DECOMPOSITION &

SCHEMA NORMALIZATION

CS 564- Fall 2016

ACKs: Dan Suciu, Jignesh Patel, AnHai Doan

HOW TO BUILD A DB APPLICATION
• Pick an application
• Figure out what to model (ER model)
– Output: ER diagram
• Transform the ER diagram to a relational schema

• Refine the relational schema (normalization)

• Now ready to implement the schema and load the

data!

CS 564 [Fall 2016] - Paris Koutris 2

DB DESIGN THEORY

• Helps us identify the “bad” schemas and improve them

1. express constraints on the data: functional
dependencies (FDs)
2. use the FDs to decompose the relations

• The process, called normalization, obtains a schema in a

“normal form” that guarantees certain properties
– examples of normal forms: BCNF, 3NF, …

CS 564 [Fall 2016] - Paris Koutris 3

SCHEMA DECOMPOSITION

CS 564 [Fall 2016] - Paris Koutris 4

WHAT IS DECOMPOSITION?
We decompose R(A1, …, An) by creating
• R1(B1, .., Bm)
• R2(C1,…, Cl)
• where {𝐵# , … , 𝐵& } ∪ {𝐶# , … , 𝐶+ } = {𝐴# , … 𝐴. }

• The instance of R1 is the projection of R onto B1, .., Bm

• The instance of R2 is the projection of R onto C1, .., Cl

CS 564 [Fall 2016] - Paris Koutris 5

EXAMPLE: DECOMPOSITION
SSN name age phoneNumber
934729837 Paris 24 608-374-8422
934729837 Paris 24 603-534-8399
123123645 John 30 608-321-1163
384475687 Arun 20 206-473-8221

SSN name age SSN phoneNumber

934729837 Paris 24 934729837 608-374-8422
123123645 John 30 934729837 603-534-8399
384475687 Arun 20 123123645 608-321-1163
384475687 206-473-8221

CS 564 [Fall 2016] - Paris Koutris 6

DECOMPOSITION DESIDERATA

What should a good decomposition achieve?

1. minimize redundancy
2. avoid information loss (lossless-join)
3. preserve the FDs (dependency preserving)
4. ensure good query performance

CS 564 [Fall 2016] - Paris Koutris 7

EXAMPLE: INFORMATION LOSS
name age phoneNumber
Paris 24 608-374-8422
John 24 608-321-1163 Decompose into:
Arun 20 206-473-8221 R1(name, age)
R2(age, phoneNumber)

name age age phoneNumber

Paris 24 24 608-374-8422
John 24 24 608-321-1163
Arun 20 20 206-473-8221

Can we put it back together?

CS 564 [Fall 2016] - Paris Koutris 8
LOSSLESS-JOIN DECOMPOSITION
R(A, B, C)
decompose (projection)

R1(A, B) R2(B, C)

recover (natural join)

R’(A, B, C)

A schema decomposition is lossless-join if for any

initial instance R, R = R’
CS 564 [Fall 2016] - Paris Koutris 9
LOSSLESS-JOIN CRITERION
• relation R(A) + set F of FDs
• decomposition of R into R1(A1) and R2(A2)

A decomposition is lossless-join if and only if at least one of

the FDs is in F+ (the closure of F) :
1. 𝑨𝟏 ∩ 𝑨𝟐 ⟶ 𝑨𝟏
2. 𝑨𝟏 ∩ 𝑨𝟐 ⟶ 𝑨𝟐

CS 564 [Fall 2016] - Paris Koutris 10

EXAMPLE
• relation R(A, B, C, D)
• FD 𝐴 ⟶ 𝐵, 𝐶

lossless-join
• decomposition into R1(A, B, C) and R2(A, D)

Not lossless-join
• decomposition into R1(A, B, C) and R2(D)

CS 564 [Fall 2016] - Paris Koutris 11

DEPENDENCY PRESERVING
Given R and a set of FDs F, we decompose R into R1
and R2. Suppose:
– R1 has a set of FDs F1
– R2 has a set of FDs F2
– F1 and F2 are computed from F

