GLOBAL OBJECT INTEGRATION INFRASTRUCTURE
SUPPLEMENTING SOA WITH DATA MANAGEMENT
Perspectives of Creation
Vladimir V. Ovchinnikov, Yuri V. Vakhromeev and Pavel A. Pyatih
Fusionsoft, May 9th St., Lipetsk, Russian Federation
Keywords: Object integration infrastructure, distributed objects, global object space.
Abstract: The paper considers a way of data-centric object integration supplementing SOA with data management.
The key aim of the proposed approach is to provide unlimited scalability of object integration into one
object space, having the classic typification mechanism, the general security and transaction management.
Declarative publishing of objects in the space make it possible to read, refer and modify any of them in the
general way, in secure and transactional manner in spite of where they are stored actually (RDBMS, XML,
etc.).
1 INTRODUCTION
The current level of IT-technologies enables
advanced solutions: complex systems
intercommunicating one with another using open
standards. IT industry has evolved a lot during the
last one-two decades; however information itself
still remains fragmented in the Net. There are mature
tools for human-convenient information
representation by means of a single program – WEB
browser. But you understand that it has nothing in
common with a consistent integration on the level of
information/data; it’s only accessing to systems
through one window. Absence of consistent
integration is apparent when a user, taking some data
from one system, tries to look for related data in an
adjacent system. This is impossible if only both
vendors do not endow the systems with appropriate
functionality initially.
If we would have said that IT society did nothing
to overcome the problem, we would be wrong.
Many things are being done:general system
intercommunication standards and architectures
(SOA, for instance) as well as particular integration
standards for specific application domains are being
developed. However, current efforts do not seem to
be enough since standard development is
complicated by two factors: considerable quantity of
interested parties and high complexity of system
integration technologies.
In this connection the question arises as to
whether the technological component can be
simplified by creating infrastructure for consistent
system integration? Of course, you can say that the
infrastructure already exists – SOA-based solutions.
Indeed, SOA may be used for high-quality
integration of a number of systems under one
program interface, but efficient solution of the task
is only possible on the basis of open standards for
the appropriate application domain, which do not
always exist. If there is no standard, each system
vendor has to rely upon their own solutions. As a
result, we acquire a set of separate systems, and
intercommunication between every pair of them
should be developed anew on both sides.
SOA gives facilities for system integration, but
integration can not always be done in sufficiently
simple and scalable manner. Real simplification of
system integration process may be achieved only
when systems being integrated start working in one
object space with common mechanisms for
accessing and referring to objects. Only this space
may eliminate boundaries between systems by
making the systems operate on common data with
no need to develop separate modules for
intercommunications of adjacent systems. The
integrated object (data) space may simplify the
system integration process significantly. In fact, an
integrated meta-system is being created by that.
272
V. Ovchinnikov V., V. Vakhromeev Y. and A. Pyatih P. (2008).
GLOBAL OBJECT INTEGRATION INFRASTRUCTURE SUPPLEMENTING SOA WITH DATA MANAGEMENT - Perspectives of Creation.
In Proceedings of the Third International Conference on Software and Data Technologies - ISDM/ABF , pages 272-279
DOI: 10.5220/0001882802720279
Copyright
c
SciTePress
Apparently, this data-oriented approach and
SOA, being behavior-oriented, do not contradict but
supplement each other. The integrated object space
brings data-oriented system integration up to the
new technological level since it hides all the data
access specifics and gives the unified interface for
object processing. The space can be extended in
flexible and transparent manner without
implementation of complex integration projects
within existing systems. By the same reasons every
SOA service benefits from the object space: services
can use common data without any restrictions.
The approach of object space creation is not new.
The very approach is used in solutions integrating
several relational schemas into one. In this case a
user interface developer can use the integrated
schema to create seamless GUI and do other data
processing in sufficiently simple manner. But these
solutions have scalability restrictions bounded by a
company or a holding company. It is impossible to
extend the set of spanned systems endlessly; it will
result in efficiency degradation for the whole
integrated system sooner or later. But we tell here
about the integrated object space with unrestricted
scalability which is able to publish any quantity of
objects and systems.
Of course the scale of tasks to be solved for
creation of such global object space raises doubts in
their principal feasibility. First, we should prove that
the space may be created technically. Intuitively, it
seems apparent that it is impossible. Second,
organizational steps of such project are infinite. For
such system to start operating in full, many vendors
should be involved, which requires enormous efforts
from the whole IT society.
