empty set. max( val ).

null as grp_b from t union all select distinct null. For example. grp_b ) != 3 order by 1. grp_b from t order by 1. 2 . Queries that use DISTINCT are typically easier to understand. 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. select grp_a. grp_b from t group by cube( grp_a. 2 . grp_b from t union all select distinct grp_a. 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. grp_b ) having grouping_id( grp_a.) GROUP BY and Ordering . GROUP BY"). but the GROUP BY approach can provide an elegant solution to otherwise cumbersome queries when more than one set of groupings is required. 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.According to Tom Kyte the two approaches are effectively equivalent (see AskTom "DISTINCT VS.

insert into t values ( 'a2' . 'd2'. changing the order in which columns appear in the GROUP BY clause has no effect on the way the result set is grouped. count(*) from t group by grp_a. grp_b . 'd2'.of the Setup topic insert into t values ( 'a2' . 'c2'. GRP_B order by grp_a. . grp_b.---------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. count(*) from t GROUP BY GRP_B. 'c2'.All other things being equal.columns have been reversed order by grp_a. grp_b. This is an illusion which is easily proved with the following snippet. 'b3' . -. count(*) from t GROUP BY GRP_A.---------b1 2 b2 3 b3 3 returns the same results as this one. select grp_a. 'c2'. 'b3' . Some programmers interpret this as meaning that GROUP BY returns an ordered result set.this time we insert rows into T using a different order from that -. truncate table t. GRP_A ---------a1 a1 a2 GRP_B COUNT(*) ---------. GRP_A -----a1 a1 a2 GRP_B COUNT(*) -----. '32' ) . grp_b. grp_b . grp_b . 'b3' . 'd2'.---------b1 2 b2 3 b3 3 The results are still ordered. What happens if we use no ORDER BY clause at all? select grp_a. this query select grp_a. Note how the same query now returns rows in a random order given new conditions. '12' ) . GRP_A ---------a1 a1 a2 GRP_B COUNT(*) ---------. For example. '22' ) . insert into t values ( 'a2' . GRP_A -.

. select grp_a.(your results may vary) The actual behaviour of GROUP BY without ORDER BY is documented in the SQL Reference Manual as follows. . 'c1'. 'c2'. upper(grp_a). '50' '40' '30' '20' '10' ) ) ) ) ) .---------b2 3 b1 2 b3 3 -. count(*) from t group by GRP_A. . count(*) from t group by grp_a.) 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. . 'd1'. Group by behavior in 10GR2 for another discussion of this issue. 'c1'. into into into into into t t t t t values values values values values ( ( ( ( ( 'a1' 'a1' 'a1' 'a1' 'a1' . select grp_a. . . To order the groupings. grp_b. . .---------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. upper(grp_a). 'c1'. GRP_A -----a1 a1 a2 GRP_B COUNT(*) -----. use the ORDER BY clause. GRP_A ---------a1 a2 UPPER(GRP_ COUNT(*) ---------. "The GROUP BY clause groups rows but does not guarantee the order of the result set.insert insert insert insert insert commit. GRP_A order by grp_a . . 'c1'. .---------a1 A1 5 . 'd1'." (See AskTom . 'd1'. . GRP_A UPPER(GRP_ COUNT(*) ---------. . count(*) from t group by GRP_A order by grp_a . 'd1'. grp_b . 'b2' 'b2' 'b2' 'b1' 'b1' . 'd1'.---------.

count(*) from t group by grp_a. select GRP_A. count(*) from t GROUP BY GRP_B . 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. 'XYZ'. SYSDATE. grp_a. grp_b order by grp_a. grp_b . 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(*) -----. like GRP_B in the example below. GRP_A COUNT(*) -----. select GRP_A.---------a1 2 a1 3 a2 3 However we may not select table columns that are absent from the GROUP BY list. select grp_a. grp_a. SELECT Lists We may group by table columns that are not in the SELECT list. select grp_a. GRP_B order by grp_a.---------b1 2 b2 3 b3 3 and we can GROUP BY constant columns that are not in the SELECT list. 'XYZ'. count(*) from t .a2 A2 3 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. As with table based columns we can include constant columns in the GROUP BY clause select 123. as with GRP_A in this example. SYSDATE. grp_b. grp_b. count(*) from t group by 123.

select 123. grp_b order by grp_a. select grp_a. This allows us to use WHERE conditions involving columns like GRP_B in the query below. SYSDATE. grp_a. HAVING When Oracle processes a GROUP BY query the WHERE clause is applied to the result set before the rows are grouped together. count(*) from t WHERE COUNT(*) > 4 group by grp_a order by grp_a .---------a1 3 a2 3 Thia does. select grp_a. grp_b .group by 123. prevent us from using conditions that involve aggregate values like COUNT(*) that are calculated after the GROUP BY clause is applied. 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(*) -----. SYSDATE.---------b1 2 b2 3 b3 3 Unlike table based columns we can select constant columns that are absent from the GROUP BY list. grp_b. count(*) from t WHERE GRP_B in ( 'b2'. which is not listed in the GROUP BY clause. For example. the following will not work. 'XYZ'. 'b3' ) group by grp_a order by grp_a . GRP_A COUNT(*) ---------. grp_a. WHERE COUNT(*) > 4 * ERROR at line 3: . GRP_B order by grp_a. however. GRP_A -----a1 a1 a2 GRP_B COUNT(*) -----. count(*) from t GROUP BY GRP_A. grp_b .---------b1 2 b2 3 b3 3 Note how all three queries returned the same number of rows. 'XYZ'.

count(*) from t WHERE GRP_A = 'a2' group by grp_a order by grp_a . count(*) from t group by grp_a HAVING GRP_A = 'a2' order by grp_a .---------a1 5 Note that the HAVING clause cannot reference table columns like VAL that are not listed in the GROUP BY clause.---------a2 3 but doing so yields the same result as using a WHERE clause. GRP_A COUNT(*) ---------. GRP_A COUNT(*) ---------. select grp_a. count(*) from t group by grp_a HAVING COUNT(*) > 4 order by grp_a .ORA-00934: group function is not allowed here For these types of conditions the HAVING clause can be used. on the other hand. select grp_a. select grp_a. HAVING VAL > 5 * ERROR at line 4: ORA-00979: not a GROUP BY expression It can.---------a2 3 . count(*) from t group by grp_a HAVING VAL > 5 order by grp_a . reference table columns like GRP_A that are in the GROUP BY clause. GRP_A COUNT(*) ---------. select grp_a.

null. For example. count(*) from t GROUP BY GROUPING SETS ( GRP_A. select grp_a.---------b1 2 b2 3 b3 3 UNION ALL could be used. grp_b. select grp_a. count(*) from t group by grp_b order by grp_b . GRP_A COUNT(*) ---------.---------a1 5 a2 3 select grp_b. count(*) from t group by grp_b order by 1.Given a choice between the last two snippets I expect using a WHERE clause provides the best performance in most. cases. count(*) from t group by grp_a UNION ALL select null. 2 . GRP_A NULL COUNT(*) ---------. count(*) from t group by grp_a order by grp_a . like this select grp_a. say we wanted to combine the results of these two queries. GROUPING SETS There are times when the results of two or more different groupings are required from a single query.---------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. GRP_B COUNT(*) ---------. With it the last query can be written as follows. if not all.---------. GRP_B ) . grp_b.

and Empty Sets Composite Columns You can treat a collection of columns as an individual set by using parentheses in the GROUPING SETS clause. grp_b . GROUPING SETS. count(*) from t GROUP BY GRP_A. grp_b ) used in the last query differs from the clause group by ( grp_a. grp_b ) in the next query. GRP_B ) order by grp_a. For example.. count(*) from t GROUP BY ( GRP_A.---------b1 2 b2 3 b3 3 select grp_a.---------a1 5 a2 3 . GRP_B order by grp_a.---------. GRP_A ---------a1 a1 a2 GRP_B COUNT(*) ---------. to write a query that returns the equivalent of these two queries select grp_a. GRP_A N COUNT(*) ---------. grp_b . null. Composite Columns. GRP_A ---------a1 a1 a2 GRP_B COUNT(*) ---------. select grp_a.---------a1 5 a2 3 b1 2 b2 3 b3 3 It is important to understand how the clause grouping sets( grp_a. grp_b. GRP_A GRP_B COUNT(*) ---------. count(*) from t GROUP BY GRP_A order by grp_a . grp_b . GRP_B). grp_b.order by grp_a.---------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_A ---------a1 a1 a2 GRP_B COUNT(*) ---------.---------b1 2 b2 3 5 b3 3 3 The term (GRP_A. GRP_B). grp_b. GRP_A ---------a1 a1 a2 GRP_B COUNT(*) ---------. grp_b . Empty Sets To add a grand total row to the result set an empty set.---------b1 2 b2 3 b3 3 returns the same results as this query . grp_b. grp_b.---------b1 2 b2 3 b3 3 8 Gotcha . GRP_B) is called a "composite column" when it appears inside a GROUPING SETS.we could use the following GROUPING SETS clause. count(*) from t GROUP BY GROUPING SETS ( (GRP_A. For example this query select grp_a. In the example below the last row is generated by the empty set grouping. () ) order by grp_a. specified as (). grp_b . GRP_B) is no different than the same expression without parentheses. GRP_A order by grp_a. GRP_A ) order by grp_a. GRP_B).Parentheses without GROUPING SETS Outside a GROUPING SETS clause (or ROLLUP or CUBE clauses) a parenthesized expression like (GRP_A. GRP_A ---------a1 a1 a1 a2 a2 GRP_B COUNT(*) ---------. grp_b . ROLLUP. can be used. count(*) from t GROUP BY GROUPING SETS ( (GRP_A. select grp_a. count(*) from t GROUP BY (GRP_A. select grp_a. GRP_B). or CUBE clause.

After all. count(*) from t GROUP BY GROUPING SETS ( GRP_A. was actually a more appropriate syntactic choice than a constant but I continued to use constants out of habit. . 0 ) order by grp_a.---------a1 5 a2 3 b1 2 b2 3 b3 3 8 The last row in the result set is generated by the "0" grouping. grp_b .---------b1 2 b2 3 b3 3 Gotcha: GROUPING SETS with Constants When I first started using GROUPING SETS I used constants to produce grand total rows in my result sets. GRP_A order by grp_a. count(*) from t GROUP BY GRP_A. both approaches seemed to produce the same results. GRP_A ---------a1 a1 a2 GRP_B COUNT(*) ---------. count(*) from t GROUP BY GRP_A. grp_b . grp_b.select grp_a. select grp_a . like this. grp_b. GRP_B.---------b1 2 b2 3 b3 3 which in turn has the same result set as this one. grp_b . GRP_A ---------a1 a1 a2 GRP_B COUNT(*) ---------. I later learnt that an empty set term.---------. select grp_a. GRP_A GRP_B COUNT(*) ---------. grp_b . GRP_B. "()". GRP_B order by grp_a.

nvl2( grp_b. 0 ) nvl2_grp_b . Readers who want to understand more about why these two queries differ can reverse engineer the two into their . I later ran into a case where the two actually produced different results. GRP_B. GRP_A GRP_B COUNT(*) ---------. nvl2( grp_b. () ) order by grp_a.---------. GRP_A -----a1 a2 (null) (null) (null) (null) GRP_B NVL2_GRP_B COUNT(*) -----. grp_b . Query 1 set null '(null)' Query 2 set null '(null)' select select grp_a grp_a . grp_b . GRP_B. count(*) from from t t GROUP BY GROUP BY GROUPING SETS ( GRP_A. grp_b . grp_b grp_a. 1. 1.---------(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.select grp_a . 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(*) ---------.---------a1 5 a2 3 b1 2 b2 3 b3 3 8 However. 0 ) order by order by grp_a. 0 ) nvl2_grp_b . grp_b . GRP_B. . grp_b .---------. () ) GROUPING SETS ( GRP_A. count(*) from t GROUP BY GROUPING SETS ( GRP_A. This is because "0" appears in both the SELECT list and the GROUP BY clause.

Readers who don't simply need to remember this rule of thumb .always use an empty set term to generate a grand total row. It uses the ROLLUP operator. 3 . count(*) from t . 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(*) ---------. Here is how the query above looks when implemented with ROLLUP. ( grp_a. () ) order by 1. grp_c . count(*) from t group by grouping sets ( ( grp_a. grp_c . 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. grp_b . 2. do not use a constant. grp_b ) .UNION ALL equivalents using the instructions at Reverse Engineering GROUPING BY Queries. select grp_a . When aggregates are required at each level a query like this can be used. grp_b. ( grp_a ) . grp_b . set null '(null)' select grp_a .---------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. grp_c ) .

() ) order by 1. as in this query. 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(*) ---------. ( grp_b ) . 3 . grp_c ) . set null '(null)' select grp_a . grp_b ) . ( grp_b. GRP_B. 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 . 2. 2.---------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.group by ROLLUP( GRP_A. grp_c ) . ( grp_a. GRP_C ) order by 1. ( grp_a. count(*) from t group by grouping sets ( ( grp_a. grp_b . grp_c . grp_b. grp_c ) . 3 . ( grp_a ) . ( grp_c ) .

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. GRP_C ) order by 1.---------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 . 3 . grp_c . count(*) from t group by CUBE( GRP_A. select grp_a . 2. GRP_B. grp_b . 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(*) ---------.

grp_c .(null) (null) (null) (null) c2 (null) 4 8 Concatenated Groupings The last type of grouping shortcut we will examine is called a Concatenated Grouping. grp_b ) .---------. grp_c ) ) order by 1. set null '(null)' select grp_a .---------- . ( grp_a. grp_b . GRP_A GRP_B GRP_C COUNT(*) ---------. 2. select grp_a . grp_c ) order by 1. count(*) from t group by grp_a . GRP_A ---------a1 a1 a1 a1 a2 a2 GRP_B GRP_C COUNT(*) ---------. 2. which effectively performs a cross-product of GRP_A with GRP_B and GRP_C. grouping sets( grp_b. grp_c .---------b1 2 b2 3 c1 4 c2 1 b3 3 c2 3 into one like this.---------.---------. 3 . count(*) from t group by grouping sets ( ( grp_a. grp_b . With it one can re-write a query like this one. 3 .

grp_c . grp_b . grp_c . grp_b . ( grp_a. grp_c . grouping sets( grp_c. grp_d .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 . count(*) from t group by grouping sets( grp_a. ( grp_b. select grp_a . ( grp_b. 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(*) ---------. grp_d ) order by 1. grp_d ) order by . 2.---------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. grp_c . grp_b ) . count(*) from t group by grouping sets ( ( grp_a.

---------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. count(*) from t GROUP BY GROUPING SETS ( GRP_A.. GRP_A.---------a1 5 a1 5 a2 3 a2 3 select grp_a. however. GRP_A ) order by grp_a . GROUP_ID Unlike a regular GROUP BY clause.---------a1 5 a1 5 a1 5 a2 3 a2 3 a2 3 . GRP_A COUNT(*) ---------. GRP_A COUNT(*) ---------. 3 GRP_B ---------(null) (null) (null) (null) (null) b1 b1 b2 b2 b2 b3 b3 GRP_C COUNT(*) ---------. 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. prove useful in data warehouse queries that deal with hierarchical cubes of data. GRP_A ) order by grp_a . including the same column more than once in a GROUPING SETS clause produces duplicate rows. Concatenated groupings can. 1. count(*) from t GROUP BY GROUPING SETS ( GRP_A. See Concatenated Groupings for more information. 2. select grp_a.

Let's now look at grouping a table that does contain null values. select grp_a.---------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.---------X1 10 X2 40 (null) 20 (null) 30 (null) 50 .---------. GRP_A COUNT(*) GROUP_ID() ---------. GRP_A ) order by grp_a. grp_b . GRP_B ) order by grp_a. group_id() . count(*).---------. GROUP_ID() from t GROUP BY GROUPING SETS ( GRP_A. set null '(null)' select * from t2 order by grp_a.---------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. GRP_A ---------A1 A1 A1 A1 A1 GRP_B VAL ---------. select grp_a. It is in such queries that GROUP_ID proves useful.---------. GRP_A. grp_b . GROUP_ID() from t GROUP BY GROUPING SETS ( GRP_A.The GROUP_ID function can be used to distinguish duplicates from each other. Note that GROUP_ID will always be 0 in a result set that contains no duplicates. count(*). There are times when more complex GROUP BY clauses can return duplicate rows however. grp_b. GRP_A GRP_B COUNT(*) GROUP_ID() ---------.

NVL and NVL2 One might expect that NVL() or NVL2 could be used to distinguish the two nulls.---------X1 1 1 X2 1 1 n/a 0 5 n/a 0 3 n/a 0 1 .grp_b. count(*) from t2 GROUP BY GROUPING SETS( (GRP_A. count(*) from t2 group by grp_a.---------X1 1 X2 1 (null) 3 (null) 1 So far so good. grp_b. grp_b .---------. grp_b.A2 (null) 60 Now consider the following GROUP BY query. grp_b order by grp_a. grp_b . Gotcha . GRP_A ) order by grp_a. count(*) from t2 GROUP BY GROUPING SETS( (GRP_A. GRP_A ---------A1 A1 A1 A2 GRP_B COUNT(*) ---------. one representing the null values stored in T2. grp_b . 1. but let's use GROUPING SETS next and see what happens. 'n/a' ) AS GRP_B . GRP_B). select grp_a. 0 ) as test . GRP_A ---------A1 A1 A1 A1 A2 GRP_B TEST COUNT(*) ---------. GRP_A ) order by grp_a.GRP_B and the other representing the set of all values in T2. like this select grp_a . GRP_A ---------A1 A1 A1 A1 A2 A2 GRP_B COUNT(*) ---------.GRP_B.GRP_B. select grp_a. NVL( t2. GRP_B).---------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. nvl2( t2.

