Difference and Union of Models,
10 years later
Ivan Porres
Åbo Akademi University
ivan.porres@abo.fi
MODELS 2013 Conference
Miami October 3, 2013
Åbo Akademi University | Domkyrkotorget 3 | 20500 Åbo | Finland
Difference and Union of Models
Marcus Alanen and Ivan Porres
UML 2003 Conference, San Francisco
Difference and Union of Models, 10 years later
2
Motivation: Fine-grained version control
with optimistic locking
Original Model
A
C
B
Designer 1
A
Designer 2
C
A
C
B
D
Final Model
A
C
D
Difference and Union of Models, 10 years later
3
The Context: 2002-2003
§
Martin Fowler: What Is the Point of the UML?, UML 2003
■
§
Jean Bézivin, MDA: From Hype to Hope, and Reality, UML
2003
■
§
“Models as sketches vs formal designs”
“Everything is a model”
Stuart Kent, "Model Driven Engineering", IFM 2002
■
Defines MDE, identifies the need for families of languages,
transformation, aligns metamodelling with formal language
engineering
Difference and Union of Models, 10 years later
4
The Problem
§
Difference of two models: M2 - M1 = Δ
§
Merge of a model and a model difference M1 + Δ = M2
§
Union of Models: Mf = M0 + (M1 - M0) + (M2 - M0 )
§
For models in any MOF-based language
Difference and Union of Models, 10 years later
5
The Problem
designer 2
Original Model
A
original model
A
C
C
A
Designer 2
C
A
C
B
D
A
C
-
B
Designer 1
d2
=
D
B
B
D
Final Model
A
designer 1
C
original model
A
D
C
d1
A
C
-
-
B
original model
A
d2
d1
A
B
-
+
D
-
Difference and Union of Models, 10 years later
B
final model
C
+
-
=
C
=
B
D
6
String-based comparison
§
Same model, serialized as two different strings
<UML:Model xmi.id = '122' name = 'Example Model'/>
...
<UML:Model name = 'Example Model' xmi.id = '122'/>
...
Difference and Union of Models, 10 years later
7
XML-based comparison
§
Same model, serialized as two different XML documents
<UML:Model xmi.id = '122' name = 'Example Model'>
<UML:Namespace.ownedElement>
<UML:Class xmi.id = '123' name = 'Customer'>
<UML:Class xmi.id = '124' name = 'Product'>
...
<UML:Model xmi.id = '122' name = 'Example Model'>
<UML:Namespace.ownedElement>
<UML:Class xmi.id = '124' name = 'Product'>
<UML:Class xmi.id = '123' name = 'Customer'>
...
Difference and Union of Models, 10 years later
8
Underlying Model Structure
§
Typed model elements, attributes store primitive values,
features store links to other model elements
Person
Company
1
+employer
*
+employee
work
namespace
Example:Model
+Name:String
+Address:String
ownedElement
association
employer:AssociationEnd
work:Association
connection
ownedElement
Compay:Class
participant
association
employee:AssociationEnd
ownedElement
connection
participant
Person:Class
owner
feature
ownedElement
Address:Attribute
feature
Name:Attribute
String:Class
type
Difference and Union of Models, 10 years later
9
Underlying Model Structure
§
Features may be ordered (sequences) or unordered (sets)
Person
Company
1
+employer
*
+employee
work
namespace
Example:Model
+Name:String
+Address:String
ownedElement
association
employer:AssociationEnd
work:Association
connection
ownedElement
Compay:Class
TheClass
participant
association
employee:AssociationEnd
ownedElement
connection
participant
Person:Class
+TheMethod(A:Integer,C:Integer)
owner
feature
ownedElement
Address:Attribute
feature
Name:Attribute
String:Class
type
Difference and Union of Models, 10 years later
10
Identification Elements
§
We assume that each model element has a universally
unique, stable identifier
§
We match elements by identity, not similarity
!
Difference and Union of Models, 10 years later
11
Representing Differences as Elementary
Operations
§
The difference of two models is not a model!
Operation
Dual
new(e,t)
del(e,t)
del(e,t)
new(e,t)
set(e,a,v0,v1)
set(e,a,v1,v0)
insert(e0,f,e1)
remove(e0,f,e1)
remove(e0,f,e1)
insert(e0,f,e1)
insertAt(e0,f,e1,i)
removeAt(e0,f,e1,i)
removeAt(e0,f,e1,i)
insertAt(e0,f,e1,i)
Difference and Union of Models, 10 years later
12
Basic Difference Algorithm
§
Mold-Mnew=Δ=[N,U,D]
§
1. Define a mapping from elements in Mold to Mnew
§
2. Define a new operation for each element in Mnew that
does not exist in Mold
§
3. Define Update Operations
§
■
attributes: define a set operation
■
unordered features: insert,remove
■
ordered features: Longest Common Subsequence algorithm
4. Define a delete operation for each element in Mold that
does not exist in Mnew
Difference and Union of Models, 10 years later
13
Example
insert(e, f , et )
remove(e, f , et )
insertAt(e, f , et , i)
removeAt(e, f , et , i)
remove(e, f , et )
insert(e, f , et )
removeAt(e, f , et , i)
insertAt(e, f , et , i)
Table1. The Map Between Operations and Dual Operations.
