Databse & Technology 2 | Connor McDonald | Managing Optimiser Statistics - A better way.pdf

itty bitty fonts in this

presentation

SQL> exec sample_font

Can you read this ?

Connor McDonald

OracleDBA

bio slide

Connor McDonald

a funny story

life was simple

(good) cost optimizer

version 7.0

optimizer_mode = CHOOSE

"we choose RULE"

SQL> analyze table SALES estimate statistics;

eventually....

it got better

added more functionality

SQL> analyze table SALES estimate statistics

2 for table

3 for columns size 10

4 for ....;

SQL syntax engine

oracle 8i

DBMS_STATS

problem....

analyze table SALES estimate statistics;SQL>

41 characters

128 characters

dbms_stats.gather_table_stats(

ownname=>'HR',

tabname=>'SALES',

cascade=>true,

estimate_percent=>20

don't get chaining

s l o w e r

"I don't think so....."

times have changed ...

SQL> desc DBMS_STATS

PROCEDURE ALTER_DATABASE_TAB_MONITORING

PROCEDURE ALTER_SCHEMA_TAB_MONITORING

PROCEDURE ALTER_STATS_HISTORY_RETENTION

PROCEDURE CLEANUP_STATS_JOB_PROC

PROCEDURE CONVERT_RAW_VALUE

PROCEDURE CONVERT_RAW_VALUE_NVARCHAR

PROCEDURE CONVERT_RAW_VALUE_ROWID

PROCEDURE COPY_TABLE_STATS

FUNCTION CREATE_EXTENDED_STATS RETURNS VARCHAR2

PROCEDURE CREATE_STAT_TABLE

PROCEDURE DELETE_COLUMN_STATS

PROCEDURE DELETE_DATABASE_PREFS

126 routines !

features

profiles

baselines

tuning sets

adaptive cursor sharing

bind peeking

better stats needed ...

... for the optimizer

hard to convince

stop using analyze

"good ol' days"

today....

optimizer is probably...

... smarter than you

SQL> select count(e.hiredate)

2 from DEPT d, EMP e

3 where e.deptno = d.deptno(+)

4 and e.sal > 10;

nested loop outer

sort merge outer

hash hash anti

nested loop anti

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

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

| 0 | SELECT STATEMENT | | 1 |

| 1 | SORT AGGREGATE | | 1 |

|* 2 | TABLE ACCESS FULL| EMP | 14 |

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

no DEPT ?

SQL> select count(e.hiredate)

2 from DEPT d, EMP e

3 where e.deptno = d.deptno(+)

4 and e.sal > 10;

not null

foreign key

key preserved

Oracle's solution....

statistics by stealth...

SQL> select owner, job_name, enabled

2 from dba_scheduler_jobs

3 where owner = 'SYS';

OWNER JOB_NAME ENABLED

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

SYS PURGE_LOG TRUE

SYS FGR$AUTOPURGE_JOB TRUE

SYS GATHER_STATS_JOB TRUE

SYS AUTO_SPACE_ADVISOR_JOB TRUE

SQL> select owner, job_name, enabled

2 from dba_scheduler_jobs

3 where owner = 'SYS';

OWNER JOB_NAME ENABLED

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

SYS AUTO_SPACE_ADVISOR_JOB FALSE

SYS BSLN_MAINTAIN_STATS_JOB TRUE

SYS DRA_REEVALUATE_OPEN_FAILURES TRUE

SYS FGR$AUTOPURGE_JOB FALSE

SYS GATHER_STATS_JOB FALSE

SYS HM_CREATE_OFFLINE_DICTIONARY FALSE

SYS ORA$AUTOTASK_CLEAN TRUE

SYS PURGE_LOG TRUE

SYS XMLDB_NFS_CLEANUP_JOB FALSE

stats STILL being collected

automatic "tasks"

"super stealth mode"

SQL> select client_name, status

2 from DBA_AUTOTASK_CLIENT;

CLIENT_NAME STATUS

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

auto optimizer stats collection ENABLED

auto space advisor ENABLED

sql tuning advisor ENABLED

most sites

stats every night

default options

in this session...

default behaviour

IDEA68

collecting statistics

IDEA70

but....

some default behaviour

IDEA73

collecting some statistics

IDEA75

Inflammatory

statements

which will

alienate the

audience

Presentation Duration

we're all hypocrites

statistical hypocrisy

"I need to change somereference data in my system"

wrong case

"nope...."

