® © 2005 IBM Corporation Access Path Selection and Physical DB Design in Relational Database Management Systems: An Overview Padua University – 10 June 2005
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
®
©2005 IB
M C
orp
ora
tion
Acce
ss P
ath
Se
lectio
n a
nd
P
hysic
al D
B D
esig
n
in R
ela
tio
na
l D
ata
ba
se
Ma
na
ge
me
nt
Syste
ms:
An O
verv
iew
����������
������� �����������
Padua U
niv
ers
ity –
10 J
un
e 2
005
IBM
Softw
are
Gro
up
2
��������
Co
nte
nts
�R
DB
MS
an
d S
QL L
angua
ge
�S
tatic S
QL a
nd D
ynam
ic S
QL
�M
app
ing t
he P
hysic
al D
esig
n
�Q
uery
Com
pile
r O
verv
iew
�Q
uery
Optim
izatio
n
�A
ccess P
ath
Sele
ction
�Q
uery
Rew
rite
�E
xam
ple
s
�T
he O
ptim
izer
Exp
lain
ed
�T
he R
ole
of th
e D
BA
IBM
Softw
are
Gro
up
3
��������
Re
lation
al M
ode
l &
Rela
tio
na
l D
ata
base M
ana
gem
ent
Syste
ms
�T
he r
ela
tio
na
l m
ode
l is
the a
pp
lication o
f
�p
red
icate
lo
gic
and
�set
theo
ry
to d
ata
base m
anag
em
ent
�T
heory
: p
red
icate
lo
gic
and s
et
math
em
ati
cs
�S
tructu
re: R
-tab
les
�In
tegrity
: d
om
ain
, co
lum
n, ta
ble
and d
ata
base inte
grity
�M
anip
ula
tion: R
-op
era
tio
ns
(restr
ict, p
roje
ct, join
, etc
.)
IBM
Softw
are
Gro
up
4
��������
R-o
pe
ratio
ns:
Th
e S
QL
La
ng
ua
ge
SELECT *
FROM Today_Orders
�Q
uery
constr
ucts
�S
et O
riente
d
�Logic
al “e
xpre
ssio
ns”
�N
O r
efe
rence t
o a
ny p
hysic
al str
uctu
res
�F
unctiona
lly in
depe
nde
nt
from
ph
ysic
al str
uctu
res
�Q
uery
perf
orm
ance
�H
ighly
(but not fu
lly)
independent fr
om
language c
onstr
ucts
�H
ighly
dependent on D
B p
hysic
al desig
n
IBM
Softw
are
Gro
up
5
��������
R-o
pe
ratio
ns:
Ma
pp
ing
th
e P
hysic
al D
esig
n
SELECT *
FROM Today_Orders
CREATE VIEW Today_OrdersAS
SELECT O.CUST#, O.ORD#, RO.PROD#, RO.QTY
FROM Orders
O
, Order_Lines RO
WHERE O.ODR# = RO.ORD#
AND O.ORDER_DATE = current date
ORDER BY O.CUST#, O.ORD#, RO.PROD#
Ind
exes
Ind
exes
IBM
Softw
are
Gro
up
6
��������
Sta
tic S
QL
�P
re-d
efin
ed s
ynta
x,
hard
-cod
ed into
a h
ost la
ng
ua
ge
(e.g
., C
ob
ol, C
, C
++
, etc
.)
�It m
ay r
efe
rence (
usually
does)
host-
pro
gra
m v
ariable
s
�E
xam
ple
EXEC SQL DECLARE CRS1CURSOR FOR
SELECT C1, C2, C3
FROM T1
WHERE C4 = :c4
OPEN CRS1
FETCH CRS1INTO :c1, :c2, :c3
…….
CLOSE CRS1
�M
ostly u
sed f
or
OLT
P a
nd B
atc
h p
rocessin
g, w
here
data
nee
ds a
re k
now
n in a
dvance
IBM
Softw
are
Gro
up
7
��������
Dyn
am
ic S
QL
�D
ata
needs tota
lly o
r part
ially
unknow
n: e.g
.
