In Chapter 3 of his PhD thesis, Kiepuszewski lists requirements for
workflow languages through workflow patterns.
Much like the software design patterns, these workflow patterns describe
recurring solutions to common problems. They are relevant to both the
implementor and the user of a workflow management system. The former uses the
workflow patterns as a common vocabulary for workflow description language
constructs and to define the semantics of a workflow model whereas the latter
uses them as a guide while formulating his workflow in the workflow system's
description language. The workflow patterns also faciliate the comparison with
other workflow systems with regard to expressiveness and power.
In Chapter 4 of his PhD thesis, Kiepuszewski maps the workflow patterns that
he identified to Petri nets to provide a formal foundation for this more
pragmatic approach to defining workflow semantics.
In this section we discuss the subset of the workflow patterns identified by
Kiepuszewski that is directly supported by the Workflow component.
The workflow patterns for basic control flow capture elementary aspects of
process control and closely match the definitions of elementary
control flow concepts provided by the Workflow Management Coalition in
[WfMC95, WfMC99].
The Sequence workflow pattern represents linear execution of workflow
steps: one action of a workflow is activated unconditionally (for example
B the figure below) after another (for example A in the figure below)
finished executing.
Use Case Example: After an order is placed, the credit card specified by the
customer is charged.
The Parallel Split workflow pattern divides one thread of execution (for example
the one that activates A in the figure below) unconditionally into multiple
parallel threads of execution (for example the ones that start in B, C, and
D in the figure below).
Use Case Example: After the credit card specified by the customer has been
successfully charged, the activities of sending a confirmation email and
starting the shipping process can be executed in parallel.
The Synchronization workflow pattern synchronizes multiple parallel threads of
execution (for example the ones that end in B, C, and D in the figure
below) into a single thread of execution (for example the one that starts in E
in the figure below).
Workflow execution continues once all threads of execution that are to be
synchronized have finished executing (exactly once).
Use Case Example: After the confirmation email has been sent and
the shipping process has been completed, the order can be archived.
The workflow patterns that have been discussed so far handle the
unconditional routing of control flow. We will now take a look at the
workflow patterns for conditional routing.
The Exclusive Choice workflow pattern defines multiple possible paths
(for example the ones that start in B, C, and D in the figure below) for
the workflow of which exactly one is chosen (for example the one that starts in
C in the figure below).
Use Case Example: After an order has been received, the payment can
be performed by either credit card or bank transfer.
The Simple Merge workflow pattern is to be used to merge the possible paths that
are defined by a preceding Exclusive Choice. It is assumed that of these
possible paths exactly one is taken (for example C in the figure below) and no
synchronization takes place.
Use Case Example: After the payment has been performed by either
credit card or bank transfer, the order can be processed further.
The workflow patterns for advanced branching and synchronization do not have
straightforward support in most [of the] workflow engines [that
Kiepuszewski evaluated]. Nevertheless, they are quite common in real-life
business scenarios.
The Multi-Choice workflow pattern defines multiple possible paths (for example
the ones that start in B, C, and D in the figure below) for the workflow
of which one or more are chosen (for example the ones that start in B and D
in the figure below). It is a generalization of the Parallel Split and Exclusive
Choice workflow patterns.
The Synchronizing Merge workflow pattern is to be used to synchronize multiple
parallel threads of execution that are activated by a preceding Multi-Choice
(for example the ones that end in B and D in the figure below).
The Discriminator workflow pattern can be applied when the assumption we made
for the Simple Merge workflow pattern does not hold. It can deal with merge
situations where multiple incoming branches may run in parallel.
It activates its outgoing node after being activated by the first incoming
branch and then waits for all remaining branches to complete before it resets
itself. After the reset the Discriminator can be triggered again.
Use Case Example: To improve response time, an action is delegated to
several distributed servers. The first response proceeds the flow, the other
responses are ignored.
The structural workflow patterns deal with restrictions that different workflow
models can impose.