A decomposition is dependency preserving if by

enforcing F1 over R1 and F2 over R2, we can enforce F
over R
CS 564 [Fall 2016] - Paris Koutris 12
GOOD EXAMPLE
Person(SSN, name, age, canDrink)
• 𝑆𝑆𝑁 ⟶ 𝑛𝑎𝑚𝑒, 𝑎𝑔𝑒
• 𝑎𝑔𝑒 ⟶ 𝑐𝑎𝑛𝐷𝑟𝑖𝑛𝑘

decomposes into
• R1(SSN, name, age)
– 𝑆𝑆𝑁 ⟶ 𝑛𝑎𝑚𝑒, 𝑎𝑔𝑒
• R2(age, canDrink)
– 𝑎𝑔𝑒 ⟶ 𝑐𝑎𝑛𝐷𝑟𝑖𝑛𝑘
CS 564 [Fall 2016] - Paris Koutris 13
BAD EXAMPLE
R1 R2
R(A, B, C)
A B A C
• 𝐴⟶𝐵 a1 b a1 c
• 𝐵, 𝐶 ⟶ 𝐴 a2 b a2 c

Decomposes into: recover

• R1(A, B) A B C
–𝐴⟶𝐵 a1 b c

• R2(A, C) a2 b c

– no FDs here!! The recovered table

violates 𝐵, 𝐶 ⟶ 𝐴
CS 564 [Fall 2016] - Paris Koutris 14
NORMAL FORMS
A normal form represents a “good” schema design:

• 1NF (flat tables/atomic values)

• 2NF
• 3NF more
restrictive
• BCNF
• 4NF
• …

CS 564 [Fall 2016] - Paris Koutris 15

BCNF DECOMPOSITION

CS 564 [Fall 2016] - Paris Koutris 16

BOYCE-CODD NORMAL FORM (BCNF)

A relation R is in BCNF if whenever 𝑋 ⟶ 𝐵 is

a non-trivial FD, then X is a superkey in R

Equivalent definition: for every attribute set X

• either 𝑋 D = 𝑋
• or 𝑋 D = 𝑎𝑙𝑙 𝑎𝑡𝑡𝑟𝑖𝑏𝑢𝑡𝑒𝑠

CS 564 [Fall 2016] - Paris Koutris 17

BCNF EXAMPLE 1
SSN name age phoneNumber
934729837 Paris 24 608-374-8422
934729837 Paris 24 603-534-8399
123123645 John 30 608-321-1163
384475687 Arun 20 206-473-8221

𝑆𝑆𝑁 ⟶ 𝑛𝑎𝑚𝑒, 𝑎𝑔𝑒

• key = {𝑆𝑆𝑁, 𝑝ℎ𝑜𝑛𝑒𝑁𝑢𝑚𝑏𝑒𝑟}
• 𝑆𝑆𝑁 ⟶ 𝑛𝑎𝑚𝑒, 𝑎𝑔𝑒 is a “bad” FD
• The above relation is not in BCNF!

CS 564 [Fall 2016] - Paris Koutris 18

BCNF EXAMPLE 2
SSN name age
934729837 Paris 24
123123645 John 30
384475687 Arun 20

𝑆𝑆𝑁 ⟶ 𝑛𝑎𝑚𝑒, 𝑎𝑔𝑒

• key = {𝑆𝑆𝑁}
• The above relation is in BCNF!

CS 564 [Fall 2016] - Paris Koutris 19

BCNF EXAMPLE 3
SSN phoneNumber
934729837 608-374-8422
934729837 603-534-8399
123123645 608-321-1163
384475687 206-473-8221

• key = {𝑆𝑆𝑁, 𝑝ℎ𝑜𝑛𝑒𝑁𝑢𝑚𝑏𝑒𝑟}

• The above relation is in BCNF!
• Q: is it possible that a binary relation is not in
BCNF?
CS 564 [Fall 2016] - Paris Koutris 20
BCNF DECOMPOSITION
• Find an FD that violates the BCNF condition
𝐴# , 𝐴M , … , 𝐴. ⟶ 𝐵# , 𝐵M , …, 𝐵&
• Decompose R to R1 and R2:

R1 R2
B’s A’s remaining
attributes

• Continue until no BCNF violations are left

CS 564 [Fall 2016] - Paris Koutris 21
EXAMPLE
SSN name age phoneNumber
934729837 Paris 24 608-374-8422
934729837 Paris 24 603-534-8399
123123645 John 30 608-321-1163
384475687 Arun 20 206-473-8221

