Towards Automated Design, Analysis and Optimization of Declarative Curation Workflows

Data curation is increasingly important. Our previous work on a Kepler curation package has demonstrated advantages that come from automating data curation pipelines by using workflow systems. However, manually designed curation workflows can be error-prone and inefficient due to a lack of user understanding of the workflow system, misuse of actors, or human error. Correcting problematic workflows is often very time-consuming. A more proactive workflow system can help users avoid such pitfalls. For example, static analysis before execution can be used to detect the potential problems in a workflow and help the user to improve workflow design. In this paper, we propose a declarative workflow approach that supports semi-automated workflow design, analysis and optimization. We show how the workflow design engine helps users to construct data curation workflows, how the workflow analysis engine detects different design problems of workflows and how workflows can be optimized by exploiting parallelism. Received 13 January 2014 | Accepted 26 February 2014 Correspondence should be addressed to Tianhong Song, Department of Computer Science, University of California, Davis. Email: [email protected] An earlier version of this paper was presented at the 9 International Digital Curation Conference. The International Journal of Digital Curation is an international journal committed to scholarly excellence and dedicated to the advancement of digital curation across a wide range of sectors. The IJDC is published by the University of Edinburgh on behalf of the Digital Curation Centre. ISSN: 1746-8256. URL: http://www.ijdc.net/ Copyright rests with the authors. This work is released under a Creative Commons Attribution (UK) Licence, version 2.0. For details please see http://creativecommons.org/licenses/by/2.0/uk/ International Journal of Digital Curation 2014, Vol. 9, Iss. 2, 111–122 111 http://dx.doi.org/10.2218/ijdc.v9i2.337 DOI: 10.2218/ijdc.v9i2.337 112 | Automated Workflow Design, Analysis and Optimization doi:10.2218/ijdc.v9i2.337 Introduction and Motivation Data curation is critical in many areas, such as the production and use of scientific data collections and repositories. For example, natural science collections data can often be rife with errors and inconsistencies, and reuse of collections data to address scientific questions imposes concerns of data quality and fitness for use upon the collection’s management community. We have continued our development of Kuration 1.0 (Dou et al., 2011), a software package and prototype for automating data curation pipelines with the Kepler scientific workflow system (Ludäscher et al., 2006). Several curation tools and services are integrated into this package as actors, enabling the construction of workflows to perform and document various data curation tasks. The typical structure of a data curation workflow includes an input actor that reads a dataset from a remote or local source, a number of data curation actors that implement different data validation methods, and an output actor that writes the result into a file or a database. Figure 1. Kuration 1.0: A Kepler/COMAD data curation workflow for collection-oriented data quality control. Figure 1 shows a data curation workflow in the Kepler workflow system (Dou et al., 2012) developed by using the COMAD workflow model (McPhillips et al., 2009). Boxes are actors (also known as processors, steps or modules) and arrows connecting them are data channels indicating the data flow between actors. COMAD and the related data assembly line approach (Zinn et al., 2009a) are an improvement over the earlier and “conventional” way of designing workflows in Kepler, as they simplify the workflow design. Practical experience with our first IJDC | General Article doi:10.2218/ijdc.v9i2.337 Tianhong Song et al. | 113 Kuration prototype, while promising (Dou et al., 2012), also yielded a number of technical challenges: 1) the use of remote services on large data collections, together with an unoptimized workflow execution model, creates scalability problems; and 2) the design, configuration and maintenance of workflows are not only time-consuming but also can be error-prone, e.g., the workflow design itself might have problems, such as incorrectly ordered actors in a workflow (see Q3 below). In order to tackle these and related issues, we propose a visionary system that (semi-) automatically detects workflow design problems and optimizes workflow design. In the first phase, the design engine selects actors from a library of existing actors based on the user’s requirements and puts them together into a “workflow story” (i.e., a sequential order of actors). Alternatively, users can provide their own workflow. In the second phase, static analysis techniques are applied before runtime, which anticipate how a workflow might behave during runtime. Using workflow graph and actor configurations, corresponding data dependency information can be captured, which in turn can be used to identify possible design problems. In the third phase, the candidate designs are further improved or optimized (e.g., exploiting parallelism) in order to achieve better performance. The flowchart in Figure 2 shows how our system works. Figure 2. Overview of the proposed workflow design, analysis and optimization system. The remainder of this paper is organized as follows: first, we introduce our workflow model; next, we describe detailed design, analysis and optimization techniques and scenarios with examples; finally, we conclude with additional implementation details and a discussion of future work.