Category Archives: oracle

dbms_random in parallel

Posted on by Posted in documentation, oracle, parallel Leave a comment

The documentation for dbms_random states:

It will automatically initialize with the date, user ID, and process ID if no explicit initialization is performed.

This phrase does not answer the question, which “process id” is going to be used for initialization in case of parallel execution. That’s why I decided to give a vivid example that shows independence of “dbms_random” generator from “process id” of slave, that is generating identical values in parallel:

with
 t  as ( select/*+ materialize */ level n from dual connect by level<=4000)
,t1 as (
         select--+ materialize parallel(t 4)
            dbms_random.string('x',4)
            ||';'
            ||(select sid||';'||process||';'||pid
               from v$session, v$process
               where sid=sys_context('USERENV','SID')
                 and PADDR=ADDR
                 and n>0
              ) f
         from t
)
,t2 as (
         select
            t1.f
           ,count(*) over(partition by regexp_substr(f,'^[^;]+')) cnt
         from t1
)
select f
      ,regexp_substr(f,'[^;]+') rnd
      ,regexp_substr(f,'[^;]+',1,2) sid
      ,regexp_substr(f,'[^;]+',1,3) process
      ,regexp_substr(f,'[^;]+',1,4) pid
from t2 
where cnt>1
order by f

Result:


F RND SID PROCESS PID
AARV;130;5472;30 AARV 130 5472 30
AARV;68;2228;29 AARV 68 2228 29
AC2R;130;5472;30 AC2R 130 5472 30
AC2R;68;2228;29 AC2R 68 2228 29
AC8O;130;5472;30 AC8O 130 5472 30
AC8O;68;2228;29 AC8O 68 2228 29
AKVZ;130;5472;30 AKVZ 130 5472 30
AKVZ;68;2228;29 AKVZ 68 2228 29
ALTQ;130;5472;30 ALTQ 130 5472 30
ALTQ;68;2228;29 ALTQ 68 2228 29
… … … … …

About the performance of exception handling

Posted on by Posted in oracle, query optimizing 1 Comment

This article is about a well-known fact about the poor performance of exception handling.

Yes, the exception handling is rather slow, however, it is not necessary to try to avoid exceptions whenever possible, and by any means. For example, I often see that people are trying to avoid them even in cases of search by primary key where probability of receiving “no_data_found” is minimal.
In general, we should analyze the possible frequency of exceptions and “overhead”, which is added by the chosen way with exception handlers.

Let me explain this with an example, which I mentioned earlier: suppose we have a code that returns a field from the table by “pk” and it returns “null” in case there is no such entry.
Test table:

create table t_test(a primary key, b)
as
select level,level from dual connect by level<=1e5;

Lets create a standard function for tests:

create or replace function f1(p in number) return number
as
  res number;
begin
  select/*+ F1 */ b into res
  from t_test t
  where t.a=p;
  return res;
exception when no_data_found then
  return null;
end;

The most common options to avoid the exceptions mechanism in these cases are the followings:

Variant 1 

create or replace function f2(p in number) return number
as
begin
  for rec in (select/*+ F2 */ b from t_test t where t.a=p) loop
    return rec.b;
  end loop;
  return null;
end;

[collapse]

By the way, don’t use this in case your cursor can not have more than one entry:
Spoiler 

create or replace function f2(p in number) return number
as
  res number;
begin
  for rec in (select/*+ F2 */ b from t_test t where t.a=p) loop
    res:=rec.b;
  end loop;
  return res;
end;

[collapse]

Otherwise, there will be attempts of the second iteration, which you can observe in the profiler.
Variant 2 

create or replace function f3(p in number) return number
as
  res number;
begin
  select/*+ F3 */ min(b) into res
  from t_test t
  where t.a=p;
  return res;
end;

[collapse]

Here I would like to propose my own option for this:
Spoiler 

create or replace function f4(p in number) return number
as
  res number;
begin
  select/*+ F4 */ 
    (select b from t_test t where t.a=p)
    into res
  from dual;
  return res;
end;