A value of "1" tells us it does. GRP_A ---------A1 A1 A1 A1 A2 A2 GRP_B COUNT(*) GROUPING_GRP_A GROUPING_GRP_B ---------. ROLLUP.---------. 1. However. grp_b . Here is one example. 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. GROUPING( GRP_A ) GROUPING_GRP_A . set null '(null)' select grp_a . GROUPING( GRP_B ) GROUPING_GRP_B from t2 group by grouping sets( (grp_a. or CUBE operation.A2 n/a 0 1 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. GROUPING can be used with DECODE to insert labels like "Total" into the result set. In the output of the following query two of the four nulls represent the set of all GRP_B values. 2 . grp_a ) order by 1 . count(*) . grp_b).-------------. not before.-------------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. grp_b ) as "Group B" . select grp_a as "Group A" . a value of "0" tells us it does not. decode ( GROUPING( GRP_B ) . count(*) as "Count" . 'Total:' . In the next topic we see how the GROUPING function can help us distinguish the two types of nulls.

count(*) from t2 group by grouping sets( (grp_a. decode( grouping( grp_b ). . grp_b .---------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. decode( grouping( grp_b ). grp_b). but Oracle interprets it as referring to the column alias. For example. To learn how aggregate functions like COUNT() and SUM() deal with null values in non-GROUP BY columns see Nulls and Aggregate Functions. grp_a ) order by grp_a . GROUPING( GRP_B ) .from t2 group by grouping sets( (grp_a. In the ORDER BY GROUPING( GRP_B ) clause one might expect the "GRP_B" term to refer to the table column. grp_b ) AS GRP_B . 'Total:'. 1. Group A ---------A1 A1 A1 A1 A2 A2 Group B Count ---------. hence the ORA00979 error. 1. select grp_a . say we attempted this query.ORA-00979 When using ORDER BY we need to be careful with the selection of column aliases. grp_b). 'Total:'. To work around the error we can either prefix the column name with its table name select . 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. grp_a ) order by grouping( GRP_B ) . Gotcha .

---------(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. 'Total:'. grp_b ) AS "Group B" . '01'. DECODE ( TO_CHAR( GROUPING( GRP_A ) ) || TO_CHAR( GROUPING( GRP_B ) ) . grp_b). or CUBE operation.---------(null) 3 X1 1 X2 1 (null) 1 Total: 5 Total: 1 or change the column alias. 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( grouping( grp_b ). 1. decode( grouping( grp_b ). grp_b). grp_a ) order by grouping( GRP_B ) . count(*) as "Count" from t2 group by grouping sets( (grp_a. 'Total:'. grp_a ) order by grouping( T2. ROLLUP.grp_a . grp_b ) AS GRP_B . count(*) from t2 group by grouping sets( (grp_a. Group A ---------A1 A1 A1 A2 A1 A2 Group B Count ---------.GRP_B ) . GRP_A ---------A1 A1 A1 A2 A1 A2 GRP_B COUNT(*) ---------. 1. select grp_a as "Group A" . 'Group "' || GRP_A || '" Total' .

GROUPING_ID( GRP_A.---------. grp_b . count(*) from t2 group by grouping sets ( grp_a. grp_b . '10'. grp_a .this column is only included for clarity .---------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. 'Group "' || GRP_B || '" Total' '11'.---------. . The following example shows how it works. grp_a .GRP_B) GRP_A GRP_B COUNT(*) ---------. select to_char( grouping( grp_a ) ) || to_char( grouping( grp_b ) ) as bit_vector -. 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. grp_b. . () order by GROUPING( GRP_A ) .---------01 1 A1 5 01 1 A2 1 10 2 X1 1 10 2 X2 1 10 2 4 . 'Grand Total' NULL AS LABEL count(*) from t2 group by grouping sets ( grp_a. grp_b. ) . GRP_B ) . GROUPING( GRP_B ) . BIT_VECTOR ---------01 01 10 10 10 11 ) LABEL COUNT(*) -----------------------. grp_a . grp_b ..-----------------------. () ) order by GROUPING_ID( GRP_A. GRP_B ) . Fortunately for us the GROUPING_ID function exists. BIT_VECTOR GROUPING_ID(GRP_A.

(GRP_A. count(*) from t2 group by grouping sets ( grp_a. grp_a . grp_b). 2 . GRP_B ) . 3 .GRP_B) -----------------------0 0 GRP_A ---------A1 A1 ) GRP_B COUNT(*) ---------.---------X1 1 X2 1 . NULL ) AS LABEL . count(*) from t2 group by grouping sets ( (grp_a. 'Group "' || GRP_B || '" Total' . grp_a . GRP_B ) . () ) order by GROUPING_ID( GRP_A. GROUPING_ID(GRP_A. grp_b. select DECODE ( GROUPING_ID( GRP_A.---------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. () order by 1 . 'Group "' || GRP_A || '" Total' . select GROUPING_ID( GRP_A. LABEL COUNT(*) -----------------------. 'Grand Total' . grp_b . GRP_B ) . 1. GRP_B). grp_b .11 3 6 Here is how we could use GROUPING_ID to streamline our original query. grp_a. is included in the GROUPING SETS clause. 2. grp_b. 3.

count(*) from t2 group by cube( grp_a. GRP_B ) != 3 order by .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. we started with a query like this one select grouping_id( grp_a. grp_a . Say. grp_b ) order by 1. We simply add a HAVING clause as follows. grp_b ) HAVING GROUPING_ID( GRP_A. grp_b ) . grp_b ) . count(*) from t2 group by cube( grp_a. grp_b . grp_b . select grouping_id( grp_a.---------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"). grp_a . 3 . 2. 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(*) ---------. for example.

GRP_B ) order by grp_a . grp_a ) as nvl_grp_a_b . 3 GRP_A ---------A1 A1 A1 A2 A1 A2 GRP_B COUNT(*) ---------. nvl( grp_b. grp_b .---------. 0 ) as nvl2_grp_b . count(*) from t GROUP BY ROLLUP ( GRP_A. set null '(null)' select grp_a . A query like this.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. 1. nvl( grp_b.---------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. for example. GRP_A ---------a1 a1 a1 a2 a2 (null) GRP_B ---------b1 b2 (null) b3 (null) (null) NVL_GRP_A_ NVL2_GRP_B COUNT(*) ---------. nvl2( grp_b. grp_b . set null (null) select grp_a .. 1. 2.---------X1 1 X2 1 3 1 5 1 X1 1 X2 1 4 GROUPING_ID(GRP_A. This is not as easy as it first may seem. grp_b . grp_a ) as nvl_grp_a_b .

nvl2( grp_b. 1. 0 ) as nvl2_grp_b . grp_a * ERROR at line 2: ORA-00979: not a GROUP BY expression As you can see. Step 1 Replace any ROLLUP or CUBE operators with their equivalent GROUPING SETS operator. 1. grp_b . grp_b . such a query produces an error because the first and second subqueries select columns that are not in the GROUP BY clause. nvl( grp_b. To determine the real equivalent UNION query we can use the following algorithm. count(*) from t GROUP BY ( GRP_A. grp_a ) as nvl_grp_a_b . nvl( grp_b. 0 ) as nvl2_grp_b . 0 ) as nvl2_grp_b . grp_a ) as nvl_grp_a_b . GRP_B ) order by . grp_a ) as nvl_grp_a_b .. grp_b . 1. count(*) from t GROUP BY ( GRP_A ) UNION ALL select grp_a . nvl2( grp_b. nvl2( grp_b. 0 ) as nvl2_grp_b . GRP_B ) order by grp_a . 1. nvl( grp_b. count(*) from t GROUP BY ROLLUP ( GRP_A. count(*) from t GROUP BY () UNION ALL select grp_a . nvl2( grp_b. In our example the query select grp_a . grp_b .

count(*) from t . 0 ) as nvl2_grp_b . grp_b . ( GRP_A. nvl( grp_b. nvl2( grp_b. nvl2( grp_b. nvl( grp_b. grp_b . 0 ) as nvl2_grp_b . which is an empty set in our example. grp_a ) as nvl_grp_a_b .---------b1 1 2 b2 1 3 a1 0 5 b3 1 3 a2 0 3 (null) 0 8 is replaced with select grp_a . 1. select grp_a . grp_a ) as nvl_grp_a_b . GRP_A ---------a1 a1 a1 a2 a2 (null) GRP_B ---------b1 b2 (null) b3 (null) (null) NVL_GRP_A_ NVL2_GRP_B COUNT(*) ---------. 1.grp_a .---------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. grp_b . grp_b .---------.---------. count(*) from t GROUP BY GROUPING SETS ( () . ( GRP_A ) . GRP_A ---------a1 a1 a1 a2 a2 (null) GRP_B ---------b1 b2 (null) b3 (null) (null) NVL_GRP_A_ NVL2_GRP_B COUNT(*) ---------. GRP_B ) ) order by grp_a .

0 ) as nvl2_grp_b . NULL ) as nvl_grp_a_b .---------. select grp_a .---------. select grp_a .---------a1 (null) a1 0 5 . 1.-----. 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 . count(*) from t GROUP BY ( GRP_A ) . 1. count(*) from t GROUP BY ( GRP_A ) . GRP_A GRP_B NVL_GRP_A_B NVL2_GRP_B COUNT(*) -----. count(*) from t GROUP BY () .-----. nvl( NULL. We therefore need to replace GRP_B with NULL. grp_a ) as nvl_grp_a_b . nvl( NULL. GRP_A GRP_B NVL_GRP_A_B NVL2_GRP_B COUNT(*) -----. grp_a ) as nvl_grp_a_b .---------(null) (null) (null) 0 8 Step 2b Now we repeat the first step using the second term in the GROUPING SETS clause. This time GRP_B is in the SELECT list but it is not in the GROUP BY list. 0 ) as nvl2_grp_b . nvl( grp_b. 0 ) as nvl2_grp_b . grp_b . 1. 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.----------. nvl2( NULL. NULL as grp_b . nvl2( NULL. ( GRP_A ).GROUP BY () . nvl2( grp_b.

1. nvl2( grp_b. grp_a ) as nvl_grp_a_b . nvl2( NULL.---------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. nvl( NULL. 1. select grp_a . 0 ) as nvl2_grp_b . count(*) from . nvl( grp_b. 0 ) as nvl2_grp_b . We can use the original SELECT list as-is. NULL ) as nvl_grp_a_b . grp_a ) as nvl_grp_a_b . nvl2( NULL. nvl( grp_b. NULL as grp_b . nvl( NULL. count(*) from t group by () UNION ALL select grp_a . nvl2( grp_b. grp_b . GRP_B ) . count(*) from t GROUP BY ( GRP_A. NULL as grp_b . count(*) from t group by ( grp_a ) UNION ALL select grp_a . GRP_A -----a1 a1 a2 GRP_B -----b1 b2 b3 NVL_GRP_A_B NVL2_GRP_B COUNT(*) ----------. grp_a ) as nvl_grp_a_b . 0 ) as nvl2_grp_b . 0 ) as nvl2_grp_b .a2 (null) a2 0 3 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. grp_b . select NULL as grp_a .---------. 1. 1.

count(*) from t group by ( grp_a. nvl2( grp_b. equivalent terms.---------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. 1. 0 ) to simpler. count(*) from t group by ( grp_a ) union all select grp_a . null as grp_b . GRP_B . 1. grp_a ) as nvl_grp_a_b . count(*) from t group by () union all select grp_a . NULL AS NVL_GRP_A_B . GRP_A AS NVL_GRP_A_B . null as grp_b . grp_b ) order by grp_a . 0 AS NVL2_GRP_B . 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(*) ----------.---------. select null as grp_a . grp_b ) ORDER BY GRP_A . GRP_A GRP_B NVL_GRP_A_B NVL2_GRP_B COUNT(*) . nvl( grp_b. 0 AS NVL2_GRP_B . grp_b . NULL ) and nvl2( NULL .t group by ( grp_a. 0 ) as nvl2_grp_b .

create table t ( grp_a varchar2(10) . 'd1'. select grp_a . . .---------. grp_d varchar2(10) . . 'c2'. grp_a ) as nvl_grp_a_b . grp_b . . grp_b . data. . 'd1'. GRP_B ) order by grp_a .---------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. GRP_A -----a1 a1 a1 a2 a2 (null) GRP_B -----b1 b2 (null) b3 (null) (null) NVL_GRP_A_B NVL2_GRP_B COUNT(*) ----------. count(*) from t GROUP BY ROLLUP ( GRP_A. 'd1'. used by the examples in this section. 'd1'. etc. 'c1'. 'c1'. val number ) . 'c1'. grp_c varchar2(10) .-----a1 a1 a1 a2 a2 (null) -----b1 b2 (null) b3 (null) (null) ----------. 'd2'. . nvl2( grp_b. 0 ) as nvl2_grp_b . '10' '20' '30' '40' '50' '12' ) ) ) ) ) ) . . . . which is repeated below for your convenience. . . . . 'c1'. 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 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. grp_b varchar2(10) . 1. 'b1' 'b1' 'b2' 'b2' 'b2' 'b3' . . . nvl( grp_b.---------. 'd1'. 'c2'. .

procedures. created in earlier parts of this section. . 'c2'. To clear session state changes (e. '32' ) . drop table t2 . 'b3' . . '22' ) . exit . etc. .insert into t values ( 'a2' . 'b3' .g. . 'X1' 'X2' null null null null . val number ) . commit . 'd2'. . . . grp_b varchar2(10) . . . . . '10' '40' '20' '30' '50' '60' ) ) ) ) ) ) . 'c2'. and VARIABLE commands) exit your SQL*Plus session after running these cleanup commands. create table t2 ( grp_a varchar2(10) . . commit . 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' . insert into t values ( 'a2' . drop table t . COLUMN. Cleanup Run the code on this page to drop the sample tables. . . . those made by SET. 'd2'.

An example of hierarchical data is shown below.Hierarchical Data This section presents various topics related to hierarchical data (also known as "tree structured" data). /nls /nls/demo /nls/mesg . Connecting Rows Say we wanted to take the following directory names from a file system and store them in a database table. 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. 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.

starting from the root rows (those with no parents) on down through to the leaf rows (those with no children). (Directory names like these would not typically be used as primary keys. . and a PARENT_KEY column. level from t START WITH parent_key is null CONNECT BY parent_key = prior key . which holds the directory name. We are bending the rules here for illustrative purposes. 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. CONNECT BY identifies all subsequent rows in the hierarchy./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.) select * from t . select key . which connects the directory to its parent directory. START WITH identifies the topmost rows in the hierarchy. The following snippet returns rows sorted hierarchically.

select lpad( ' '. KEY_INDENTED LEVEL --------------. level-1 ) || key as key_indented . It can be used outside the CONNECT BY clause if required. level-1 ) || key as key_indented .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 PRIOR operator in hierarchical queries gives us access to column information from the parent of the current row. level from t START WITH parent_key is null CONNECT BY parent_key = prior key . The topmost level is assigned a LEVEL of 1. To better illustrate hierarchical relationships the LEVEL column is commonly used to indent selected values. select lpad( ' '. like this.-----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.

select lpad( ' '. simply choose a leaf row as the starting point and apply the PRIOR operator to the PARENT_KEY column instead of the KEY column.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 . level from t START WITH KEY = 'delx' connect by key = PRIOR PARENT_KEY .-----delx 1 data 2 ctx 3 server 4 Gotchas . level-1 ) || key as key_indented . KEY_INDENTED LEVEL --------------. from leaf to root. 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.

CONNECT BY conditions are not applied to rows in level 1 of the hierarchy. KEY_INDENTED LEVEL --------------. . level-1 ) || key as key_indented . In the following snippet note how the KEY <> 'delx' condition did not filter out the row with a KEY value of 'delx'.-----delx 1 data 2 ctx 3 server 4 Order of Operations The clauses in hierarchical queries are processed in the following order. level-1 ) || key as key_indented . Filter Condition in CONNECT BY select lpad(' '. CONNECT BY clause 4. START WITH clause 3. Filter Condition in WHERE select lpad(' '. WHERE clause conditions that are not joins. join conditions (either in the FROM clause or the WHERE clause) 2. 1. level from t start with key = 'delx' connect by key = PRIOR PARENT_KEY and KEY <> 'delx' . level-1 ) || key as key_indented . level from t WHERE LEVEL != 3 start with key = 'server' connect by parent_key = prior key -. 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. select lpad( ' '. level from t --start with key = 'server' CONNECT BY parent_key = prior key and LEVEL != 3 .

-----server 1 bin 2 config 2 ctx 2 Sorting Since START WITH and CONNECT BY apply a hierarchical sorting scheme to your data. 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. Regular Query select key from t ----ORDER BY NAME . For example. in your hierarchical queries.KEY_INDENTED LEVEL --------------. . you should generally not use any features that apply other sorting schemes. such as ORDER BY or GROUP BY. Doing so would negate the need for START WITH and CONNECT BY in the first place. Hierarchical Query select key from t start with parent_key is null connect by parent_key = prior key ORDER BY NAME .-----server 1 bin 2 config 2 ctx 2 delx 4 enlx 4 eslx 4 KEY_INDENTED LEVEL --------------. It yields the same results as if CONNECT BY was not used at all.

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 --------------server ctx mig data eslx enlx delx admin config log bin nls mesg . Ascending Siblings select lpad(' '. KEY_INDENTED --------------nls demo mesg server bin config log ctx admin data delx enlx eslx Descending Siblings select lpad(' '. The following examples show how ORDER SIBLINGS BY can be used to sort siblings in ascending and descending order respectively.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. 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 . 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").

exception when no_data_found then return( null ). p_separator varchar2 default '/' ) return varchar2 is v_parent_key t. recursive PL/SQL function can be used in place of ORDER SIBLINGS BY. -----------------------------------------------------------create or replace function KEY_PATH ( p_key t. -.etc.key%type . p_separator ) || p_separator || v_key end if. hierarchical loops in the data. Ascending Siblings Descending Siblings select select lpad(' '.This function is only for demonstration purposes.appearing in KEY values. level-1) || key as key_indented as key_indented .In a real application more robust code would be needed -.to guard against things like separator characters -. / show errors No errors.key%type .Note: --. v_key t key = p_key . key v_parent_key. level-1) || key lpad(' '. For earlier versions a custom. end. else return ( KEY_PATH( v_parent_key. ).mig demo Oracle 8i and Earlier The ORDER SIBLINGS BY clause is only available in Oracle version 9i or greater. if v_parent_key is null then return ( p_separator || v_key ).parent_key%type . -. begin select into from where parent_key. v_key t. ------------------------------------------------------------.

