Jarrar: RDF Stores: Challenges and Solutions

Mustafa Jarrar
Lecture Notes, Web Data Management (MCOM7348)
University of Birzeit, Palestine
1st Semester, 2013

RDF Stores
Challenges and Solutions
Dr. Mustafa Jarrar
University of Birzeit
mjarrar@birzeit.edu
www.jarrar.info
Jarrar © 2013

1

Watch this lecture and download the slides from
http://jarrar-courses.blogspot.com/2013/11/web-data-management.html

Jarrar © 2013

2

Lecture Outline
Part I: Querying RDF(S P O) tables using SQL
Part 2: Practical Session (RDF graphs )
Part 3: SQL-based RDF Stores

Keywords: RDF, RDF Stores, querying SPO Table, Graph Databases, Querying Graph, self joins,
RDF3X, Oracle Semantic Technology, C-Store, vertical patronizing, MashQL, Graph
Signature, Semantic Web, Data Web,

Jarrar © 2013

3

Part 1
Querying RDF(S P O) tables using SQL

Jarrar © 2013

4

RDF as a Data Model
RDF (Resource Description Framework) is the Data Model
behind the Semantic/Data Web.
RDF is a graph-based data model.
In RDF, Data is represented as triples: <Subject, Predicate,
Object>, or in short: <S,P,O>.

S

P
Jarrar © 2013

O
5

RDF as a Data Model
RDF’s way of representing data as triples is more
elementary than Databases or XML.
–  This enables easy data integration and interoperability
between systems (as will be demonstrated later).
•  EXAMPLE: the fact that the movie Sicko (2007) is directed
by Michael Moore from the USA can be represented by the
following triples which form a directed labeled graph:
<
<
<
<
<
<

M1,
M1,
M1,
D1,
D1,
C1,

name, Sicko >
year, 2007 >
directedBy, D1 >
name, Michael Moore >
country, C1 >
name, USA>

2007
Sicko

name

M1

directe
dBy

Jarrar © 2013

D1

name

Michael
Moore
6

RDF graphs as an SPO table
Typically, an RDF Graph is stored
as one <S,P,O> Table in a
Relational Database Management
System.
2007
Sicko
M1

directedBy

Michael Moore
Emmy Awards

D1
P1

2009

M2

1995

Capitalism
1995
M3

directedBy
actedIn

Washington DC

D2
P2
Brave Heart

M4

2007
directedBy
actedIn

USA

C1

Oscars

Mel Gibson

1996

Nadine Labaki
D3
Caramel

Lebanon

C2
location
C3

P3

Beirut
name

Sweden

Stockholm

Stockholm Festival

2007

Jarrar © 2013

S
M1
M1
M1
M2
M2
M2
M3
M3
M3
M4
M4
M4
D1
D1
D1
D2
D2
D2
D2
D3
D3
D3
D3
…

P
year
Name
directedBy
directedBy
Year
Name
Year
directedBy
Name
Year
directedBy
Name
Name
hasWonPrizeIn
Country
Counrty
hasWonPrizeIn
Name
actedIn
Name
Country
hasWonPrizeIn
actedIn
…

O
2007
Sicko
D1
D1
2009
Capitalism
1995
D2
Brave Heart
2007
D3
Caramel
Michael Moore
P1
C1
C1
P2
Mel Gibson
M3
Nadine Labaki
C2
P3
M4
…

7

RDF graphs as an SPO table
How do we query graph-shaped data
stored in one relational table?
How do we traverse a graph stored
in one relational table?
à Self Joins!

Jarrar © 2013

S
M1
M1
M1
M2
M2
M2
M3
M3
M3
M4
M4
M4
D1
D1
D1
D2
D2
D2
D2
D3
D3
D3
D3
…

P
year
Name
directedBy
directedBy
Year
Name
Year
directedBy
Name
Year
directedBy
Name
Name
hasWonPrizeIn
Country
Counrty
hasWonPrizeIn
Name
actedIn
Name
Country
hasWonPrizeIn
actedIn
…

O
2007
Sicko
D1
D1
2009
Capitalism
1995
D2
Brave Heart
2007
D3
Caramel
Michael Moore
P1
C1
C1
P2
Mel Gibson
M3
Nadine Labaki
C2
P3
M4
…

8

How to Query RDF data stored in SPO table?
S
M1
M1
M1
M2
M2
M2
M3
M3
M3
M4
M4
M4
D1
D1
…

P
year
Name
directedBy
directedBy
Year
Name
Year
directedBy
Name
Year
directedBy
Name
Name
hasWonPrizeIn
…