[collapse]

And now lets carry out a basic test by executing these functions for the test table:

declare
  v       integer;
  v_start integer:= 1;
  v_end   integer:= 100000;
    l_timer integer := dbms_utility.get_time;
    procedure print(msg varchar2) is
    begin
      dbms_output.put_line(to_char((dbms_utility.get_time-l_timer)/100,'9990.00')||' '||msg);
      l_timer:=dbms_utility.get_time;
    end;
     
begin
  print('start');
  for i in v_start..v_end loop
    v:=f1(i);
  end loop;
  print('1');
  for i in v_start..v_end loop
    v:=f2(i);
  end loop;
  print('2');
  for i in v_start..v_end loop
    v:=f3(i);
  end loop;
  print('3');
  for i in v_start..v_end loop
    v:=f4(i);
  end loop;
  print('4');
end;

As a result, we get the following ratio:

Variant Time(sec)
Variant 1(with exception) 3.03
Variant 2(with cycle) 3.62
Variant 3(with min) 3.34
Variant 4(scalar subquery) 3.10

As you can see, the original query is the fastest in case the exceptions are not called! Lets now check it with different percents of exceptions: exceptions will be for queries с i<=0, the total number of calls will be 100001, I will change v_start and v_end in pairs: (-5000, 95000), (10000, 90000), (-50000, 50000), (-90000, 10000): 

declare
  v       integer;
  v_start integer:=-50000;
  v_end   integer:= 50000;
    l_timer integer := dbms_utility.get_time;
    procedure print(msg varchar2) is
    begin
      dbms_output.put_line(to_char((dbms_utility.get_time-l_timer)/100,'9990.00')||' '||msg);
      l_timer:=dbms_utility.get_time;
    end;
     
begin
  print('start');
  for i in v_start..v_end loop
    v:=f1(i);
  end loop;
  print('1');
  for i in v_start..v_end loop
    v:=f2(i);
  end loop;
  print('2');
  for i in v_start..v_end loop
    v:=f3(i);
  end loop;
  print('3');
  for i in v_start..v_end loop
    v:=f4(i);
  end loop;
  print('4');
end;
/

Summary table of multiple comparisons:

Variant 0% ~5% ~10% ~50% ~90%
Variant 1(with exception) 3.04 3.12 3.16 3.82 4.51
Variant 2(with cycle) 3.18 3.21 3.20 3.51 3.85
Variant 3(with min) 3.37 3.34 3.29 3.25 3.18
Variant 4(scalar subquery) 3.12 3.06 3.03 2.98 2.94

What conclusions can be drawn from this:

  • As you can see, 5% of exceptions is a kind of turning point for this table, when the standard option with exception becomes less effective than the option with subquery (by about ~4.5% to be precise), and about ~10% than the remaining two.
  • Options with “min” and a cycle as a whole are worse than the option with a subquery.
  • Options with subquery and “min” becomes faster when the number of “empty” queries increases.

A funny fact about collect

Posted on by Posted in collect, oracle, parallel, query optimizing 1 Comment

Many people know that oracle creates domain types on its own when necessary, for example when using a type declared in a package (before 11g they could be observed in dba_objects with the name like ‘PLSQL%’).

Fact #1

It acts in the same way when calling an aggregate function “collect”.

-- Firstly we check if there are such types
DB11G/XTENDER> select t.type_name,t.type_name,t.typecode 
 2 from dba_types t 
 3 where t.type_name like 'SYSTP%';
 
no rows selected
 
 
-- Executing a query with collect
DB11G/XTENDER> select collect(level) from dual connect by level<=10;
 
COLLECT(LEVEL)
-------------------------------------------------------------------------
 
SYSTPZvGjVQTySRSjYVlHXyEE2Q==(1, 2, 3, 4, 5, 6, 7, 8, 9, 10)
 
1 row selected.
 