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. 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. 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. does not exist in values under T. select lpad(' '. 0 otherwise.from t start with parent_key is null connect by parent_key = prior key ORDER BY KEY_PATH( KEY. '~' ) DESC . '/' ). sorts lower than all characters that exist in T.KEY For descending siblings the code RPAD( KEY_PATH( KEY. CONNECT_BY_ISLEAF from t start with key = 'server' connect by parent_key = prior key . '/' ) ASC . 50. Violating these rules can result in incorrectly sorted output.KEY ("~" in this example). '~' ) 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. 50. '/' ).KEY 2. level-1 ) || key as key_indented .

causing CONNECT_BY_ISLEAF to be 0 for "config" and "ctx" even though those rows look like leaf nodes in the end result. KEY_INDENTED CONNECT_BY_ISLEAF --------------. The following example illustrates this.. connect_by_isleaf . not that of the underlying table data.----------------server 0 bin 1 config 1 ctx 1 As we saw happen with the LEVEL column in the preceding tutorial. connect_by_isleaf from t start with key = 'server' CONNECT BY parent_key = prior key and LEVEL <= 2 -. They are filtered out by the " LEVEL <= 2 " condition. level-1 ) || key as key_indented . In the following query however. not the CONNECT BY clause as above. the order of evaluation of the CONNECT BY and WHERE clauses can also affect the behaviour of the CONNECT_BY_ISLEAF pseudo column. 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.----------------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. level-1 ) || key as key_indented . select lpad(' '. In it. 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. select lpad(' '.filters out descendents of "config" and "ctx" . KEY_INDENTED CONNECT_BY_ISLEAF --------------. " LEVEL <= 2 " is placed in the WHERE clause.

key .this time using the WHERE clause start with key = 'server' connect by parent_key = prior key . level-1 ) || key as key_indented . KEY_INDENTED CONNECT_BY_ISLEAF --------------.filter out descendents of "config" and "ctx". CONNECT_BY_ROOT key as root_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 .from t WHERE -.----------------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. level . CONNECT_BY_ROOT name as root_name from t start with parent_key is null connect by parent_key = prior key . select lpad(' '. KEY_INDENTED LEVEL KEY --------------. LEVEL <= 2 -.-----. name .

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. in practice.---------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. which already operates on the root row itself. level-1 ) || key as key_indented . level-1 ) || key name . actually works in some cases (as tested in Oracle 10g). select lpad(' '. 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' ) . as key_indented . While there would be little use for CONNECT_BY_ROOT in the START WITH condition. 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. KEY_INDENTED LEVEL ROOT_KEY --------------. level .-----. select lpad(' '. using CONNECT_BY_ROOT in the CONNECT BY condition can be useful and. 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.The manual page for CONNECT_BY_ROOT states You cannot specify this operator in the START WITH condition or the CONNECT BY condition. .

In a real application more robust code would be needed -. v_key t. -----------------------------------------------------------create or replace function KEY_PATH ( p_key t.SYS_CONNECT_BY_PATH( key . -. '/' ) SYS_CONNECT_BY_PATH( name.etc.key%type .appearing in KEY values. .to guard against things like separator characters -. begin select parent_key. recursive PL/SQL function can be used in place of SYS_CONNECT_BY_PATH. -. 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 . '/' ) from t start with parent_key is null connect by parent_key = prior key . 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. key from t into where v_parent_key. hierarchical loops in the data.Note: --. ------------------------------------------------------------. For earlier versions a custom. p_separator varchar2 default '/' ) return varchar2 is v_parent_key t. v_key key = p_key .parent_key%type .This function is only for demonstration purposes.key%type .

select lpad(' '. name . were required. '/' ) as KEY_PATH from t start with parent_key is null connect by parent_key = prior key . end. like NAME. select key . name . If no hierarchical information. 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.g.). 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. LEVEL. else return ( KEY_PATH( v_parent_key. e. . p_separator ) || p_separator || v_key end if. / show errors No errors. if v_parent_key is null then return ( p_separator || v_key ). KEY_PATH( key. exception when no_data_found then return( null ). level-1 ) || key as key_indented .

'config'. 'NLS' . varchar2(10) . 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'. 'data' . etc. . create table ( key parent_key name ). . . 'delx' . . . 'enlx' . 'nls' null 'server' 'server' 'config' 'server' 'ctx' 'ctx' 'data' 'data' 'data' 'ctx' . ). 'admin' . ). 'log' . 'DATA' . 'nls' . null . ). . '/' ) as KEY_PATH from t order by KEY_PATH . used by the examples in this section. ). 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. ). ). 'ctx' . 'bin' . varchar2(10) . . . ). ). 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 Run the code on this page in SQL*Plus to create the sample tables. ). . ). ). t varchar2(10) . 'mig' . 'SERVER' 'BIN' 'CONFIG' 'LOG' 'CTX' 'ADMIN' 'DATA' 'DELX' 'ENLX' 'ESLX' 'MESG' ). data. . ).KEY_PATH( key. 'eslx' . 'DEMO' . . ).

and VARIABLE commands) exit your SQL*Plus session after running these cleanup commands. To clear session state changes (e.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 Run the code on this page to drop the sample tables.g. exit . procedures. created in earlier parts of this section. drop function KEY_PATH . etc. COLUMN. those made by SET. drop table t .

refresh groups. Materialized views can be used for many purposes. on the other hand. The power of materialized views comes from the fact that. once created. After completing it you should have enough information to use materialized views effectively in simple applications. With materialized views. You may still see this term in some Oracle 11g materialized view error messages. This tutorial explores materialized view basics. Terminology With relational views the FROM clause objects a view is based on are called "base tables". thus remaining consistent with relational view terminology. partitioning. updatable materialized views. but should not be used in new code. This keyword is supported for backward compatibility. For more complex applications links at the end of the tutorial will point to information on advanced features not covered here. Materialized Views were originally known as "Snapshots" in early releases of Oracle. 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". including: • • • • Denormalization Validation Data Warehousing Replication. Oracle can automatically synchronize a materialized view's data with its source information as required with little or no programming effort. .g. e. these objects are either called "detail tables" (in data warehousing documentation) or "master tables" (in replication documentation and the Oracle Database SQL Reference guide).Materialized Views A Materialized View is effectively a database table that contains the results of a query.

materialized views allow you to store the definition of a query in the database. select * from T . KEY ---------1 2 3 4 VAL ----a b c Materialized View create materialized view mv as select * from t . KEY ---------1 2 3 4 VAL ----a b c select * from V . Table View create view v as select * from t . KEY ---------1 2 3 4 VAL ----a b c .Views vs Materialized Views Like its predecessor the view. select * from MV .

Table select rowid from T order by rowid . KEY ---------1 2 3 4 VAL ----A B C select * from V . Table select * from T . select * from T . Now that the materialized view has been refreshed its data matches that of its base table. ROWID -----------------AAAgZFAAEAAADyEAAA AAAgZFAAEAAADyEAAB AAAgZFAAEAAADyEAAC AAAgZFAAEAAADyEAAD The difference between views and materialized views becomes even more evident than this when table data is updated. . The rowids of the materialized view. after the update. as below. KEY ---------1 2 3 4 VAL ----A B C View select * from V .refresh( 'MV' ). 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. ROWID -----------------AAAgY9AAEAAAAVfAAA AAAgY9AAEAAAAVfAAB AAAgY9AAEAAAAVfAAC AAAgY9AAEAAAAVfAAD Materialized View select rowid from MV order by rowid . KEY ---------1 2 3 4 VAL ----a b c View Materialized View Note how. KEY ---------1 2 3 4 VAL ----A B C execute dbms_mview. Data in materialized views must be refreshed to keep it synchronized with its base table. however. Refreshing can either be done manually. the view data matches the table data but the materialized view data does not. Table update t set val = upper(val). This indicates the materialized view is returning a physically separate copy of the table data. ROWID -----------------AAAgY9AAEAAAAVfAAA AAAgY9AAEAAAAVfAAB AAAgY9AAEAAAAVfAAC AAAgY9AAEAAAAVfAAD View select rowid from V order by rowid . KEY ---------1 2 3 4 VAL ----A B C select * from MV . or automatically by Oracle in some cases. on the other hand.Unlike views. indicating the view returns the exact same data stored in the table. KEY ---------1 2 3 4 VAL ----A B C Materialized View select * from MV . differ from those of the table.

update t set val = lower(val). 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. 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. REFRESH COMPLETE There are various ways to refresh the data in a materialized view. In such cases it would be better to process only the changed rows. drop view v . 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".Cleanup drop materialized view mv . select key. 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.REFRESH( LIST => 'MV'. the simplest way being a complete refresh.REFRESH procedure to initiate it. even when the new result set is identical to the old one. The REFRESH COMPLETE clause tells Oracle to perform complete refreshes by default when a materialized view is refreshed. val. We will explore this type of refresh next. KEY ---------1 2 3 4 VAL ----a b c ROWID -----------------AAAWgHAAEAAAAIEAAA AAAWgHAAEAAAAIEAAB AAAWgHAAEAAAAIEAAC AAAWgHAAEAAAAIEAAD execute DBMS_MVIEW. METHOD => 'C' ). We will use the DBMS_MVIEW. val. even though the data in table T was unchanged throughout. . select key. Let's see a complete refresh in action now. for Complete refresh. rowid from mv . create materialized view mv REFRESH COMPLETE as select * from t . rowid from mv . This is because complete refreshes create a whole new set of data. commit.

describe MLOG$_T Name Null? -------------------------------------------. Fortunately there is a way to refresh only the changed rows in a materialized view's base table. In practice developers other than Dizwell never actually need to reference this table.KEY column mirrors the base table's primary key column T.------------------------------------KEY NOT NULL NUMBER VAL VARCHAR2(5) create materialized view log on t .-------KEY SNAPTIME$$ DMLTYPE$$ OLD_NEW$$ CHANGE_VECTOR$$ Type ---------------------NUMBER DATE VARCHAR2(1) VARCHAR2(1) RAW(255) The MLOG$_T. This is because a table can only ever have one materialized view log related to it at a time. To see what a materialized view log looks like we can examine the table used to implement it.KEY. The other MLOG$ columns are system generated. Before a materialized view can perform a fast refresh however it needs a mechanism to capture any changes made to its base table. complete refreshes of materialized views can be expensive operations. Note how the materialized view log is not given a name. like this. describe T Name Null? Type -------------------------------------------. T. so a name is not required. no rows selected The query above shows that a materialized view log is initially empty upon creation. This mechanism is called a Materialized View Log. Materialized View Logs As mentioned earlier. UPDATE t set val = upper( val ) where KEY = 1 . . select * from MLOG$_T . Rows are automatically added to MLOG$_T when base table T is changed. This is called fast refreshing. We can create a materialized view log on our test table.Cleanup drop materialized view mv . but showing it here helps illustrate materialized view log behaviour.

. dmltype$$ from MLOG$_T . val ) values ( 5. no rows selected WITH PRIMARY KEY To include the base table's primary key column in a materialized view log the WITH PRIMARY KEY clause can be specified. KEY ---------1 5 DMLTYPE$$ ---------U I If the changes affecting T are rolled back. because WITH PRIMARY KEY is the default option when no WITH clause is specified. which did not have a WITH clause.-------KEY SNAPTIME$$ DMLTYPE$$ OLD_NEW$$ CHANGE_VECTOR$$ Type --------------------NUMBER DATE VARCHAR2(1) VARCHAR2(1) RAW(255) Note how MLOG$_T contains T's primary key column. create materialized view log on t WITH PRIMARY KEY . so are the changes to MLOG$_T. drop materialized view log on t . This materialized view log is equivalent to the one created earlier in this topic.KEY. WITH ROWID To include rowids instead of primary keys WITH ROWID can be specified. T. drop materialized view log on t .INSERT into t ( KEY. 'e' ). dmltype$$ from MLOG$_T . desc mlog$_t Name Null? -------------------------------------------. create materialized view log on t WITH ROWID . select key. column dmltype$$ format a10 select key. rollback . Rollback complete.

------------------------------------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. insert. drop materialized view log on t . dmltype$$ from mlog$_T . create materialized view log on t2 WITH SEQUENCE . INSERT into T values ( 5. UPDATE T2 set amt = 333 where key = 60 . desc mlog$_t Name Null? -------------------------------------------. which contains rowids from table T. PRIMARY KEY .g. SEQUENCE$$ KEY DMLTYPE$$ . 3. commit. WITH SEQUENCE A special SEQUENCE column can be include in the materialized view log to help Oracle apply updates to materialized view logs in the correct order when a mix of Data Manipulation (DML) commands. create materialized view log on t WITH ROWID. update and delete. e. key. create materialized view log on t WITH SEQUENCE .-------M_ROW$$ 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$$ column. UPDATE T set val = upper(val) where key = 5 . drop materialized view log on t . INSERT into T2 values ( 60. are performed on multiple base tables in a single transaction. 300 ). 'e' ).desc mlog$_t Name Null? -------------------------------------------. select SEQUENCE$$. A materialized view log can also be created with both a rowid and a primary key column.

dmltype$$ from mlog$_T2 .---------60081 5 I 60083 5 U select SEQUENCE$$. column old_new$$ format a10 . SEQUENCE$$ KEY DMLTYPE$$ ---------. In fact. "Oracle recommends that the keyword SEQUENCE be included in your materialized view log statement unless you are sure that you will never perform a mixed DML operation (a combination of INSERT. KEY ---------1 2 3 4 5 VAL ----a b c E Type NUMBER VARCHAR2(5) DATE VARCHAR2(1) VARCHAR2(1) RAW(255) UPDATE t set val = 'f' where key = 5 . or DELETE operations on multiple tables). UPDATE. key.---------.------------------------------------KEY VAL SNAPTIME$$ DMLTYPE$$ OLD_NEW$$ CHANGE_VECTOR$$ select * from t . In the next snippet we include the VAL column. drop materialized view log on t .from Creating Materialized Views: Materialized View Logs" WITH Column List The WITH clause can also contain a list of specific base table columns." -.---------60082 60 I 60084 60 U Since mixed DML is a common occurrence SEQUENCE will be specified in most materialized view logs.---------. create materialized view log on t WITH ( VAL ). desc mlog$_t Name Null? -------------------------------------------. Oracle recommends it.---------.

update t set val = 'g' where key = 5 . ( VAL ). column old_new$$ format a9 select sequence$$. create materialized view log on t with sequence. drop materialized view log on t . which Oracle uses internally during refreshes) to the old value.---------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. key. The OLD_NEW$$ column identifies the value as either an old or a new value. SEQUENCE$$ KEY VAL ---------. it helps to have both the old value and the new value explicitly saved in the materialized view log. VAL. aka the "old" value.----. create materialized view log on t with sequence. create materialized view log on t with sequence ( VAL ) INCLUDING NEW VALUES . KEY VAL OLD_NEW$$ ---------. We can do that using the INCLUDING NEW VALUES clause.----60085 5 f 60086 5 g OLD_NEW$$ --------O N Note how both the old and the new values are stored in the same column.Commas The syntax diagrams for the CREATE MATERIALIZED VIEW LOG command indicate a comma is required between each component of the WITH clause.g. old_new$$ from mlog$_t order by sequence$$ . 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$$.---------. "( VAL )". val. e. Gotcha . In some situations. drop materialized view log on T .select key. However this does not appear to be the case when the component is a column list. val. which we will identify in later topics. like this. . old_new$$ from mlog$_t . ( VAL ). primary key . primary key * ERROR at line 1: ORA-00922: missing or invalid option Omitting the comma before the column list works better.

insert into t2 select * from t2_backup . . 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. commit. nor are any column transformations. Gotcha . primary key. No horizontal or vertical subsetting is permitted.from Oracle® Database Administrator's Guide 10g Release 2 (10.e. drop materialized view log on t2 .from Oracle® Database Administrator's Guide 11g Release 1 (11. insert into t select * from t_backup . refreshes where only the individual materialized view rows affected by base table changes are updated.create materialized view log on t with sequence ( VAL ). • For materialized view logs and queue tables. In Oracle 10g these restrictions are: • Tables with materialized view logs defined on them cannot be redefined online. nor are any column transformations.DBMS_REDEFINITION The DBMS_REDEFINITION package has certain restrictions related to materialized view logs. This is also called "incremental" refreshing.Restrictions for Online Redefinition of Tables In Oracle 11g they are: • After redefining a table that has a materialized view log. Earlier in this tutorial we saw how the rowids for each row in a materialized view changed after a complete refresh. online redefinition is restricted to changes in physical properties. delete t .2) . • For materialized view logs and queue tables. The only valid value for the column mapping string is NULL. -. Now let's see what happens to a materialized view's rowids after a fast refresh. Materialized view log created. No horizontal or vertical subsetting is permitted. -. The only valid value for the column mapping string is NULL. First we use the REFRESH FAST clause to specify that the default refresh method should be fast. i.1) . online redefinition is restricted to changes in physical properties. the subsequent refresh of any dependent materialized view must be a complete refresh.Restrictions for Online Redefinition of Tables Cleanup delete t2 . drop materialized view log on t .

Thus. with a fast refresh the materialized view data is not touched when no changes have been made to the base table. KEY ---------1 2 3 4 VAL ----a b XX ROWID -----------------AAAWm+AAEAAAAaMAAA AAAWm+AAEAAAAaMAAB AAAWm+AAEAAAAaMAAC AAAWm+AAEAAAAaMAAD . val.refresh( list => 'MV'.create materialized view log on t with sequence . create materialized view mv REFRESH FAST as select * from t . unlike a complete refresh where each row would have been created anew. method => 'F' ). select key. update t set val = 'XX' where key = 3 . rowid from mv . The "F" value for the "method" parameter ensures the refresh will be a Fast one. commit. select key. rowid from mv . Now let's update a row in the base table. KEY ---------1 2 3 4 VAL ----a b c ROWID -----------------AAAWm+AAEAAAAaMAAA AAAWm+AAEAAAAaMAAB AAAWm+AAEAAAAaMAAC AAAWm+AAEAAAAaMAAD Now we refresh the materialized view. rowid from mv .refresh( list => 'MV'. select key. val. method => 'F' ). val. execute dbms_mview. execute dbms_mview. KEY ---------1 2 3 4 VAL ----a b c ROWID -----------------AAAWm+AAEAAAAaMAAA AAAWm+AAEAAAAaMAAB AAAWm+AAEAAAAaMAAC AAAWm+AAEAAAAaMAAD The rowids did not change.

we'll learn more about this later). 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.Still no change in the rowids. rowid from mv . val. val. telling us that row 3 was updated during the refresh. drop materialized view mv . val. even though MV was created above with the REFRESH FAST clause. In row 3 we can see that VAL changed from "c" to "XX" though. KEY ---------1 2 VAL ----a b ROWID -----------------AAAWnBAAEAAAAaMAAA AAAWnBAAEAAAAaMAAB . execute dbms_mview. rowid from mv . In the following example note how. 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.refresh( list => 'MV'. select key. select key. KEY ---------1 2 3 4 VAL ----a b XX ROWID -----------------AAAWnBAAEAAAAaMAAA AAAWnBAAEAAAAaMAAB AAAWnBAAEAAAAaMAAC AAAWnBAAEAAAAaMAAD execute dbms_mview. method => 'C' ). method => 'F' ). A materialized view created with REFRESH FAST can still be refreshed completely if required though. This indicates that a complete refresh was performed. select key.refresh( list => 'MV'. create materialized view mv REFRESH COMPLETE as select * from t . all its rowids change after the refresh.