We should note that further reasoning about the
integrated object space is not only theoretical. Ideas
discussed here underlie the product EntryService of
Fusionsoft Company which is accessible here:
http://www.fusionsoft-online.com/entryservice.php.
Let’s consider the main doubts arising when we
start talking about the integrated object space on a
global scale. The first doubt is how to ensure
sufficient scalability for such space if there are so
many, inestimable, objects to publish? Whether the
space preserves its efficiency during extension? The
apprehension can only be dispelled in more detailed
discussion below. Here we can give the short
answer: hierarchical organization of objects and
systems, like it is done in X.500/LDAP and DNS,
gives all necessary means to solve the scalability
problem.
The second doubt is what to do with those
objects which are already stored in different data
bases and used within existing systems? For the
objects to be used in creation of new integrated
systems, they should be logically mapped to the
integrated object space without physical copying. In
this case, actions over the objects done within the
space should be immediately mapped to actions over
the stored objects, ensuring information consistency.
This approach lets create new systems which are
initially integrated with other systems without a
separate integration project.
The third doubt is whether it is possible to ensure
the sufficient level of information security for such
integrated object space? The short answer is yes
since objects can be published at an information
owner’s place, and so both special security facilities
as well as standard ones (VPN, etc.) can be
employed.
And the main doubt is how can the object space
be useful for resolving integration problems? If two
systems were created separately, then the answer is
not comforting: they can not communicate one with
another without revising. Revision is necessary
regardless of an integration technology used since
data relations should be created between objects of
different systems.
Then why we introduce some object space if
existing systems can be only integrated using
replication, ETL and other means as before? First,
the approach being discussed is mainly oriented at
development of new systems integrated initially by
the fact of creation. Systems created newly and
working with objects through the object space are
seamlessly integrated with all systems worked with
the same objects before. This is apparent from the
fact that any system may read, modify and refer to
objects regardless of how and where they are stored,
and so end users can get any interesting information
from any system without restrictions.
The above does not mean that the approach can
not be effectively employed for integration of
existing systems which was not built over the
integrated object space initially. Besides data
integration for some systems itself, new user
interfaces are required to work with integrated data
as a whole, only then end users can benefit from it.
Of course, modules for existing applications,
working with data of adjacent system directly, can
be created, but this functionality should be added
twice for every pair of systems being integrated, to
applications of both systems. It is often so even if a
user interface is built for WEB browsers.
This problem can be solved by creation of the
integrated object space over the systems and by
development of user interfaces and data processing
GLOBAL OBJECT INTEGRATION INFRASTRUCTURE SUPPLEMENTING SOA WITH DATA MANAGEMENT -
Perspectives of Creation
273
algorithms over the space. Doing so, we achieve
integrity of user interfaces and make it simple to
extend the cluster of integrated systems. When other
systems are added into the cluster, their objects are
to be published in the object space for end users to
become able to work with them. If new objects are
of the types being already shown in GUI, then no
additional programming is necessary. Of course, if
some types are not covered, then GUI should be
revised once for every new object type added.
There is no need in the integrated object space
when isolated systems are created. But when we
concern integration of several systems into one
meta-system, a consistent object space, based on one
or another technology, becomes needed, especially
when systems being integrated are created at
different time. Having no such integrated object
space, solutions will be scrappy to a greater or lesser
extent.
There may be a lot of different approaches to
organization of an object space. Some of them are
implemented as high-quality products (for instance,
approaches to relational schema integration). Further
we will discuss our vision of this complex question
and try to disclose motives for every architectural
decision made, not claming to uniqueness of
solution.
The key property of our approach is its
unrestricted scalability which can be useful within
one company as well as within the Net as a whole.
Organizational problems concerning the global
character of the object space should be considered
separately and are not discovered here.
2 SCALABILITY OF THE
INTEGRATED OBJECT SPACE
Unrestricted scalability of the integrated object
space was declared above. How can we achieve that
extension of the space will not result in its
degradation in time? Such scalability is possible
only in case of complete diversification of
information flows between the object space and
clients using it. It means that if all data flows go
through one or bounded set of network nodes, then
scalability of such solution will be limited by
bandwidth to these nodes. The limit will be reached
sooner or later. So we conclude that there should not
be such nodes. But how this decision goes with
object space connectedness? Does not it mean that
the object space will become fragmented?