�Q
uery
too
ls
�Q
uery
typed b
y e
nd-u
ser
on a
bro
wser
page a
nd b
uffere
d b
y
applic
ation p
rogra
m
�E
xam
ple
EXEC SQL PREPARE STMTINTO :MYSQLDA FROM :SQLBUFFER
EXEC SQL DECLARE CRS1CURSOR FOR STMT
EXEC SQL OPEN CRS1USING :C
EXEC SQL FETCH CRS1INTO :C1, :C2, :C3
….
EXEC SQL CLOSE CRS1
SQLBUFFER = ‘SELECT C1, C2, C3 FROM T1 WHERE C4 = ?’
IBM
Softw
are
Gro
up
8
��������
Qu
ery
Co
mp
iler
Ove
rvie
w
SQ
L Q
uery
Pars
er
Glo
bal Q
uery
Sem
antics
Query
Re-W
rite
Pla
nO
ptim
ization
Code
Genera
tio
n
Qu
ery
Gra
ph
Mo
del
Query
Pla
n
Query
Exp
lain
Pla
nE
xp
lain
Executa
ble
Pla
n
IBM
Softw
are
Gro
up
9
��������
Ste
ps in
Qu
ery
Co
mp
ilatio
n
�P
ars
ing
�A
na
lyze "
text"
of S
QL q
uery
�D
ete
ct synta
x e
rrors
�C
reate
inte
rna
l qu
ery
re
pre
senta
tio
n
�S
em
antic C
heckin
g
�V
alid
ate
SQ
L s
tate
ment
�V
iew
ana
lysis
�In
corp
ora
te c
onstr
ain
ts,
trig
gers
, etc
.
�Q
uery
Optim
ization
�M
od
ify q
uery
to im
pro
ve
perf
orm
ance (
Query
Re
write
)
�C
hoose the m
ost effic
ient
"access p
lan"
(Query
O
ptim
ization)
�C
ode G
enera
tion: genera
te
code that is
�E
xecuta
ble
�E
ffic
ient
Query
Com
pila
tion e
xecute
d d
uring
�A
pplic
ation P
rogra
m B
IND
pro
cess (
Sta
tic S
QL)
�E
xecution o
f S
QL P
repare
(D
ynam
ic S
QL)
IBM
Softw
are
Gro
up
10
��������
Qu
ery
Op
tim
iza
tio
n:
Acce
ss P
ath
Se
lectio
n
�B
ase
d o
n e
stim
ating Q
uery
execu
tion c
ost
…
�C
PU
�I/O
�E
lapsed T
ime
�…
of A
ccess P
ath
variations b
ased o
n
�U
sin
g d
iffe
rent in
dex(e
s)
/ no index a
t all
�U
sin
g d
iffe
rent jo
in s
equence
�U
sin
g d
iffe
rent jo
in m
eth
od
�U
sin
g q
uery
para
llelis
m
�…
.
�Less e
xpe
nsiv
e A
cce
ss P
ath
chosen
IBM
Softw
are
Gro
up
11
��������
Wh
at’s a
n I
nd
ex
�A
n Index is a
physic
al
str
uctu
re a
llow
ing d
irect (a
nd
ord
ere
d)
access to r
ecord
s
matc
hin
g the v
alu
es o
f th
e
Index K
ey o
r a p
ort
ion o
f it
�A
n Index k
ey c
an b
e
�sin
gle
-colu
mn o
r m
ulti-colu
mn
�U
niq
ue o
r no
n-u
niq
ue
�A
dditio
nal o
ptions m
ight
ap
ply
(e
.g.