Information

Technology

Infrastructure

Library

servicecall

helpdesk

problemrecordDONE !

phew....

corrected

same site....

"Every night, I would like to potentially change the

performance characteristics of every single SQL statement in

the database"

"no problem...."

collecting stats = risk

Ensor's paradox

performance group

bstat/estat

tkprof

BMC patrol

afiedt.buf

"It is only safe to gather statistics ..."

"...when to do so will make no difference"

recommendation #1

unless things are bad...

do not change statistics

"surely it can't hurt?"

options=>'GATHER STALE'

options=>'GATHER EMPTY'

options=>'GATHER AUTO'

"no plans changed"

"no queries ran slower"

still problems

problem # 1

dbms_stats

the goal of statistics

minimise expensive SQL....

added more expensive SQL !

SQL> alter session set sql_true = true;

Session altered.

SQL> begin

2 dbms_stats.gather_table_stats(

3 'DEMO',

4 'PEOPLE);

5 end;

SQL> select

2 count(distinct GENDER),

3 min(GENDER),

4 max(GENDER),

5 count(distinct NAME),

6 min(NAME),

7 max(NAME)

21 from PEOPLE;

hard !

problem #2

lingering pain

invalidation

library cache

select *

from PEOPLE

insert into PEOPLE

select * from ...select ...

from PEOPLE,

where ...

delete from T

where X in

( select PID

from PEOPLE )

declare

v people.name%type;

SQL> begin

3 'DEMO',

4 'PEOPLE);

5 end;

gathering statistics

really hard core SQL

belted CPU with hard parsing

everything ran the same

oracle 9

SQL> desc DBMS_STATS

PROCEDURE GATHER_TABLE_STATS

Argument Name Type In/Out Default?

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

OWNNAME VARCHAR2 IN

TABNAME VARCHAR2 IN

PARTNAME VARCHAR2 IN DEFAULT

ESTIMATE_PERCENT NUMBER IN DEFAULT

BLOCK_SAMPLE BOOLEAN IN DEFAULT

METHOD_OPT VARCHAR2 IN DEFAULT

DEGREE NUMBER IN DEFAULT

GRANULARITY VARCHAR2 IN DEFAULT

CASCADE BOOLEAN IN DEFAULT

STATTAB VARCHAR2 IN DEFAULT

STATID VARCHAR2 IN DEFAULT

STATOWN VARCHAR2 IN DEFAULT

NO_INVALIDATE BOOLEAN IN DEFAULT

STATTYPE VARCHAR2 IN DEFAULT

FORCE BOOLEAN IN DEFAULT

NO_INVALIDATE BOOLEAN IN DEFAULT

no_invalidate => false | true

oracle 10

no_invalidate => auto

SQL> select

2 x.ksppinm name,

3 y.kspftctxvl value

4 from

5 sys.x$ksppi x,

6 sys.x$ksppcv2 y

7 where

8 x.indx+1 = y.kspftctxpn and

9 x.ksppinm = '_optimizer_invalidation_period'

NAME VALUE

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

_optimizer_invalidation_period 18000

SQL> exec dbms_stats.gather_table_stats(

user,'PEOPLE');

select *

from PEOPLEinsert into PEOPLE

select * from ...

select ...

from PEOPLE,

where ...

delete from T

where X in

( select PID

from PEOPLE )

declare

v people.name%type;

better ?

hard parse storm avoided

5 hours before problems

unless things are bad...

DO NOT CHANGE STATISTICS

Inflammatory

statements

which will

alienate the

audience

we don't know if things are bad

until its too late

SQL's are slow

users fed up

stats out of date

2 hours to gather stats

collect stats

in readiness

pending statistics

"pending" - private

"published" – optimizer uses

SQL> begin

2 dbms_stats.set_schema_prefs(

3 ownname=>'DEMO',

4 pname=>'PUBLISH',

5 pvalue=>'FALSE');

6 end;

database,

SQL> alter session set

2 optimizer_use_pending_statistics = true;

SQL> begin

2 dbms_stats.publish_pending_stats(

3 ownname=>'DEMO',

4 tabname=>null);

5 end;

recommendation #2

when collecting statistics

PUBLISH = false

always....

