2020 December 3
Daniel Katz recently wrote a piece entitled “Software and Software Metadata” summarizing some of his current thoughts about the software metadata and what it would take to achieve a sort of “FAIR-ness” for software. As I read the article, I realized that — while I’m 100% on board with the goals — I’ve actually convinced myself that the implementation strategy we need is pretty different from what Katz outlines.
I’ve got some bigger thoughts on this topic, but here I just want to focus on the idea of having software authors provide citation metadata by putting special files, like CITATION.cff or codemeta.json, in their source repositories. This idea is super natural and super appealing, but I think it’s not the right way to go.
Here are some questions that illustrate my reasoning:
What do you do about monorepos? Sometimes repositories contain more than one
project, and it is entirely reasonable for some of those projects to want to be
independently citeable. You could certainly sprinkle multiple
files around the repo, of course. But if you break the assumption of a 1:1
mapping between repos and citeable projects, many seemingly straightforward
architectural decisions become much more complicated (as my recent experience
What do you do about the time axis? Say that I create a
and, as a good citizen, update it with every new release of my software. If
someone is using an older version of my software and gets the citation data from
my repo, they’ll get the wrong information. But if I tell them to get citation
data from a file bundled with the software they’re using, how do I fix mistakes?
How long do I maintain citation instructions for old releases? How do I embed
the DOI of a release into its code if I can only obtain the DOI after making the
release? One can devise answers to these questions, but I think they point at a
more fundamental issue that I’ll draw out below.
What do you do about closed-source software? Now, I bow to no one in terms of eagerness to promote open scientific software. But closed-source tools are out there, and a good metadata architecture should handle them. In particular, while I’m more than happy for a system to favor open codes, it’s a very worrisome “smell” to me if open- and closed-source software have to be handled very differently.
Extending this point, consider the metadata ecosystem for journal articles. Do we ask authors to include metadata in their “source code”? Absolutely not. In fact, the metadata that are most important for citeability (e.g., DOIs, year/volume/page numbers) are not even allowed into the author’s hands — they’re created by publishers, during the publication process, not the writing/coding process. I think this comparison point is really worth paying attention to — the empirically successful article citation system emphatically rejects the paradigm that we’re talking about for software!
The example of article citation helps draw out the flaws in the “metadata in repo” paradigm. We should be citing published objects. When we talk about software, the published object is not the source repository — it’s the release. Citation metadata are data about releases, which are derived from snapshots of code repositories. Trying to keep citation metadata in the source repository introduces fundamental, unavoidable sequencing and synchronization problems: it creates a circularity where there should be none. You can come up with tactics to deal with the specific issues raised above, but they’re all symptoms of this fundamental issue.
The specific case of citation metadata isn’t the only one where you run into problems storing data about releases in their source repositories. I’ve convinced myself that even the everyday task of assigning version numbers suffers from the same flaws. With my Cranko project I’ve built a tool that aims to address versioning in a way that avoids these issues, and now that I’ve been using it regularly I’m completely convinced that this insight is deeply important. Whenever I try to talk to people about it, I can tell that I haven’t found the words to convey just why I feel that way, but Cranko has truly changed my (software) life.
If this profound flaw is lurking out there, why are we trying to put software metadata in source repositories anyway? Software and articles are different in important ways — but what’s the specific issue at play here? I believe that it’s a factor that’s not often mentioned.
It’s not that software gets updated (we have living articles now) and it’s not that software has source code (TeX in Git works quite nicely, thank you very much). It’s that software is self-published. With a few high-volume publishers, gathering metadata is easy; with thousands of small publishers, it’s not. Try another thought experiment: if somehow every software release had to go through the US Library of Congress, would we be talking about citation metadata files in repositories? Would software citation even be considered a problem to be solved? I think not.
While metadata-in-the-repo attempts to solve the problem of distributed metadata gathering, I believe that the circularity that it introduces means that it is ultimately a flawed foundation that can’t be safely built upon. And this perspective motivates a rethinking of what we need to do to fix software citation — but that will have to wait for a later post.