PURGE_LOG can be used. COUNT(*) ---------0 DBMS_MVEW. create materialized view log on t . In the example below note how the log table is empty after the refresh. indicating a fast refresh. Cleanup drop materialized view mv . commit . update t set val = 'c' where key = 3 . Purging Materialized View Logs Oracle automatically purges rows in the materialized view log when they are no longer needed. select count(*) from mlog$_t . method => 'F' ).3 XX 4 AAAWnBAAEAAAAaMAAC AAAWnBAAEAAAAaMAAD Note how none of the rowids in MV changed. COUNT(*) ---------0 insert into t values ( 5.refresh( list => 'MV'. commit. select count(*) from mlog$_t . 'e' ) . select count(*) from mlog$_t . create materialized view mv refresh fast as select * from t . drop materialized view log on t .PURGE_LOG If a materialized view log needs to be purged manually for some reason a procedure called DBMS_MVEW. COUNT(*) ---------1 execute dbms_mview. .

COUNT(*) ---------0 The "num" and "flag" parameters can be used to partially purge the log. See the PURGE_LOG manual page for further details. select count(*) from mlog$_t . method => 'F' ).select count(*) from mlog$_t . execute dbms_mview. KEY ---------1 2 3 4 5 VAL ----a b c e execute dbms_mview. line 2712 ORA-06512: at line 1 Such materialized views will need to be refreshed completely. Once a materialized view log has been purged any materialized views dependent on the deleted rows cannot be fast refreshed. END. select * from mv .PURGE_LOG( master => 'T'.DBMS_SNAPSHOT". flag => 'delete' ) .refresh( list => 'MV'. line 2743 ORA-06512: at "SYS. select * from mv . . COUNT(*) ---------1 execute DBMS_MVIEW.DBMS_SNAPSHOT". commit."T" younger than last refresh ORA-06512: at "SYS.refresh( list => 'MV'. COUNT(*) ---------0 update t set val = 'X' where key = 5 . method => 'F' ). method => 'C' ). * ERROR at line 1: ORA-12034: materialized view log on "SCOTT". select count(*) from mlog$_t . num => 9999. Attempting a fast refresh will raise an error. line 2537 ORA-06512: at "SYS.DBMS_SNAPSHOT".refresh( list => 'MV'. BEGIN dbms_mview.

1. Materialized views based on T cannot therefore be fast refreshed. create materialized view MV REFRESH FAST as select * from t2 . It can never be fast refreshed. drop materialized view mv ."T2" does not have a materialized view log For the second case materialized views are created without error. It can be fast refreshed after certain kinds of changes to the base table but not others. It can always be fast refreshed. 3.KEY ---------1 2 3 4 5 VAL ----a b c X Cleanup delete from t where key = 5 . KEY T_KEY AMT ---------. In the example below table T does not have a materialized view log on it. obviously. 2.---------.---------10 1 100 20 1 300 30 1 200 40 2 250 . REFRESH FAST Categories There are three ways to categorize a materialized view's ability to be fast refreshed. The next example demonstrates why. as select * from t2 * ERROR at line 3: ORA-23413: table "SCOTT". and will always be fast refreshed unless a complete refresh is explicitly requested. If we attempt to create such a materialized view we get an error. drop materialized view log on t . The third case is a little trickier. 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. select * from t2 . commit.

delete from t2 where key = 5 .refresh( list => 'MV'.---------AAAhMzAAEAAAEG8AAA 1 300 AAAhMzAAEAAAEG8AAB 2 250 So far everything works as expected. amt_max from mv . sequence ( t_key. line ORA-06512: at line 1 unsupported after deletes/updates 2255 2461 2430 . 500 ).---------AAAhMzAAEAAAEG8AAA 1 300 AAAhMzAAEAAAEG8AAB 2 500 Again.50 2 150 create materialized view log on t2 with primary key.DBMS_SNAPSHOT". We created a materialized view log and created a materialized view with fast refresh as its default refresh method. execute dbms_mview. select rowid. line ORA-06512: at "SYS. * ERROR at line 1: ORA-32314: REFRESH FAST of "SCOTT".DBMS_SNAPSHOT". create materialized view mv REFRESH FAST as select t_key.REFRESH command) and the materialized view correctly shows 500 as the maximum value for rows with T_KEY = 2. it worked as expected. Now let's try deleting a row from the base table. amt ) including new values . method => 'F' ). ROWID T_KEY AMT_MAX -----------------. select rowid.refresh( list => 'MV'.DBMS_SNAPSHOT". insert into t2 values ( 5."MV" ORA-06512: at "SYS.---------. t_key. amt_max from mv . commit.refresh( list => 'MV'. Let's try inserting a row into the base table. rowid. The view was fast refreshed (the rowid's did not change after the DBMS_MVIEW.---------. method => 'F' ). commit. 2. method => 'F' ). line ORA-06512: at "SYS. execute dbms_mview. BEGIN dbms_mview. max( amt ) amt_max from t2 group by t_key . ROWID T_KEY AMT_MAX -----------------. END. t_key.

it is only fast refreshable for inserts and direct loads.refresh( list => 'MV'. not updates or deletes.This time we received an error when we attempted a fast refresh. i. • There are even more restrictions for materialized views containing joins. etc. DBMS_MVIEW. 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. • • • • • • CREATE MATERIALIZED VIEW . method => 'C' ).EXPLAIN_MVIEW. amt_max from mv . UNION ALL. or never? One way would be to learn all the documented restrictions for fast refreshable materialized views. (We will see why it is an insert-only view in the next topic.e. The reason is because this type of materialized view is an "insert-only" materialized view. select rowid. aggregates.---------. ALL. sometimes.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 . ROWID T_KEY AMT_MAX -----------------. The following links can help you find them if required though. execute dbms_mview. Here are some of them.) To synchronize an insert-only materialized view after a delete we need to do a complete refresh. t_key. 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. subqueries.---------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. They are documented in various sections of a few different manuals and are too numerous and complex to repeat here.

DBMS_MVIEW. deleting old rows from MV_CAPABILITIES_TABLE 2. MV_CAPABILITIES_TABLE There are two ways to get the output from DBMS_MVIEW. predicting whether or not a materialized view is fast refreshable can be complicated.EXPLAIN_MVIEW utility can simplify this task however. documented CREATE TABLE command for MV_CAPABILITIES_TABLE can be found on UNIX systems at $ORACLE_HOME/rdbms/admin/utlxmv. Here is an abridged version. mvowner varchar(30) .sql. msgno integer . mvname varchar(30) . It uses the DBMS_MVIEW. The full. Cleanup drop materialized view mv . related_text varchar(2000) . seq number ) .EXPLAIN_MVIEW .• • • 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. capability_name varchar(30) .EXPLAIN_MVIEW As we saw in the preceding topic. simpler alternative for determining whether a materialized view is fast refreshable or not. related_num number . The DBMS_MVIEW. It is also available in Oracle's documentation at Oracle Database Data Warehousing Guide .Basic Materialized Views . msgtxt varchar(2000) .EXPLAIN_MVIEW with the table output method typically involves 1.Using MV_CAPABILITIES_TABLE (see Gotcha for a related bug). VARRAY Output Using DBMS_MVIEW. possible character(1) . Full details on how the utility works are available at the preceding link. The material below will help you use the utility effectively. via a table or via a varray. create table MV_CAPABILITIES_TABLE ( statement_id varchar(30) .EXPLAIN_MVIEW utility which we will explore next. To use the table method the current schema must contain a table called MV_CAPABILITIES_TABLE.EXPLAIN_MVIEW. drop materialized view log on t2 . running DBMS_MVIEW.

From http://www. v_capabilities sys.com/en/topic-12884. CREATE MATERIALIZED VIEW command text.sqlsnippets.Typical Usage: --set long 5000 -select my_mv_capabilities( 'MV_NAME' ) as mv_report from dual . create or replace function my_mv_capabilities ( p_mv in varchar2 . selecting new rows from MV_CAPABILITIES_TABLE. PCT. v_nl constant char(1) := unistr( '\000A' ). REWRITE. .new line v_previous_possible char(1) := 'X' .EXPLAIN_MVIEW's varray output option instead and supplement it with a custom function called MY_MV_CAPABILITIES. p_include_pct_capabilities in varchar2 default 'N' . To save time in this tutorial we will use DBMS_MVIEW.3. -.html --. any value big enough to contain the -report output will do -------------------------------------------------------------------------------pragma autonomous_transaction .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 --.Parameters: --p_mv -o this value is passed to DBMS_MVIEW.capabilities like REFRESH_FAST_PCT are included in the report -N .ExplainMVArrayType . p_capability_name_filter in varchar2 default '%' .EXPLAIN_MVIEW's "mv" parameter -o it can contain either a query. -or a materialized view name --p_capability_name_filter -o use either REFRESH. p_linesize in number default 80 ) return clob as -------------------------------------------------------------------------------. or the default --p_include_pct_capabilities -Y . --o the value 5000 is arbitraty.

related_text . ------------------------------------------------------------.possible . end if.possible <> v_previous_possible then v_output := v_output || v_nl || case v_capability. mvname .v_output clob . for v_capability in ( select capability_name .print section body -----------------------------------------------------------declare ' ' of: ' of: ' || ':' .possible when 'T' then 'Capable of: when 'Y' then 'Capable of: when 'F' then 'Not Capable when 'N' then 'Not Capable else v_capability. begin dbms_mview.explain_mview( mv => p_mv. msg_array => v_capabilities ) .print section heading -----------------------------------------------------------if v_capability. seq ) loop ------------------------------------------------------------.possible end || v_nl . 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 . possible . v_previous_possible := v_capability. possible desc .

'(.' || v_indented_line_size || '} |.' || v_indented_line_size || '} |.print capability name indented 2 spaces v_output := v_output || v_nl || ' ' || v_capability. return( v_output ). end.{1.related_text || ' ' . ' \1' || v_nl ) . end.{1.{1. . -.' || v_indented_line_size || '})' . commit . end loop. -. begin -.capability_name || v_nl .5 ).{1.msgtxt || ' ' . '(.print related text indented 4 spaces and word wrapped if v_capability. ' \1' || v_nl ) .print message text indented 4 spaces and word wrapped if v_capability.v_indented_line_size varchar2(3) := to_char( p_linesize .' || v_indented_line_size || '})' . end if. end if.msgtxt is not null then v_output := v_output || regexp_replace ( v_capability.related_text is not null then v_output := v_output || regexp_replace ( v_capability.

set long 5000 select my_mv_capabilities( 'SELECT * FROM T'. This completes our preparations.) The EXPLAIN_MVIEW output above shows that fast refresh is not possible in this case because T has no materialized view log. A list of messages and related text is available at Table 8-8 MV_CAPABILITIES_TABLE Column Details.EXPLAIN_MVIEW can analyze three different types of materialized view code: 1.EXPLAIN_MVIEW With a Query DBMS_MVIEW. Here is an example that explains a simple query which could appear as the defining query in a CREATE MATERIALIZED VIEW command. a defining query 2. MV_REPORT ------------------------------------------------------------------------------Capable of: REFRESH_COMPLETE Not Capable of: REFRESH_FAST REFRESH_FAST_AFTER_INSERT SCOTT. Now let's see DBMS_MVIEW. 'REFRESH' ) as mv_report from dual .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. an existing materialized view.EXPLAIN_VIEW in action. a CREATE MATERIALIZED VIEW command 3. DBMS_MVIEW./ show errors No errors. .

max( amt ) amt_max from t2 group by t_key . create materialized view mv refresh fast as select t_key. DBMS_MVIEW. amt ) including new values . rowid. rewrite. the insert-only one we saw in the preceding topic REFRESH FAST Categories. create materialized view log on t . Rewrite capabilities will be covered in Query Rewrite Restrictions and Capabilities.EXPLAIN_MVIEW With Existing Materialized View For our last example we will explain an existing materialized view. DBMS_MVIEW. 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. sequence ( t_key. For now we will only examine refresh capabilities. create materialized view log on t2 with primary key. select my_mv_capabilities ( 'CREATE MATERIALIZED VIEW MV REFRESH FAST AS SELECT * FROM T' .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.Note that DBMS_MVIEW. and partition change tracking (PCT) capabilities. 'REFRESH' ) as mv_report from dual .EXPLAIN_MVIEW can report on a materialized view's refresh.

CREATE TABLE MV_CAPABILITIES_TABLE . It does not mean the materialized view is fast refreshable in all cases.sql file and the CREATE TABLE command at Oracle Database Data Warehousing Guide . -. commit. execute dbms_mview. Gotcha Both the $ORACLE_HOME/rdbms/admin/utlxmv.POSSIBLE will either be "T" or "F". In actual use we can see the values are really "Y" and "N".T = capability is possible -.explain_mview( 'select * from t' ). POSSIBLE CHARACTER(1). select my_mv_capabilities( 'MV'. delete from mv_capabilities_table . 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.. column possible format a8 ..F = capability is not possible ...Using MV_CAPABILITIES_TABLE state the values in MV_CAPABILITIES_TABLE. 'REFRESH' ) as mv_report from dual . Note also that the "REFRESH_FAST" capability will appear whenever at least one of the other REFRESH_FAST_% capabilities is available. but not other types of DML.Basic Materialized Views ..

max( amt ) amt_max from t2 group by t_key .EXPLAIN_MVIEW output is saved to a varray. insert into t2 values ( 5.---------. t_key. drop materialized view log on t2 . 2. sequence ( t_key. select rowid. however. rowid.refresh( list => 'MV' ). amt_max from mv . 500 ). used when DBMS_MVIEW. 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. POSSIBLE -------Y N The values "T" and "F" are. . create materialized view mv REFRESH FORCE as select t_key. execute dbms_mview. Cleanup set long 80 drop materialized view mv .---------AAAWpLAAEAAAAaMAAA 1 300 AAAWpLAAEAAAAaMAAB 2 250 First let's try an insert and a refresh. otherwise it performs a COMPLETE refresh. It performs a FAST refresh if possible. commit.select distinct POSSIBLE from mv_capabilities_table . ROWID T_KEY AMT_MAX -----------------. amt ) including new values . create materialized view log on t2 with primary key. drop materialized view log on t .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. REFRESH FORCE In REFRESH FAST Categories and DBMS_MVIEW.

. FAST or COMPLETE.refresh( list => 'MV' ). drop materialized view log on t2 . Instead Oracle performed a COMPLETE refresh (note how the rowids for each row changed).---------. This time with REFRESH FORCE we did not.---------. t_key. select rowid. create materialized view mv NEVER REFRESH as select * from t . ROWID T_KEY AMT_MAX -----------------. Cleanup drop materialized view mv . t_key. NEVER REFRESH If for some reason we need to prevent refresh operations of any sort. Now let's try a delete followed by a refresh. execute dbms_mview."MV" unsupported after deletes/updates" error at this point. 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.select rowid. commit.---------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. amt_max from mv . select * from mv .---------AAAWpLAAEAAAAaMAAC 1 300 AAAWpLAAEAAAAaMAAD 2 250 In the REFRESH FAST Categories topic we received an "ORA-32314: REFRESH FAST of "SCOTT". on our materialized views we can use the NEVER REFRESH method. delete from t2 where key = 5 . ROWID T_KEY AMT_MAX -----------------. amt_max from mv .

) NEVER REFRESH can come in handy though when refresh operations on a materialized view need to be prevented temporarily during maintenance or debugging operations. BEGIN dbms_mview. To refresh ON DEMAND materialized views we explicitly call one of the following procedures. * ERROR at line 1: ORA-23538: cannot explicitly refresh a NEVER REFRESH materialized view ("MV") ORA-06512: at "SYS. . commit .refresh( 'MV' ).refresh( 'MV' ).DBMS_SNAPSHOT". Cleanup drop materialized view mv . commit . (If you know of any please let me know using the Comments link below. line 2537 ORA-06512: at "SYS.REFRESH command. drop materialized view mv . I cannot see a practical reason for having a materialized view with NEVER REFRESH set at all times. update t set val = lower(val) . line 2743 ORA-06512: at "SYS. line 2712 ORA-06512: at line 1 Oracle prevented the refresh by raising an error. In this case the materialized view's refresh mode can be changed to NEVER REFRESH using the ALTER MATERIALIZED VIEW command.DBMS_SNAPSHOT". create materialized view mv REFRESH ON DEMAND as select * from t . execute dbms_mview. In other words this create materialized view mv as select * from t . is equivalent to this. END.DBMS_SNAPSHOT". This is know as ON DEMAND refreshing and it is the default refresh mode when none is specified in the CREATE MATERIALIZED VIEW command.update t set val = upper(val) . ON DEMAND Up to this point in the tutorial we have always refreshed our materialized views manually with the DBMS_MVIEW.

