Consolidation of Interacting BPEL Process Models with Fault Handlers Sebastian Wagner, Oliver Kopp, and Frank Leymann Institute of Architecture of Application Systems, University of Stuttgart, Germany {wagnerse,kopp,leymann}@iaas.uni-stuttgart.de Abstract The interaction behavior between processes of organizations and their suppliers can be modeled by using choreographies. When an organization decides to gain more control about their suppliers and to minimize transaction costs they may decide to insource these companies. This also requires the integration of the partner processes into the organization. In previous work we proposed an approach to merge (consolidate) interacting BPEL process models of different partners into a single process model by deriving control flow links between the process models from their interaction specification. In this work we are detailing this consolidation approach. Thereby, special attention is turned on extending the consolidation operations in a way that process models with fault handlers can be merged. 1 Introduction To reduce transaction costs or to gain more control companies often decide to integrate suppliers into their organization (in-sourcing, mergers, and acquisitions). This requires the integration of the processes and the organizational structure of the two companies. In this work we focus on the integration on the process-level. More precisely, we want to merge (or consolidate) complementing process models whose interaction behavior is described by a choreography. Process modeling languages such as BPEL [10] or BPMN [11] offer different language constructs to raise and handle faults that work similar to throw-catch constructs in traditional programming languages such as Java. As fault handling constructs can also cause message exchanges between interacting processes they are also affected by the consolidation. In this paper we want to describe a technique to merge BPEL process models that communicate via fault handlers. Moreover, we describe an extension of the merge operations proposed in [13] and [14]. We use BPEL4Chor [2] to model BPEL choreographies as this language provides a means to define message links between the communication activities of the interacting BPEL processes. We assume that the reader is familiar with BPEL. Nevertheless, we give a brief overview about BPEL’s fault handling concepts in Sect. 2. In Sect. 3 an overview on the merge operations is provided and extensions to them are discussed. Then, Sect. 4 presents a technique to merge processes that communicate via BPEL fault handlers. After discussing related work in Sect. 5, Sect. 6 concludes this paper and provides an outlook on future work. 2 BPEL Fault Handling Basics BPEL offers three language constructs to repair faulty situations during process execution, namely fault handlers, compensation handlers and termination handlers. If a fault occurs within a scope all running activities within this scope are terminated and its fault handlers are called. A fault handler is represented by a catch or catchAll block. Thereby, multiple catch blocks can be defined for a scope. Each catch block catches a particular fault that may be thrown during execution of the scope and contains BPEL activities to handle this fault. A catchAll block contains logic to catch all other faults that do not match to a particular catch block. If no explicit fault handlers are defined for a scope it has an implicit default fault handler attached to it. If any kind of fault occurs during the execution of the scope the default fault handler triggers compensation handling for its child scopes (see below) and finally rethrows the fault to its parent scope. If this scope does not provide a fault handler for this fault either, it is propagated up to its parent scope and so on until the process scope is reached. If the process scope cannot catch the fault the process fails and is terminated. Compensation handlers contain activities to undo work that was successfully per- formed by a scope they are attached to, e. g., canceling a flight that was booked by the activities of the scope. Hence, they are only executed if their associated scope has completed successfully. To control the termination of a scope that is still running a termination handler can be attached to it. Within the termination handler activities can be defined that are performed before the actual termination of the scope. If no explicit termination handler was defined for a scope its default termination handler compensates its child scopes. A more detailed discussion about fault and compensation handling concepts in BPEL was provided by Khalaf et al. [4]. Asynchronous Merge 3 Asynchronous and Synchronous Consolidation We introduced the consolidation op- Process A Process B Process Pmerged def: vrc erations to merge asynchronous and syn- def: vin def: vrc Root chronous communicating process models • • S A S B S RC in [13]. The aim of the consolidation is Opaque Opaque def: vin S • RC • that the atomic activities of the different S m RC Opaque Opaque Invoke Receive and participants in the merged process model v →min m→v SYN rc SSYN RC Assign Empty have the same control flow relations as S • RC • v →v in rc Opaque Opaque in the original choreography. The basic S • RC • Opaque Opaque idea behind the consolidation algorithm is SCOPE SCOPE that the message links imply control flow Flow relations between the activities of the com- Choreography CAB Merged Process Model municating process models. The message Figure 1. Asynchronous Merging Operation link m in Fig. 1 implies for instance that the successor RC• of the receive activity is always performed after the predecessor