My wish-list for the next YAML July 28, 2021 on Drew DeVault's blog

YAML is both universally used, and universally reviled. It has a lot of problems, but it also is so useful in solving specific tasks that it’s hard to replace. Some new kids on the block (such as TOML) have successfully taken over a portion of its market share, but it remains in force in places where those alternatives show their weaknesses.

I think it’s clear to most that YAML is in dire need of replacement, which is why many have tried. But many have also failed. So what are the key features of YAML which demonstrate its strengths, and key weaknesses that could be improved upon?

Let’s start with some things that YAML does well, which will have to be preserved.

• Hierarchical relationships emphasized with whitespace. There is no better way of representing a hierarchical data structure than by actually organizing your information visually. Note that semantically meaningful whitespace is not actually required — the use of tokens like { is acceptable — so long as, by convention, hierarchies are visually apparent.

• Defined independently of its implementation. There should not be a canonical implementation of the format (though a reference implementation is, perhaps, acceptable). It should not be defined as “a config library for $language”. Interoperability is key. It must have a specification.

• Easily embeds documents written in other formats. This is the chief reason that YAML still dominates in CI configuration: the ability to trivially write scripts directly into config file, without escaping anything or otherwise molesting the script.

  tasks:
  - configure: |
      jit_flags=""
      if [ "$(uname -m)" != "x86_64" ]
      then
          jit_flags=--without-jit
      fi
      ./configure \
          --prefix=/usr \
          $jit_flags    
  - build: |
      make    
  - test: |
      make check    
  

• Both machine- and human-editable. It’s very useful for both humans and machines to collaborate on a YAML file. For instance, humans write build manifests for their git.sr.ht repos, and then the project hub adds steps to download and apply patches from mailing lists before submitting them to the build driver. For the human’s part, the ability to easily embed scripts (see above) and write other config parameters conveniently is very helpful — everyone hates config.json.

• Not a programming language. YAML entities are a problem, but we’ll talk about that separately. In general, YAML files are not programs. They’re just data. This is a good thing. If you want, you can use a separate pre-processor, like jsonnet.

What needs to be improved upon?

• A much simpler grammar. No more billion laughs, please. Besides this, 90% of YAML’s features go un-used, which increases the complexity of implementations, not to mention their attack surface, for little reason.

• A means of defining a schema, which can influence the interpretation of the input. YAML does this poorly. Consider the following YAML list:

  items:
  - hello
  - 24
  - world
  

  Two of these are strings, and one is a number. Representing numbers and strings plainly like this makes it easier for humans to write, though requiring humans to write their values in a format which provides an unambiguous type is not so inconvenient as to save this trait from the cutting room floor. Leaving the ambiguity in place, without any redress, provides a major source of bugs in programs that consume YAML.

• I don’t care about JSON interoperability. Being a superset of JSON is mildly useful, but not so much so as to compromise any other features or design. I’m prepared to yeet it at the first sign of code smells.

Someday I may design something like this myself, but I’m really hoping that someone else does it instead. Good luck!