KEY VAL ---------. Here is an example. 'e' ).REFRESH.REFRESH DBMS_MVIEW. val from mv . 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. commit.• • • DBMS_MVIEW.REFRESH_DEPENDENT Here is an example that uses DBMS_MVIEW. select * from mv where key = 5 . select rowid. key. 'e' ).REFRESH( 'MV' ). create materialized view log on t . This is possible using the ON COMMIT refresh mode. . ROWID KEY VAL -----------------. create materialized view mv REFRESH FAST ON COMMIT as select * from t .----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. select * from mv where key = 5 . insert into t values ( 5. insert into t values ( 5.---------. no rows selected execute DBMS_MVIEW.----5 e Cleanup drop materialized view mv . commit. delete from t where key = 5 .REFRESH_ALL_MVIEWS DBMS_MVIEW.

because the materialized view contains an Oracle-supplied type create materialized view mv2 REFRESH FAST ON COMMIT as select key. val. ROWID KEY VAL -----------------. In the following example materialized view MV (created at the top of this page) was created with REFRESH FAST. -.---------. No call to DBMS_MVIEW.REFRESH was required. The materialized view cannot contain object types or Oracle-supplied types.this materialized view is not fast refreshable -. commit. as select key. select rowid. ROWID KEY VAL -----------------. will therefore raise an error. Let's issue a COMMIT. sys_xmlgen( val ) as val_xml from t . key. . 2.---------. 1.select rowid. T. 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. Restrictions Materialized views can only refresh ON COMMIT in certain situations.----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. val from mv . key. The first case produces an error during the CREATE MATERIALIZED VIEW command. Attempting a distributed transaction on its base table. val. val from mv .----AAAXNGAAEAAAAasAAA 1 a AAAXNGAAEAAAAasAAB 2 b AAAXNGAAEAAAAasAAC 3 c AAAXNGAAEAAAAasAAD 4 Nothing happend yet. The base tables will never have any distributed transactions applied to them.

Gotcha The SQL Language Reference manual says this about the ON COMMIT clause. commit. select * from t . alter materialized view mv refresh ON DEMAND . create materialized view mv2 REFRESH COMPLETE ON COMMIT as select key.cleanup test data in preparation for next section delete from t where key >= 5 . . commit. commit * ERROR at line 1: ORA-02050: transaction 5. val from T@REMOTE . "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.) ON DEMAND materialized views have no such restriction. val from T@REMOTE . I also assumed that specifying "REFRESH ON COMMIT" is equivalent to specifying "REFRESH FAST ON COMMIT".insert into t select key+10.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. insert into t select key+10.21. commit . val from t . KEY ---------1 2 3 4 5 11 12 13 14 15 VAL ----a b c e a b c e -. 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." -. The following examples prove neither is correct however. as the following snippet demonstrates.5632 rolled back.

ROWID KEY VAL -----------------.----AAAXNMAAEAAAAakAAA 1 a AAAXNMAAEAAAAakAAB 2 b AAAXNMAAEAAAAakAAC 3 c AAAXNMAAEAAAAakAAD 4 insert into t values ( 5. The next example examines the behavior of "REFRESH ON COMMIT" without a specified refresh method. val from mv2 . val from mv2 . commit . drop materialized view log on t .because it has no materialized view log drop materialized view mv2 . "REFRESH ON COMMIT" is not therefore equivalent to "REFRESH FAST ON COMMIT". not FAST.---------. commit . key. .fast refreshable materialized views on T can no longer be created on T -.---------. delete from t where key >= 5 . create materialized view mv2 REFRESH ON COMMIT as select key. 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". select rowid.----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. when no REFRESH method is specified the default behaviour is "REFRESH FORCE" regardless of whether ON COMMIT is used or not. select rowid. was specified with ON COMMIT. -. key. Cleanup drop materialized view mv . val from t . drop materialized view mv2 . In fact.As we can see the CREATE MATERIALZIED view command succeeded even though COMPLETE. 'e' ). ROWID KEY VAL -----------------.

search_condition . val from t . sequence ( t_key. constraint_type.--------------. not to be confused with a base table). the table containing the results of the query. create materialized view mv refresh fast on commit as select t_key.--------------SYS_C0019948 P SYS_C0019948 In the next example Oracle automatically adds a check constraint. count(*) row_count from t2 group by t_key . rowid. constraint_type. CONSTRAINT_NAME CONSTRAINT_TYPE INDEX_NAME -------------------. drop materialized view mv .e. which is part of the materialized view also called "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. amt ) including new values .Constraints System Generated Constraints When a materialized view is created Oracle may add system generated constraints to its underlying table (i. create materialized view mv as select key. In the following example note how Oracle automatically adds a primary key constraint to the table called "MV". column constraint_name format a20 column constraint_type format a15 column index_name format a15 select constraint_name. column search_condition format a30 select constraint_name. index_name from user_constraints where TABLE_NAME = 'MV' .

-----------------------------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. When the materialized view is in ON COMMIT mode these constraints effectively constrain the materialized view's base tables.-----------------------------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.MY_CONSTRAINT) violated Implementing multirow validation rules such as this one properly is not possible using check constraints on regular tables. We will learn more about this powerful multirow validation approach in a . commit.---------. 1. 500 ). commit * ERROR at line 1: ORA-12008: error in materialized view refresh path ORA-02290: check constraint (SCOTT. Let's see this in action by creating a check constraint on MV. CONSTRAINT_NAME -------------------SYS_C0019949 MY_CONSTRAINT CONSTRAINT_TYPE SEARCH_CONDITION --------------. select * from t2 . constraint_type.---------10 1 100 20 1 300 30 1 200 40 2 250 50 2 150 alter table mv -. KEY T_KEY AMT ---------. Implementing them using triggers can be difficult if not impossible. search_condition from user_constraints where table_name = 'MV' .note we used "alter table" here add CONSTRAINT MY_CONSTRAINT CHECK ( ROW_COUNT <= 3 ) DEFERRABLE .--------------. insert into T2 values ( 5. CONSTRAINT_NAME CONSTRAINT_TYPE SEARCH_CONDITION -------------------. select constraint_name.from where user_constraints table_name = 'MV' . With materialized views they are declared using a few lines of code and are virtually bullet proof when applied correctly.

e. the table containing the results of the query.column_name from user_indexes i inner join user_ind_columns ic . select constraint_name. Constraints. not to be confused with a base table). ALTER MATERIALIZED VIEW mv add constraint my_second_constraint check ( row_count < 4 ) deferrable . Cleanup drop materialized view mv . search_condition from user_constraints where table_name = 'MV' . In the following example note how Oracle automatically adds an index to implement the system generated primary key we saw in the preceding topic. column index_name format a15 column column_name format a15 select index_name . constraint_type. Gotcha When we created MY_CONSTRAINT above we use an ALTER TABLE command. drop materialized view log on t2 .future SQL Snippets tutorial so stay tuned! In the mean time Ask Tom "Declarative Integrity" has some good information on the subject.uniqueness . Indexes When a materialized view is created Oracle may add system generated indexes to its underlying table (i. CONSTRAINT_NAME -------------------SYS_C0019949 MY_CONSTRAINT MY_SECOND_CONSTRAINT CONSTRAINT_TYPE SEARCH_CONDITION --------------. i. create materialized view mv as select key. Curiously enough an ALTER MATERIALIZED VIEW command would have worked too. Until the documentation says this is legal it is best to use ALTER TABLE.-----------------------------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. val from t . ic.

uniqueness . ie. i.--------------. rowid.--------------SYS_C0019959 UNIQUE KEY In the next example Oracle automatically generates a function based index.using ( index_name ) where i. create materialized view mv refresh fast on commit as select t_key. In the following example we will add an index on the T_KEY column. INDEX_NAME UNIQUENES COLUMN_NAME COLUMN_EXPRESSION --------------. sequence ( t_key.----------------------------------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.--------. amt ) including new values . COUNT(*) ROW_COUNT from t2 group by t_key . .table_name = 'MV' .column_name .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.) Adding Your Own Indexes We can add out own indexes to MV just as we would a regular table. INDEX_NAME UNIQUENES COLUMN_NAME --------------. See Nulls and Equality: SQL Only for additional info. create materialized view log on t2 with primary key. ic.table_name = 'MV' . drop materialized view mv .--------. column column_expression format a35 select index_name .

column_name from user_indexes i inner join user_ind_columns ic using ( index_name ) where i.uniqueness . i.---------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 .create index MY_INDEX on mv ( T_KEY ) . T_KEY ROW_COUNT ---------. 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 .access("T_KEY"=2) . ic. select index_name .table_name = 'MV' .

select t_key.---------1 300 2 250 faster than its equivalent query. amt ) including new values . drop materialized view log on t2 . Cleanup drop materialized view mv . sequence ( t_key.Note ----. MAX( AMT ) AMT_MAX from t2 group by t_key . T_KEY AMT_MAX ---------. ENABLE QUERY REWRITE Materialized views can be useful for pre-calculating and storing derived values such as AMT_MAX in the following snippet. amt_max FROM MV order by t_key . create materialized view log on t2 with primary key. T_KEY AMT_MAX ---------.---------1 300 2 250 . max( amt ) as amt_max FROM T2 group by t_key order by t_key . create materialized view mv refresh fast on commit as select t_key. rowid. Such materialized views make queries like this select t_key.dynamic sampling used for this statement Note how the optimizer chose an INDEX RANGE SCAN from MY_INDEX in step 2.

set autotrace on explain set linesize 95 select t_key. alter materialized view mv ENABLE QUERY REWRITE . To see it in action we first need to make the materialized view available to Query Rewrite like this. T_KEY AMT_MAX ---------. Finally we can confirm Oracle will use the materialized view in queries by turning SQL*Plus's Autotrace feature on.---------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 | ------------------------------------------------------------------------------------- . 'MV' ) . Next we collect statistics on the materialized view to help Oracle optimize the query rewrite process.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.) Note that materialized views which do not include the ENABLE QUERY REWRITE clause will have Query Rewrite disabled by default.gather_table_stats( user. (See Gotcha .ORA-00439 below if you encounter an ORA-00439 error at this step. execute dbms_stats. max( amt ) as amt_max FROM T2 group by t_key order by t_key .

count(*) as row_count . .dynamic sampling used for this statement Note how the optimizer chose to access T2 this time. Gotcha . 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. alter session set QUERY_REWRITE_ENABLED = FALSE . Without the Query Rewrite feature the execution plan would look like this.ORA-00439 The materialized view query rewrite feature is not available in Oracle XE and some other Oracle configurations. Each time this query is executed it has to recalculate MAX(VAL) from the information in T2 for each group. max( amt ) as amt_max FROM T2 group by t_key order by t_key . a more expensive approach than simply selecting pre-calculated column values from MV is. create materialized view mv2 refresh fast on commit ENABLE QUERY REWRITE as select t_key . T_KEY AMT_MAX ---------. count(amt) as amt_count from t2 group by t_key . select t_key.---------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 ----.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.

an expression containing the USER pseudo column or the SYSTIMESTAMP function. 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. single table materialized view with query rewrite disabled. e. drop materialized view log on t2 .g. val.from t2 * ERROR at line 9: ORA-00439: feature not enabled: Materialized view rewrite Cleanup alter session set query_rewrite_enabled = true . USER from t . create materialized view mv DISABLE QUERY REWRITE as select key. as select key. 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. val from t . set autotrace off drop materialized view mv . MV_REPORT . Attempting to violate these restrictions results in an error. In EXPLAIN_MVIEW we used a utility called MY_MV_CAPABILITIES to explore a materialized view's refresh capabilities. 'REWRITE' ) as mv_report from dual . val. In the snippets below we will use this same utility to explore rewrite capabilities. create materialized view mv ENABLE QUERY REWRITE as select key. set long 5000 select my_mv_capabilities( 'MV'. First lets look at a simple.

------------------------------------------------------------------------------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. select my_mv_capabilities( 'MV'. 'REWRITE' ) as mv_report from dual . 'REWRITE' ) as mv_report from dual . create materialized view mv enable query rewrite as select key.) Enabling query rewrite on the materialized view changes this. MV_REPORT ------------------------------------------------------------------------------Capable of: REWRITE REWRITE_FULL_TEXT_MATCH REWRITE_PARTIAL_TEXT_MATCH REWRITE_GENERAL Now all rewrite capabilities are available. MV_REPORT ------------------------------------------------------------------------------- . select my_mv_capabilities( 'MV'. but not others. If the materialized view happened to referenced a remote table then some rewrite capabilities would be available. drop materialized view mv . val from T@REMOTE . (Descriptions of each capability name are available at Table 8-7 CAPABILITY_NAME Column Details. alter materialized view mv ENABLE QUERY REWRITE .

KEY T_KEY AMT ---------. KEY ---------1 2 3 4 VAL ----a b c select * from t2 .---------.---------10 1 100 20 1 300 30 1 200 40 2 250 50 2 150 create materialized view mv as select t.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 . . t2.key t_key . Join Queries So far in this tutorial we have only seen materialized views based on a single table.key t2_key .amt t2_amt from t. Here is a simple example.t_key . t2. Materialized views can also be created on multi-table queries to store the pre-calculated results of expensive join operations. t2 where t. select * from t .key = t2. t.val t_val .

The Prototype Applying these restrictions and recommendations to our test case above yields the following prototypical materialized view with joins. create materialized view log on t with rowid. 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. sequence .) to be fast refreshable certain restrictions beyond the General Restrictions on Fast Refresh must be met. create materialized view log on t2 with rowid. t2_amt from mv .---------. you must collect statistics on it using the DBMS_STATS package. 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. t_val.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. then having an index on each of the join column rowids in the detail table will enhance refresh performance greatly. drop materialized view mv . because this type of materialized view tends to be much larger than materialized views containing aggregates. -. sequence . "If a materialized view contains joins but no aggregates." -. subqueries. t2_key. They are as follows." -.from Refreshing Materialized Views: Tips for Refreshing Materialized Views Without Aggregates "After you create the materialized view. create materialized view mv . unions.---------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.from CREATE MATERIALIZED VIEW. T_KEY ---------1 1 1 2 2 T_VAL T2_KEY T2_AMT ----. etc. Oracle Database needs the statistics generated by this package to optimize query rewrite.select t_key.

t2 where t. .---------.amt t2_amt t. .val t_val t2. here are MV's initial contents. create index mv_i2 on mv ( t2_row_id ) .rowid t2_row_id from t. Whenever we create a fast refreshable view we should use our EXPLAIN_MVIEW utility.rowid t_row_id t2. select t_key. T_KEY ---------1 1 1 2 T_VAL T2_KEY T2_AMT ----. t2_key. t2_amt from mv .refresh fast on commit enable query rewrite as select t. 'REFRESH' ) as mv_report from dual .key = t2.---------a 10 100 a 20 300 a 30 200 b 40 250 . .key t_key t. to confirm it can be refreshed in all required situations. . 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. . 'MV' ) . MY_MV_CAPABILITIES. execute dbms_stats. create index mv_i1 on mv ( t_row_id ) . First.t_key .key t2_key t2. t_val.gather_table_stats( user. set long 5000 select my_mv_capabilities( 'MV'.

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. commit. insert into t2 values ( 60. t2_amt from mv . 3.2 b 50 150 Now let's do some DML on both base tables and see the effect on MV. update t set val = upper(val) . Query Rewrite Materialized views containing joins can be used by the query rewrite facility (see ENABLE QUERY REWRITE). 'REWRITE' ) as mv_report from dual . T_KEY ---------1 1 1 2 2 3 T_VAL T2_KEY T2_AMT ----. create materialized view mv2 refresh fast as . t_val. select t_key. MV_REPORT ------------------------------------------------------------------------------Capable of: REWRITE REWRITE_FULL_TEXT_MATCH REWRITE_PARTIAL_TEXT_MATCH REWRITE_GENERAL Gotcha . t2_key.---------A 10 100 A 20 300 A 30 200 B 40 250 B 50 150 C 60 300 Both changes are reflected in MV.---------. 300 ) . select my_mv_capabilities( 'MV'. as expected.

rowid t2_row_id from T INNER JOIN T2 ON ( T.key t_key .KEY = T2.KEY = T2.rowid t2_row_id from T INNER JOIN T2 ON . .key t_key t.key t2_key .key t2_key t2.1 explains that it is really an undocumented limitation of fast refresh materialized views. . t. select t.rowid t_row_id t2.T_KEY )' .amt t2_amt t. 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 . .. . ( T. 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. t2.T_KEY ) T INNER JOIN T2 ON ( T.val t_val . t.val t_val t2. Metalink note 420856. t2. t2.amt t2_amt . 'REFRESH_FAST_AFTER_INSERT' ) as mv_report from dual .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.rowid t_row_id . select my_mv_capabilities ( 'create materialized view mv2 refresh fast as select t.KEY = T2.

select * from t2 order by key .) to be fast refreshable certain restrictions beyond the General Restrictions on Fast Refresh must be met. select * from mv order by t_key . Here is a simple example. etc. insert into t2 select * from t2_backup . subqueries. These .Cleanup drop materialized view mv . SUM(AMT) AMT_SUM from t2 group by t_key . drop materialized view log on t .---------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 .---------1 600 2 400 REFRESH FAST For a materialized view with only aggregates (no joins.---------. delete t . delete t2 . insert into t select * from t_backup . unions. materialized views containing aggregate functions are also possible. drop materialized view log on t2 . Aggregate Queries In addition to materialized views based on join queries. KEY T_KEY AMT ---------. T_KEY AMT_SUM ---------. commit.

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. MIN. COUNT.additional restrictions are fully documented at Restrictions on Fast Refresh on Materialized Views with Aggregates. STDDEV." o "Specify the SEQUENCE clause if the table is expected to have a mix of inserts/direct-loads. The table below summarized these requirements. AVG." • aggregates in the defining query must be either SUM. Recommendations . VARIANCE. For our current test case the most significant restrictions are these. deletes. and updates. 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. • all base tables must have materialized view logs that: o "Contain all columns from the table referenced in the materialized view.) Oracle recommends including the Optional Aggregates expressions to obtain the most efficient and accurate fast refresh of the materialized view." o "Specify with ROWID and INCLUDING NEW VALUES.