activity •S of the invoke activity S in process model A as RC• cannot be performed before S completed. However, no statement can be made about the execution sequence © Sebastian Wagner 4 between S• and RC•, e. g., if they have to be performed simultaneously or if S• is per- formed before RC• and so on. This is different from the synchronous scenario depicted in Fig. 2. There RC• is always performed before S•. As activity S does not complete until it received a response message from RP that is performed after RC•. Given these implicit control flow dependencies, the sending and receiving activities can act as merge points. Therefore the consolidation operation materializes the implicit control flow to explicit control flow relations between the activities. The consolidation algorithm to merge an arbitrary number of process models is described in the following. As a prerequisite we assume that the choreography is modeled correctly [3] and deadlock free [8]. Moreover, we assume there exists just one instance of each participant per choreography instance, i. e., interaction patterns involving multiple instances of one participant such as one-to-many send/receive [1] are not supported yet by the consolidation algorithm. Another restriction we make is that a repeatable constructs such as a BPEL ForEach loop do not contain any communication activities that are replaced by control flow links between the process models to be merged. As this would violate the BPEL restriction that repeatable constructs must not be crossed by control flow links [SA00070] 1 . In a first step a new process model Pmerged is created that contains a flow as root activity. For each of the process models P1 to Pn to be merged a separate scope is created in the flow activity of Pmerged . This ensures that the scope activities are performed simultaneously. Each scope contains the root activity (along with its child activities) of one of the process models P1 to Pn . The purpose of the scope is to isolate the activities of the process models from each other as they were also isolated in the original choreography. To avoid that an uncaught fault thrown in one scope causes the other scope to crash (as uncaught faults are propagated up to the process scope) a catchAll fault handler is added to the scopes as shown in Fig. 2. The catchAll contains a compensate activity to emulate the default compensation that would have been triggered in an original process without an explicit fault handler. In case an explicit fault handler was defined in a catchAll block on the original process scope nothing is changed. If a process scope of a process to be merged has already a a catch block defined simply the catchAll block is added to this fault handler. Then the message links are materialized to control flow links. In the asynchronous case the invoke activity S is replaced by an assign activity SY N S and the receive activity RC by an empty activity SY N RC . SY N S emulates the message transfer between between the former invoke and receive activity, i. e., it copies the message from the input variable vs of the invoke to the variable vrc of the receive activity where the message was copied to before. To perform the assignment the declaration of variable vrc is lifted to the parent scope that encloses the two scopes that contain the participant activities. Otherwise, SY N S could not access vrc . The empty activity SY N RC replaces the former receive RC. To avoid name clashes between variables it might be necessary to adapt the variable names accordingly during the consolidation. The incoming and outgoing links of S and RC are mapped to SY N S and SY N RC , respectively. An additional link from SY N S to SY N RC is created. This link ensures that SY N RC is not started before SY N S was executed. 1 Static Analysis (SA) Fault Codes defined in the BPEL specification [10] Synchronous Merge Process A Process B Process Pmerged def: vout def: vrc def: vin def: vrc FHA Root S  RC def: vin def: vrp Opaque Opaque S  RC faultName=“F1”/> Opaque Opaque S RC and m Invoke Receive SYNS SYNRC vin → m m →vrc Assign Empty m' →vout vin → vrc RC RC Opaque and S Opaque Opaque RP SYNSR SYNRP Reply Empty Assign vrp → vout vrp → m' RP S RP Opaque Opaque Opaque SCOPE SCOPE Flow Figure 2. Synchronous Merging Operation The synchronous merge is sketched in5 Fig. 2. There additionally the reply activity © Sebastian Wagner RP is replaced by the reply activity SY N RP to emulate the transfer of the response message sent via message link m0 . SY N RP copies the value of the former reply variable vrp to the output variable of the former invoke activity vout . The declaration of vout has to be lifted to the parent scope as well to make this variable accessible for SY N RP . The empty activity SY N SR is added for the same reason SY N RC was added. The control links of RC are mapped to SY N RP and the outbound links of S are mapped to SY N RC . Moreover, a new link is created to connect SY N S and SY N SR and another one between SY N RP and SY N SR to ensure that the successors of the former invoke activity are not started before. 