Clu
ste
r)
�M
ost com
mon Indexes h
ave a
B
-tre
e s
tructu
re (
see fig
ure
)
�In
dexes p
rovid
e p
erf
orm
ance
advanta
ge for
data
retr
ieval,
but in
cre
ase the c
ost of D
ele
te
/ In
sert
/ U
pdate
opera
tions
IBM
Softw
are
Gro
up
12
��������
Acce
ss P
ath
Qu
iz –
1:
Sin
gle
Ta
ble
Acce
ss
�T
1 =
Rela
tional T
able
(not a V
iew
)
�S
ingle
table
in s
ingle
ph
ysic
al file
�T
CA
RD
=
1,0
00,0
00
�In
de
x I
X1 o
n (
C1),
with D
up
licate
s
�In
de
x I
X2 o
n (
C2),
with D
up
licate
s
�In
de
x I
X3 o
n (
C4),
Uniq
ue
�W
hic
h Index m
ight pro
vid
e the b
est A
ccess P
ath
?
�A
ny s
uggestions for
impro
vin
g the A
ccess P
ath
?
SELECT C1, C2, C3, C4
FROM T1
WHERE C1 = ?
AND C2 = ?
ORDER BY C3
IBM
Softw
are
Gro
up
13
��������
Acce
ss P
ath
Qu
iz 1
: F
ilte
r F
acto
rs
�F
F(C
ol=
litera
l) =
1 / C
ol-C
AR
D
�C
ol-C
AR
D
= N
um
ber
of dis
tinct valu
es for
Col
�T
CA
RD
= N
um
ber
of tu
ple
sin
Table
�E
xam
ple
:
�C
1-C
AR
D=
10,0
00; C
2-C
AR
D =
100,0
00
�F
F(C
1 =
litera
l) =
1 / 1
0,0
00 =
0,0
001 (i.e
. 0,0
1%
)
�F
F(C
2 =
litera
l) =
1 / 1
00,0
00 =
0,0
0001 (
i.e. 0,0
01%
)
SELECT C1, C2, C3, C4
FROM T1
WHERE C1 = ?
AND C2 = ?
ORDER BY C3
IBM
Softw
are
Gro
up
14
��������
Acce
ss P
ath
Qu
iz 1
: C
om
pa
rin
g I
nd
ex A
cce
ss
�U
sin
g IX
1 (
C1)
�Q
ualif
yin
g T
uple
s (
estim
ate
) =
10**
6*
10**
(-4
)=
100
�R
etr
ieve a
ll 10
0 q
ualif
yin
g T
uple
s,
app
ly p
redic
ate
on C
2,
sort
(O
RD
ER
B
Y C
3)
�U
sin
g IX
2 (
C2)
�Q
ualif
yin
g T
uple
s (
estim
ate
) =
10**
6*
10**
(-5
)=
10
�R
etr
ieve a
ll 10 q
ualif
yin
g T
uple
s,
apply
pre
dic
ate
on C
1,
sort
�W
hat about S
kew
ing?
�W
hat if IX
4(C
1,C
2)?
�S
uppose n
um
ber
of
Dis
tinct
Ke
ys in I
X4 =
50
0,0
00
�F
F(C
1 =
litera
l A
ND
C2 =
litera
l) =
4 *
10**
(-6
)
�C
an t
he R
DB
MS
derive F
F(C
1 =
litera
l, C
2 =
litera
l) b
y its
kn
ow
ledge o
f F
F(C
1 =
litera
l) a
nd F
F(C
2 =
litera
l) ?
�A
ny a
dvanta
ge, if a
bove h
old
s tru
e, w
ith IX
4(C
1,C
2,C
3)
?
IBM
Softw
are
Gro
up
15
��������
Acce
ss P
ath
Qu
iz 1
: S
ke
win
g
y
1000y
Colu
mn V
alu
e
Frequency
Avera
ge
Actu
al
Actu
al
Actu
al
IBM
Softw
are
Gro
up
16
��������
Ske
win
g:
No
tes
�F
ilter
Facto
r assum
es U
niform
Dis
trib
ution o
f V
alu
es
Fre
que
ncy
�T
hat’s u
su
ally
not
the c
ase in r
ea
l lif
e s
ituations
�R
DB
MS
Optim
izer
needs to k
now
more
on v
alu
es fre
quency
(e.g
. dis
trib
ution s
tatistics)
�Q
uery
perf
orm
ance w
ill b
e h
igh
ly d
epen
de
nt
on
valu
es s
pe
cifie
d in p
red
icate
s
�F
or
better
optim
ization, R
DB
MS
Optim
izer
must know
com
parison v
alu
es in p
redic
ate
at com
pile
tim
e
�U
se D
ynam
ic S
QL w
ith
out
para
mete
r m
ark
ers
; or
..