Oracle Database needs the statistics generated by this package to optimize query rewrite. set long 5000 select my_mv_capabilities( 'MV'. 'REFRESH' ) as mv_report from dual . MV_REPORT ------------------------------------------------------------------------------Capable of: . to confirm it can be refreshed in all required situations. create index mv_i1 on mv ( t_key ) . execute dbms_stats. 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. sum(amt) as amt_sum .gather_table_stats( user. count(*) as row_count . Additionally we also expect that our GROUP BY columns will often be specified in WHERE or JOIN clauses." -. drop materialized view mv . create materialized view mv refresh fast on commit enable query rewrite as select t_key . count(amt) as amt_count from t2 group by t_key . Whenever we create a fast refreshable view we should use our EXPLAIN_MVIEW utility. 'MV' ) . To improve the performance of such queries we will therefore add indexes to our materialized view's GROUP BY columns. MY_MV_CAPABILITIES. you must collect statistics on it using the DBMS_STATS package. The Prototype Applying these restrictions and recommendations to our test case above yields the following prototypical materialized view with aggregates.from CREATE MATERIALIZED VIEW.The recommendation about gathering statistics that we saw in the Join Queries topic also applies to materialized views with aggregates. amt ) including new values . sequence ( t_key. create materialized view log on t2 with rowid. "After you create the materialized view.

---------1 600 3 3 2 0 2 2 3 300 1 1 Both changes are reflected in MV.---------.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.---------1 600 3 3 2 400 2 2 Now let's do some DML on the base table and see the effect on MV. as expected. insert into t2 values ( 60. T_KEY AMT_SUM ROW_COUNT AMT_COUNT ---------. select * from mv order by t_key .---------. commit. 'REWRITE' ) as mv_report from dual . here are MV's initial contents. MV_REPORT ------------------------------------------------------------------------------- . 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 . 3. select * from mv order by t_key . update t2 set amt = 0 where t_key = 2 . T_KEY AMT_SUM ROW_COUNT AMT_COUNT ---------.---------.---------. First. select my_mv_capabilities( 'MV'. 300 ) .

Capable of: REWRITE REWRITE_FULL_TEXT_MATCH REWRITE_PARTIAL_TEXT_MATCH REWRITE_GENERAL Gotcha . 900 ) . select * from mv2 order by t_key . insert into t2 values ( 70. but what happens if we do not include these columns? Let's find out.---------1 600 2 0 3 300 Let's try an INSERT.Insert-Only Materialized Views On Commit We know that COUNT(*). T_KEY AMT_SUM ---------. select * from mv2 order by t_key . and sometimes COUNT(expr). must be included in our materialized views for them to be fast refreshable in all cases. sum(amt) as amt_sum -count(*) as row_count . -count(amt) as amt_count from t2 group by t_key . 3. commit . create materialized view mv2 refresh fast on commit enable query rewrite as select t_key . T_KEY AMT_SUM ---------.---------1 600 2 0 3 1200 .

The view was fast refreshed after the transaction committed. except this time we won't get any warnings or errors if a commit fails to trigger a fast refresh.Looks good. 3 rows deleted. T_KEY AMT_SUM ---------. 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. So we've confirmed we have another insert-only materialized view. delete from t2 where t_key = 1 . .REFRESH( 'MV2'. When I first learned materialized views I stumbled across this behaviour by accident and found it puzzling. T_KEY AMT_SUM ---------. commit. Let's try synchronizing MV2 manually using DBMS_MVIEW. The materialized view did not refresh on commit and no errors were generated. Commit complete.---------2 0 3 1200 That's a little better. 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. all the rows for T_KEY = 1 were deleted from T2 but the group still appears in MV2. The manual page for CREATE MATERIALIZED VIEW did not mention insert-only materialized views so I had no clue to their existence. select * from mv2 order by t_key . After all. 'c' ) select * from mv2 order by t_key . until I re-read the page a third time and followed up on this seemingly inconsequential little comment.---------1 600 2 0 3 1200 Oops. execute DBMS_MVIEW.REFRESH.

EXPLAIN_MVIEW to see whether or not it is an "insert-only" materialized view. 'REFRESH' ) as mv_report from dual .EXPLAIN_MVIEW utility is when working with fast refreshable materialized views. 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. For instructions on actually implementing the refresh. delete t2 . refer to Oracle Database Advanced Replication and Oracle Database Data Warehousing Guide. .EXPLAIN_MVIEW has to say about MV2. So the lesson here is do not assume materialized views created with REFRESH FAST ON COMMIT will always refresh fast on commit. drop materialized view log on t2 . drop materialized view mv2 ." -.from CREATE MATERIALIZED VIEW This eventually lead me to learn about insert-only refreshing and how indispensable the DBMS_MVIEW. Always check it with DBMS_MVEW. Let's see what DBMS_MVIEW. This is how to recognize an insert-only materialized view. but not after other types of DML."(The REFRESH clause) only sets the default refresh options. insert into t2 select * from t2_backup . Cleanup drop materialized view mv . select my_mv_capabilities( 'MV2'. commit.

max(amt) amt_max .EXPLAIN_MVIEW) tells us about our query.---------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. sequence ( key. amt ) including new values . max(key) keep ( dense_rank last order by amt ) as t2_key_of_amt_max from t2 group by t_key . create materialized view log on t2 with rowid . T_KEY AMT_MAX T2_KEY_OF_AMT_MAX ---------. given this base table select t_key.Nested Materialized Views Sometimes a single materialized view will not meet our requirements. T_KEY AMT KEY ---------. amt.---------.---------. select t_key t_key . amt.----------------1 300 20 2 250 40 (T2_KEY_OF_AMT_MAX identifies the KEY value associated with the highest AMT value in each group. For example.) As always the first step is to create a materialized view log on T2. key . set long 5000 select . t_key. Now let's see what the MY_MV_CAPABILITIES utility (created in topic DBMS_MVIEW. key from t2 order by t_key.

max(amt) amt_max . max(amt) from t2 group by t_key . STDDEV. Let's try writing the query using a subquery instead of LAST.my_mv_capabilities( 'select t_key t_key . t2. 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. AVG. MIN and MAX (see Restrictions on Fast Refresh on Materialized Views with Aggregates). max(key) as t2_key_of_amt_max from t2 where ( t_key. 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. select my_mv_capabilities( 'select t_key t_key . VARIANCE. 'REFRESH' ) as mv_report from dual . max(amt) amt_max .amt ) in ( select t_key. max(key) keep ( dense_rank LAST order by amt ) as t2_key_of_amt_max from t2 group by t_key' .

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. 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 .) group by t_key' . max( amt ) over ( partition by t_key ) as amt_max . MV_REPORT ------------------------------------------------------------------------------Capable of: REFRESH_COMPLETE . Perhaps an analytic approach will work? select my_mv_capabilities( 'select distinct t_key t_key . 'REFRESH' ) as mv_report from dual .

before creating a type of materialized view we have not tried before we must be aware of its restrictions. Fortunately Oracle allows for a materialized view like MV3 and automatically manages the refresh order when all three views are refreshable on commit. However the materialized view for step 3. Since deriving the desired result set is conceptually a three step process 1. Restrictions and Recommendations As always. will need to be based on MV1 and MV2 and will need to refresh after they do. • The defining query must contain joins or aggregates. find the highest KEY value per AMT 3. • If REFRESH FAST is specified then all materialized views in any chain related to the materialized view must also specify REFRESH FAST. Materialized views like MV3 are called "Nested Materialized Views". which we will call MV1 and MV2. • The base materialized views must contain joins or aggregates. For nested materialized views they are these. can be based on table T2 and can be refreshed independently of each other. whether they are tables or materialized views. which we will call MV3. Note the term "Nested Materialized View" does not refer to MV1 and MV2. 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 is a bit of a relief actually since the technique is rather crass. even though they could be thought of as being "nested" within MV3. find the highest AMT value 2. must each have materialized view logs. Note that all base objects in a nested materialized view. are treated as tables.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. • All base objects. regardless of whether they are tables or materialized views. . We need to rethink our approach.

create materialized view log on mv2 with rowid . count(*) row_count from t2 group by t_key.We are now ready to craft our three step solution. 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 . count(amt) amt_count . create materialized view MV1 refresh fast on commit as select t_key . amt . count(*) row_count from t2 group by t_key . amt_count. sequence ( t_key. 'REFRESH' ) as mv_report from dual . amt_max. max(amt) amt_max . amt . sequence . create materialized view log on mv1 with rowid . select my_mv_capabilities( 'MV1'. row_count ) including new values . max(key) max_key_per_amt .

select my_mv_capabilities( 'MV3'.amt_max . mv1.amt_max = mv2. mv1. mv2 where mv1.( t_key.t_key . mv2. max_key_per_amt. MV_REPORT ------------------------------------------------------------------------------Capable of: REFRESH_COMPLETE REFRESH_FAST REFRESH_FAST_AFTER_INSERT REFRESH_FAST_AFTER_ONETAB_DML REFRESH_FAST_AFTER_ANY_DML .max_key_per_amt as t2_key_of_amt_max . 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 = mv2.amt . 'REFRESH' ) as mv_report from dual .t_key and mv1.rowid mv2_rowid from mv1.rowid mv1_rowid . row_count ) including new values . select my_mv_capabilities( 'MV2'. 'REFRESH' ) as mv_report from dual . mv2.

the nested one. key from t2 order by t_key. Let's confirm that MV3. update t2 set amt = 300 where key = 30 . commit.----------------1 300 30 2 250 40 3 650 60 It does.We finally have a fast refreshable materialized view solution. amt_max. Now let's put all three materialized views through their paces. t2_key_of_amt_max from mv3 order by t_key . insert into t2 values ( 70. It matches the results returned by the first query we tried which used the LAST function. 3. 3. contains the correct results. amt. commit. Mission accomplished. select t_key. amt_max. t2_key_of_amt_max from mv3 order by t_key .---------. First we perform a few mixed DML transactions. T_KEY AMT KEY ---------. amt. T_KEY AMT_MAX T2_KEY_OF_AMT_MAX ---------. 550 ). 450 ).---------. T_KEY AMT_MAX T2_KEY_OF_AMT_MAX ---------. select t_key. delete from t2 where key = 70 .----------------1 300 20 2 250 40 Good. insert into t2 values ( 60. update t2 set amt = 650 where key = 60 .---------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. Cleanup .---------. key . select t_key.

drop materialized view mv2 . create table t2_backup as select * from t2 . 30.drop materialized view mv1 . val varchar2(5) ) . etc. 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. drop materialized view mv3 . create table t_backup as select * from t . Cleanup . Setup Run the code on this page in SQL*Plus to create the sample tables. drop materialized view log on t2 . 1. t_key number . 1. data. insert into t2 select * from t2_backup . . amt number ) . 3. 2. 100 300 200 250 150 ) ) ) ) ) . ). 20. 50. 1. 2. . . primary key not null references t not null values values values values values ( ( ( ( ( 10. 'a' 'b' 'c' null ). commit. ). 2. create table t ( key number . create table t2 ( key number . insert insert insert insert commit. 4. delete t2 . 40. used by the examples in this section. ).

do not drop it from your schema -. etc. COLUMN.unless you specifically created it for this tutorial and no longer wish to -. exit . drop table t2_backup . drop table t2 .g. To clear session state changes (e. drop table t_backup . and VARIABLE commands) exit your SQL*Plus session after running these cleanup commands. drop table t .WARNING!! --.use it ------------------------------------------------------------------------------drop table mv_capabilities_table . -------------------------------------------------------------------------------.Run the code on this page to drop the sample tables. those made by SET. created in earlier parts of this section.MV_CAPABILITIES_TABLE is an Oracle table. procedures. drop function my_mv_capabilities .

'a2'] . null ] = sum(num_val)[any.'n/a'. set null "" select case when key like 'Total%' then key else null end as total .any] + sum(num_val)[any.MODEL Clause This section presents tutorials on the MODEL clause of the SELECT command. nvl( group_2. Introduced in Oracle 10g.n/a'. 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). num_val[ 'Total 4 . num_val from t model dimension by ( cast(key as varchar2(20)) as key . null ] = .any] . null. nvl( group_1. null ] = sum(num_val)[any. 'n/a' ) as group_1 .A + a2'. group_1 . null ] = sum(num_val)[any.------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.a1 + a3'. num_val[ 'Total 3 . It also adds procedural features to SQL previously available only through PL/SQL calls.'C').any] . group_2 . 'n/a' ) as group_2 ) measures( num_val ) rules ( num_val[ 'Total 1 . num_val[ 'Total 2 . null. For example. create a report containing ad-hoc totals like this. null.A + C'. with a single command.'A'. null. 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 ---------.group_1 in ('A'.group_1 <> 'A'.

total nulls first . 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 ) -.800 (This last technique is explained fully in another section of SQL Snippets at Rows to String: MODEL Method 1.---------A A A B B B C C n/a n/a Total 1 .A + a2 Total 3 . group_2 . GROUP_1 ---------A B C (null) STRING ---------------------------------------100.' || string[iteration_number+1] ) order by group_1 . or impossible to do with a non-MODEL SELECT command.200.------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.'a3')] ) order by group_1 .300 100 200.group_2 in ('a1'.Note 1 rules upsert iterate( 6 ) until ( presentv(string[iteration_number+2]. inefficient.1. Here is an example.n/a Total 4 .A + C Total 2 . TOTAL GROUP_1 -------------------.a1 + a3 GROUP_2 NUM_VAL ---------. set null "(null)" column string format a40 select group_1.) .300 100.any.sum(num_val)[any. substr( string.0) = 0 ) ( string[0] = string[0] || '.

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

num_val . num_val from t model DIMENSION BY ( KEY AS KEY_3 ) measures ( num_val ) rules () order by key_3 . 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.. 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. select 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. num_val from t model DIMENSION BY ( ROWNUM AS KEY_3 ) measures ( num_val ) rules () order by key_3 . KEY_3 NUM_VAL ---------.from t model DIMENSION BY ( KEY * 10 AS KEY_3 ) measures ( num_val ) rules () order by key_3 . num_val from . if your DIMENSION BY columns do not give you a unique key for your result set you will get an error. select group_2.------10 100 20 200 30 300 40 50 300 60 100 70 100 80 90 200 100 800 select KEY_3.

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

KEY NUM_VAL NUM_VAL_2 DATE_VAL_2 NOTE -----.---------. date_val_2.------. SYSDATE AS DATE_VAL_2 . date_val. KEY -----1 2 3 4 5 6 7 8 9 10 DATE_VAL NUM_VAL ---------.7 8 9 10 100 200 800 If we want to include more measure columns we do it like this. note from t model dimension by ( key ) MEASURES ( num_val .------------ . num_val from t model dimension by ( key ) MEASURES ( DATE_VAL. select key.------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.---------. 'A BRIEF NOTE' AS NOTE ) rules( ) order by key . num_val. NUM_VAL ) rules () order by key . num_val_2. NUM_VAL * 10 AS NUM_VAL_2 . select key.

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 .

MAX( VAL ) MIN( VAL ) COUNT( * ) COUNT( VAL ) COUNT( DISTINCT VAL ) from t1 group by group_key order by group_key .---------. min_val . You can use the NVL function in the argument to an aggregate function to substitute a value for a null. 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 . but returns either a number or zero. Ignoring Nulls According to the SQL Reference Manual section on Aggregate Functions: All aggregate functions except COUNT(*) and GROUPING ignore nulls. select group_key . like these. then the function returns null.Nulls Nulls and Aggregate Functions This tutorial demonstrates how aggregate functions deal with null values. and COUNT will return values that for the most part ignore nulls.---------. COUNT never returns null.-------------.-----------------Group-1 (null) (null) 2 0 0 . Techniques for generating results that ignore nulls and results that include nulls are highlighted. count_val . max_val .---------. count_distinct_val GROUP_KEY MAX_VAL MIN_VAL COUNT_ALL_ROWS COUNT_VAL COUNT_DISTINCT_VAL ---------. or contains only rows with nulls as arguments to the aggregate function. MIN . For all the remaining aggregate functions. This means that. if the data set contains no rows. count_all_rows .

even though Group-2 contains null VAL values and Group-3 does not. 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. which we can use as follows to give us MAX and MIN results that include nulls. and COUNT(DISTINCT . MIN( VAL ) KEEP ( DENSE_RANK FIRST ORDER BY VAL ) min_val from t1 group by group_key order by group_key . select group_key.. The other two versions of COUNT() ignored null values.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. For aggregate functions like MAX. and SUM including nulls in the calculation is of little practical use..) however. MAX( VAL ) KEEP ( DENSE_RANK LAST ORDER BY VAL ) max_val . GROUP_KEY MAX_VAL MIN_VAL . Note also that only COUNT_ALL_ROWS returned a count that included null values. we sometimes need results that take nulls into account. Two of them are the FIRST and LAST functions. Fortunately for us there are aggregate functions in addition to COUNT(*) and GROUPING which do not ignore nulls. 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. Including Nulls For mathematical aggregate functions like AVG. For example. MEDIAN. MIN.

-------------------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. Fortunately.0. COUNT( DISTINCT VAL ) + MAX( NVL2(VAL.1) ) count_distinct_val_2 from t1 group by group_key order by group_key .---------(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. For example. GROUP_KEY COUNT_DISTINCT_VAL_1 COUNT_DISTINCT_VAL_2 ---------. DENSE_RANK and RANK Two more aggregate functions where including nulls in the calculation may be necessary are the DENSE_RANK and RANK functions.-------------.-------------------. If the VAL column contained values whose DUMP output is truncated then the results can be incorrect.---------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. as with FIRST and LAST. select group_key . select group_key. DENSE_RANK and RANK include nulls by default. 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 ---------. . COUNT( DISTINCT DUMP(VAL) ) count_distinct_val_1 .

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. 'z' ) ) max_val . GROUP_KEY ---------Group-1 Group-2 Group-3 MAX_VAL ---------z z Z MIN_VAL COUNT_DISTINCT_VAL ---------. min( NVL( VAL. NULL. . like this. 'z' ) ) count_distinct_val from t1 group by group_key order by group_key .-----------------z 1 a 2 A 2 Note how none of the columns above contain the desired results which. '~' ) ). max( NVL( VAL. DECODE( MAX( NVL( VAL. GROUP_KEY ---------Group-1 Group-2 Group-3 MAX_VAL ---------(null) (null) Z MIN_VAL COUNT_DISTINCT_VAL_1 ---------. Taking the NVL idea a little further programmers sometimes employ more complex solutions such as this one. count( distinct NVL( VAL. as you will recall. select group_key . '~'. select group_key . For example.---------------------------. A simple application of this logic can lead to trouble however. RANK( NULL ) WITHIN GROUP ( ORDER BY VAL ) null_rank_within_group from t1 group by group_key order by group_key . 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.-------------------(null) 1 a 3 A 2 A simple application of NVL clearly will not do then. You can use the NVL function in the argument to an aggregate function to substitute a value for a null. MAX( VAL ) ) max_val . 'z' ) ) min_val . are these.DENSE_RANK( NULL ) WITHIN GROUP ( ORDER BY VAL ) null_dense_rank_within_group . say we choose to substitute all null values with a 'z'.

