Optimal Task Ordering in Chain Data Flows: Exploring the Practicality of Non-scalable Solutions

Modern data flows generalize traditional Extract-TransformLoad and data integration workflows in order to enable end-to-end data processing and analytics. The more complex they become, the more pressing the need for automated optimization solutions. Optimizing data flows comes in several forms, among which, optimal task ordering is one of the most challenging ones. We take a practical approach; motivated by real-world examples, such as those captured by the TPC-DI benchmark, we argue that exhaustive non-scalable solutions are indeed a valid choice for chain flows. Our contribution is that we thoroughly discuss the three main directions for exhaustive enumeration of task ordering alternatives, namely backtracking, dynamic programming and topological sorting, and we provide concrete evidence up to which size and level of flexibility of chain flows they can be applied.