• The FD 𝑆𝑆𝑁 ⟶ 𝑛𝑎𝑚𝑒, 𝑎𝑔𝑒 violates BCNF

• Split into two relations R1, R2 as follows:
R1 R2
name
SSN phoneNumber
age

CS 564 [Fall 2016] - Paris Koutris 22

 EXAMPLE CONT’D

R1 R2
name
SSN phoneNumber
age

𝑆𝑆𝑁 ⟶ 𝑛𝑎𝑚𝑒, 𝑎𝑔𝑒

SSN name age SSN phoneNumber
934729837 Paris 24 934729837 608-374-8422
123123645 John 30 934729837 603-534-8399
384475687 Arun 20 123123645 608-321-1163
384475687 206-473-8221

CS 564 [Fall 2016] - Paris Koutris 23

BCNF DECOMPOSITION PROPERTIES

BCNF decomposition:
– removes certain types of redundancy
– is lossless-join
– is not always dependency preserving

CS 564 [Fall 2016] - Paris Koutris 24

BCNF IS LOSSLESS-JOIN
Example:
R(A, B, C) with 𝐴 ⟶ 𝐵 decomposes into:
R1(A, B) and R2(A, C)

• BCNF decomposition satisfies the lossless-join

criterion!

CS 564 [Fall 2016] - Paris Koutris 25

BCNF IS NOT DEPENDENCY PRESERVING

R(A, B, C)
• 𝐴⟶𝐵
• 𝐵, 𝐶 ⟶ 𝐴

The BCNF decomposition is:

• R1(A, B) with FD 𝐴 ⟶ 𝐵
• R2(A, C) with no FDs

There may not exist any BCNF decomposition

that is FD preserving!
CS 564 [Fall 2016] - Paris Koutris 26
BCNF EXAMPLE (1)
Books (author, gender, booktitle, genre, price)
• 𝑎𝑢𝑡ℎ𝑜𝑟 ⟶ 𝑔𝑒𝑛𝑑𝑒𝑟
• 𝑏𝑜𝑜𝑘𝑡𝑖𝑡𝑙𝑒 ⟶ 𝑔𝑒𝑛𝑟𝑒, 𝑝𝑟𝑖𝑐𝑒

What is the candidate key?

• (author, booktitle) is the only one!

Is is in BCNF?
• No, because the left hand side of both (not trivial) FDs
is not a superkey!

CS 564 [Fall 2016] - Paris Koutris 27

BCNF EXAMPLE (2)
Books (author, gender, booktitle, genre, price)
• 𝑎𝑢𝑡ℎ𝑜𝑟 ⟶ 𝑔𝑒𝑛𝑑𝑒𝑟
• 𝑏𝑜𝑜𝑘𝑡𝑖𝑡𝑙𝑒 ⟶ 𝑔𝑒𝑛𝑟𝑒, 𝑝𝑟𝑖𝑐𝑒

Splitting Books using the FD 𝑎𝑢𝑡ℎ𝑜𝑟 ⟶ 𝑔𝑒𝑛𝑑𝑒𝑟:

• Author (author, gender)
FD: 𝑎𝑢𝑡ℎ𝑜𝑟 ⟶ 𝑔𝑒𝑛𝑑𝑒𝑟 in BCNF!
• Books2 (authos, booktitle, genre, price)
FD: 𝑏𝑜𝑜𝑘𝑡𝑖𝑡𝑙𝑒 ⟶ 𝑔𝑒𝑛𝑟𝑒, 𝑝𝑟𝑖𝑐𝑒 not in BCNF!

CS 564 [Fall 2016] - Paris Koutris 28

BCNF EXAMPLE (3)
Books (author, gender, booktitle, genre, price)
• 𝑎𝑢𝑡ℎ𝑜𝑟 ⟶ 𝑔𝑒𝑛𝑑𝑒𝑟
• 𝑏𝑜𝑜𝑘𝑡𝑖𝑡𝑙𝑒 ⟶ 𝑔𝑒𝑛𝑟𝑒, 𝑝𝑟𝑖𝑐𝑒

Splitting Books using the FD 𝑎𝑢𝑡ℎ𝑜𝑟 ⟶ 𝑔𝑒𝑛𝑑𝑒𝑟:

• Author (author, gender)
FD: 𝑎𝑢𝑡ℎ𝑜𝑟 ⟶ 𝑔𝑒𝑛𝑑𝑒𝑟 in BCNF!

