Automatically Composing Data Workflows with Relational Descriptions and Shim Services

Many scienti c problems can be represented as computational work ows of operations that access remote data, integrate heterogeneous data, and analyze and derive new data. Even when the data access and processing operations are implemented as web or grid services, work ows are often constructed manually in languages such as BPEL. Adding semantic descriptions of the services enables automatic or mixed-initiative composition. In most previous work, these descriptions consists of semantic types for inputs and outputs of services or a type for the service as a whole. While this is certainly useful, we argue that is not enough to model and construct complex data work ows. We present a planning approach to automatically constructing data processing work ows where the inputs and outputs of services are relational descriptions in an expressive logic. Our work ow planner uses relational subsumption to connect the output of a service with the input of another. This modeling style has the advantage that adaptor services, so-called shims, can be automatically inserted into the work ow where necessary.