O
2007
Sicko
D1
D1
2009
Capitalism
1995
D2
Brave Heart
2007
D3
Caramel
Michael Moore
P1
…

2007
Sicko
M1

directedBy

Michael Moore
Emmy Awards

D1
P1

2009

M2

1995

Capitalism
1995
M3

directedBy
actedIn

Washington DC

D2
P2
Brave Heart

M4

2007
directedBy
actedIn

USA

C1

Oscars

Mel Gibson

1996

Nadine Labaki
D3
Caramel

Exercises:

Lebanon

C2
location
C3

P3

Beirut
name

Sweden

Stockholm Festival

2007

Stockholm
(1) What is the name of director D3?
(2) What is the name of the director of the movie M1?
(3) List all the movies who have directors from the USA and their directors.
(4) List all the names of the directors from Lebanon who have won prizes
and the prizes they have won.

Jarrar © 2013

9


Exercise 1: A simple query.
–  What is the name of director D3?
–  Consider the table MoviesTable(s,p,o).
–  SQL:
Select o
From MoviesTable T
Where T.s = ‘D3’ AND T.p = ‘Name’;

Answer: Nadine Labaki
Jarrar © 2013

S
M1
M1
M1
M2
M2
M2
M3
M3
M3
M4
M4
M4
D1
D1
D1
D2
D2
D2
D2
D3
D3
D3
D3
…

P
year
Name
directedBy
directedBy
Year
Name
Year
directedBy
Name
Year
directedBy
Name
Name
hasWonPrizeIn
Country
Counrty
hasWonPrizeIn
Name
actedIn
Name
Country
hasWonPrizeIn
actedIn
…

O
2007
Sicko
D1
D1
2009
Capitalism
1995
D2
Brave Heart
2007
D3
Caramel
Michael Moore
P1
C1
C1
P2
Mel Gibson
M3
Nadine Labaki
C2
P3
M4
…
10


Exercise 2: A path query with 1 join.
–  What is the name of the director of the
movie M1.
–  Consider the table MoviesTable (s,p,o).
–  SQL:
Select T2.o
From MoviesTable T1, MoviesTable T2
Where {T1.s = ‘M1’ AND
T1.p = ‘directedBy’ AND
T1.o = T2.s AND
T2.p = ‘name’};

Answer: Michael Moore

Jarrar © 2013

S
M1
M1
M1
M2
M2
M2
M3
M3
M3
M4
M4
M4
D1
D1
D1
D2
D2
D2
D2
D3
D3
D3
D3
…

P
year
Name
directedBy
directedBy
Year
Name
Year
directedBy
Name
Year
directedBy
Name
Name
hasWonPrizeIn
Country
Counrty
hasWonPrizeIn
Name
actedIn
Name
Country
hasWonPrizeIn
actedIn
…

O
2007
Sicko
D1
D1
2009
Capitalism
1995
D2
Brave Heart
2007
D3
Caramel
Michael Moore
P1
C1
C1
P2
Mel Gibson
M3
Nadine Labaki
C2
P3
M4
…
11

Exercise 3: A path query with 2 joins.
–  List all the movies who have directors from
the USA and their directors.
–  Consider the table MoviesTable(s,p,o).
–  SQL:
Select T1.s, T1.o
From MoviesTable T1, MoviesTable T2,
MoviesTable T3
Where {T1.o = T2.s AND
T2.o = T3.s AND
T1.p = ‘directedBy’ AND
T2.p = ‘country’ AND
T3.p = ‘name’ AND
T3.o = ‘USA’};
Answer: M1 D1; M2 D1; M3 D2

Jarrar © 2013

S
M1
M1
M1
M2
M2
M2
M3
M3
M3
…
D1
D1
D1
D2
D2
D2
D2
…
C1
C1
C2
C2
…

P

O

year
Name
directedBy
directedBy
Year
Name
Year
directedBy
Name
…
Name
hasWonPrizeIn
Country
Counrty
hasWonPrizeIn
Name
actedIn
…
Name
Capital
Name
Capital
…

2007
Sicko
D1
D1
2009
Capitalism
1995
D2
Brave Heart
…
Michael Moore
P1
C1
C1
P2
Mel Gibson
M3
…
USA
Washington DC
Lebanon
Beirut
…
12