even if you immediately publish

atomic publication

and don't wait 5 hours

_optimizer_invalidation_period

for most tables

dbms_stats.lock_table_stats

collecting system stats

different story

absolutely vital…

…to have them

maximize I/O throughput

SQL> set autotrace traceonly explain

SQL> select * from T;

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

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

| 0 | SELECT STATEMENT | | 53459 | 4489K| 113 (1)|

| 1 | TABLE ACCESS FULL| T | 53459 | 4489K| 113 (1)|

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

SQL> select value

2 from v$spparameter

3 where name = 'db_file_multiblock_read_count';

-----------

SQL> alter session set events

2 = '10046 trace name context forever, level 8';

SQL> select * from T;

PARSING IN CURSOR #22 len=15 dep=0 uid=124 oct=3 lid=124 ...

select * from T

END OF STMT

PARSE #22:c=10000,e=108,p=0,cr=0,cu=0,mis=0,r=0,dep=0,og=4

EXEC #22:c=0,e=65,p=0,cr=0,cu=0,mis=0,r=0,dep=0,og=4

WAIT #22: nam='db file scattered read' ela= 3387

file#=5 block#=13122 blocks=16 obj#=199963

SSTIOMAX

db_block_size = 8192

db_file_multiblock_read_count = 128

the solution ?

system stats

SQL> select pname, pval1

2 from sys.aux_stats$

PNAME PVAL1

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

SREADTIM 4.065

MREADTIM 6.173

CPUSPEED 567

MBRC 12

MAXTHR 48203776

SLAVETHR -1

use MBRC to cost

use db_file_multiblock_read_count to read

Inflammatory

statements

which will

alienate the

audience

forget about system statistics

do not use them

absolutely vital…

…to have them

do not “use”* them

* = collect, change, set

problem #1

problem #2

monitoring for bad....

optimizing for bad

recommendation #3

gather

into STAT table

monitor

set once

the defaults are probably fine

SQL> select pname, pval1

2 from sys.aux_stats$

PNAME PVAL1

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

IOSEEKTIM 10 (ms)

IOTFRSPEED 4096 (bytes/ms)

SREADTIM = IOSEEKTIM + db_block_size / IOTFRSPEED = 10

MREADTIM = IOSEEKTIM + mbrc * db_block_size / IOTFRSPEED = 26

10.2 and above

db_file_multiblock_read_count

_db_file_exec_read_count

_db_file_optimizer_read_count

dbms_resource_manager

to calibrate

so far...

not changing statistics

eventually....

the time will come...

one possible reason

even if stats unchanged …

… your plans might !

SQL> create table T as

2 select to_date('01-AUG-2011')+

3 trunc(dbms_random.value(0,7)) dte,

4 rpad('padding',20) padding

5 from dual

6 connect by level <= 100000;

Table created.

SQL> select dte, count(*)

2 from t

3 group by dte

4 order by 1;

DTE COUNT(*)

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

01-AUG-11 14334

02-AUG-11 14222

03-AUG-11 14167

04-AUG-11 14510

05-AUG-11 14346

06-AUG-11 14349

07-AUG-11 14072

15,000 rows per day always

SQL> exec dbms_stats.gather_table_stats(user,'T')

PL/SQL procedure successfully completed.

SQL> create index IX on T ( dte ) ;

Index created.

SQL> select * from t where dte = '03-AUG-2011';

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

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

| 0 | SELECT STATEMENT | | 14286 | 404K|

|* 1 | TABLE ACCESS FULL| T | 14286 | 404K|

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

a week later…

SQL> insert into T

2 select to_date('08-AUG-2011')+

3 trunc(dbms_random.value(0,7)) dte,

4 rpad('padding',20) padding

5 from dual

6 connect by level <= 100000;

100000 rows created.