4 Consolidation in the Context of Fault Handlers In this section we discuss the challenges that arise when materializing the control flow from message links between communication activities that reside within BPEL fault handlers. Thereby, we focus on the cross boundary link constraint imposed by the BPEL specification [SA00071]. This constraint specifies that no control link must point to a target activity within a fault handler from outside the fault handler, i. e., no link must point into a catch or catchAll block. In the following we distinguish between three different scenarios (i) fault handlers without communicating activities (ii) fault handlers with only outgoing message links and (iii) fault handlers with at least one incoming message link. The first scenario is trivial as there is no communication between the fault handlers of the two process models that have to be merged. Consequently, they can be simply merged with the consolidation operations introduced in 3. The second is scenario is depicted in Fig. 3. In process model A the fault handler FH A is attached to the scope SA . FH A contains an asynchronous invoke activity A4 that is related to the corresponding receive activity B2 in process model B via message Scenario 2: FHs with Outbound Links link m. Note, that for simplicity reasons the flow activity containing SA and SB is not explicitly depicted in Fig. 3 and in the following figures. Process A Process B Process Pmerged SA SB SA SB … FHA … … FHA … A3 A3 A1 A4 m B2 A2 A4 B2 Opaque Invoke Receive Opaque Assign Empty … … … … A2 A3 A5 A5 B3 B3 Opaque Opaque Opaque Opaque Opaque Opaque Flow … Flow … … Figure 3. Scenario 2: Message Link pointing from a Fault Handler To merge the process models in a first step the merge operations introduced in Choreography Sect. 3 are applied. CABin process model P This results merged . As the new control flow links materialized from the message links leave the fault handler boundaries outbound only, the cross boundary link constraint is not violated. The scenario © Sebastian Wagner in Fig. 4 is very similar to the previous 7 one except that invoke activity A4 is synchronous, hence, a second message link from the reply activity points back Scenario 3: FHs with Inbound Links to A4. The synchronous consolidation operation creates from the message link m2 the control flow link l 2 . This link crosses the fault handler boundary of FH A inbound in order to realize that A5 is performed after B2 or B3 respectively. This violates the cross boundary link constraint. Process A Process B def: vout Process Pmerged def: vrc def: vrc FHA def: vrp def: vrp def: vout FHA SB SA def: va SB SA def: va … def: vin A3 … def: vin A3 l1 … A4 m1 B1 … A4 Assign B1 Empty Invoke Receive vin → m1 vin → vrc A1 m2 →vout A1 B2 Opaque B2 Opaque Opaque Opaque l2 A5 A6 A2 A2 B3 Opaque Opaque B3 Opaque Empty Assign Reply vrp → vout Flow vrp → m2 A5 … … … Opaque Flow Figure 4. Scenario 3: Message Link pointing into a Fault Handler © Sebastian Wagner 8 Scenario 3: FHs with Inbound Links - Solution To develop a solution to overcome this Process Pmerged def: v out def: v problem control flow links pointing into a S def: v a def: v rc rp FH def: v fault handler have to be avoided, hence, the in A8 control flow has to be modified accordingly. S FH … A A To resolve the incoming link issue we can A7 S B either remove the fault handler completely or Empty A1 we move the activities with incoming links Opaque … A3 out of fault handler. Removing the fault han- A2 A4 l B1 1 Opaque Assign dler completely is not an option as a fault in v →v rc Empty handler can be activated any time during run- … B2 A6 Opaque l 2 time of a scope if a fault is thrown. This Empty ^ B3 A5 behavior cannot be emulated by using other Opaque Assign v →v rp out BPEL constructs. Hence, we suggest a solu- Flow … Flow tion where the fault handler is kept and the fault handling activities are factored out of Figure 5. Merged Process Model with outfac- it. tored Fault Handler Logic A scope S is given that© Sebastian has several Wagner fault 9 handlers FH 1 S to FH n S represented by catch or catchAll blocks. After asynchronous and synchronous merges were performed for each fault handler FH i S , it is checked if it contains activities that are target of a link originating from outside of FH i S . If this is the case for at least one fault handler FH i S , a new scope SFH is created that contains a flow activity. SFH is required to ensure that the activities have still access to the data context of the fault handler they were moved from (see below). The flow activity within SFH serves as container for the scope S and its fault handlers. Then for each fault handler FH i S with an incoming link its root activity root i FH is moved to the scope SFH and replaced by a new empty activity ei FH . Between ei FH and root i FH a control link is created where ei FH acts as source. This ensures that the fault handling logic contained in root i FH is always performed when FH i S is activated. All fault handlers that have no incoming control flow link remain unchanged. Figure 5 shows the merged process model Pmerged where the root activity A3 of the fault handler was factored out. The new empty activity A7 acts as source for the control link pointing to A3. The activities and control flow links that were moved out of the fault handler cannot access the local data context (local variable, partner link declarations etc.) that was either defined in scope S or in the catch block anymore as they reside in the parent scope SFH . To make the data context visible again, the defined data have to be moved to the parent scope SFH of the fault handling activities. In Fig. 5 this affects the variables va and vin , hence, they are lifted from scope S to SFH . The solution described above keeps the control flow relations between the (non- communicating) basic activities as defined in the choreography. If SA in Fig. 5 completes successfully the fault handler FH A is uninstalled and all links originating from activity A7 are marked as dead. Thus, activity A3, its child activities and also all activities within SB are not activated either (dead path elimination). This is the same behavior as modeled in the choreography. In case a fault