 
-- We check it again and observe that a new type SYSTP% has appeared
DB11G/XTENDER> select t.type_name,t.type_name,t.typecode 
 2 from dba_types t 
 3 where t.type_name like 'SYSTP%';
 
TYPE_NAME                      TYPE_NAME                      TYPECODE
------------------------------ ------------------------------ ------------
SYSTPZvGjVQTySRSjYVlHXyEE2Q==  SYSTPZvGjVQTySRSjYVlHXyEE2Q==  COLLECTION

And also we select data on it from sys.obj$. We will need it later:

DB11G/XTENDER> select obj#,type#,ctime,mtime,stime,status
  2  from sys.obj$ o$
  3  where o$.name = 'SYSTPZvGjVQTySRSjYVlHXyEE2Q==';
 
  OBJ#  TYPE# CTIME               MTIME               STIME              
------ ------ ------------------- ------------------- -------------------
103600     10 28.04.2012 01:02:35 28.04.2012 01:02:35 28.04.2012 01:02:35

This is a natural result as you can not return information to the client without having described it.

Fact #2

Now as we know that the type is created, it is interesting to know what will happen to this type: whether it will be removed either after fetch or after disconnecting the client? For example, the domain type was automatically dropped when the packet type was removed. Are we going to have a similar automatic remove here?

According to Bug 4033868: COLLECT FUNCTION LEAVES TEMPORARY SYS TYPES BEHIND this bug is fixed in “11.0”, but I am testing it on 11.2.0.1 and a basic check after disconnection showed, that this type exists until instance restart. In fact, it stays there even after that, but it is not displayed in dba_objects.

I will drop it manually, so I won’t have to restart the instance once again. This is absolutely similar to how oracle “removes” this type in 11.2:

DB11G/XTENDER> drop type "SYSTPZvGjVQTySRSjYVlHXyEE2Q==";
 
Type dropped.

And now lets check it:

DB11G/XTENDER> select * from dba_types
   where type_name='SYSTPZvGjVQTySRSjYVlHXyEE2Q==';
 
no rows selected

It seems like we have removed it, but what if we look in the sys.obj$:

DB11G/XTENDER> select obj#,type#,ctime,mtime,stime,status
  2  from sys.obj$ o$
  3  where o$.name = 'SYSTPZvGjVQTySRSjYVlHXyEE2Q==';
 
  OBJ#  TYPE# CTIME               MTIME               STIME              
------ ------ ------------------- ------------------- -------------------
103600     10 28.04.2012 01:02:35 28.04.2012 01:40:37 31.12.4712 23:59:59

As you can see, the object is still there, but with type#=10 and with “stime” equal to the last date of year 4712, and before that it was type#=13 and stime=mtime=ctime, and in 10.2 after manual drop of this type no entries were left. I will explain the correspondence of the fields from “sys.obj$” and “dba_objects” to clarify this: obj# – object_id, type# ~ type code, ctime,mtime,stime – created, last_ddl_time, timestamp respectively. By the “dba_objects” view code we will see that type# = 10 is supposedly “NON-EXISTENT” and displaying it is not necessary.

and (o.type# not in (1  /* INDEX - handled below */,
                      10 /* NON-EXISTENT */)

And the date being set to 31.12.4712 23:59:59 indicates its irrelevance – this future is too distant 🙂

Fact #3

“Collect” is a pretty buggy thing in general, as I have repeatedly noticed, and Metalink has information about lots of bugs related to “collect” (for example, Bug 11906197 “Parallel query with COLLECT function fails with ORA-7445/ORA-600.”,”Bug 8912282: COLLECT+UNIQUE+ORDER DOES NOT REMOVE DUPLICATES”, “Bug 6145841: ORA-600[KOLOGSF2] ON CAST(COLLECT(..)) CALL”,”Bug 11802848: CAST/COLLECT DOES NOT WORK IN VERSION 11.2.0.2 WITH TYPE SYS.DBMS_DEBUG_VC2COLL”, “Bug 6996176: SELECT COLLECT DISTINCT GROUP BY STATEMENT RETURNS DUPLICATE VALUES”)