We use the following terminology in our
solution. A part of the integrated object space placed
at one host is an object service providing the
program interface being general for all such
services. Every service manages its own part of the
object space and implements functions for accessing,
modification and referring to objects published with
it. Integrity of the object space goes from
hierarchical service organization with a single root,
which helps to find any service when necessary.
The service hierarchy has part-whole relationship
semantics: a service may be terminal one working
with stored objects directly or intermediate one
combining some other services. Terminal services
manages their objects exclusively, one object can
not be managed by several terminal services.
Intermediate services combine subordinate services
just logically and do all work on them through
subordinate services rather than directly. One
subordinate service may be part of an only
intermediate service.
So all services are parts of the root service
directly or indirectly. Does it mean that we violate
our requirement and the root service is the one
participating in all information flows between clients
and the object space? No, it does not. Existence of
the root service does not forbid using the service
containing an interesting object directly. Then does
it mean that a client should know the service
managing the interesting object before using it? It’s
not true as well.
The program interface of all services has two
parts: inter-service and client. The inter-service part
backs service hierarchy and inter-service
communication, while the client part does not
concern inter-service organization mechanisms at
all. From the client’s point of view every service
gives means to use any object of the object space
with no any restriction: it is sufficient to connect to
any service to get access to any object published.
How can we achieve both accessibility of all objects
and absence of a single node transmitting all
information flows through?
Accessibility of objects through any service is
backed by inter-service communication when an
unknown object requested by a client is to be found.
After an object is found, it is accessed through the
terminal service where the object is stored, which
gives significant diversification of access ways to
objects with known placement. Therefore the object
search algorithm gets critical importance: scalability
of the whole object space depends on diversification
of information flows in the search process.
ICSOFT 2008 - International Conference on Software and Data Technologies
274
One of solutions could be creation of object
registries for every service. Terminal services could
have the registry of objects stored in it directly, and
intermediate services could have the registry of
objects stored in nested services. But this solution
does not provide the necessary level of
diversification for access ways, and here is the
reason.
On the one hand, such registries could ensure for
an object to be found in sufficiently efficient manner
since we could go up the service hierarchy to the
service where the placement of the object is
registered (like searching an address for a domain in
DNS). As a result, the root service would be used
only when objects to be found and the client are in
“opposite” sides of the object space.
On the other hand, this approach does not give
the necessary level of information flow
diversification. If the “opposite” part of the object
space contain a considerable amount of objects
looked for, then density of data flows to the root
service and its nearby services will rise significantly.
As a result, when clients show noticeable activity,
then the object space will degrade inevitably.
Moreover, objects are not as permanent as
associations between domains and addresses.
Objects can be created, deleted, or modified
actively. And having taken into account that this
approach requires to propagate object registries from
terminal services toward the root one, we conclude
that this propagation itself will result in absolute
efficiency degradation for the object space
regardless of any other aspects.
Therefore this approach, based on object
registries, can not be used in the integrated object
space being global. But is there any alternative
which does not result in the object space’s
degradation during extension? Yes, there is.
The key to alternative solution is such way of
object identification within the integrated object
space which ensures efficiency of looking for the
terminal service storing the object. Of course, the
simplest decision here is to use a terminal service’s
address as a part of object identification. But this
decision will impose strict restrictions on object
space flexibility. For instance, if one decides to split
one service on two or to move some objects between
services, then it will require modification of all
references to the objects being moved, which is
impossible in general case since the amount of
references may be enormous and their placement is
not always known. Therefore we can not use the
terminal service as a part of object identification.
But how else can we structure object identifiers
to find their terminal services efficiently? The
answer is usage of hierarchical identification where
all object identifiers form a tree from one root, and
every node of the tree knows the service managing
objects identified by this node and nodes nested to it
directly or indirectly.
How efficient can be search in case of
hierarchical identifier, will it degrade like the
version with object registries? First, the search is
done once for every identification tree node and not
for every object covered by the node, which reduces
information flows significantly. And second, nodes
rarely change association to the services backing
them. It enables local caching of this information
within any service placed on the way between a
client and the root service. Therefore we can
conclude that services can be found using the similar
algorithm as looking for domain in DNS. So service
search efficiency is comparable with domain search
efficiency for DNS.
So the hierarchical identification solution gives
diversification of information flows during search
for a service as well as during work with final
objects . The level of diversification is sufficient for
the integrated object space to function efficiently
with unrestricted scalability.
