The Web service composition problem involves the creation of a choreographer that provides the interaction between a set of component services, to realize a goal service. It is desirable to have an automated or semi-automated composition process that accepts a business requirement specification in the form of a goal web service and automatically creates a composite service from a set of available component web services.

A one-step composition process is inadequate in scenarios where the complete requirements are not known or the user is unaware of the functionality provided by the component services. In such cases, composition of the goal web service might fail, because of an incomplete specification of the goal service or because the specified goal service cannot be composed using the available component services. An iterative approach towards web service composition would help to address failures that can occur during composition, by providing the user with feedback as to the cause of the failure and possible recovery solutions to the failure. The user can then change the goal service based on this feedback to arrive at a successful composition.

This work addresses the analysis of failures that occur during composition and the type of feedback to be provided to the user. The approach first identifies the cause of failure and explores all possible recovery options for the failure. If possible, suggestions are made to modify the goal service based on the solution provided by every recovery option. The composition process is then simulated on the modified goal services to detect future failures. From amongst these modified goal services, the service which has the least failures is then provided to the user as feedback.

The main contribution of this work is an approach towards analysis and recovery from failures that occur during web service composition. The web service composition algorithm is explained, the goal of failure analysis and recovery is described and the approach towards failure analysis is then described in detail. This is followed by a demonstration of the implementation of the algorithm in two case studies.