Test table 
create table test_parallel parallel 8 as
select mod(level,8) a, level b 
from dual 
connect by level<=1000;
create index IX_TEST_PARALLEL on TEST_PARALLEL (A);

[collapse]

Errors 

DB11G/XTENDER> select/*+ PARALLEL(2)*/ cast(collect(a) as number_table) from test_parallel ;
select/*+ PARALLEL(2)*/ cast(collect(a) as number_table) from test_parallel
*
ERROR at line 1:
ORA-12801: error signaled in parallel query server P000
ORA-21710: argument is expecting a valid memory address of an object
 
Elapsed: 00:00:00.12
DB11G/XTENDER> select cast(collect(b) as number_table) from test_parallel group by a;
select cast(collect(b) as number_table) from test_parallel group by a
*
ERROR at line 1:
ORA-12805: parallel query server died unexpectedly
 
Elapsed: 00:00:17.57

[collapse]

I don’t experience such errors while using my slow aggregate. Here is a common dilemma: whether to use an unstable but speedy “collect”, or a slow aggregate of my own…

An example of an aggregate 

create or replace type ncollect_type as object
(
  
  data            sys.ku$_objnumset,
   
  static function ODCIAggregateInitialize
    ( sctx in out ncollect_type )
    return number ,
 
  member function ODCIAggregateIterate
    ( self  in out ncollect_type ,
      val   in     number
    ) return number ,
     
  member function ODCIAggregateDelete
    (  self in out  ncollect_type, 
       val  in      number
    ) return number ,
  member function ODCIAggregateTerminate
    ( self        in  ncollect_type,
      returnval   out sys.ku$_objnumset,
      flags in number
    ) return number ,
     
  member function ODCIAggregateMerge
    ( self in out ncollect_type,
      ctx2 in     ncollect_type
    ) return number
)
/
create or replace type body ncollect_type is
 
  static function ODCIAggregateInitialize
  ( sctx in out ncollect_type )
  return number
  is
  begin
    sctx := ncollect_type( sys.ku$_objnumset()) ;
    return ODCIConst.Success ;
  end;
 
  member function ODCIAggregateIterate
  ( self  in out ncollect_type ,
    val   in     number
  ) return number
  is
  begin
    self.data:=self.data multiset union sys.ku$_objnumset(val);
    return ODCIConst.Success;
  end;
 
  member function ODCIAggregateDelete
  (  self in out  ncollect_type, 
     val  in      number
  ) return number
  is
  begin
    self.data:=self.data multiset except sys.ku$_objnumset(val);
    return ODCIConst.Success;
  end;
 
  member function ODCIAggregateTerminate
  ( self        in  ncollect_type ,
    returnval   out sys.ku$_objnumset ,
    flags       in  number
  ) return number
  is
  begin
    returnval:=self.data;
    return ODCIConst.Success;
  end;
   
  member function ODCIAggregateMerge
  ( self in out ncollect_type ,
    ctx2 in     ncollect_type
  ) return number
  is
  begin
    self.data := self.data multiset union ctx2.data;
    return ODCIConst.Success;
  end;
end;
/

[collapse]

And the results:

Variant Time(sec)
select/*+ NO_PARALLEL*/ cast(collect(b) as number_table) from test_parallel group by a; 0.03
select/*+ NO_PARALLEL*/ ncollect(b) from test_parallel group by a 0.08
select ncollect(b) from test_parallel group by a; 0.07
select/*+ NO_PARALLEL*/ collect(a) from test_parallel; 0.02
select/*+ NO_PARALLEL*/ ncollect(a) from test_parallel 0.18
select ncollect(a) from test_parallel; 0.19
select/*+ NO_PARALLEL*/ collect(b) from test_parallel; 0.02
select/*+ NO_PARALLEL*/ ncollect(b) from test_parallel 0.18
select ncollect(b) from test_parallel; 0.06