And now the question arises: how to implement
the identifier hierarchy in the most rational way, so
that efforts for its creation could be minimized and
identifiers could remain stable when moving
objects? When answering the question, we use one
apparent observation: objects and object properties
form the part-whole hierarchy. For instance, an
object of the type “person” can have the property
“passport” which in its turn is an object and can
have its own property “passport number”. So any
object can be considered as a node being parent for
all its nested properties, and each property, in such
interpretation, is an object, and so it can be a parent
node for other properties. And so forth.
If we build such hierarchy from one root object,
then hierarchical identifier of an arbitrary object may
be formed as the path from the root object to the
given one. How should we identify each step оf the
path in that case? If we talk about objects and their
properties, then it is natural to use property name
since it is well-known that properties have unique
names within objects, and so they identify each
branch of the tree unambiguously. Therefore, the
unique identifier for the object “passport number”
may be /People/Vladimir/Passport/Number, where
“People” is the property of the root object,
GLOBAL OBJECT INTEGRATION INFRASTRUCTURE SUPPLEMENTING SOA WITH DATA MANAGEMENT -
Perspectives of Creation
275
“Vladimir” is the property of the object “People”,
and so on.
Such approach let us reduce the identifier
hierarchy to the object hierarchy and to avoid
artificiality of the identifier hierarchy. Identifiers
acquire substantial meaning and may be used for
referring by systems as well as by people. Efforts for
creation of such hierarchy are equal to efforts for
creation of the object structure, no more.
So, taking into account these additions
concerning identifier organization, we can ascertain
that the whole object space, being the object
hierarchy now, can be split on several sub-trees
having its root object associated with some service
used to store this root and nested objects. Sub-trees
can be extracted in arbitrary way with the single
restriction: every object should be associated with
one of services directly or through its ancestors.
Therefore, it is possible the situation when sub-trees
form another tree, where some branches give rise to
other sub-trees for other services. As a result, the
same service may be intermediate, having some
subordinate services, as well as terminal, storing
some objects directly. It gives significant flexibility
of distribution of objects among services, and so it
gives load-balancing flexibility and high scalability
of the object space as a whole.
Sub-trees may be joined and split on parts
without changes in object identification. As a result,
identifiers remain stable when moving objects. It is
sufficient to assign a new service to some higher
object to get access to it and all subordinate objects
through this service, without changes of identifiers.
By that we acquire higher flexibility of object
distribution among services.
Also it is important consequence that the object
hierarchy can be used to solve research tasks and not
only getting of specific objects by known identifiers.
Thus a user can browse the object tree investigating
what objects exist and what properties they have.
Moreover, the hierarchy can be overbuilt with a
query language like X-Path/X-Query. Efficient
execution of queries of such language is a separate
question, and it requires of course corrections in
light of global character of the object space.
3 OBJECT TYPIFICATION
The object space can have no typification
theoretically. But how to understand that two objects
describe the same entity, for instance, people, in this
case? Since every object has properties, we can try
to extract an entity from the fact of presence of some
typical properties, for example, for the entity
“Person”, the properties could be “Name” and
“Age”. But stars have names and ages too, which
does not mean that stars are people. Therefore
properties are not reliable for this.
There is only method to resolve the problem of
associating objects to entities: we should introduce
object typification. Every object in this case should
be related to some type which should
unambiguously determine the appropriate entity of
the real world. For instance, objects of the type
“Person” are not starts, but are people, and vice
versa. Object typification defines the way of
interpretation for all objects of some type. For
example, the task of analysis of social bounds has
sense for people, and the task of distance analysis
has sense for stars, but not vice versa.
So, to understand entities related to objects, we
should define a type for every object existing in our
space. But it is not enough. Correct interpretation of
an object means correct interpretation of its
properties, which requires that the same property
should be equally named for all objects of the same
type. Therefore, every type description should
contain the structure of typical object of the type.
Only in this case all objects of one type may be
processed in common way, reasoning from
commonness of their structure.
How should we store object types in the object
space to access them efficiently, not less efficient
than for objects themselves? The solution is simple:
we should represent object types as a special kind of
objects. Object types, being objects simultaneously,
relate to the special type corresponding to the entity
“Type”. Such recursive type definition let process
them as any other object.
What if an entity is a particular case of some
other entity, for instance, the entity “Sportsman” is a
particular case of the entity “Person”: what type
should be defined for the particular entity in this
case? The problem is that we can not relate one
object to two types simultaneously, but any object of
the type “Sportsman” should be related to the type
“Person”, what can we do? To resolve the
contradiction the concept of type inheritance should
be introduced.