Exercise 4: Star query.
–  List all the names of the directors from Lebanon who
have won prizes and the prizes they have won.
–  Consider the table MoviesTable (S,P,O). S
P
Select T1.o, T4.o
From MoviesTable T1, MoviesTable T2,
MoviesTable T3, MoviesTable T4
Where {T1.p = ‘name’ AND
T2.s = T1.s AND
T2.p = ‘country’ AND
T2.o = T3.s AND
T3.p = ‘name’ AND
T3.o = ‘Lebanon’ AND
T4.s = T1.s AND
T4.p = ‘hasWonPrizeIn’};
Answer: ‘Nadine Labaki’ , P3

Jarrar © 2013

D1
D1
D1
D2
D2
D2
D2
D3
D3
D3
D3
…
C1
C1
C2
C2
…

O
Name
Michael Moore
hasWonPrizeIn P1
Country
C1
Counrty
C1
hasWonPrizeIn P2
Name
Mel Gibson
actedIn
M3
Name
Nadine Labaki
Country
C2
hasWonPrizeIn P3
actedIn
M4
…
…
Name
USA
Capital
Washington DC
Name
Lebanon
Capital
Beirut
…
…
13

Part 2:
Querying SPO Tables using SQL

Jarrar © 2013

14

Practical Session
Given the RDF graph in the next slide, do the following:
(1) Insert the data graph as <S,P,O> triples in a 3-column table in any
DBMS.
Schema of the table: Books (Subject, Predicate, Object)
(2) Write the following queries in SQL and execute them over the
newly created table:
•  List all the authors born in a country which has the name
Palestine.
•  List the names of all authors with the name of their affiliation
who are born in a country whose capital’s population is 14M.
Note that the author must have an affiliation.
•  List the names of all books whose authors are born in Lebanon
along with the name of the author.

Jarrar © 2013

15

Palestine
Said

BK1

BK2

CN1

AU1
CU
Author

BK3

The
P rophet

BK4

Author

Name

Name

Jerusalem

Colombia
University
14 .0M

India

CN2
AU3

CA1

Viswanathan

AU2

Wamadat

Author

Capital

7.6K

Capital

CA2

Name

New
Delhi

Naima

Lebanon
Gibran

CN3

AU4

2.0M

CA3

Beirut

This data graph is about books. It talks about four books (BK1-BK4).
Information recorded about a book includes data such as; its author,
affiliation, country of birth including its capital and the population of its
capital.
Jarrar © 2013

16

Practical Session
• Each student should work alone.
• The student is encouraged to first answer each query mentally from the
graph before writing and executing the SQL query, and to compare the
results of the query execution with his/her expected results.
• The student is strongly recommended to write two additional queries,
execute them on the data graph, and hand them along with the required
queries.
• The student must highlight problems and challenges of executing queries
on graph-shaped data and propose solutions to the problem.
• Each student must expect to present his/her queries at class and, more
importantly, he/she must be ready to discuss the problems of querying
graph-shaped data and his/her proposed solutions.
• The final delivery should include (i) a snapshot of the built table, (ii) the
SQL queries, (iii) The results of the queries, and (iv) a one-page discussion
of the problems of querying graph-shaped data and the proposed possible
solutions. These must be handed in a report form in PDF Format.
Jarrar © 2013

17

Part 3
SQL-based RDF Stores
Querying Graph-Shaped Data using SQL

Jarrar © 2013

18

Complexity of querying graph-shaped data
Where does the complexity of querying graph-shaped data
stored in a relational table lie?
The complexity of querying a data graph lies in the many selfjoins that are to be performed on the table.
Accessing multiple properties for a resource requires subjectsubject joins (Star-shaped queries).
Path expressions require subject-object joins.
In general, a query with n edges on an SPO table requires
n-1 self-joins of that table.
Jarrar © 2013

19

Solution-1: Subject-Property Matrixes
Clustered Property Table:

àOracle Solution [3]
implemented in 11g

•  A clustered property table, contains clusters of
properties tend to be deﬁned together.
•  Multiple property tables with different clusters of
properties may be created to optimize queries.
•  A key requirement for this type of property table is
that a particular property may only appear in at most
one property table.
•  Advantage: Some queries can be answered directly
from the property table with no joins.
- Ex: retrieve the title of the book authored in 2001.
Jarrar © 2013

20

Solution-2: Vertically Partitioning
Six Vertically Partitioned Tables:

•  Triples table is rewritten into n two column tables
where n is the number of unique properties.
•  In each of these tables, the ﬁrst column contains the
subjects that deﬁne that property and the second
column contains the object values for those subjects.
•  An experimental implementation of this approach
was performed using an open source columnoriented database system (C-Store).