�R
e-o
ptim
ize S
QL q
uery
at
execution t
ime
IBM
Softw
are
Gro
up
17
��������
Acce
ss P
ath
Qu
iz –
2:
Jo
in
�T
1, T
2 =
Rela
tional T
able
s (
not a V
iew
s)
�T
1-C
AR
D =
500,0
00
�T
2-C
AR
D =
5,0
00,0
00
�F
ilter
facto
rs
�F
F(T
1.C
2 =
litera
l) =
10**
(-4
)
�F
F(T
2.C
3 =
litera
l) =
10**
(-6
)
�N
o Indexes: H
ow
would
you m
anage the join
?
�W
hic
h Indexes w
ould
you r
ecom
mend?
SELECT T1.C1, T1.C2, T2.C3
FROM T1
, T2
WHERE T1.C1 = T2.C1
AND T1.C2 = ?
AND T2.C3 = ?
ORDER BY T1.C1
IBM
Softw
are
Gro
up
18
��������
Acce
ss P
ath
Qu
iz –
2:
Mo
st C
om
mo
n J
oin
Me
tho
ds
�N
este
d L
oop J
oin
(N
LJ)
�F
or
each q
ualif
yin
g tuple
of T
x, sele
ct m
atc
hin
g tuple
sfr
om
Ty
�U
sually
, an Index o
n T
y.J
Cs
used
�M
erg
e S
ca
n J
oin
(M
SJ)
�S
ort
qualif
yin
g tuple
sfr
om
Tx
on join
colu
mns
�S
ort
qualif
yin
g tuple
sfr
om
Ty
on join
colu
mns
�M
erg
e tuple
sm
atc
hin
g join
pre
dic
ate
s
�H
ash J
oin
(H
J)
�S
imila
r to
NLJ
�H
ashin
g u
sed to s
peed u
p r
etr
ieval of m
atc
hin
g tuple
sfr
om
Ty
IBM
Softw
are
Gro
up
19
��������
Qu
ery
Re
wri
te
�R
ew
riti
ng
a g
iven
SQ
L q
uery
in
to a
sem
an
tically e
qu
ivale
nt
form
th
at
�m
ay b
e p
rocessed m
ore
eff
icie
ntly
�giv
es t
he O
ptim
izer
more
latitu
de
�W
hy
�S
am
e q
uery
ma
y h
ave m
ultip
le r
epre
se
nta
tio
ns in S
QL
�C
om
ple
x q
ueri
es o
ften r
esult in r
edun
dancy,
especia
lly w
ith v
iew
s
�Q
uery
genera
tors
�o
ften
pro
du
ce s
ub
op
tim
al
qu
eri
es t
hat
do
n't
perf
orm
well
�d
on
't p
erm
it "
han
d o
pti
miz
ati
on
"
�D
B2 c
ap
ab
ilit
ies is b
ased
on
Sta
rbu
rst
Qu
ery
Rew
rite
�R
ule
-base
d q
uery
re
wri
te e
ng
ine
�T
ransfo
rms legal Q
GM
into
more
eff
icie
nt
QG
M
�T
erm
inate
s w
hen n
o r
ule
s e
ligib
le o
r b
udg
et
excede
d
IBM
Softw
are
Gro
up
20
��������
Qu
ery
Re
wri
te:
Exa
mp
les
�E
quiv
ale
nt pre
dic
ate
s
�N
AM
E L
IKE
‘a%
’
�S
UB
ST
R(N
AM
E,1
,1)
= ‘a’
�T
ransfo
rm S
ubsele
ctin
to J
oin
SELECT T1.*
FROM T1
WHERE EXISTS
(SELECT *
FROM T2
WHERE T1.K = T2.K)
SELECT DISTINCT T1.*
FROM T1
, T2
WHERE T1.K = T2.K
�V
iew
Merg
e
�A
llow
s a
dditio
nal jo
ins o
rder
�C
an e
limin
ate
redun
dant
join
s
�R
edundant jo
in e
limin
ation
�S
atisfies m
ultip
le r
efe
rences t
o t
he s
am
e t
ab
le w
ith a
sin
gle
scan
IBM
Softw
are
Gro
up
21
��������
Qu
ery
Re
wri
te:
Su
bq
ue
ry M
ad
ne
ss!