Inheritance of some derived type from some base
type means that every object of the derived type is
also an object of the base type by definition, and so
it has all the properties of the base type. In our
example, we should introduce inheritance of the type
“Sportsman” from the type “Person”. When
inheriting, there is no necessity to repeat the
structure of the base type within its derived types:
ICSOFT 2008 - International Conference on Software and Data Technologies
276
inheritance of properties is implied. Therefore, in
spite of the fact that the type “Sportsman” has only
additional properties, like a sport kind and
achievements, every object of the type “Sportsman”
also has all properties of the type “Person”, like a
name and age.
So to find out is an object related to some type, it
is insufficient to get the object’s own type: all
ascendants of the object’s own type should be
analyzed too. If the ascendant type set contains the
requested type or the requested type is equal to the
object’s own type, then the object is considered to be
related to this type regardless of its own type can be
not the same. In our example, according to this rule,
every object of the type “Sportsman” is an object of
the type “Person” inevitably.
May the object space contain several types for
the same entity? Yes, it is possible if types are
developed by different companies with no
coordination. In this case, objects being of the same
entity but of different types will be interpreted
differently. The object space, being integrated
physically, starts splitting on parts unconnected one
with another logically.
There are two solutions for the problem:
organizational and technical. The organizational
solution includes preliminary standardization of
types by some standardization institution. As a
result, different companies use the same set of types
initially. This approach saves resources for applied
system development dramatically, but it is not
always possible since standard development requires
noticeable amount of time and coordination among
many interested parties. So, standards are created, as
a rule, after several alternative solutions enter the
market.
The technical solution may be applied even after
system creation, but it is much more complicated.
According to this approach, a standard or a private
treaty among companies is created on the basis of
existing scattered types. The standard defines a new
set of standard types covering the appropriate
application domain in full. This organizational step
is inevitable if we wish to get complete system
integration.
Further, every system using its own types should
inherit the types from standard ones; a module
backing the inheritance should be added. This task
does not require modifying a system itself, it can be
solved on the level of the service integrating the
system to the object space by providing access to
objects of the system. As a result, every special
system uses its own types to manage its objects, but
it is also possible to use the objects out of the system
with standard types. It allows us decouple next
integration steps for different systems.
The third step is the most complicated one: a
new version, based on standard types only, is to be
created for every system. Of course, there is an
alternative: to convert external objects of standard
types to internal types of the system (using ETL or
“on the fly”). But this solution has limited
scalability, timeliness, and efficiency since the
quantity of objects to convert may grow
uncontrollably. Efficiency of the system may
degrade completely in time, and administration of
the converting process may become overly
expensive. This approach, of course, can be used in
some special cases, when, for instance, the system is
not expected to evolve, but it’s more long-sighted
and scalable to issue a new version of the system.
At the end of the third step, after we have used
standard types, we should find objects which
instances were multiplied in different systems,
reduce the instances to one for every object by
removing surplus ones, and replace references to
surplus instances by references to chosen master
instances. As a result, we restore integrity of the
object space: every object is represented with one
instance of a standard type, and objects are
interpreted in the same way in all systems.
Here the question arises: could such integration
result in reducing efficiency of existing systems
because we start using remote objects? On the one
hand, of course, yes: integration requires some
additional resources, it is inevitable. On the other
hand, efficiency problems may be resolved after, for
instance, a local service caching often used objects
may be installed. As a result, somewhat reduction of
remote object access efficiency becomes apparent
only in case of active modifications of the objects by
the remote side.
It is evident that the approach described above is
allied to approaches of object-oriented
programming, conceptual schema design, relational
schema integration, ontology design and fusion, but
we can not affirm that it is one of them. The
approach absorbed all that is necessary for
qualitative system integration within one integrated
object space.
4 SECURITY OF THE OBJECT
SPACE
The integrated object space may be used for solving
critical tasks only if it gives flexible and reliable
GLOBAL OBJECT INTEGRATION INFRASTRUCTURE SUPPLEMENTING SOA WITH DATA MANAGEMENT -
Perspectives of Creation
277
technology for ensuring object access security. First
of all, it requires permission management for
reading, referring and modifying objects.
There are three levels of permission
management: object space, intersystem and network.
The object space’s security management enables for
authorized people to do certain activity over some
objects. When necessary, permissions can be
logically propagated down the object tree to simplify
administration. It lets us describe the most general
permissions once on higher levels of the object tree
(as a rule, permissions of reading and referring to
public objects), and special permissions on lower
levels for nested objects.
