open force field

An open and collaborative approach to better force fields

Science Updates

We can have nice things

Posted on 19 Jan 2022 by Josh Mitchell

I think in scientific computing we can be so focused on the problems we’re solving that we’re slow to pick up new tooling that merely makes our lives easier. It’s like we imagine that, since we’re scientists, we have to make do with what we’ve got because we don’t have time to get anything better. But we can have nice things too; in this day and age, we are fortunate to have an enormous amount of high-quality tooling that means that many errors are caught before they make it to a release: GitHub Actions, conda-forge, PyTest, Black, Mamba, MyPy. With a little diligence, a lot of the frustration of building and maintaining working code can be mitigated.

I should say, this frustration can be mitigated as long as it happens in code. Documentation tooling has not had the same attention. I firmly believe we can have nice things in documentation as well. I’d like to tell you about what we’ve done with documentation in the last year, and where we could take OpenFF’s documentation tooling in 2022.

The Book and The Reference

I’ve noticed two patterns that software documentation seems to fall into. I call these patterns “The Book” and “The Reference”. I’m not the first person to draw this distinction, but I think it’s helpful, so let me describe it.

The Book explains the design of the software and gently introduces the reader, starting with the simplest or most common use cases and building up. It is designed to be read cover-to-cover, or at least for sections or chapters to be read in their entirety. It is linear in structure. It’s very helpful for users when they’re getting started with a piece of software. It must have a structure that facilitates a gentle learning curve and so cannot mirror the structure of the codebase.

The Reference describes the entire public API of a piece of software and everything it can do. It is designed to be easy to navigate so that the reader can quickly and easily find the precise bit of info they need. It repeats information liberally so that the user can find what they’re looking for easily. It is hierarchical in structure. It needs to be easier to read than the source code or else it’s pretty pointless. It’s very helpful when users are familiar with an API but just need to find the signature of that one function, or when users only need an isolated behavior from a library and don’t want to read The Book in its entirety to find that behavior. It can mirror the structure of the codebase itself, which is already organized by behavior.

The popular documentation tools in the Python ecosystem — Sphinx and Mkdocs — are designed around writing The Book. They provide tools to document the API as well, but the natural result of their design produces docs that are hybrids of The Book and The Reference. The Python Standard Library docs are a great example, as is the documentation for Sphinx itself. I think this causes a trade-off between having a logically coherent narrative with a gentle learning curve on one hand, and on the other having a structured document that experienced users can navigate quickly. Once you get used to them, the Python docs are fine, but when you first read them it can be confusing: Do I want the str() function, or the str class, or the string module? Probably parts from all three, as well as the section on collections. By contrast, when you first read the Sphinx docs they tell a clear story; but when I’m experienced and want to understand an extension, where are the docs for the Sphinx class? Oh yeah, under the “Application API” heading. Naturally.

Automating documentation

Book-style documentation is organized to be read, not to be referenced. Reference-style documentation is organized to be referenced, not read. But our users need to both read and reference our documentation, and we don’t have time to maintain two parallel sets of documentation — unless we can automate one of them.

A sign of good Python code is that it is well documented with docstrings. Each API endpoint, whether method, attribute, function, class, or module, provides a textual description of what it is and how it should be used. Code that has well-written docstrings and correct type annotations and is in a module that makes sense already has well organized, complete documentation that is even partially tested in CI. All we need to do is format it nicely, add cross-references, and put it on the web!

In 2021, we experimented with fully automated API docs in two new projects, Interchange and BespokeFit. Both projects already use type annotations extensively, so we get clear, cross-referenced documentation, even between projects, even with only minimal docstrings. We use the builtin Sphinx extensions Autosummary and Autodoc, the fantastic third party extension autodoc_pydantic, as well as a few custom templates to produce these API docs directly from our library code, and I’ve also been able to contribute code back upstream. With solid, well-maintained tools, an error in the API reference reflects an error in our code, whether it be a missing type annotation, a misleading docstring, or a semantically private function missing a leading underscore.

I’ve also written Book-style docs for BespokeFit. I don’t think The Book will be automated in the near future, and I’m not sure it should be. Automating The Reference works because it’s about describing what the code does to a human. The Book is much more about why; its about developers (or expert users, or some chump who nags the developers a lot) explaining to users why the code is a certain way, or why they should do this and not that. The machine just does what its told and doesn’t ask why, and so that’s what the machine can teach the user. Maybe one day AI will be able to infer the developers motivations, but I can’t imagine that being better than hearing it from the developer themselves.