is thrown and caught by FH A activity A7 is executed and its outgoing links are activated. This causes the former root activity of FH A A3 and its children to be executed. This in turn causes all activities in SB to be performed. When SA completes successfully but SFH was not completed yet by the process engine (which usually happens immediately after SA was completed) and SFH terminated due to an error in its parent scope the default termination handler compensates all successfully executed child scopes of SFH . In case a fault is thrown during the execution of the fault handler FH A this fault is simply thrown to its parent scope. Also this behavior is kept as SFH does not catch any fault, i. e., it simply rethrows the fault to its parent scope that used to be the parent scope scope of SA . This also happens when SA throws a fault that is not caught by fault handler FH A . 5 Related Work Compared to many other techniques that merge processes that are semantically equivalent such as different variants of the same process, we aim to merge collaborating processes. Mendling and Simon [9] propose for instance an approach where semantically equivalent events and functions of Event Driven Process Chains [12] are merged. Küster et al. [5] describe how change logs can be employed to merge different process variants that were created from the same original process. Instead of directly generating a BPEL orchestration out of a BPEL4Chor choreogra- phy, an intermediate format may be used. There is currently no approach keeping the structure of the generated orchestration close to the structure of the original choreography. For instance, Lohmann and Kleine [7] do not generate BPEL scopes out of Petri nets, even if the formal model of Lohmann [6] generates a Petri net representation of BPEL scopes. 6 Conclusion and Outlook In this work we extended the process consolidation approach presented in [13] and [14]. We have shown, how to isolate the activities of the different partners from each other by using BPEL scopss and we also extended the asynchronous and synchronous merge operations to reduce the number of control flow links that may be created during the consolidation operation. The main contribution of this work is a technique to merge process models that interacted via fault handlers before they were merged. To satisfy the constraint that no control links must point into a fault handler we have shown a technique to factor the fault handling activities out of the handler. In future works we also have to propose a way to merge process models that interact via compensation handlers and event handlers. This is even more challenging as they allow neither inbound nor outbound control flow links. Another issue we have to address is that our current merge operations create process models that violate the peer-scope- dependency rule. Basically, this rule states that two scopes enclosed within the same parent scope must have no cyclic control-flow dependencies, otherwise the compensation order of these scopes cannot be determined. However, in practice this rule is not enforced by engines such as the Apache ODE2 or BPEL-g3 . 2 http://ode.apache.org/ 3 http://code.google.com/p/bpel-g/ In this paper, we informally argued that the consolidation approach is correct. A first approach to provide a more formal validation has been presented in [14]. Our ongoing work is to evaluate the Petri net formalizations with respect to formal foundations for our merging approach. Acknowledgments This work was partially funded by the BMWi project Migrate! (01ME11055) and the BMWi project CloudCycle (01MD11023). References 1. Barros, A., Dumas, M., ter Hofstede, A.: Service Interaction Patterns. In: BPM. Springer (2005) 2. Decker, G., Kopp, O., Leymann, F., Weske, M.: Interacting services: From specification to execution. Data & Knowledge Engineering 68(10), 946–972 (Apr 2009) 3. Decker, G., et al.: Non-desynchronizable Service Choreographies. In: ISCOC 2008 4. Khalaf, R., Roller, D., Leymann, F.: Revisiting the Behavior of Fault and Compensation Handlers in WS-BPEL. In: OTM 2009 5. Küster, J., Gerth, C., Förster, A., Engels, G.: A Tool for Process Merging in Business-Driven Development. In: Proceedings of the Forum at the CAiSE (2008) 6. Lohmann, N.: A Feature-Complete Petri Net Semantics for WS-BPEL 2.0. In: WS-FM’07: Web Services and Formal Methods, 4th International Workshop (2007) 7. Lohmann, N., Kleine, J.: Fully-automatic Translation of Open Workflow Net Models into Simple Abstract BPEL Processes. In: Modellierung. Gesellschaft für Informatik e. V. (2008) 8. Lohmann, N., Kopp, O., Leymann, F., Reisig, W.: Analyzing BPEL4Chor: Verification and Participant Synthesis. In: WS-FM’07: Web Services and Formal Methods, 4th International Workshop (2007) 9. Mendling, J., Simon, C.: Business Process Design by View Integration. In: BPM Workshops. Springer (2006) 10. OASIS: Web Services Business Process Execution Language Version 2.0 – OASIS Standard (2007) 11. Object Management Group (OMG): Business Process Model and Notation (BPMN) Version 2.0 (2011), OMG Document Number: formal/2011-01-03 12. Scheer, A.W., Thomas, O., Adam, O.: Process Aware Information Systems: Bridging People and Software Through Process Technology, chap. Process Modeling Using Event-Driven Process Chains. Wiley-Interscience (2005) 13. Wagner, S., Kopp, O., Leymann, F.: Towards Choreography-based Process Distribution In The Cloud. In: Proceedings of the 2011 IEEE International Conference on Cloud Computing and Intelligence Systems (2011) 14. Wagner, S., Kopp, O., Leymann, F.: Towards Verification of Process Merge Patterns with Allen’s Interval Algebra. In: ZEUS. CEUR, Bamberg (2012)