A common obstacle in the development of scientific software is that it is typically carried out by researchers doubling as software engineers, a profession in which they usually lack formal training or experience. While they may try to educate themselves via online courses or other materials, historically those materials have typically been targeted at software developers in industry, where different best practices may apply. Moreover, software quality and reusability are generally secondary considerations at best since development itself is only indirectly funded and rewarded through the scientific results it produces. Thus, scientists funded to achieve certain scientific goals face difficult design decisions with respect to the scope and generality of their software.
Scientific software developers tend to extremes
Experience shows that many scientists in this position fall into one of two extremes. On the one hand, scientists for whom code is purely a means to an end spend minimal time educating themselves on best practices in software development, instead opting for the minimal scope, generality, and overall quality that gets the job done. On the other hand, scientists who enjoy software development educate themselves on good development practices and tend to write more readable, reusable, and generalizable code; but they expend unnecessary effort on generalizing software well beyond the needs of any demonstrated practical application. The former strategy is expedient in the short term but highly inefficient in the long run since it ultimately results in many researchers simply reinventing the wheel. Conversely, the latter strategy is usually inefficient in the short term since the time and effort put into development hinder research progress; and without suitable evidence that this effort will pay off in the long term, such development cannot be justified on scientific grounds. Furthermore, efficient generalization requires a lot of experience and foresight that are usually hard to obtain.
Finding the middle path
While organizations such as The Carpentries have emerged to help researchers gain the skills needed for proper software development, the problem of finding the right balance between these two development extremes is usually still left to researchers. In an ideal scenario, researchers would develop software that solves their immediate needs while maintaining just enough flexibility to enable later generalization if needed. In practice, however, achieving this balance is notoriously difficult. In our article for Computing in Science & Engineering we proposed the lazy refactoring technique as a practical approach to finding that balance.
In short, lazy refactoring advocates that new development should remain at a prototype stage until the specific scientific goal is achieved, but it should be refactored into more general, reusable code once a third use for it is found. This ensures that the majority of development time is spent on solving the problem that motivated the development.
The lazy refactoring approach depends on adhering to certain minimal code standards to ensure that prototypes can be generalized with reasonable effort once their reuse potential has been identified. Specifically, prototype code should do the following:
- Be under version control to track the logical development, for example, with git.
- Provide minimal documentation: a README file is a must. As a benchmark, the documentation should be sufficient such that one can digest the logical flow of the code without significant effort.
- Adhere to some common sense standard applicable for the programming language used (e.g., PEP8 for Python or the C++ Core Guidelines).
- Be validated (ideally against known results) in the context of the scientific application.
- Include a license (ideally permissive) such that people other than the original author are able to reuse the code at a later stage.
Lazy refactoring has several advantages. First, it ties all development to specific, tangible research goals, removing the burden of needing to justify development efforts to funding sources that are interested primarily in scientific outcomes. Second, it defers any generalization to a stage at which the problem and software use cases are better understood, which helps in designing a better API. Third, it reduces the risk that the additional development effort spent on producing generalized code is completely wasted. This situation could happen when, for example, a method is implemented within a general software package but is quickly superseded by a superior method before being used enough to justify the implementation effort.
Applying lazy refactoring to new software or new feature development
Since we originally proposed this method, we have found it suitable as a general guideline for making decisions on how to prototype and generalize software. It is especially easy to apply in the development of completely new software. Initially, it may seem less applicable in the context of mature software projects, but the basic concept remains the same: any new feature should still be justified by immediate need and be prototyped before being generalized.
Eventually, some projects reach a level of maturity at which there are a large number of users and use cases and it becomes increasingly challenging to tie all development efforts to specific scientific research needs. At this stage, lazy refactoring can still be applied to the development of new features, but regular maintenance efforts such as reducing technical debt or resolving dependency incompatibilities need not be vetted through this process. Ensuring that such software remains usable and amenable to development within a rapidly changing hardware and software environment is justified by the existing applications of the software, and maintainers should consider seeking dedicated funding.
Balancing immediate needs with a path forward
In summary, lazy refactoring helps developers of scientific software in academic contexts effectively prioritize software development efforts by focusing on immediate scientific needs balanced with a clear path toward general-purpose utilization. This process helps justify development efforts to funding agencies while effectively improving overall code quality.
Acknowledgments
This article is based on a paper in the IEEE Computing in Science and Engineering special issue on Accelerating Scientific Discovery with Reusable Software: Carl S. Adorf, Vyas Ramasubramani, Joshua A. Anderson, and Sharon C. Glotzer, How to Professionally Develop Reusable Scientific Software -- And When Not To, Computing in Science and Engineering 21, 66-79 (2019). DOI: 10.1109/MCSE.2018.2882355
Image Attribution
Image by Dave Taylor from Boulder, CO, is licensed under CC BY 2.0.
Author bios
Dr. Carl Simon Adorf is a postdoctoral researcher in the group of Prof. Nicola Marzari at École polytechnique fédérale de Lausanne (EPFL). He obtained his Ph.D. at the University of Michigan under the supervision of Sharon C. Glotzer in 2019. Dr. Adorf is an expert in employing machine learning algorithms for identifying and analyzing crystallization pathways of colloidal crystals and has made a huge impact on the scientific computing community by inventing and leading the development of the signac framework (signac.io). He is currently a developer for the AiiDA computing infrastructure (aiida.net) and the Materials Cloud open science platform (materialscloud.org) as part of his involvement with the Horizon 2020 MARKETPLACE project.
Vyas Ramasubramani is a Ph.D. candidate in the group of Sharon Glotzer at the University of Michigan, Ann Arbor, where he studies particle self-assembly, especially of proteins, at the nanoscale and microscale. Vyas is a maintainer and lead developer for both the signac data management framework and the freud simulation analysis toolkit, and he created the rowan package for quaternion mathematics. He is also a core developer for HOOMD-blue, a popular particle simulation package that is highly optimized for GPUs.