A common restriction workflow models impose is that arbitrary cycles, ie. one
or more activities are done repeatedly, are not supported. As an alternative,
special loop constructs that mark the start and end point of a structured cycle
are offered.
The execution of the workflow is (successfully) terminated when there are no
activated activities left and no other activity can be activated. This implicit
termination of workflow execution can be used in addition to explicit end
activities.
The execution of a workflow instance is cancelled.
Use Case Example: An order is cancelled.
The workflow model is activity-based. The activities that are to be completed
throughout the workflow and the transitions between them are mapped to the
nodes and edges of a directed graph. This choice was made to faciliate the
application of the Graph-Oriented Programming paradigm for the implementation
of the Workflow component. Using a directed graph as the foundation for the
workflow model makes it possible to define the syntax of the workflow
description language using the formalism of graph grammars.
The execution of a workflow starts with the graph's only Start node. A
graph may have one or more End nodes that explicitly terminate the
workflow execution.
After a node has finished executing, it can activate one or more of its
possible outgoing nodes. Activation adds a node to a set of nodes that
are waiting for execution. During each execution step, a node from this set
is executed. When the execution of a node has been completed, the node is
removed from the set.
The workflow execution is implicitly terminated when no nodes are activated
and no more nodes can be activated (see the Implicit Termination workflow
pattern that was discussed above).
The workflow model supports state through the concept of workflow variables.
Such a variable can either be requested as user input (from an Input node) or
be set and manipulated through the VariableSet, VariableAdd, VariableSub,
VariableMul, VariableDiv, VariableIncrement, and VariableDecrement nodes.
While a VariableSet node may set the value of a workflow variable to any type
that is supported by the underlying programming language, the other variable
manipulation nodes only operate on numbers.
Variables are bound to the scope of the thread in which they were defined. This
allows parallel threads of execution to use variables of the same name without
side effects.
When the execution of a workflow reaches an Input node (see above), the
execution is suspended until such time when the user input has been provided and
the execution can be resumed.
The control flow semantics of the workflow model draws upon the workflow
patterns that were discussed above. The Sequence, Parallel Split,
Synchronization, Exclusive Choice, Simple Merge, Multi-Choice,
Synchronizing Merge, Discriminator, and Cancel Case workflow patterns are
all directly supported by the workflow model.
Exclusive Choice and Multi-Choice nodes have branching conditions attached
to them that operate on workflow variables to make their control flow decisions.
A special node, named Loop, to conveniently express loops is also available.
So far we have only discussed nodes that control the flow and that can
manipulate workflow variables. We are still missing a type of nodes that
actually performs an activity. This is where the Action node comes
into play.
When the execution of a workflow reaches an Action node, the business logic of
the attached Service Object is executed. Service Objects "live" in the domain
of the application into which the workflow engine is embedded. They have read
and write access to the workflow variables to interact with the rest of the
workflow.
The workflow model supports sub-workflows to break down a complex workflow
into parts that are easier to conceive, understand, maintain, and which can
be reused.
A sub-workflow is started when the respective Sub-Workflow node is reached
during workflow execution. The execution of the parent workflow is suspended
while the sub-workflow is executing. It is resumed once the execution of the
sub-workflow has ended.
This section discusses the design and implementation of the Workflow component.
The Workflow engine has been designed and implemented as four loosely coupled
components. The Workflow component provides an object-oriented framework to
define workflows and an execution engine to execute them. The
WorkflowDatabaseTiein and WorkflowEventLogTiein components tie the
Database and EventLog components into the main Workflow component for
persistence and monitoring, respectively. The WorkflowSignalSlotTiein
component leverages the SignalSlot component for the Workflow
component's plugin system.
A workflow can be defined programmatically by creating and connecting objects
that represent control flow constructs. The classes for these objects are
provided by the Workflow Definition API. This API also provides the
functionality to save workflow definitions (ie. object graphs) to and load
workflow definitions from a data storage. Two data storage backends have been
implemented, one for relational database systems and another for XML files.