SQL> select dte, count(*)

2 from t

3 group by dte

4 order by 1;

DTE COUNT(*)

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

01-AUG-11 14334

07-AUG-11 14072

08-AUG-11 14261

09-AUG-11 14106

10-AUG-11 14410

11-AUG-11 14289

12-AUG-11 14358

13-AUG-11 14252

14-AUG-11 14324

statistics unchanged

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

| 1 | TABLE ACCESS BY INDEX ROWID| T | 2381 |

|* 2 | INDEX RANGE SCAN | IX | 2381 |

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

| 1 | TABLE ACCESS BY INDEX ROWID| T | 1 |

|* 2 | INDEX RANGE SCAN | IX | 1 |

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

SQL> select low_value, high_value

2 from user_tab_columns

3 where table_name = 'T'

4 and column_name = 'DTE';

LOW_VALUE HIGH_VALUE

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

786F0801010101 786F0807010101

SQL> set serverout on

SQL> declare

2 res date;

3 begin

4 dbms_stats.convert_raw_value(

5 '786F0801010101',res);

6 dbms_output.put_line(result);

7 dbms_stats.convert_raw_value(

8 '786F0806010101',res);

9 dbms_output.put_line(result);

10 end;

01-AUG-11

06-AUG-11

01-AUG-11 06-AUG-11

when the time comes…

as accurate as possible

as cheaply as possible

statistics accurately

extended statistics

cardinality is everything

same with Oracle

some real(ish) data

SQL> desc VEHICLE

Name Null? Type

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

ID NUMBER

MAKE VARCHAR2(6)

MODEL VARCHAR2(6)

SQL> select count(*)

2 from VEHICLE;

COUNT(*)

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

2,157,079

default stats not enough

2 from VEHICLE

3 where MAKE = 'HOLDEN';

COUNT(*)

----------

415387

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

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

| 0 | SELECT STATEMENT | | 1 | 7 | 138|

| 1 | SORT AGGREGATE | | 1 | 7 | |

|* 2 | INDEX RANGE SCAN| MAKE_IX | 55310 | 378K| 138|

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

histogram

SQL> begin

2 dbms_stats.gather_table_stats(user,'VEHICLE',

3 method_opt=>'for all columns size 1,'||

4 'for columns MAKE size 254,'||

5 'for columns MODEL size 254');

6 end;

2 from VEHICLE

3 where MAKE = 'HOLDEN';

COUNT(*)

----------

415387

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

| 1 | SORT AGGREGATE | | 1 | 7 | |

|* 2 | INDEX RANGE SCAN| MAKE_IX | 418K| 2859K| 1024|

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

make AND model

2 from VEHICLE

3 where MAKE = 'HOLDEN'

4 and MODEL = 'COMMODORE';

COUNT(*)

----------

214468

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

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

| 0 | SELECT STATEMENT | | 1 | 14 |

| 1 | SORT AGGREGATE | | 1 | 14 |

| 2 | BITMAP CONVERSION COUNT | | 39527 | 540K|

| 3 | BITMAP AND | | | |

| 4 | BITMAP CONVERSION FROM ROWIDS| | | |

| 6 | BITMAP CONVERSION FROM ROWIDS| | | |

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

50% holdens are commodores

two things

241241

no correlation

10g and before

SQL> select

2 DBMS_STATS.CREATE_EXTENDED_STATS(

3 user, 'VEHICLE','(MAKE,MODEL)') tag

4 from dual;

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

SYS_STU8QPK2S$PEWHARK2CP3#1F#G

SQL> begin

2 dbms_stats.gather_table_stats(user,'VEHICLE',

3 method_opt=>

4 'for columns (make,model) size 254');

5 end;

2 from VEHICLE

COUNT(*)

----------

214468

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

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

| 1 | SORT AGGREGATE | | 1 | 14 | |

|* 2 | TABLE ACCESS FULL| VEHICLE | 220K| 3018K| 1956|

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

composite indexes

11.2 column groups

SQL> exec DBMS_STATS.SEED_COL_USAGE(NULL,NULL,60);

implemented via virtual columns

various implications

accurate stats not always possible...

you can't have stats

on everything !

when stats wont do...

... use the real data

dynamic sampling

SQL> drop table TOUGH_DATA purge;

Table dropped.

SQL> create table TOUGH_DATA nologging as

2 select

3 rownum pk,

4 dbms_random.string('U',10) str

5 from dual

6 connect by level < 1000000

Table created.

user,'TOUGH_DATA')

2 from TOUGH_DATA

3 where str like '%XX'

COUNT(*)

----------

2 from TOUGH_DATA

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

|* 2 | TABLE ACCESS FULL| TOUGH_DATA | 50000 |

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

5% assumption

SQL> select /*+ dynamic_sampling(t 2) */ count(*)

2 from TOUGH_DATA t

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

|* 2 | TABLE ACCESS FULL| TOUGH_DATA | 1252 |

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

simple queries as well

SQL> create table EASY_DATA nologging as

2 select

3 rownum pk,

4 chr(65+trunc(rownum/40000)) str

5 from dual

6 connect by level < 1000000

Table created.