• Splitting Books2 (author, booktitle, genre, price):

– BookInfo (booktitle, genre, price)
FD: 𝑏𝑜𝑜𝑘𝑡𝑖𝑡𝑙𝑒 ⟶ 𝑔𝑒𝑛𝑟𝑒, 𝑝𝑟𝑖𝑐𝑒 in BCNF!
– BookAuthor (author, booktitle) in BCNF!

CS 564 [Fall 2016] - Paris Koutris 29

THIRD NORMAL FORM (3NF)

CS 564 [Fall 2016] - Paris Koutris 30

3NF DEFINITION

A relation R is in 3NF if whenever 𝑋 ⟶ 𝐴, one

of the following is true:
• 𝐴 ∈ 𝑋 (trivial FD)
• X is a superkey
• A is part of some key of R (prime attribute)

BCNF implies 3NF

CS 564 [Fall 2016] - Paris Koutris 31

3NF CONT’D
• Example: R(A, B, C) with 𝐴, 𝐵 ⟶ 𝐶 and 𝐶 ⟶ 𝐴
– is in 3NF. Why?
– is not in BCNF. Why?

• Compromise used when BCNF not achievable: aim

for BCNF and settle for 3NF
• Lossless-join and dependency preserving
decomposition into a collection of 3NF relations is
always possible!
CS 564 [Fall 2016] - Paris Koutris 32
3NF ALGORITHM

1. Apply the algorithm for BCNF decomposition until all

relations are in 3NF (we can stop earlier than BCNF)
2. Compute a minimal basis F’ of F
3. For each non-preserved FD 𝑋 ⟶ 𝐴 in F’, add a new
relation R(X, A)

CS 564 [Fall 2016] - Paris Koutris 33

 3NF EXAMPLE (1)
Start with relation R (A, B, C, D) with FDs:
• 𝐴⟶𝐷
• 𝐴, 𝐵 ⟶ 𝐶
• 𝐴, 𝐷 ⟶ 𝐶
• 𝐵⟶𝐶
• 𝐷 ⟶ 𝐴, 𝐵

Step 1: find a BCNF decomposition

• R1 (B, C)
• R2 (A, B, D)

CS 564 [Fall 2016] - Paris Koutris 34

 3NF EXAMPLE (2)
Start with relation R (A, B, C, D) with FDs:
• 𝐴 ⟶𝐷
• 𝐴, 𝐵 ⟶ 𝐶
• 𝐴, 𝐷 ⟶ 𝐶
• 𝐵⟶𝐶
• 𝐷 ⟶ 𝐴, 𝐵

Step 2: compute a minimal basis of the original set of FDs:

• 𝐴⟶𝐷
• 𝐵⟶𝐶
• 𝐷⟶𝐴
• 𝐷⟶𝐵

CS 564 [Fall 2016] - Paris Koutris 35

 3NF EXAMPLE (3)
Start with relation R (A, B, C, D) with FDs:
• 𝐴⟶𝐷
• 𝐴, 𝐵 ⟶ 𝐶
• 𝐴, 𝐷 ⟶ 𝐶
• 𝐵⟶𝐶
• 𝐷 ⟶ 𝐴, 𝐵

Step 3: add a new relation for any FD in the basis

that is not satisfied:
• all the dependencies in F’ are satisfied!
• the resulting decomposition R1, R2 is also BCNF!

CS 564 [Fall 2016] - Paris Koutris 36

IS NORMALIZATION ALWAYS GOOD?

• Example: suppose A and B are always used

together, but normalization says they should be in
different tables
– decomposition might produce unacceptable
performance loss
• Example: data warehouses
– huge historical DBs, rarely updated after creation
– joins expensive or impractical

CS 564 [Fall 2016] - Paris Koutris 37

RECAP
• Bad schemas lead to redundancy
• To “correct” bad schemas: decompose relations
– lossless-join
– dependency preserving
• Desired normal forms
– BCNF: only superkey FDs
– 3NF: superkey FDs + dependencies with prime
attributes on the RHS

CS 564 [Fall 2016] - Paris Koutris 38