SE
LE
CT
*
FR
OM
( S
ELE
CT
F
LA
G,
TO
_N
UM
BE
R (
NU
M )
NU
M
FR
OM
SU
BT
ES
T
WH
ER
EF
LA
G =
'N' )
WH
ER
EN
UM
> 0
;
�E
RR
OR
: O
RA
-01722: in
valid
num
ber
�R
eason: query
rew
rite
SE
LE
CT
F
LA
G, T
O_N
UM
BE
R (
NU
M )
NU
M
FR
OM
SU
BT
ES
T
WH
ER
EF
LA
G =
'N'
AN
D T
O_N
UM
BE
R (
NU
M )
> 0
;
and p
redic
ate
sequence c
hanged!!!
�“M
odel vs. im
ple
menta
tion is o
ne o
f th
e g
reat lo
gic
al d
iffe
ren
ces”
(C.
Date
)
Zdv
C
428
N
23
N
Abc
C
100
N
NU
MF
LA
G
IBM
Softw
are
Gro
up
22
��������
Op
tim
ize
r E
xp
lain
ed
–1
ase
lect
co
d_n
dg
_co
ntr
op
arte
, co
un
t(*)
fr
om
rai
nb
ow
.co
inte
staz
ion
e ct
wh
ere
exis
ts (
sele
ct *
fr
om
rai
nb
ow
.co
ntr
op
arte
cp
w
her
e cp
.idn
_co
ntr
op
arte
=
ct.id
n_c
oin
test
atar
ioan
dcp
.idn
_cr_
tip
o_c
on
tro
par
te =
421
9)
gro
up
by
cod
_nd
g_c
on
tro
par
te
hav
ing
co
un
t(*)
> 1
0 o
rder
by
2 d
esc
SE
LE
CT
Q4.
$C1
AS
"C
OD
_ND
G_C
ON
TR
OP
AR
TE",
Q4.
$C0
FR
OM
(S
EL
EC
T C
OU
NT
(* )
, Q3.
$C0
FR
OM
(S
ELE
CT
Q2.
CO
D_N
DG
_CO
NT
RO
PA
RT
E
FR
OM
RA
INB
OW
.CO
NTR
OP
AR
TE
AS
Q1
, R
AIN
BO
W.C
OIN
TE
STA
ZIO
NE
A
S Q
2 W
HE
RE
(Q1.
IDN
_CO
NT
RO
PA
RT
E =
Q
2.ID
N_C
OIN
TE
STA
TAR
IO)
AN
D (
Q1.
IDN
_CR
_TIP
O_C
ON
TR
OP
AR
TE
= 42
19))
AS
Q3
GR
OU
P B
Y Q
3.$C
0) A
S Q
4 W
HE
RE
(10
< Q
4.$C
0)
OR
DE
R B
Y Q
4.$C
0 D
ES
C
IBM
Softw
are
Gro
up
23
��������
Op
tim
ize
r E
xp
lain
ed
–1
b (
afte
r R
eo
rg)
sele
ct
cod
_nd
g_c
on
tro
par
te, c
ou
nt(
*)
fro
m r
ain
bo
w.c
oin
test
azio
ne
ctw
her
e ex
ists
(se
lect
*
fro
m r
ain
bo
w.c
on
tro
par
te c
p
wh
ere
cp.id
n_c
on
tro
par
te =
ct
.idn
_co
inte
stat
ario
and
cp.id
n_c
r_ti
po
_co
ntr
op
arte
= 4
219)
g
rou
p b
y co
d_n
dg
_co
ntr
op
arte
h
avin
g c
ou
nt(
*) >
10
ord
er b
y 2
des
c
SE
LE
CT
Q4.