SQL> exec dbms_stats.gather_table_stats(user,'EASY_DATA')

A,A,A,A,A,B,B,B,B,.....1,2,3,4,5,6......

2 from EASY_DATA

3 where str = 'F'

4 and pk > 900000;

COUNT(*)

----------

2 from EASY_DATA

3 where str = 'F'

4 and pk > 900000;

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

|* 2 | TABLE ACCESS FULL| EASY_DATA | 4000 |

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

SQL> select /*+ dynamic_sampling(t 2) */ count(*)

2 from EASY_DATA t

3 where str = 'F'

4 and pk > 900000;

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

parallel queries (maybe) sampled

2 from EASY_DATA t

3 where str = 'F'

4 and pk > 900000;

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

| 2 | PX COORDINATOR | | |

| 3 | PX SEND QC (RANDOM) | :TQ10000 | 1 |

| 5 | PX BLOCK ITERATOR | | 4 |

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

- dynamic sampling used for this statement (level=4)

low frequency

high cost

statistics cheaply

SQL> desc DBA_TABLES

Name Null? Type

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

OWNER NOT NULL VARCHAR2(30)

TABLE_NAME NOT NULL VARCHAR2(30)

COLUMN_NAME NOT NULL VARCHAR2(30)

NUM_ROWS NUMBER

SQL> desc DBA_TAB_COLS

Name Null? Type

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

OWNER NOT NULL VARCHAR2(30)

TABLE_NAME NOT NULL VARCHAR2(30)

COLUMN_NAME NOT NULL VARCHAR2(30)

NUM_DISTINCT NUMBER

LOW_VALUE RAW(32)

HIGH_VALUE RAW(32)

SQL> desc PEOPLE

Name Null? Type

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

PID NUMBER

GENDER CHAR(1)

NAME VARCHAR2(47)

AGE NUMBER

DEATH_RATE NUMBER

SQL> alter session set sql_true = true;

Session altered.

SQL> begin

3 'DEMO',

4 'PEOPLE);

5 end;

select

/*+ no_parallel(t)

no_parallel_index(t)

dbms_stats

cursor_sharing_exact

use_weak_name_resl

dynamic_sampling(0)

no_monitoring

no_substrb_pad */

count(*),

count("GENDER"),

count(distinct "GENDER"),

substrb(dump(min("GENDER"),16,0,32),1,120),

substrb(dump(max("GENDER"),16,0,32),1,120),

from "MCDONAC"."PEOPLE" t

count("GENDER")

count(distinct "GENDER")

min("GENDER")

max("GENDER")

count("GENDER")

count(distinct "GENDER")

min("GENDER")

max("GENDER")

one pass

SQL> select count(*) from PEOPLE;

COUNT(*)

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

500000000

Elapsed: 00:04:43.73

SQL> begin

3 user,

4 'PEOPLE',

5 estimate_percent=>25);

6 end;

ERROR at line 1:

ORA-01652: unable to extend temp

segment by 128 in tablespace TEMP

Elapsed: 00:16:37.12

one pass NDV

12 6 3 12 3 7 11 12 33 6 11 12 6 45 15 7 15 17 45 6 17

12 6 3 7 11 33 45 15 17

NDV = 9

12 6 3 7 11 33 45 15 17

what about large NDV ?

"Magic" HASH

remaining hashes x

2^number of splits

SQL> create table ONE_PASS nologging

2 as select substr(text,1,1) single_char

3 from DBA_SOURCE;

Table created.

SQL> select count(distinct single_char) ndv

2 from one_pass;

----------

16 buckets...

