> Everyone hates YAML. Everyone writes a lot of YAML.

I don’t know a single engineer (ops or not) who enjoys writing YAML, yet it is utterly unavoidable.

I’ve lost count of the amount of bugs and broken deploys that have happened because of YAML, or because of a type error caused by it.

Yeah, what I want is JSON with comments and nice multi-line string support. I don't like how much syntactic magic YAML does (I don't need or want the country code for Norway to be parsed as a boolean False value). I still don't know what the exclamation point does (e.g., !Ref).

And clearly YAML is the wrong tool for infra-as-code since CloudFormation has to build a macro system, conditionality, referencing, and a couple different systems for defining and calling functions (templates being one and their implicit functions being another). We also see tools like Troposphere and CDK which are effectively different ways to generate CloudFormation YAML via programming languages (or more precisely programming languages that were designed for humans).

And it's not just limitations inherent to CloudFormation--Helm has long had templates for generating YAML, but those also weren't sufficiently powerful/expressive/ergonomic so Helm3 is supporting Lua as well. And as I understand it, Terraform is constantly adding more and more powerful features into HCL.

So what's the solution? It's pretty simple--we should keep the YAML around, but it should be the intermediate representation (IR), not the human interface. The human interface should be something like a functional language[^1] (or an imperative language that is written in a functional style) that evaluates to that YAML IR layer. The IR is then passed to something like Kubernetes or Terraform or CloudFormation which understand it, but it's not the human interface.

As for the high-level language, something like [Starlark][0] would work well. It's purpose-built for being an evaluated configuration language. However, I would argue that a static type system (at least an optional static type system) is important--it's easy enough to imagine someone extending Starlark with type annotations and building a static type checker (which is much easier for Starlark since it's a subset of Python which is intended to be amenable to static analysis).

This, I think, is the proper direction for infrastructure-as-code tooling.

[^1]: Functional in that it is declarative instead of imperative--not necessarily that the syntax should be as hard to read as OCaml or Haskell. Also, while YAML is also declarative, it doesn't have a notion of evaluation or variables.

[0]: https://docs.bazel.build/versions/master/skylark/language.ht...

We’ve been using starlark for kubernetes configs at my prev gig and I quite liked it. Open-sourced some of that stuff as https://github.com/cruise-automation/isopod (which is based on https://github.com/stripe/skycfg). I hear stripe are also using their thing with terraform although not sure to which extent.