A key issue in supporting the synthesis of data intensive computation and data management is to liberate users from low-level parallel programming, by specifying applications functionally independent of the underlying server infrastructure, and further, by providing high-level primitives to express the control flow of applying functions to data partitions. Currently only few such primitives, e.g. Map-Reduce and Cross-Apply, are available, and their expressive power is limited to “flat parallel computing”. To deal with “structured parallel computing” where a function is applied to multiple objects with execution order dependencies, a general framework for creating and combining such primitives is required.