64 127

96 127

SQL> set serverout on

SQL> declare

2 type t_bucket is table of varchar2(1);

3 l_synopsis t_bucket;

4 l_splits number := 0;

5 l_hash int;

6 l_min_val int := 0;

7 l_synopsis_size int := 16;

8 begin

9 for i in ( select single_char from one_pass ) loop

10 l_hash := ascii(i.single_char);

12 if l_synopsis.count = l_synopsis_size then

13 l_min_val :=

14 case

15 when l_min_val = 0 then 64

19 end;

20 l_splits := l_splits + 1;

23 for j in 1 .. l_min_val loop

24 if l_synopsis.exists(j) then

25 l_synopsis.delete(j);

26 end if;

27 end loop;

28 end if;

30 if l_hash > l_min_val then

31 l_synopsis(l_hash) := 'Y';

32 end if;

33 end loop;

34 dbms_output.put_line(l_synopsis.count *

35 power(2,l_splits));

36 end;

Splitting, keeping entries above 64

the reality

16384 bucket limit

18,446,744,073,709,551,616 hash range

SQL> begin

3 user,

4 'PEOPLE',

5 estimate_percent=>25);

6 end;

Elapsed: 00:05:39.82

recommendation #4

estimate_percent

use DEFAULT size (AUTO)

only for columns

without histograms

Inflammatory

statements

which will

alienate the

audience

histograms...

...SUCK

not really but...

SQL> select * from MY_TABLE_100K_ROWS

2 where r = :b1

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

| 1 | INDEX UNIQUE SCAN| PK | 1 |

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

~ num rows / num distinct

SQL> select * from MY_TABLE_100K_ROWS

2 where r < :b1

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

| 1 | INDEX RANGE SCAN| PK | 5000 |

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

cool ?

method_opt

for all columns size auto

sys.col_usage$

SQL> desc SYS.COL_USAGE$

Name Null? Type

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

OBJ# NUMBER

INTCOL# NUMBER

EQUALITY_PREDS NUMBER

EQUIJOIN_PREDS NUMBER

NONEQUIJOIN_PREDS NUMBER

RANGE_PREDS NUMBER

LIKE_PREDS NUMBER

NULL_PREDS NUMBER

TIMESTAMP DATE

SQL> create table T (

2 skew varchar2(10),

3 even number);

Table created.

SQL> insert into T

2 select

3 case

4 when rownum > 99995 then 'SPECIAL'

5 else dbms_random.string('U',8)

6 end,

7 mod(rownum,200)

8 from dual

9 connect by level <= 100000

100000 rows created.

5 special

values

distribution

user,'T', estimate_percent=>null);

SQL> select COLUMN_NAME,NUM_DISTINCT,DENSITY,

2 ( select count(*)

3 from user_tab_histograms

5 and column_name = c.column_name ) hist_cnt

6 from user_tab_cols c

8 order by table_name,COLUMN_ID

COLUMN_NAME NUM_DISTINCT DENSITY HIST_CNT

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

SKEW 99996 .00001 2

EVEN 200 .005 2

SQL> select * from T where skew = 'SPECIAL';

SKEW VAL

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

SPECIAL 196

SPECIAL 197

SPECIAL 198

SPECIAL 199

SPECIAL 0

5 rows selected.

SQL> select * from T where even = 5;

TAG VAL

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

IBRXGVIE 5

[snip]

500 rows selected.

user,'T', estimate_percent=>null);

SQL> select COLUMN_NAME,NUM_DISTINCT,DENSITY,

2 ( select count(*)

3 from user_tab_histograms

5 and column_name = c.column_name ) hist_cnt

6 from user_tab_cols c

8 order by table_name,COLUMN_ID

COLUMN_NAME NUM_DISTINCT DENSITY HIST_CNT

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

SKEW 99996 .00001 2

EVEN 200 .000005 200

recommendation #5

set_schema_prefs

for all columns size 1

explicit control for each table

set_table_prefs

for column CCC size nnn

one more useful tool

SQL> desc CONFERENCE_ATTENDEES

Name Null? Type

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

PID NUMBER

GENDER CHAR(1)

NAME VARCHAR2(40)

AGE NUMBER

DEATH_RATE NUMBER

SQL> select

2 count(distinct GENDER) NDV

3 from

4 CONFERENCE_ATTENDEES

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

SQL> select

2 ntile(GENDER) over ( ... )

3 from

4 CONFERENCE_ATTENDEES

set_table_stats

num_rows

avg_row_len

set_column_stats

high_value

num_distinct

when all else fails ...