$C1
AS
"C
OD
_ND
G_C
ON
TR
OP
AR
TE",
Q4.
$C0
FR
OM
(S
EL
EC
T C
OU
NT
(* )
, Q3.
$C0
FR
OM
(S
ELE
CT
Q2.
CO
D_N
DG
_CO
NT
RO
PA
RT
E
FR
OM
RA
INB
OW
.CO
NTR
OP
AR
TE
AS
Q1
, R
AIN
BO
W.C
OIN
TE
STA
ZIO
NE
A
S Q
2 W
HE
RE
(Q1.
IDN
_CO
NT
RO
PA
RT
E =
Q
2.ID
N_C
OIN
TE
STA
TAR
IO)
AN
D (
Q1.
IDN
_CR
_TIP
O_C
ON
TR
OP
AR
TE
= 42
19))
AS
Q3
GR
OU
P B
Y Q
3.$C
0) A
S Q
4 W
HE
RE
(10
< Q
4.$C
0)
OR
DE
R B
Y Q
4.$C
0 D
ES
C
IBM
Softw
are
Gro
up
24
��������
Th
e O
ptim
ize
r: S
um
ma
ry
�V
ery
cle
ver
pie
ce o
f code
�B
eh
avio
r heavily
dep
ende
nt
on lo
gic
al a
nd p
hysic
al
desig
n
�N
um
ber
of jo
ins (
logic
al desig
n)
�In
dexes a
nd their c
hara
cte
ristics
�A
lso d
ep
en
dent
on
�R
DB
MS
configura
tion (
e.g
. buffer
siz
e, sort
heap s
ize, etc
.)
�N
um
ber
of C
PU
and C
PU
pow
er
�I/O
configura
tion
IBM
Softw
are
Gro
up
25
��������
DB
A R
ole
: M
ost
rele
va
nt
pe
rfo
rma
nce
kn
ob
s
�R
DB
MS
co
nfigura
tion s
ett
ings
�H
ighly
dependent on n
am
ed R
DB
MS
�Log
ical d
esig
n
�B
e a
ware
of cost of over-
norm
aliz
ation
�H
igh
er
num
ber
of
join
s
�H
igh
er
num
ber
of
join
ed t
ab
les
�P
hysic
al desig
n
�N
um
ber
of in
dexes
�In
dex k
ey
�B
e a
ware
of
the m
ain
ten
ance
cost
of
each inde
x
�G
oal: b
ala
nce t
hro
ug
hput
vs.
sin
gle
query
pe
rform
ance
IBM
Softw
are
Gro
up
26
��������
DB
A R
ole
: A
Me
tho
d A
pp
roa
ch
to
SQ
L O
ptim
iza
tio
n
�E
nsure
the e
ntire
Data
Access L
ogic
is p
art
of th
e S
QL
sta
tem
ent
�A
void
ha
ndlin
g p
redic
ate
s o
r re
latio
nal o
pera
tors
(e.g
. jo
in)
inapplic
atio
n c
od
e,
unle
ss r
eq
uir
ed b
y a
kno
wn
pro
duct
limitation
�R
un O
ptim
izer
Expla
in
�U
nders
tand the p
rovid
ed info
rmation
�U
nders
tand w
heth
er
there
is r
oom
for
impro
vem
ent
�A
t th
e S
QL s
yn
tax level
�A
t th
e D
B d
esig
n leve
l
�U
nders
tand the c
ost of im
ple
menta
tion a
nd p
ote
ntial im
pact on
exis
ting a
pplic
ations.