àC-Store Solution [4]

Jarrar © 2013

21

Solution-3: Indexes and Dictionaries
Build Indexes.
–  On each column: S, P, O
–  On Permutations of two columns: SP, SO, PS, PO, …
–  On Permutation of three columns: SPO, SOP, PSO, POS,
OSP, OPS, …
Dictionary Encoding String Data?
–  Replacing all literals by IDs.
–  Triple stores are compressed.
–  Allows for faster comparison and matching operations.

à RDF3X Solution [5]

Jarrar © 2013

22

Discussion
Advantages and Disadvantages of storing RDF in
other schemas?

What do you prefer?

Any other suggestions to optimize queries over
graph-shaped data?

Jarrar © 2013

23

MashQL (Query Formulation & indexing Graphs)
(Research started at the University of Cyprus, then at Birzeit University)

What about summarizing the data graph, and instead of querying the
original data, we query the summary and get the same results?
What about summarizing 15M SPO triples into less then 500,000?
This is called Graph Signature Indexing. It is currently in research at
Birzeit University.
Initial comparison
between GS performance
and Oracle’s Semantic
Technologies

Jarrar © 2013

24

MashQL (Query Formulation & indexing Graphs)
A graphical query formulation
language for the Data Web
A general structured-data
retrieval language (not merely
an interface)

RDF Input
From:
http://www.site1.com/rdf

Query
Everything
Artist “^Dion”
YearProdYear > 2003
Title

RAlbumTitle

Addresses all of the following challenges:
ü  The user does not know the schema.
ü  There is no offline or inline schemaontology.
ü  The query may involve multiple sources.
ü  The query language is expressive (not a single-purpose interface)
Jarrar © 2013

25

MashQL Example
Celine Dion’s
Albums after
2003?

www.site1.com/rdf
:A1
:A1
:A1
:A1
:A2
:A2

RDF Input
From:

:Title “Taking Chances”
:Artist “Dion C.”
:ProdYear 2007
:Producer “Peer Astrom”
:Title “Monodose”
:Artist “Ziad Rahbani”

www.site2.com/rdf

Query

:B1
:B1
:B1
:B1
:B1
:B2
:B2
:B2
:B2

Everything
Artist “^Dion”
YearProdYear > 2003
Title

RAlbumTitle

rdf:Type bibo:Album
:Title “The prayer”
:Artist “Celine Dion”
:Artist “Josh Groban”
:Year
2008
rdf:Type bibo:Album
:Title “Miracle”
:Author “Celine Dion”
:Year
2004

§  Interactive Query Formulation.
§  MashQL queries are translated into and executed as SPARQL.
Jarrar © 2013

26

MashQL Example
Celine Dion’s
Albums after
2003?

www.site1.com/rdf
:A1
:A1
:A1
:A1
:A2
:A2

RDF Input
From:

:ProdYear 2007

www.site2.com/rdf

Query

:B1
:B1
:B1
:B1
:B1
:B2
:B2
:B2
:B2

Everything

Jarrar © 2013

rdf:Type bibo:Album
:Year
2008
rdf:Type bibo:Album
:Year
2004

27

MashQL Example
Celine Dion’s
Albums after
2003?

www.site1.com/rdf
:A1
:A1
:A1
:A1
:A2
:A2

RDF Input
From:

:ProdYear 2007

www.site2.com/rdf

Query

:B1
:B1
:B1
:B1
:B1
:B2
:B2
:B2
:B2

Everything
Everything
Album
A1
A2
B1
Types Individual
s

rdf:Type bibo:Album
:Year
2008
rdf:Type bibo:Album
:Year
2004

Background queries
SELECT X WHERE {?X ?P ?O}
Union
SELECT X WHERE {?S ?P ?X}
SELECT X WHERE {?S rdf:Type ?X}

Jarrar © 2013

28

MashQL Example
Celine Dion’s
Albums after
2003?

www.site1.com/rdf
:A1
:A1
:A1
:A1
:A2
:A2

RDF Input
From:

:ProdYear 2007

www.site2.com/rdf

Query
Everything
Artist
Artist
Author
Title
ProdYear
Type
Year

Contains Dion
Equals
Equals
Contains
Contains
OneOf
Not
Between
LessTha
n
MoreTha
n

:B1
:B1
:B1
:B1
:B1
:B2
:B2
:B2
:B2

rdf:Type bibo:Album
:Year
2008
rdf:Type bibo:Album
:Year
2004

Background query
SELECT P WHERE {?Everything ?P ?O}

Jarrar © 2013

29

MashQL Example
Celine Dion’s
Albums after
2003?

www.site1.com/rdf
:A1
:A1
:A1
:A1
:A2
:A2

RDF Input
From:

:ProdYear 2007

www.site2.com/rdf

Query
Everything
Artist “^Dion
Year ProdYear
Artist
Author
Title
ProdYear
Type
Year

MoreTh 2003
Equals
Equals
Contains
OneOf
Not
Between
LessTha
MoreThan
n
MoreTha
n

:B1
:B1
:B1
:B1
:B1
:B2
:B2
:B2
:B2

rdf:Type bibo:Album
:Year
2008
rdf:Type bibo:Album
:Year
2004

Background query

Jarrar © 2013

30

MashQL Example
Celine Dion’s
Albums after
2003?

www.site1.com/rdf
:A1
:A1
:A1
:A1
:A2
:A2

RDF Input
From:

:ProdYear 2007

www.site2.com/rdf

Query
Everything
Artist “^Dion
YearProdYear > 2003
Title
Artist
Author
Title
Title
ProdYear
Type
Year

AlbumTitle

:B1
:B1
:B1
:B1
:B1
:B2
:B2
:B2
:B2

rdf:Type bibo:Album
:Year
2008
rdf:Type bibo:Album
:Year
2004

Background query

Jarrar © 2013

31

MashQL Example
Celine Dion’s
Albums after
2003?
RDF Input
From:

Query
Everything
Artist “^Dion”
YearProdYear > 2003
Title

RAlbumTitle

PREFIX S1: <http://site1.com/rdf>
PREFIX S2: <http://site1.com/rdf>
SELECT ?AlbumTitle
FROM <http://site1.com/rdf>
FROM <http://site2.com/rdf>
WHERE {
{?X S1:Artist ?X1} UNION {?X S2:Artist ?X1}
{?X S1:ProdYear ?X2} UNION {?X S2:Year ?X2}
{{?X S1:Title ?AlbumTitle} UNION {?X
S2:Title ?AlbumTitle}}
FILTER regex(?X1, “^Dion”)
FILTER (?X2 > 2003)}

Jarrar © 2013

32

MashQL Query Optimization
•  The major challenge in MashQL is interactivity!
•  User interaction must be within 100ms.
•  However, querying RDF is typically slow.
•  How about dealing with huge datasets!?
•  Implemented Solutions (i.e., Oracle) proved to perform
weakly specially in MashQL background queries.

A query optimization solution was developed.,
called Graph Signature Index.

Jarrar © 2013

33

MashQL Query Optimization: Graph Signature
‘Graph Signature’ optimizes RDF queries by :
Summarizing the RDF dataset
(Algorithms were developed)

Then, executing the queries on the summary
instead of the original dataset
(A new execution plan was developed)

Jarrar © 2013

34

Experiments
•  Experiments performed on 15M SPO triples (rows) of the Yago Dataset.
•  Graph Signature Index of less than 500K was built on the data.
•  Two sets of queries were performed, and the results were as follows:
Extreme-case linear queries of depth 1-5:
Query
GS
Oracle

A1
0.344
110.390

A2
1.953
302.672

A3
5.250
525.844

A4
10.234
702.969

A5
24.672
>20min

Queries that cover all MashQL Background queries some queries that
might occur as the final queries generated in MashQL:
Query
Q1
Q2
Q3
Q4
Q5
Q6
Q7
Q8
Q9
GS
0.441 0.578 0.587 0.556 0.290 0.291 0.587 0.785 2.469
Oracle 25.640 29.875 29.593 29.641 19.922 21.922 62.734 64.813 32.703
Jarrar © 2013

35

References
1.  Anton Deik, Bilal Faraj, Ala Hawash, Mustafa Jarrar:
Towards Query Optimization for the Data Web - Two Disk-Based
algorithms: Trace Equivalence and Bisimilarity.
2.  Mustafa Jarrar and Marios D. Dikaiakos:
A Query Formulation Language for the Data Web. IEEE Transactions
on Knowledge and Data Engineering.IEEE Computer Society.
3.  Chong E, Das S, Eadon G, Srinivasan J: An efficient SQL-based RDF
querying scheme. VLDB’05.
4.  Abadi et al, Scalable Semantic Web Data Management Using Vertical
Partitioning. In VLDB 2007.
5.  Neumann T, Weikum G: RDF3X: RISC style engine for RDF.
VLDB’08.

Jarrar © 2013

36

Jarrar: RDF Stores: Challenges and Solutions

More Related Content

Viewers also liked (16)

Similar to Jarrar: RDF Stores: Challenges and Solutions (20)

Recently uploaded (20)

Jarrar: RDF Stores: Challenges and Solutions