AClass
Δ = [[ new(Class, u2 ),
new(Generalization, u3 )],
[ insert(u3 , namespace, u0 ),
insert(u3 , parent, u1 ),
insert(u3 , child, u2 ),
insert(u1 , specialization, u3 ),
insert(u0 , ownedElement, u2 ),
insert(u0 , ownedElement, u3 ),
insert(u2 , namespace, u0 ),
insert(u2 , generalization, u3 ),
set(u2 , name, “”, “Sub”)],
[]
]
AClass
Sub
Figure4. Difference Between Two Simple Models.
Difference and Union of Models, 10 years later
composition, due to bidirectionality constraints. The new class u 2 is connected to the
14
Basic Merge Algorithm
Operation O
Dual operation Õ
new(e,t)
del(e,t)
del(e,t)
new(e,t)
set(e, f , vo , vn )
set(e, f , vn , vo )
insert(e, f , et )
remove(e, f , et )
remove(e, f , et )
insert(e, f , et )
insertAt(e, f , et , i) removeAt(e, f , et , i)
removeAt(e, f , eApply
t , i) insertAt(e,
Δ f , et , i)
Table1. The Map Between Operations and Dual Operations.
§
Δ=[N,U,D],
M1 + Δ = M2
■
1. Apply all new operations
■
2. Apply all update operations
■
3. Apply all delete operations
AClass
Δ = [[ new(Class, u2 ),
new(Generalization, u3 )],
[ insert(u3 , namespace, u0 ),
insert(u3 , parent, u1 ),
insert(u3 , child, u2 ),
insert(u1 , specialization, u3 ),
insert(u0 , ownedElement, u2 ),
insert(u0 , ownedElement, u3 ),
insert(u2 , namespace, u0 ),
insert(u2 , generalization, u3 ),
set(u2 , name, “”, “Sub”)],
[]
]
AClass
Sub
Figure4. Difference Between Two Simple Models.
Apply dual(Δ)
composition, due to bidirectionality constraints. The new class u 2 is connected to the
old class u1 via the Generalization element, using its specialization and generalization
features. Since all features start with their respective “zero” value, the name feature of
the new class is also set by the difference.
XMI has facilities for representing arbitrary differences between two models, using an XML element called XMI.difference. The positive operations new, insert and
Difference and Union
of Models,
later in a XMI document using the XMI.add element,15while the
insertAt
can 10
beyears
described
negative operations del, remove and removeAt can be specified using the XMI.delete
element. The set operation can be represented using the XMI.replace element. XMI
Union of Two Models
Original Model
A
C
B
Designer 1
A
Designer 2
C
A
C
B
D
Final Model
A
C
D
§
Assume two developers update a model simultaneously
■
§
M1-Mbase = Δ1, M2-Mbase = Δ2
We want to merge both changes into the base modelHowever, the order of application of differences matters
■
(Mbase + Δ1 ) + Δ2 ≠ (Mbase + Δ2 ) + Δ1
Difference and Union of Models, 10 years later
16
[A, B,C, D]
Example
Figure7. Example of a merge of an ordered sequence. Note how the insertion of B at index 1
pushes the insertion of D from index 2 to index 3.
[A,C]
F1 = [insertAt(B, 1)] .
[A, B,C]
F2 = [insertAt(D, 2)] #
[A, B, D,C]
& F2 = [insertAt(D, 2)]
[A,C, D]
# F1 = [insertAt(B, 1)]
[A, B,C, D]
Difference and Union of Models, 10 years later
17
Δ is naturally preferred. The modification is necessary to avoid errors. We define the
difference minimisation operator ⊗:
Union of Two Models
Δ"2 = ⊗(Δ2 , Δ1 ),
Δ"1 = ⊗(Δ1 , Δ2 )
Now the equation becomes:
§
Solution: a difference
minimization
operator ⊗
"
"
M∪ = Δ1 (Δ2 (Mbase )) = Δ2 (Δ1 (Mbase ))
■ M1-Mbase = Δ2, M2-Mbase = Δ2
This principle■ is
illustrated
in
Figure
5., Without
loss of generality, this paper
Δ2also
’ = ⊗(Δ
Δ
’
=
⊗(Δ
Δ
)
2, Δ1),
1
1
2
discusses only the calculation of Δ"2 = ⊗([C2 , F2 , D2 ], [C1 , F1 , D1 ]).
■ Conflict resolution: (Mbase + Δ1 ) + Δ2’ = (Mbase + Δ2 ) + Δ1’
■
Δ1’ may contain less operations than Δ1
Δ"2 !
"
M∪
!
M1
"
$
# Δ1
#
M2
$
#
Δ1 #
"
!