�E
.g.
addin
g ind
exes
�R
e-o
rderin
g in
de
x c
olu
mns
�Im
ple
ment changes
IBM
Softw
are
Gro
up
27
��������
Th
e I
mp
act
of P
oo
r O
ptim
iza
tio
n
�In
the p
ast, d
ue t
o lim
ited m
em
ory
siz
e
�I/O
bottle
necks
�Long e
lapsed tim
es
�T
oday,
with
very
pow
erf
ul C
PU
s a
nd v
ery
larg
e
mem
ory
siz
e
�H
igh C
PU
utiliz
ation
�T
hro
ughput lo
wer
than e
xpecte
d
�Lockin
g c
onte
ntion
We a
ll te
nd to o
verlook the im
pact of share
d r
esourc
es, due to the
incre
dib
le a
mount of re
sourc
es a
vaila
ble
on o
ur
ow
n s
ingle
-user
PC
IBM
Softw
are
Gro
up
28
��������
Co
nte
nts
Re
vie
w
�R
DB
MS
an
d S
QL L
angua
ge
�S
tatic S
QL a
nd D
ynam
ic S
QL
�M
app
ing t
he P
hysic
al D
esig
n
�Q
uery
Com
pile
r O
verv
iew
�Q
uery
Optim
izatio
n
�A
ccess P
ath
Sele
ction
�Q
uery
Rew
rite
�E
xam
ple
s
�T
he O
ptim
izer
Exp
lain
ed
�T
he R
ole
of th
e D
BA
IBM
Softw
are
Gro
up
29
��������
IBM
Softw
are
Gro
up
30
��������
Bib
liogr
aphy
-1
1.
Selin
ger
et
al.,
“Access P
ath
Sele
ctio
n in a
Rela
tiona
l D
ata
base
Manag
em
ent
Syste
m”,
Sig
mo
d 1
979,
htt
p:/
/ww
w.c
s.b
erk
ele
y.e
du/~
bre
wer/
cs26
2/A
ccessP
ath
2.
Ioannid
is a
nd K
ang,
“Ran
do
miz
ed A
lgorith
ms f
or
Optim
izin
g L
arg
e J
oin
Queries”,
Sig
mod
199
0,
htt
p:/
/ww
w.c
c.g
ate
ch.e
du/c
om
puting/D
ata
base/r
eadin
ggro
up
/art
icle
s/p
31
2-
ioan
nid
is.p
df
3.
Ham
id P
irahesh, T
. Y
. C
liff
Leung,
Waqar
Ha
san,
"A R
ule
Engin
e f
or
Query
T
ransfo
rmation in S
tarb
urs
t an
d I
BM
DB
2 C
/S D
BM
S",
IC
DE
1997,
pp.
391-4
00,
htt
p:/
/ww
w.d
iku
.dk/u
nderv
isnin
g/2
003f/
72
9/p
apers
/rule
_en
gin
e.p
df
4.
Pete
r G
assner,
Gu
y M
. Lohm
an,
K.
Bern
hard
Schie
fer,
Yu
nW
ang,
"Query
Optim
ization in
the I
BM
DB
2 F
am
ily",
Data
Engin
eeri
ng B
ulle
tin 1
6(4
): 4
-18
(1993),
ftp
://f
tp.r
esearc
h.m
icro
soft
.com
/pub/d
ebull/
93D
EC
-CD
IBM
Softw
are
Gro
up
31
��������
Bib
liogr
aphy
-2
1.
Subqu
ery
Mad
ness! at
htt
p:/
/fiv
e.p
airlis
t.net/
pip
erm
ail/
ora
cle
-
art
icle
/20
04/0
00012.h
tml;
htt
p:/
/dba-
ora
cle
.com
/ora
cle
_n
ew
s/2
004
_9_1
4a_
200
4.h
tm;
htt
p:/
/ww
w.d
bd
ebunk.c
om
/pa
ge/p
ag
e/1
35
13
81.h
tm
2.
Cost C
ontr
ol: Insid
e the O
racle
Optim
izer,
htt
p:/
/ww
w.o
racle
.com
/techno
log
y/o
ram
ag/w
ebcolu
mns/2
00
3/t
echart
icle
s/b
urleso
n_cb
o_p
t1.h
tml
3.
DB
2 R
ed
books:
htt
p://w
ww
.re
dbooks.ibm
.co
m/
4.
Ora
cle
White P
apers
:
htt
p:/
/ora
cle
.itt
oolb
ox.c
om
/do
cum
ents
/defa
ult.a
sp?S
ection=
White+
Papers
Related Documents