Before permissions can be verified, a service
should know the user, on behalf of which some
activity is being carried out. For that, the user should
connect to the object space and go through the
process of authentication. But where to store user
information, taking into account that storing user
names and passwords at one place is not convenient,
scalable, or reliable solution? In this case, by
increasing the quantity of users, their names will
tend to be duplicated. Moreover, access to user
information will be done though a single node.
Loosing connection with this node will result in
unavailability of the object space for all users in the
whole. So, the solution with a single node is not
appropriate at all.
In our solution, users of the global object space
are described as objects of the special type, and the
objects can be stored at any part of the space. Of
course, the solution does no exclude usage of one or
several services for centralized user management.
But at the same time, any company, which decides
to manage user accounts by any reason, can do that
within its own services. One of motives of such
decision may be ensuring accessibility of the object
space for some users, for instance, for employees of
the company.
It is evident that user passwords should be
encrypted by means of one of known digest
algorithms proved their effectiveness. In our
solution, for instance, a user can choose the
algorithm in case he (she) wants to choose.
To simplify administration, the object space
should let group users into roles: groups of users
solving similar tasks. It allows permission
management on the level of roles, not separate users,
which increases flexibility of the administration
process. So, for an administrator to give some
standard permission set to a user, it is sufficient to
give the role having the permissions.
The second level of permission management is
the systems, having services operating over them,
themselves. The systems may have their own
mechanisms of authentication and authorization, as a
rule, in case of DBMS usage. The object space
should give the possibility to use the mechanisms as
another security layer.
At this level, a name and a password to access a
certain external system may be assigned for a role or
a user. To make administration simpler, the
assignment can be done automatically with a special
tool. Then an administrator can manage permissions
for the role/user using system-dependent tools on the
level of the system itself, for instance, DBMS. If a
permission set should be given to any user
connected to a certain external system, then a guest
account should be provided for the system, which is
to be used for unregistered users.
And the last level of security is the network
level. Let some company consider necessary to
create a system as a part of the integrated object
space, so that it can refer and use external objects
published globally, but intersystem objects are to be
accessible within local network (or VPN) only, for
example, because of financial character of the
system. In this case, access to the local service can
be restricted not only on the level of the object space
or DBMS, but also on the network level by
forbidding any external connections to the object
space service. As a result, the local service will be
operating as a part of the global object space, but it
will not be accessible from the outside.
5 TRANSACTIONS IN THE
OBJECT SPACE
The object space can be considered as completely
integrated only if it provides a single transaction
management mechanism. A user (or a client
program working on behalf of the user) should have
possibility to make a sequence of changes in one
transaction regardless of whether objects are within
one or several services. Neither user nor client
developer should be concerned about whether a
transaction is distributed or local: it is to be
ascertained by actual distribution of objects being
changed.
Therefore, all services of the object space use a
family of transaction managers comparable each
with other, having common rules of
intercommunication, including the rules of
transaction identification, two-phase distributed
ICSOFT 2008 - International Conference on Software and Data Technologies
278
transaction commitment, etc. By initiating a
sequence of object changes using some service, the
client activates the transaction manager of the
service. If objects of different services were affected,
transaction managers of the services get,
transparently, the context of the distributed
transaction the changes are within, and manage the
local changes as a part of the distributed transaction.
Distributed transaction management is to be done
according to the model X/Open DTP. The model
says that if a system storing objects has its own
transaction manager, for instance, DBMS, then
transaction control over the objects may be
delegated to the manager. In this case, the
appropriate service’s transaction manager may do no
work itself, but delegate it to the transaction
manager of the subordinate system.
Questions of reliability and transaction durability
may be solved on the level of a service’s transaction
manager if it does full management, journaling etc.,
or else on the level of the subordinate system’s
transaction manager.
6 SUMMARY
We hope that we proved technical practicability of
creation of the global object space. Ideas stated here
were checked in practice and were used in the
product EntryService of the Fusionsoft company,
which you can get here: http://www.fusionsoft-
online.com/entryservice.php. We are using the
product for creation of the object space now. We
welcome joining us in this task: info@fusionsoft-
online.com.
GLOBAL OBJECT INTEGRATION INFRASTRUCTURE SUPPLEMENTING SOA WITH DATA MANAGEMENT -
Perspectives of Creation
279