! Δ2
Mbase
Difference and Union of Models, 10 years later
18
Figure5. The principle of calculating the union of two models, given their base model. Either
Conflict Resolution
§
§
Model Elements
■
new,del: not possible since ids are unique
■
del,del: automatic resolution (no actual conflict)
■
del, update: manual or metamodel specific resolution
Attributes
■
§
set,set: manual or metamodel specific resolution if values are
different
Features
■
update, update: automatic resolution in some cases based on
longest common subsequence, manual or metamodel
specific if update breaks multiplicity constraints
Difference and Union of Models, 10 years later
19
at the same index. Of importance is also to remember that the resulting sequence
F2! (F1 (Forig )) cannot contain the same element twice. The algorithm does not check
Ordered Features
for such possibilities when interleaving the operations, so the resulting sequence must
be inspected for such occurrences, and the user must choose which elements to remove,
before modifying the feature e. f .
Figure 7 shows an example of the algorithm for interleaving ordered features.
[A,C]
F1 = [insertAt(B, 1)] !
" F2 = [insertAt(D, 2)]
F2! = [insertAt(D, 3)] "
!
[A, B,C] [A,C, D]
[A, B,C, D]
Figure7. Example of a merge of an ordered sequence. Note how the insertion of B at index 1
pushes the insertion of D from index 2 to index 3.
4.2 A Version Control System
Difference and Union of Models, 10 years later
20
A Complete Model Union System
1. Model Difference calculation
2. Difference minimization
■
Automatic metamodel-independent conflict resolver
■
■
Automatic metamodel-dependent conflict resolver
■
■
Algorithm presented in the article
Uses knowledge specific to each modeling language
Manual resolution
■
Requires a proper difference visualization and inspection
3. Model union calculation
Difference and Union of Models, 10 years later
21
Limitations
§
Unique, stable, universal model element identifiers across
model revisions
§
Alternative: element matching based on similarity
■
U. Kelte, J. Wehren, J. Niere, A Generic Difference Algorithm
for UML Models
Difference and Union of Models, 10 years later
22
Limitations
§
Metamodel is assumed to be static
§
This is not a realistic assumption, specially for domainspecific modeling languages
■
B. Gruschko, D. Kolovos, R. Paige, Towards synchronizing
models with evolving metamodels
Difference and Union of Models, 10 years later
23
Limitations
§
Metamodel independent algorithms are generally applicable
but do not exploit useful knowledge of the modeling
language in the mapping, conflict resolution and merge
steps
■
S. Nejati et al., Matching and Merging of Statecharts
Specifications
■
J. Kü̈ster, C Gerth, A. Förster, G. Engels, Detecting and
Resolving Process Model Differences in the Absence of a
Change Log
Difference and Union of Models, 10 years later
24
Limitations
§
Standardized representation of differences is not discussed
■
§
A. Cicchetti, D. Di Ruscio, and A. Pierantonio, A Metamodel
Independent Approach to Difference Representation
Manual conflict resolution requires inspection of base
model and differences
■
H. Störrle, Making sense to modelers: Presenting UML class
model differences in prose
Difference and Union of Models, 10 years later
25
Limitations
§
No actual tools are discussed
■
Y. Lin, J.Gray, F. Jouault, DSMDiff: a differentiation tool for
domain-specific models
■
C. Brun, A. Pierantonio, Model Differences in the Eclipse
Modelling Framework, Eclipse EMF Compare
Difference and Union of Models, 10 years later
26
Future work (in 2013)?
§
Better conflict detection
§
Better difference reporting
§
Better tool support for multiple languages
§
Better conflict resolution
§
K. Altmanninger, P. Brosch, G. Kappel, P. Langer,M. Seidl, K.
Wieland, M. Wimme, Why Model Versioning Research is
Needed!? An Experience Report
Difference and Union of Models, 10 years later
27
10 year most influential paper?
§
It describes a problem and solves it
■
§
The problem is relevant and general
It also describes, sometimes implicitly, many other
interesting problems to solve
■
It brings the problem to the attention of a research
community
Difference and Union of Models, 10 years later
28
10 year most influential paper?
§
The right time
■
§
Anybody could have stumbled into and solved the problem
before us
The right forum
■
Alanen and Porres, Difference and union of models, Proc. of
UML 2003, Springer: 240 citations
■
Alanen and Porres, Version control of software models,
Chapter in Advances in UML and XML-Based Software
Evolution, 2004, Idea Group Publishing: 9 citations
Difference and Union of Models, 10 years later
29
10 year most influential paper?
§
Would it have been accepted to MODELS 2013?
■
Of course not, it was published 10 years ago
■
No validation, experimental tool, evaluation results
Difference and Union of Models, 10 years later
30
The 2003 debate, 10 years later
“Models are sketches”
Martin Fowler: What is
the Point of the UML?
■
“Models are formal artifacts”
vs
Jean Bézivin, MDA: From
Hype to Hope, and Reality
We were right (all of us)!
■ UML is the lingua franca for software design sketches
■ Empirical studies show great success in the use of
UML and domain-specific languages in the industry.
Difference and Union of Models, 10 years later
31
Acknowledgments
§
Dr. Marcus Alanen
§
Prof. Ralph Johan Back, Prof. Johan Lilius
§
The MODELS community
Difference and Union of Models, 10 years later
32