Through the Workflow Execution API the execution of a workflow definition can
be started, resumed, and cancelled. The figure below shows the conceptual
architecture for the workflow engine.
The idea that a workflow system should be comprised of loosely coupled
components is discussed, for instance, in [DAM01, DG95, PM99].
Manolescu states that an object-oriented workflow architecture must provide
abstractions that enable software developers to define and enact how the work
flows through the system [DAM01].
The component-based workflow architecture Micro-Workflow encapsulates
workflow features in separate components. This architecture follows the
Microkernel pattern which applies to software systems that must be able to
adapt to changing system requirements. It separates a minimal functional core
from extended functionality and customer-specific parts. The microkernel also
serves as a socket for plugging in these extensions and coordinating their
collaboration [FB96].
This section proposes a so-called workflow virtual machine as the executing
component of a component-based workflow architecture.
Given the fact that standardization efforts, e.g. XPDL [WfMC05] proposed by
the WfMC, have essentially failed to gain universal acceptance [WA04], the
problem of developing a [workflow system] that supports
changes in the [workflow description language] needs to be addressed.
Fernandes et. al. propose to split the [workflow system] into two layers: (1) a
layer implementing a Workflow Virtual Machine, which is responsible for most
of the [workflow system] activities; and (2) a layer where the different
[workflow description languages] are handled, which is responsible for making
the mapping between each [workflow description language] and the Workflow
Virtual Machine [SF04].
A workflow virtual machine isolates the executing part of a workflow management
system, the backend, from the parts that users interact with, the frontend.
This isolation allows for the definition of a backend language to describe
exactly the workflows that are supported by the executer and its underlying
workflow model. This backend language is not the workflow description language
users use to define their workflows. They use frontend languages that can be
mapped to the system's backend language.
The manual of JBoss jBPM [JBOSS], a platform for multiple process languages
supporting workflow, business process management, and process orchestration,
introduces Graph-Oriented Programming [as a] new implementation technique that
serves as a basis for all graph-based process languages.
Graph-Oriented Programming implements the graphical representation and the
wait states of a process language in an object-oriented programming language.
The former can be achieved by providing a framework of node classes. Objects of
these classes represent the nodes in the process graph, relations between these
objects represent the edges. Such an object graph can then be traversed for
execution. These executions need to be persistable, for instance in a relational
database, to support the wait states.
The aforementioned node classes implement the Command design pattern and
encapsulate an action and its parameters.
The executing part of the workflow engine is implemented in an Execution class.
An object of this class represents a workflow in execution. The execution object
has a reference to the current node. When the execution of a workflow is started,
a new execution object is created and the current node is set to the workflow's
start node. The execute() method that is to be provided by the node classes is
not only responsible for executing the node's action, but also for propagating
the execution: a node can pass the execution that arrived in the node [to] one
of its leaving transitions to the next node.
Like Fowler in [MF05], the authors of the JBoss jBPM manual acknowledge the
fact that current software development relies more and more on domain specific
languages. They see Graph-Oriented Programming as a means to implement domain
specific languages that describe how graphs can be defined and executed on top
of an object-oriented programming language.
In this context, a process language (such as a workflow description language) is
nothing more than a set of Node implementations. The semantics of each node
are defined by the implementation of the execute() method in the respective node
class. This language can be used as the backend language of a Workflow Virtual
Machine (see above). In this lanugage, the workflow is represented as a graph of
command objects. The workflow patterns (see above) make up the requirements for
and can be mapped to the respective classes.
One of the advantages of using a domain specific language that Fowler gives in
[MF05] regards the involvement of lay programmers: domain experts who are not
professional programmers but program in domain specific languages as part of the
development effort. In essence this means that a software system that provides
a domain specific language can be customized and extended without knowledge of
the underlying programming language that was used to implement it.