COUNT( distinct NVL( VAL.-------------------~ 2 ~~~ 2 but using the NVL approach gives us these. 'Group-5' ) group by group_key order by group_key . 'Group-5'. given this data insert into t1 values ( 10. '~' ) ) count_distinct_val from t1 where group_key in ( 'Group-4'. '~~~' ). decode( min( nvl(val. GROUP_KEY ---------Group-1 Group-2 Group-3 MAX_VAL ---------(null) (null) Z MIN_VAL COUNT_DISTINCT_VAL ---------. '~' ) ) count_distinct_val from t1 group by group_key order by group_key . max( val ) ) max_val . null. NULL. null ). '~'. select group_key .DECODE( MIN( NVL( VAL. incorrect results. null. '~' ) ). MIN( VAL ) ) min_val . '~' ) ). ). '~'. '~'. these solutions will fail if such values are ever inserted into the table. For example. count( distinct nvl( val. 'Group-4'.-----------------(null) 1 a 3 A 2 Without some mechanism to ensure '~' and strings that sort higher than '~' never appear in VAL however. min( val ) ) min_val . '~' ) ).-----------------(null) 1 (null) 2 To avoid these gotchas simply use the non-NVL alternatives presented under "Including Nulls" above. 'Group-4'. insert into t1 values ( 10. 'Group-5'. GROUP_KEY ---------Group-4 Group-5 MAX_VAL ---------(null) ~~~ MIN_VAL COUNT_DISTINCT_VAL ---------. null insert into t1 values ( 10. the results should be GROUP_KEY ---------Group-4 Group-5 MAX_VAL ---------(null) (null) MIN_VAL COUNT_DISTINCT_VAL_1 ---------. . decode( max( nvl(val. insert into t1 values ( 10. '~' ).

this basic solution is the easiest to understand and implement. If this is the behavior you need.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). then read no further. SQL + PL/SQL These techniques work in both SQL and PL/SQL. However. select * from t where c2 = c3 . Consider a table where two of its columns can contain null values. 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. C1 --1 4 C2 ---------A (null) C3 ---------A (null) begin for r in ( select * from t ) loop .---------1 A A Row 4 is not returned because. in SQL.---------. select * from t. if you need a query that returns row 1 and row 4 then try one of the solutions in the subtopics to follow. C1 C2 C3 --. select * from t where ( C2 = C3 OR ( C2 IS NULL AND C3 IS NULL ) ). a null is not considered to be equal to or unequal to any value (including another null). OR with IS NULL While a bit cumbersome.

. If we look at the table definition for T desc t Name Null? Type ---------------------------------------------. 'x' ) . might be inserted into columns C2 or C3 some day.' ). 'x'. Row 4 contains matching values.C3 IS NULL ) ) then dbms_output.if ( R. end loop.C2 = R. or whatever value you choose to use. 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. end. / Row 1 contains matching values. 'x' ) .put_line( 'Row ' || r. C1 --1 4 C2 ---------A (null) C3 ---------A (null) One problem with this solution is that the replacement value "x".-------------------------------------C1 NUMBER C2 VARCHAR2(10) .C3 OR ( R. 'x' ) = nvl( c3. null ). 'x' ) = nvl( c3. NVL The following NVL approach is a popular one. insert into t values( 5.c1 || ' contains matching values. commit. select * from t where nvl( c2. This would cause a SELECT statement that has been working properly until that day to all of a sudden start returning the wrong answer. end if. select * from t where nvl( c2. like this.C2 IS NULL AND R.

put_line( 'Row ' || r. .C2. create function SAME( p_1 in varchar2. 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). 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. end. C3 ) = 'Y' .C3 VARCHAR2(10) we see that values in C2 and C3 can be at most 10 characters long. R. '12345678901' ) . 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 ). .c1 || ' contains matching values. C1 --1 4 C2 ---------A (null) C3 ---------A (null) begin for r in ( select * from t ) loop if SAME( R. / select * from t where SAME( C2. select * from t where nvl( c2. end loop. end.C3 ) = 'Y' then dbms_output. '12345678901' ) = nvl( c3.' ). end if.

' as result from dual where DUMP( LPAD( 'A'. DECODE One approach uses the DECODE function. select * from t where DUMP(C2) = DUMP(C3) . To return rows where two columns contain the same value we can therefore use a command like the following. RESULT -------------------------------------Oops! This row should not be returned. 4000 ) ) . SQL Only The following techniques. If the values being compared produce truncated DUMP output then the comparison can produce false positives. select 'Oops! This row should not be returned. not PL/SQL commands. Here is an example. DECODE treats two nulls as equivalent./ Row 1 contains matching values. Unlike the "=" operator. C1 --1 4 C2 ---------A (null) C3 ---------A (null) DUMP Another approach uses the DUMP function. 'N' ) = 'Y' . while more compact than the solutions presented in SQL + PL/SQL . With this approach however one function is required for comparing NUMBER values. C1 --1 4 C2 ---------A (null) C3 ---------A (null) This approach has a couple of limitations however. the output of the DUMP function is truncated at 4000 characters. . Row 4 contains matching values. one for DATE values. select * from t where DECODE( C2. 'Y'. One. 4000) ) = DUMP( LPAD( 'B'. C3. etc. one for VARCHAR2 values. unfortunately only work in SQL commands.

C2 and C3 must be the exact same datatype for the comparison to work.Two. SYS_OP_MAP_NONNULL does not work in PL/SQL. 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. 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 'Oops! This row should be returned. Comparing compatible datatypes such as VARCHAR2 and CHAR will fail to match any rows. but it is not. C1 --1 4 C2 ---------A (null) C3 ---------A (null) As with the other solutions on this page. CHAR. select * from t where SYS_OP_MAP_NONNULL( C2 ) = SYS_OP_MAP_NONNULL( C3 ) . column varchar2_val format a30 column char_val format a30 fold_before select dump( c2 ) as varchar2_val .' as result from t where dump( c2 ) = dump( 'A' ) . dump( 'A' ) as char_val from t where c1 = 1 . begin for r in ( select * from t ) loop . no rows selected Examining the output of DUMP shows why this occurs.

. 4. c2 varchar2(10) . .put_line( 'Row ' || r. ). their behavior or availability can change at any time making them a risky thing to include in your code. column 5: PL/SQL: Statement ignored It also has a length limitation. c3 varchar2(10) ). 'A' 'A' null null . insert insert insert insert into into into into t t t t values( values( values( values( 1.C2 ) = SYS_OP_MAP_NONNULL( R.' ). ). ). . 'A' 'B' 'A' null ). 3.C2 ) = SYS_OP_MAP_NONNULL( R. 4000 ) ) = SYS_OP_MAP_NONNULL( LPAD( 'B'. They also make support and maintenance harder for others who need to work with your code and are not familiar with the feature.C3 ) ) then dbms_output. 4000 ) ) . used by the examples in this section. end loop. / if ( SYS_OP_MAP_NONNULL( R. end. Be sure to read Using SQL Snippets ™ before executing any of these setup steps.if ( SYS_OP_MAP_NONNULL( R. end if.c1 || ' contains matching values. data. select * from dual where SYS_OP_MAP_NONNULL( LPAD( 'A'.C3 ) ) * ERROR at line 6: ORA-06550: line 6. 2. etc. column 10: PLS-00201: identifier 'SYS_OP_MAP_NONNULL' must be declared ORA-06550: line 6. . Setup Run the code on this page in SQL*Plus to create the sample tables. create table t ( c1 number . . select * * ERROR at line 1: ORA-01706: user function result value was too large While undocumented features such as this one are compelling.

Be sure to read Using SQL Snippets ™ before executing any of these setup steps. procedures. To clear session state changes (e. those made by SET. and VARIABLE commands) exit your SQL*Plus session after running these cleanup commands.commit. drop table t . drop function same . set null '(null)' set numformat 99 set serveroutput on Cleanup Run the code on this page to drop the sample tables.g. exit . COLUMN. etc. created in earlier parts of this section.

select day_of_week.---------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. For example.---------1 100 3 300 4 400 5 500 and you want a report that looks like this DAY_OF_WEEK VAL ----------. having a way to create a series of integers greatly simplifies certain queries. DAY_OF_WEEK VAL ----------. if your data looks like this: select * from t . t. . 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.val from days_of_the_week d left outer join t using ( day_of_week ) order by day_of_week .Integer Series Generators Sometimes.

end.. Fortunately there are flexible. Integer Table Method This tutorial demonstrates how to generate a series of integers using a generic integer table. end loop. The feature graph below will help you decide which method is best for you.Occassionally however. 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. or for a system you do not have CREATE TABLE privileges on. . / PL/SQL procedure successfully completed.Small Numbers and Performance Comparison . To load the table a simple loop like the following will do the trick. you may need a different set of integers just for one specific query. 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 . generic techniques for generating integers. The tutorials in this section demonstrate a few of them. Other techniques are discussed in the topics listed in the menu to the left. Since this table only has a single indexed column we specified organization index to make this an Index-Organized table and save storage space. commit.Large Numbers pages. for adhoc reports. 10 loop insert into integers values ( i ). In these cases it may be impractical or impossible to create a dedicated table that meets your needs. You can create such a table like this. begin for i in -5 . Table created. One of the most straightforward ways to generate a series of integers is by adding a generic integer table to your application. create table integers ( integer_value integer primary key ) organization index .

---------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.val from integers i left outer join t on ( i. when you need a specific series of integers you can use the INTEGERS table like this.integer_value . select i. (You can learn more about MODEL at SQL Features Tutorials: MODEL .We used -5 and 10 as the limits of our series in this example. INTEGER_VALUE -------------5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10 Later. select * from integers .integer_value between 0 and 6 order by i. t. In practice you would choose limits that anticipate the smallest and largest integers you will ever need.day_of_week ) where i.integer_value = t.integer_value as day_of_week . DAY_OF_WEEK VAL ----------.

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

select rownum from all_objects where rownum <= 10 . 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. 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. ( 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. 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.1 as day_of_week . t.val from ( select rownum . like this. Other techniques are discussed in the tutorials listed in the menu to the left. select day_of_week . I.e.( integer_value[ for key from :v_first_key to :v_last_key INCREMENT :v_increment ] = cv(key) ) . The data dictionary view ALL_OBJECTS is a popular choice for this method. 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.

Applying the technique to a table with two rows.from all_objects where rownum <= 7 ) i left outer join t using ( day_of_week ) order by day_of_week . "a" and "b". yields some insight into how such queries work. With this technique you can generate a series of integers starting at "1" using a query like this. 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. break on level duplicates skip 1 column path format a10 .---------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?". select level from dual connect by level <= 10 . Other techniques are discussed in the tutorials listed in the menu to the left. DAY_OF_WEEK VAL ----------.

The effect also proves useful in situations where more than one integer series is required from a single query. sys_connect_by_path( key. '/' ) as path. Note there is some debate about whether queries without PRIOR in the CONNECT BY clause are legal or not. . negative.select from connect by order by LEVEL ---------1 1 2 2 2 2 3 3 3 3 3 3 3 3 level. This is discussed further in the "Gotchas" section below. See Multiple Integer Series: CONNECT BY LEVEL Method for more details and an example. clear breaks variable v_total_rows number execute :v_total_rows := 0 select level from dual connect by level <= :v_total_rows . This effect may be useful where an exponentially increasing number of output rows is required. 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. or null however. If the number of rows is set with a bind variable whose value can be 0. It always generates at least one row in these cases. the technique may not work as expected. key t4 level <= 3 level. 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.

A simple WHERE clause fixes this behaviour. execute :v_total_rows := null PL/SQL procedure successfully completed.LEVEL ---------1 execute :v_total_rows := -5 PL/SQL procedure successfully completed. / LEVEL ---------1 1 row selected. / no rows selected execute :v_total_rows := null . no rows selected execute :v_total_rows := -5 PL/SQL procedure successfully completed. select from WHERE connect level dual :V_TOTAL_ROWS >= 1 by level <= :v_total_rows . / LEVEL ---------1 1 row selected. execute :v_total_rows := 0 PL/SQL procedure successfully completed.

/ no rows selected execute :v_total_rows := 3 PL/SQL procedure successfully completed. That is the Question . t.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 ----------.val from ( select level .PL/SQL procedure successfully completed.---------0 1 100 2 3 300 4 400 5 500 6 Gotchas To Use PRIOR or Not to Use PRIOR. 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 . / LEVEL ---------1 2 3 3 rows selected.

one expression in condition must be qualified with the PRIOR operator to refer to the parent row. select level from dual connect by level <= 10 AND PRIOR DUMMY = DUMMY . do not produce the desired 10 rows of output (as tested in Oracle 10g). The fact the CONNECT BY clause works without error in Oracle 10g and some 9i versions somewhat supports this view. LEVEL ---------1 2 3 4 5 6 . Note that the following queries. 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. as dictated by this statement in the SQL Reference Manual "in a hierarchical query.VALUE IS NOT NULL .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.Volder).2) Others argue that this statement is a documentation bug. select level from dual connect by level <= 10 and PRIOR DBMS_RANDOM. ERROR: ORA-01436: CONNECT BY loop in user data select level from dual connect by level <= 10 AND PRIOR 1 = 1 . seemingly equivalent to the CONNECT BY LEVEL <= 10 solution. 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" .Oracle® Database SQL Reference 10g Release 2 (10." -.

Laurent Schneider and other posters at these threads. Until there is a definitive answer. Acknowledgements Mikito Harakiri.Weird Results (bug?) Ask Tom "how to display selective record twice in the query?". Tom Kyte. Ed In Oracle 9i. select * from (select level from dual connect by level < 10) . be aware there is a risk the technique may not work in future versions. may help (I have not tested this). • • • • Ask Tom "Can there be an infinite DUAL?" .minor simplification CONNECT BY Generator Rules | Ask Mr.Weird results Ask Tom "Can there be an infinite DUAL?" .7 8 9 10 The jury is still out on using CONNECT BY LEVEL to generate integers. if you try the CONNECT BY LEVEL technique and get a single row when expecting muliple rows. LEVEL ---------1 2 3 4 5 6 7 8 9 9 rows selected. like this select level from dual connect by level < 10 . I have not tested these myself but here are some posts that describe problems. as in this snippet.2. have been reported in Oracle versions earlier than 10. or variations of it. LEVEL ---------1 putting the query in an inline view. Issues Issues with this technique. • • Ask Tom "how to display selective record twice in the query?" Ask Tom "Can there be an infinite DUAL?" .

Other techniques are discussed in the tutorials listed in the menu to the left.CUBE Method This tutorial demonstrates how to generate a series of integers using the CUBE clause of the SELECT statement. ROWNUM ---------1 2 3 4 5 6 7 8 To return 16 rows (2^4): ROWNUM ---------1 2 3 4 select rownum 5 from 6 ( 7 select 1 8 from dual 9 group by cube( 1. 4 ) 10 ) . 2. ROWNUM ---------1 2 3 4 To return 8 rows (2^3): select rownum from ( select 1 from dual group by cube( 1. 2. To return 4 rows (2^2): select rownum from ( select 1 from dual group by cube( 1. 2 ) ) . 11 12 13 14 15 16 To return 9 rows: select rownum from ( select 1 ROWNUM ---------1 2 . 3. 3 ) ) . Here are some examples that generate a series of integers using CUBE.

otherwise you will not get the correct number of rows. 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). 3 4 5 6 7 8 9 We can apply this technique to our day of the week scenario with this query. DAY_OF_WEEK VAL ----------. 3 ) ) where rownum <= 7 ) select day_of_week . 2. 3.1 as day_of_week from ( select 1 from dual group by cube( 1.val from days_of_the_week left outer join t using ( day_of_week ) order by day_of_week . select rownum as integer_value from ( select 1 from t2 group by cube ( 1.Explained. t. 4 ) ) where rownum <= 9 . 2. 2 ) -.will only generate 2^2 rows . as in this example which attempts to generate 7 integers but only succeeds in generating 4.from dual group by cube( 1.---------0 1 100 2 3 300 4 400 5 500 6 For more details about how this method works see CUBE . with days_of_the_week as ( select rownum .

select rownum from t2 group by cube( 1. 6. 2. . cube( 1..-----. 3. C2 . c2. 2 ) . .---------x y 42 . C1 C2 C3 -----. INTEGER_VALUE ------------1 2 3 4 Inline View Attempting to use rownum without the inline view will cause errors or incorrect results.-----. 7.---------x y 42 C1 C2 SUM_C3 -----. sum( c3 ) as sum_c3 from t2 GROUP BY C1. 2 ) . and then turn it into an integer series generator for the values 1 to 4. select c1.) where rownum <= 7 -. (Read the Oracle manual page on the CUBE grouping operation first if you are not already familiar with this feature. or 1 in this query. transform it into a query that uses CUBE the traditional way.) set null "(null)" select * from t2 . not 5. 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. select rownum * ERROR at line 1: ORA-00979: not a GROUP BY expression select rownum from t2 group by rownum..can only be <= 4.

etc. c2. 1 ---------1 1 1 1 1 select 1 ---------from t2 1 group by CUBE ( 1.---------(null) 42 y 42 (null) 42 y 42 C2 ANY_LITERAL -----.7 because it is . Two literals will give you four rows (2^2).3.'b' and still get the same number of rows.select c1.see Note 1 1 . select rownum as integer_value from ( select 1 from t2 group by cube ( 1.2. c2. I like to use arguments like 1. C1 -----(null) (null) x x C1 -----(null) (null) x x C2 SUM_C3 -----. 2 ) ). sum( c3 ) as sum_c3 from t2 group by CUBE( c1. 2 ) ) where ROWNUM <= 3 .6. c2 ).----------(null) 1 y 1 (null) 1 y 1 select c1. three literals will give you eight rows (2^3). You could use arguments like 1. 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. 1 AS ANY_LITERAL from t2 group by cube( c1. four literals will give you sixteen rows (2^4). 1 or 'a'.5. 1 1 select ROWNUM AS INTEGER_VALUE from ( select 1 from t2 group by cube ( 1. c2 ). 2 ) -. SELECT 1 from t2 group by cube ( c1.4. c2 ) . The important part is how many literals you include.