In 2021, I’ve learned a few things about writing The Book and automating The Reference:

  • Automate what you reliably can, and no more.
  • Try not to repeat info from The Reference in The Book, just link to it. This helps keep The Book correct and in date, and also keeps the focus of The Book on understanding the software rather than detailing all of its behavior.
  • A lot of documentation decisions become code decisions when you automate the API reference, and managing privacy and re-exports is very important. I think this can lead to a clearer API structure, but changing the API without breaking it is always fraught.
  • People prefer Markdown to reStructuredText.

In 2022, I want to take Reference automation further. In 2021 I added support for the Python __all__ attribute to Autosummary so that devs can have their public API reflected in the docs; this year I want to come up with a similar way to specify whether a module or package should be documented or private without requiring devs to rename it. I also want to move forward in automating the API reference for the Toolkit, which may involve rethinking how we approach organising that API. Finally, I’d love to be able to write docstrings in Markdown; Autodoc’s output is currently hard-coded as rST, and extending this is a big enough project that it hasn’t been attempted yet, though some initial experiments are taking place in the MyST project. I’m hoping to spend a few cycles investigating this issue more deeply.

Testing documentation

Documentation goes out of date. It seems to go out of date proportionately to how tightly it is coupled to code:

  • Function signatures: Out of date means there’s a bug
  • Code examples that are tested in CI: Usually correct, but not always the best method
  • Type annotations: Commonly too lax or too strict, but usually correct
  • Docstrings: Slowly drift out of date, but fairly easy to correct and usually checked in code review
  • Prose in The Book: Probably needs to be re-read and revised at least every year
  • Books published more than a year or two ago: Wrong

A lot of this is just life, and someone needs to do the work to keep it in date. Even then, it would be fantastic to develop tooling to help remind authors to check The Book. Habitually linking to the API reference for code that’s being explained provides a warning when that link goes out of date, but perhaps it could also give a hint to the author if the code itself has been changed since then. Perhaps the markup could include a timeline over which the author expects to have to check it, and CI could buzz the author or raise a GitHub issue when it needs checking.

But some documentation can actually be tested. The classic example of this is code examples. If the docs include example code, it seems clear that the examples should be tested in CI. Python provides doctest for exactly this purpose, and pytest supports it. Unfortunately, doctest is designed to mimic an interactive Python session. Every line of code must be prepended with a prompt, all output must be included, test sessions operate on the level of files rather than text blocks, and the author has very little control over the presentation. To get the benefit of tested code examples, every code block must follow all of these conventions. If the prompt is forgotten, the test passes silently.

Maybe I’m spoiled by rustdoc, but I want all of my code examples to be tested, whether the author formats them a certain way or not. I want to be able to write code examples in the way that best facilitates understanding, which I think is usually not a prompt. I don’t want my examples to be poisoned by whatever other examples are on the page. And I want to be able to decide to hide a boring import line or ignore the exception I’m demonstrating.

In 2021, I got started writing a doctest framework of my own for Python. In 2022, I hope to complete it, give it a better name, and be able to rely on it in production. It currently supports all the above features, and I’m using it experimentally on the in-progress BespokeFit docs in CI.

An alternative approach that already exists is to write pages with code examples as Jupyter notebooks. Tools like MyST-NB and nbsphinx allow notebooks to be written in more-or-less ordinary Markdown and be executed or rendered to the web as needed. We already use this approach in the Toolkit docs, but my goal with my doctest framework is for it to work on every code block automatically, not just the ones that have been specially prepared.

The other stuff

OpenFF’s documentation has had some other nice things this year. We’ve moved to a new look that’s unified across the project. This is maintained as our own Sphinx theme. I want to integrate the ReadTheDocs versioning system into the theme in 2022 for a uniform look. We also introduced a couple of refreshed examples demonstrating use of the Toolkit and its software ecosystem. In 2022, I expect to refresh, prune and write new examples featuring new projects like Interchange and BespokeFit!

We’ve also been talking about some extra special new things in 2022, like unifying the documentation under the domain and even adding some video documentation. Stay tuned!