I'm a big fan of Dhall in general, because:
- It's an aggressively simple language, so there's not that much to learn
- It's pure and functional (all it does is compute config, it's not for writing programs)
- It's strongly typed, which hugely increases confidence compared to editing pages of YAML
In addition, Dhall turns out to be an excellent way to manage CI boilerplate.
The problem with boilerplate is that you can't abstract it away. We typically end up copy/pasting or using generators, both of which miss out on later improvements made to the source. Ideally we could do what we normally do in languages, which is use libraries.
But when you're talking about github workflow files, travis configuration files, Dockerfiles, cloudbuild files and all the other CI configuration, it's impractical to do that. You'd have to pick a language, and make sure that's installed. Then probably a package manager. Then you'd have to put together a bit of a framework for generating various files, and sort out package publication for anything you want to share between projects.
We need something that's simple, appropriate for the task and supports super lightweight dependency management. Dhall is exactly that:
- Frictionless dependency management: you can import dhall code from the internet (safely, and with efficient caching)
- Lightweight and simple to install (a couple static binaries totalling a dozen mb)
- Purpose-built: dhall's stated goal is to replace YAML. Which is (mostly) what we're doing!
When run, it will replace whatever is currently present in the output directory (generated by default) with generated files. You should never set dhall-render's output directory to something which already contains useful files, they will be deleted.
Pointing this to an existing directory (effectively deleting it) is an unfortunate footgun to enable, but unfortunately it's a byproduct of an otherwise effective system - it's simple and successfully garbage collects unwanted files regardless of how you got into your current state.
The idea is that we can have one big dhall expression to define "all the generated files". This is super powerful - thanks to dhall's remote imports, you can use whatever types and abstractions you need by importing straight from the internet (it even supports private Github repositories).
The expression contains a map of path -> file expression, each of which is a dhall value (containing the file contents as well as some metadata).
We render this tree of files (into a generated/ directory for cleanliness), then install symlinks in the workspace, pointing to the various generated files.
There's options to control the generation - you can mark files as executable, or use a plain file instead of a symlink (e.g. to workaround Github actions choking on symlinks).
And it's self-hosting. You can use the SelfInstall module to add this tool as a generated file in your repo, eliminating manual setup. You can even pre-bake the default file path to use, resulting in a custom script which doesn't need any arguments.
The quickest way is with this one-liner:
curl -sSL https://raw.githubusercontent.com/timbertson/dhall-render/master/bootstrap.sh | bash
That'll get you set up with a default dhall/files.dhall (a copy of bootstrap/files.dhall), run the initial render to install dhall-render itself, and pin your dhall-render version to a specific commit.
If you'd rather set it up manually, make a copy of bootstrap/files.dhall in dhall/files.dhall, then run:
echo '(./dhall/files.dhall).files.`dhall/render`.contents' | dhall text | ruby
You should then run ./dhall/fix --freeze ./dhall/files.dhall to pin your dhall-render version.
Yes, typically. One of the main uses is for files which must be in the repository, as a contract (e.g. .travis.yml or .github/workflows/*.yml). So for those you have no choice.
If you mark files as "linguist-generated" (see [./gitattributes][]), they'll be hidden by default in Github pull request diffs, which can be convenient.
In order to make sure your generated files remain in sync with the source expressions, you can follow a process outlined in the dhall manual.
You can use the exe attribute to get the default dhall-render executable (ruby script).
If you don't use files.dhall as your file expression, you can use SelfInstall.makeExe SelfInstall::{ path = "path/to/files.dhall" } to create a self-install with a different default path.
There's also the fix attribute, which is the (not stable but handy) [./maintenance/fix][] script. This walks the current directory (or arguments) and by default evaluates and formats any .dhall files it finds. You can pass --lint, --freeze etc to perform other operations instead.
dhall-render simply runs dhall-to-json on your files.dhall expression, and processes the results. If you have type errors, you can run dhall --file files.dhall --explain to get more details.
dhall-to-json and ruby
Honestly, it's liberating. Until I started down this path, I didn't realise how much mental friction I had with CI boilerplate. I was always unmotivated to fix things or improve my workflow, because it was tedious and I knew I'd have to completely redo it for every repo where I wanted the benefit.
Now, I feel like it's actually worth making proper reusable abstractions for this stuff because I can actually leverage it across as many repos as I need, and only have to maintain it once.
If you only have a couple of repos, probably not. I have a great deal of personal and work repos that I deal with, and for me the thought that each one doesn't have to reinvent the wheel is very satisfying.
It's not zero-dependency, obviously -- you can't get around your contributors needing dhall in order to change things. Many work environments have ways to make this simple (e.g. shared Docker tooling). But for public repos it depends what environment / tooling you can assume your users have.
I wrote it quickly, and it's only got a handful of automated tests. But it's very simple, the examples work, and I use it daily.
It was expedient, because I want to use it in a workplace where ruby is ubiquitous. Maybe one day I'll rewrite it in Haskell, but then I'd need to set up more infrastructure in order to actually automate and distribute binaries.