6. Other techniques are discussed in the tutorials listed in the menu to the left. If you do not have privileges to create a type like this see Setup . select column_value from table( integer_table_type( 1.1.10 ) ) .5. like 10 or 20.4. COLUMN_VALUE -----------1 1 4 4 .1.1.1. you can use a simple query like this one.8. select column_value from table( integer_table_type( 1.1. desc integer_table_type integer_table_type TABLE OF NUMBER(38) The Solution If you need a manageable number of integers.easier to tell there are 7 arguments (which produce 2^7 rows) with this approach than with an argument list like 1.3.9.10 ) ) .1.4.4.7.8.1.4. 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. 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.Note 1.2. Type Constructor Expression Method This tutorial demonstrates how to generate a series or set of integers using Type Constructor Expressions for collection types.

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.
desc integer_table_type integer_table_type TABLE OF NUMBER(38)

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.
desc integer_table_type integer_table_type TABLE OF NUMBER(38)

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

end. return.7) ) . t. select i. select * from table( integer_series(-5.6) ) i left outer join t on ( i.loop pipe row(i).day_of_week ) order by i.val from table( integer_series(0.---------0 1 100 2 3 300 4 400 5 500 6 .column_value as day_of_week . 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.column_value .column_value = t. end loop. here is how you use the INTEGER_SERIES function. / The Solution Now that we have our prerequisites in place. DAY_OF_WEEK VAL ----------.

072 2.Performance Comparison .608 65.076 session uga memory 0 0 0 0 0 65.464 65.464 See Statistics Descriptions for a description of each metric.144 327.081 2. See the log file from these tests for more details. • 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.-----------.071 2.072 131.684 2. Statistics The following table shows database statistics where values for one method differ by more than 100 from another method.071 2.684 2.684 2. Integer ROWNUM CONNECT BY Type Type Constructor Pipelined METRIC_NAME Table MODEL + Big Table LEVEL CUBE Constructor + Cartesian Product Function ---------------------------------------.680 redo size 2.144 262. .072 327.536 131.680 session pga memory 196.684 2.071 2.464 65.-----------.072 196.144 262.744 2.144 262.-----------.684 2.144 262.608 65.Small Numbers The following tables show performance metrics for one run each of the eight integer series generation techniques described in the preceeding tutorials.144 262.199 2. Latch Gets The following table shows total latch gets for each method.640 2.071 2.-----------.-----------------------------Elapsed Time (1/100 sec) 3 3 3 2 5 262 3 4 session pga memory max 262.----------------------.536 131.144 262.684 sorts (rows) 2.

-----------.-----------.-----------.-----------------------------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 .-----------.Integer ROWNUM CONNECT BY Type Type Constructor Pipelined METRIC_NAME Table MODEL + Big Table LEVEL CUBE Constructor + Cartesian Product Function ---------------------------------------.----------------------.

242 632 986 .-----------------------------sum 757 575 709 575 559 3.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 -----------.----------------------.-----------.-----------.-----------.dml lock allocation 1 1 1 1 FOB s.

• 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 100.031.Large Numbers The following tables show performance metrics for one run each of six integer series generation techniques described in the preceeding tutorials. Results will even differ from one set of test runs to the next on the same machine.072 0 262.-----------.144 4.533.964 session pga memory 196. Run your own tests and average the results from multiple runs before making performance decisions.680 session logical reads 6.888 6 39 72 + Big .608 65.Techniques that use a small number of latches scale better than techniques that use a large number of latches.128 262.016.616 262.680 session uga memory max 261. Integer ROWNUM CONNECT BY Type Constructor Pipelined METRIC_NAME Table MODEL Table LEVEL + Cartesian Product Function ---------------------------------------.840 6 6.-----------.----------------------.000. See the log file from these tests for more details.964 2. Statistics The following table shows database statistics where values for one method differ by more than 100 from another method.------------------.784.-----------Elapsed Time (1/100 sec) 59 349 68 67 57 544 session pga memory max 262.964 261.144 327. Warning: Results on your own systems with your own data will differ from these results.927 45 78 111 consistent gets 6.536 131. 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.879 45 6.964 4.144 327. Performance Comparison .252 261.144 2. The CUBE Method technique was excluded from this test because it failed to complete in under 10 minutes.984 261.

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 65.375 13.-----------. Integer ROWNUM CONNECT BY Type Constructor Pipelined METRIC_NAME Table MODEL Table LEVEL + Cartesian Product Function ---------------------------------------.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 20.658 29 29 29 0 6 0 0 311 309 309 0 See Statistics Descriptions for a description of each metric.376 13. Latch Gets The following table shows total latch gets for each method.946 836 13.464 37 34 34 34 65.375 13.------------------.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 13.375 13.-----------.-----------cache buffers chains 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 .927 163 196 258 session idle bit 13.822 5 6.----------------------.888 6 39 no work .consistent gets from cache 6.840 72 6.consistent read gets 6.464 6.376 13.

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.883 16.633 13.-----------sum 28.136 Techniques that use a small number of latches scale better than techniques that use a large number of latches. Results will even differ from one set of test runs to the next on the same machine.-----------.447 28.089 48.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 -----------.----------------------. .------------------.879 14. Run your own tests and average the results from multiple runs before making performance decisions. Warning: Results on your own systems with your own data will differ from these results.

---------a (null) b 0 c 1 d 2 e 3 we sometimes need queries that generate results like these KEY QTY INTEGER_VALUE --. For example. Join Method Many of the single series solutions presented earlier can be easily adapted to generate multiple series with a simple join. The subtopics in this section demonstrate various ways to accomplish this. create table integers ( integer_value integer primary key ) organization index .2. Sometimes many different integer series are required in the same query.---------. insert into insert into insert into insert into insert into commit. Here is an example of how this is done using the Integer Table Method technique. ). ). .------------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 "1.3". ).Multiple Integer Series The preceding topics showed how to generate a single series of integers. given a data table like this KEY QTY --. integers integers integers integers integers values values values values values ( ( ( ( ( 1 2 3 4 5 ).2" but the row where KEY='e' needs the series "1. ).

select key.qty ) order by key.3.integer_value <= t3. qty.4. integer_value from t3 left outer join integers on ( integers.------------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. 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. .column_value <= t3. column_value as integer_value from t3 left outer join table( integer_table_type(1.---------. integer_value .---------.5) ) integers on ( integers.2.qty ) order by key. qty. KEY QTY INTEGER_VALUE --.set null "(null)" break on key duplicates skip 1 select key.

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 . integer_value from t3 model PARTITION BY ( KEY ) dimension by ( 1 as key2 ) measures ( qty. integer_value .MODEL Method With MODEL queries there is no need to use the join technique described in Join Method. set null "(null)" break on key duplicates skip 1 select key. key2. qty.------------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. 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.---------.---------. cast( null as integer ) as integer_value ) rules ( integer_value[ FOR KEY2 FROM 1 TO QTY[1] INCREMENT 1 ] = cv(key2) ) order by key. KEY KEY2 QTY INTEGER_VALUE --.

sys_connect_by_path( key.------------. Gotchas .select key.---------. 'b' ).------------.qty order by key.qty order by key.---------.value is not null and level <= t3.---------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. sys_connect_by_path( key. integer_value . qty. level as integer_value.---------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'. integer_value . level as integer_value. '/' ) as path from t3 where qty >= 1 connect by KEY = CONNECT_BY_ROOT KEY and level <= t3. KEY QTY INTEGER_VALUE PATH --. KEY QTY INTEGER_VALUE PATH --. qty. '/' ) as path from t3 where qty >= 1 connect by KEY = PRIOR KEY and prior dbms_random.

qty. This is because it violates two restrictions documented at Hierarchical Query Operators: 1. select key. one expression in the CONNECT BY condition must be qualified by the PRIOR operator.trc file: ORA-07445: exception encountered: core dump [ACCESS_VIOLATION] [__VInfreq__msqopnws+2740] [PC:0x30E2580] [ADDR:0x2C] [UNABLE_TO_READ] [] ." 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. "You cannot specify (CONNECT_BY_ROOT) in the START WITH condition or the CONNECT BY condition.qty order by key. 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). level as integer_value from t3 start with qty >= 1 connect by KEY = CONNECT_BY_ROOT KEY and level <= t3. "In a hierarchical query.The CONNECT_BY_ROOT technique may not work without error in all cases and it may not work in Oracle versions beyond 10g. ERROR at line 1: ORA-03113: end-of-file on communication channel ERROR: ORA-03114: not connected to ORACLE From the ." 2. integer_value .

Begin Open c1. --process rows. Rec1 c1%rowtype. In order to reduce the number of these context switches we can use a feature named bulk binding. They declare a cursor. insert. Too many context switches may degrade performance dramatically. Fetch c1 into r1. One of the things i usuallly come accross is that developers usually tend to use cursor for loops to process data. Declare Cursor c1 is select column_list from table_name>. 6 loop . Bulk binding is available for select. Bulk collect is the bulk binding syntax for select statements.. 2 3 rec1 c1%rowtype. End. SQL> create table t_all_objects as select * from all_objects. Table created. delete and update statements. COUNT(*) ---------213248 SQL> declare cursor c1 is select object_name from t_all_objects.. fetch from it row by row in a loop and process the row they fetch. ---replicated a couple of times SQL> select count(*) from t_all_objects. open it. Exit when c1%notfound. Loop. 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. 4 begin 5 open c1. SQL> r 1* insert into t_all_objects select * from t_all_objects 6664 rows created. SQL> insert into t_all_objects select * from t_all_objects. End loop. 3332 rows created. Bulk binding lets us to transfer rows between the sql engine and the plsql engine as collections.Bulk Collect Executing sql statements in plsql programs causes a context switch between the plsql engine and the sql engine.

7 loop 8 fetch c1 bulk collect into rec1 limit 200. Elapsed: 00:00:44. 3 type c1_type is table of c1%rowtype.count loop 10 null. 12 exit when c1%notfound..07 . 13 end loop. 4 rec1 c1_type. 3 type c1_type is table of c1%rowtype.32 As can be clearly seen. 14 15 16 end.rec1. end loop.7 8 9 10 11 12 13 14 fetch c1 into rec1. exit when c1%notfound. process those rows and fetch again. / PL/SQL procedure successfully completed. 7 8 fetch c1 bulk collect into rec1. 5 begin 6 open c1. Elapsed: 00:00:05. 17 / PL/SQL procedure successfully completed. 9 10 11 end. 9 for i in 1. bulk collecting the rows shows a huge performance improvement over fetching row by row. 5 begin 6 open c1. we can limit the number of rows to bulk collect.75 SQL> declare 2 cursor c1 is select object_name from t_all_objects. Otherwise process memory gets bigger and bigger as you fetch the rows. 12 / PL/SQL procedure successfully completed. SQL> declare 2 cursor c1 is select object_name from t_all_objects. end. The above method (which fetched all the rows) may not be applicable to all cases. 11 end loop. When there are many rows to process. null. 4 rec1 c1_type. Elapsed: 00:00:04.

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.-----Parse 1 Execute 1 Fetch 0 ------.55 11.-------.---------. END LOOP. However.---------.amount_local. If you have the luxury of time.45 0 0 1 rows -------0 1 0 -------1 . you can test your code both with and without bulk binds.40 0 0 0 0. r.-------10.000 rows from a cursor or performs that many similar UPDATE statements will most likely benefit from bulk binds. No universal rule exists to dictate when bulk binds are worthwhile and when they are not.-------0.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.05 0. The same goes for a program that issues five or six UPDATE statements.04 0 0 1 10. call count ------. We'll look at TKPROF reports that demonstrate the impact bulk binds can have.-----total 2 cpu elapsed disk query current -------. Running both versions of the code through SQL trace and TKPROF will yield reports from which you may derive a wealth of information. A PL/SQL program that reads a dozen rows from a cursor will probably see no noticeable benefit from bulk binds. the cost of adding a few lines of code is so slight that I would lean toward using bulk binds when in doubt.-------. However. v_amount_usd NUMBER.00 0 0 0 -------. so you need to ask yourself if the improved runtime performance will justify the expense. 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.order_id. amount_local /* no bulk bind */ FROM open_orders.-------.00 0.-------.60 11. a program that reads 1. END.currency_code). COMMIT. BEGIN FOR r IN c_orders LOOP v_amount_usd := currency_convert (r. UPDATE open_orders /* no bulk bind */ SET amount_usd = v_amount_usd WHERE order_id = r. currency_code.

) The PL/SQL engine used 10. TYPE t_char_array IS TABLE OF VARCHAR2(10) INDEX BY BINARY_INTEGER.393 logical reads and 1.---------. v_currency_codes t_char_array. currency_code. BEGIN OPEN c_orders. v_row_count NUMBER := 0.08 1.32 1 60576 31022 0. TYPE t_num_array IS TABLE OF NUMBER INDEX BY BINARY_INTEGER. please recognize it is for illustrative purposes only.-------. Now consider the following excerpts from another TKPROF report: ************************************************************************ DECLARE CURSOR c_orders IS SELECT order_id.19 7.-----total 30287 cpu elapsed disk query current -------.08 1.00 0. LOOP FETCH c_orders BULK COLLECT INTO v_order_ids.---------. this is a very simple program that does not use bulk binds.19 7.-------.287 fetch calls against the cursor.-------1. v_order_ids t_num_array.19 CPU seconds.-----total 30289 cpu elapsed disk query current -------.00 0. amount_local /* no bulk bind */ FROM open_orders call count ------. The UPDATE statement was executed 30. using 7.-------. (The code borders on being silly.-------.55 CPU seconds to run this code (this figure does not include CPU time used by the SQL engine).286 times.00 0 0 0 7. There were 30.00 0. v_amounts_local t_num_array. EXIT WHEN v_row_count = c_orders%ROWCOUNT.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 ------. requiring 30.---------.-------.-------.00 0 0 0 1.00 0 0 0 0.-------0.************************************************************************ SELECT order_id.08 CPU seconds.-------. v_amounts_usd t_num_array.10 0 30393 0 -------.---------. v_currency_codes.00 0 0 0 -------.-------.-------7.-------0. . amount_local /* bulk bind */ FROM open_orders. currency_code.33 1 60576 31022 rows -------0 30286 0 -------30286 As you can see.-----Parse 1 Execute 30286 Fetch 0 ------. v_amounts_local LIMIT 100.00 0.-----Parse 1 Execute 1 Fetch 30287 ------.

---------.-------3. COMMIT.-------. bringing logical reads down from 30. amount_local /* bulk bind */ FROM open_orders call count ------. END LOOP.00 0. The UPDATE statement was executed only 303 times instead of 30.-------0.48 seconds.00 0 0 0 3.-------.66 0 0 0 rows -------0 1 0 -------1 ************************************************************************ SELECT order_id. v_currency_codes(i)).55 seconds to 0.-------.-----total 304 cpu elapsed disk query current -------.63 0.-------.00 0 0 0 -------.62 0 0 0 0.-------.60 0. CLOSE c_orders. . FOR i IN 1.-----Parse 1 Execute 1 Fetch 0 ------.00 0.38 0 30895 31021 0.00 0.-------. call count ------.-------0.-------. FORALL i IN 1. but works with data 100 rows at a time instead of one row at a time.03 0.815 and CPU time down from 1.19 seconds to 3.-------0.00 0.60 seconds.03 0 0 0 0.-----total 2 cpu elapsed disk query current -------.00 0 0 0 -------. END.---------.-------0.-----Parse 1 Execute 303 Fetch 0 ------.286..-------.75 8.v_order_ids. We can see that CPU time used by the PL/SQL engine has reduced from 10.59 0 4815 0 -------.---------.-----Parse 1 Execute 1 Fetch 303 ------.v_row_count := c_orders%ROWCOUNT.count LOOP v_amounts_usd(i) := currency_convert (v_amounts_local(i).59 0 4815 0 rows -------0 0 30286 -------30286 ************************************************************************ UPDATE open_orders /* bulk bind */ SET amount_usd = :b1 WHERE order_id = :b2 call count ------.75 8..287.75 seconds. currency_code.-------.count UPDATE open_orders /* bulk bind */ SET amount_usd = v_amounts_usd(i) WHERE order_id = v_order_ids(i).08 seconds to 0.38 0 30895 31021 rows -------0 30286 0 -------30286 This code uses bulk binds to do the same thing as the first code sample.---------.-----total 305 cpu elapsed disk query current -------.---------.00 0.393 to 4.00 0 0 0 0.48 0.-------.---------.-------. reducing CPU time from 7.48 0.00 0 0 0 0. END LOOP.-------.v_order_ids. There were only 303 fetch calls against the cursor instead of 30.-------0.

. Conclusion Bulk binds allow PL/SQL programs to interact more efficiently with the SQL engine built into Oracle. execute. Thus using bulk binds in your PL/SQL programs can certainly be worth the effort. the benefit is not truly as rosy as it appears in these TKPROF reports. Just remember that SQL trace can inflate the perceived benefit. Since bulk binds reduce the number of SQL calls. The Oracle Call Interface has supported array processing for 15 years or more. the savings were about 25% when SQL trace was not enabled.In this example it would appear that bulk binds were definitely worthwhile – CPU time was reduced by about 75%. While these TKPROF reports suggest that in this example bulk binds shaved about 50% off of the elapsed time. and can offer significant performance improvements for certain types of programs. and the increased efficiency it brings is well understood. elapsed time by 50%. It is nice to see this benefit available to PL/SQL programmers as well. SQL trace adds much less overhead to code that uses bulk binds. PL/SQL bulk binds are not hard to implement. The SQL trace facility imparts an overhead that is proportional to the number of parse. enabling your PL/SQL programs to use less CPU time and run faster. and fetch calls to the SQL engine. This is still a significant savings. Although bulk binds are indeed beneficial here. and logical reads by 50%.

Sign up to vote on this title
UsefulNot useful