My side project Mantle is an infra-as-code tool built with Rust which is configured with YAML files. The key thing my docs need to communicate past the getting started phase to my users is the format for these config files.

The first iteration of my configuration docs were written by hand in a markdown file in the docs site repo. They weren't terrible, but I had a few issues with them.

Choose the format for the job

When I started writing these docs, I organized them by complex types. At the top of the file, I listed each of the top-level properties in the config file along with their type. If the type was complex, I created a new heading in the document and linked to it.

This worked, but it's a format that's better suited to documenting the classes in a library than the properties in a config file.

I once watched someone read the document for the first time, and it was unclear to them how the information they were seeing related to the config file itself, even with the smattering of example blocks I included.

Manual changes are tedious

Because I was writing the document by hand, if I ever wanted to make a style change to the page, I had to make it manually for every object and property. This was not ideal.

Separate from code, separate from tooling

Documentation in a standalone markdown file is fine for humans to read, but not so good for computers. If I ever wanted to surface this documentation in other mediums (e.g. inline in VSCode or with an interactive CLI), I would need to rethink things.

Enter, JSON Schemas

JSON Schemas are a format for specifying what makes JSON documents valid. Schemas themselves are JSON documents which makes them easy to read and write. Since YAML is basically just JSON with cleaner syntax, you can use a JSON schema to validate YAMl files too!

Now I could have replaced my manually written markdown file with a manually written JSON schema file, but this would have been very unpleasant. I still want to format my docs with markdown, and considering JSON has no multiline strings that would have been very painful. It also would have meant manually documenting all of my type information.

Fortunately, many languages have tooling for generating JSON schemas from their type systems, and Rust is no exception. I discovered schemars which can be easily used to generate JSON schemas from my Rust structs and enums just by adding a #[derive(JsonSchema)] attribute to them. And the great thing is, schemars integrates with serde so that it understands serde attributes and produces a schema which matches serde's validation!

I did experience a couple of challenges using schemars, however. schemars uses Rust's doc comments to capture the JSON schema description property, but it does some mangling to the text first which means that it breaks markdown formatting. To get around this I am working off of a fork of the library, but I hope to merge in an option to address this properly in the future.

Another challenge I ran into was wanting to add extension properties to my JSON schema. schemars does support this through its Visitor API, but there is currently no attribute to add an extension property to a Rust type directly. My solution was to add a special line at the top of my descriptions when I wanted to add a custom property and then I parsed these out in a Visitor. I also hope to add this as an attribute to the core library when I get a chance.

Generating some docs

Now I just needed to generate some docs. Coming back to the format question from earlier, I decided to try and emulate some examples of other projects documenting large config files like GitHub Actions, GitLab CI, and Rollup.js. Basically, instead of organizing by classes, these docs are just a nested list of all the properties in the schema. Each property includes the full path name (e.g. jobs.<job-id>.name). This makes it very clear to the reader how the documentation can be applied to the config file itself!

I took a look at some existing JSON schema docs generators but none of them really looked the way I wanted them to. Most of them used a similar organization to what I started with, so I decided to roll my own.

I wrote a TS script to load a JSON schema and flatten the properties into a big list, along with some additional metadata like the full property name, and whether it is a required property.

Then, I wrote a simple Handlebars template which iterated over the properties and printed their docs in a consistent format.

Autocomplete and in-editor docs

To take advantage of my schemas even more, I tried loading it into VSCode. With the YAML extension installed, it was as easy as adding "yaml.schemas": { "mantle.yml": "schema.json" } to add autocomplete and in-editor docs to my config files.

But, they looked... bad. The formatting was all off.

It turns out VSCode interprets the description property of schemas as plaintext, and if you want it to look correct, you need to use a markdownDescription property. I was also using some non-standard markdown syntax which my docs platform (Docusaurus) supports (like admonitions and code block titles).

To resolve this, I wrote another TS script to transform a schema into something VSCode can properly interpret. I used remark to parse the markdown into an AST, then modified the AST.

My biggest challenge here was related to Node.js's ESM support. Recently, the remark ecosystem has "upgraded" to ESM-only, but currently TS has very bad interop with these modules. In the end, I decided to just use older versions of these packages which use CommonJS modules.

Tying it all together

Once I was happy with the schema I was generating from my types, I updated my deploy GitHub Action to generate the schema and upload it with my GitHub release assets. Now whenever I bump my version, a GitHub Action builds my project, creates a Release, and uploads the binaries and schema.

Then I updated my docs generation script to start by downloading all schemas from past releases. I use the latest one to generate my docs page, then I run all of them through my transformer to make them VSCode-ready and save them to my site's static directory. Now when I deploy the site, these schemas are hosted with it (e.g. my v0.11.0 schema).

My users can now just add "yaml.schemas": { "mantle.yml": "https://mantle-docs.vercel.app/schemas/v0.11.0/schema.json" } to add autocomplete to their editor!

TODO: Conclusion