... lie and cheat

analyze ... validate structure

SQL> desc INDEX_STATS

Name Null? Type

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

HEIGHT NUMBER

BLOCKS NUMBER

NAME VARCHAR2(30)

PARTITION_NAME VARCHAR2(30)

OPT_CMPR_COUNT NUMBER

OPT_CMPR_PCTSAVE NUMBER

wrap up

don't collect stats .... unless

don't collect system stats ... unless

don't collect histograms ... unless

default estimate size (NDV)

lie and cheat

Connor McDonald

OracleDBA

ORA-00041

www.oracledba.co.uk

“active time limit exceeded - session terminated”

340340

actual versus estimate

SQL> select /*+ GATHER_PLAN_STATISTICS */ count(*)

2 from VEHICLE

COUNT(*)

----------

214468

SQL> SELECT *

2 FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR(

3 NULL, NULL, 'ALLSTATS LAST'));

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

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

| 1 | SORT AGGREGATE | | 1 | 1 | 1 |

|* 2 | TABLE ACCESS FULL| VEHICLE | 1 | 220K| 214K|

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

343343

employ someone....

344344

check every query...

SQL> select /*+ GATHER_PLAN_STATISTICS */ count(*)

2 from VEHICLE

COUNT(*)

----------

214468

SQL> SELECT *

2 FROM TABLE(DBMS_XPLAN.DISPLAY_CURSOR(

3 NULL, NULL, 'ALLSTATS LAST'));

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

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

| 1 | SORT AGGREGATE | | 1 | 1 | 1 |

|* 2 | TABLE ACCESS FULL| VEHICLE | 1 | 220K| 214K|

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

346346

347347

but just maybe ....

348348

... someone already is

SQL> create table EMP as

2 select rownum empno,

3 mod(rownum,10) jobid,

4 mod(rownum,10)*1000 salary,

5 mod(rownum,50)+1 deptno

6 from dual

7 connect by rownum < 100000;

SQL> create table DEPT as

2 select rownum deptno,

3 'dept'||rownum dname

4 from dual

5 connect by rownum <= 100;

100,000 rows

100 rows

SQL> exec dbms_stats.gather_table_stats(user,'EMP');

SQL> exec dbms_stats.gather_table_stats(user,'DEPT');

SQL> create index EMP_IX on EMP ( deptno );

SQL> create index DEPT_IX on DEPT ( deptno );

SQL> select e.empno, d.dname

2 from emp e, dept d

3 where d.deptno = e.deptno

4 and e.jobid = 1

5 and e.salary > 5000;

no rows selected

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

| 0 | SELECT STATEMENT | | | |

| 1 | MERGE JOIN | | 4444 | 104K |

| 2 | TABLE ACCESS BY INDEX ROWID| DEPT | 100 | 1000 |

| 3 | INDEX FULL SCAN | DEPT_IX | 100 | |

|* 4 | SORT JOIN | | 4444 | 62216 |

|* 5 | TABLE ACCESS FULL | EMP | 4444 | 62216 |

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

4 and e.jobid = 1

hard to

optimize

re-run the query

no anythingchanges to

SQL> select e.empno, d.dname

2 from emp e, dept d

3 where d.deptno = e.deptno

4 and e.jobid = 1

no rows selected

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

| 0 | SELECT STATEMENT | | | |

|* 1 | HASH JOIN | | 89 | 2136 |

| 2 | TABLE ACCESS FULL| DEPT | 1 | 10 |

|* 3 | TABLE ACCESS FULL| EMP | 4444 | 62216 |

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

the optimizer knows what "hard" is

cardinality feedback loop

its not continuous

several restrictions

but still very cool

Databse & Technology 2 | Connor McDonald | Managing Optimiser Statistics - A better way.pdf

Documents

Ophthalmic data & Databse

Hubspot_Introduciong Marketing the Agile Way.pdf

WEB DATABSE INTERFACING

Recovery Databse dengan SQL Server

United Way.pdf

01 - Prepara The Way.pdf

Understanding the Mi'kmaq Way.pdf

98.Learn Russian the Fast and Fun Way.pdf

introduction of databse

ANOVA ONE WAY.pdf

The Kyokushin Way.pdf

Comprehensive Company Databse

How to Draw Cars the Hot Wheels Way.pdf

Learn Japanese the Fast and Fun Way.pdf

Databse Tend Window

Nene Way.pdf