I've been a Professional Config File Wrangler for two decades, and I can tell you that it's always nicer to have a config file that's built to task rather than being forced to tie yourself into knots when somebody didn't want to write a parser.

The difference between a data format and a configuration file is the use case. JSON and YAML were invented to serialize data. They only make sense if they're only ever written programmatically and expressing very specific data, as they're full of features specific to loading and transforming data types, and aren't designed to make it easy for humans to express application-specific logic. Editing them by hand is like walking a gauntlet blindfolded, and then there's the implementation differences due to all the subtle complexity.

Apache, Nginx, X11, RPM, SSHD, Terraform, and other programs have configuration files designed by humans for humans. They make it easy to accomplish tasks specific to those programs. You wouldn't use an INI file to configure Apache, and you wouldn't use an Apache config to build an RPM package. Terraform may need a ton of custom logic and functions, but X11 doesn't (Terraform actually has 2 configuration formats and a data serialization format, and Packer HCL is different than Terraform HCL). Config formats minimize footguns by being intuitive, matching application use case, and avoiding problematic syntax (if designed well). And you'd never use any of them to serialize data. Their design makes the programs more or less complex; they can avoid complexity by supporting totally random syntax for one weird edge case. Design decisions are just as important in keeping complexity down as in keeping good UX.

Somebody could take an inventory of every configuration format in existence, matrix their properties, come up with a couple categories of config files, and then plop down 3 or 4 standards. My guess is there's multiple levels of configuration complexity (INI -> "Unixy" (sudoers, logrotate) -> Apache -> HCL) depending on the app's uses. But that's a lot of work, and I'm not volunteering...

I quite like CUELang (https://cuelang.org/), although it not yet widely supported.

It has a good balance between expressivity and readability, it got enough logic to be useful, but not so much it begs for abuses, it can import/export to yaml and json and features an elegant type system which lets you define both the schema and the data itself.

I hope it gains traction.

toml is pretty much the best one I have seen so far. At least for small to medium size config files.

Toml has some hairy bits. Lists of objects, lists of lists of objects, objects of lists of objects. Complex objects with top level fields...

Yep,

Although I do feel like there is a case to be made that if you need a Turing complete configuration language then in most cases you failed your users by pushing too many decisions on to them instead of deciding on sensible defaults.

And if you are dealing with one of the rare cases where Turing complete configuration is desirable then maybe use Lua or something like that instead.

I'm not defending YAML. YAML is terrible. It's even worse with logic and/or templates (looking at you, Ansible). Toml is certainly better but I'm still baffled as to why we don't have a "better YAML". YAML could almost be okay.

There's also Lua, which is a full Turing complete language but is still pretty nice for writing config files, and is easy to embed.

Followup to my own post: don't forget about Scheme! Same nice properties as Lua, but you get some extra conveniences from using s-expressions (which can represent objects somewhat more flexibly, like XML, than Lua, which is more or less 1:1 with JSON).

There's StrictYAML[1][2]. Can't say I've used it as let's face it, most projects bind themselves to a config language - whether that be YAML, JSON, HCL or whatever - but I'd like to.

[1] https://hitchdev.com/strictyaml/

[2] https://github.com/crdoconnor/strictyaml

Yeah, I think it's because nobody sat down and methodically created it.

People create config languages that work for their use case and then it is just a happy accident if it works for other things.

I don't think anyone has put serous effort into designing a configuration language. And by that I mean collect use cases, study how other config languages does things, make drafts, and test them. etc...

Terraforms HCL is well designed.

I know a lot of people hate it but I find it to be the only configuration language that makes any sense for moderately large configs.

It’s short, readable, unambiguous, great IDE support. Got built in logic, variables, templates, functions and references to other resources - without being Turing complete imperative language, and without becoming a xml monstrosity.

Seriously there is nothing even close to it. Tell me one reasonable alternative in wide use that’s not just some preprocessor bolted onto yaml, like Helm charts or Ansible jinja templates.

There's a world of difference between "simple configuration needs" and "complex configuration needs".

I will take a kubernetes deployment manifest as an example that you would want to express in a hypothetically perfect configuration language. Now, eventueally, you end up in the "containers" bit of the pod template inside the deployment spec.

And in that, you can (and arguably should) set resources. But, in an ideal world, when you set a CPU request (or, possibly, limit, but I will go with request for now) for an image that has a Go binary in it, you probably also want to have a "GOMAXPROCS" environment variable added that is the ceiling of your CPU allocation. And if you add a memory limit, and the image has a Java binary in it, you probably want to add a few of the Java memory-tuning flags.

And it is actually REALLY important that you don't repeat yourself here. In the small, it's fine, but if you end up in a position where you need to provide more, or less, RAM or CPU, on short notice (because after all, configuration files drive what you have in production, and mutating configuration is how you solve problems at speed, when you have an outage), any "you have to carefully put the identical thing in multiple places" is exactly how you end up with shit not fixing themselves.

So, yeah, as much hate as it gets, BCL may genuinely be better than every other configuration language I have had the misfortune to work with. And one of the things I looked forward to, when I left the G, was to never ever in my life have to see or think about BCL ever again. And then I saw what the world at large are content with. It is bloody depressing is what it is.

Cuelang?

Yeah absolutely. I think there are four corners to the square: "meant to be written by humans/meant to be written by computers" and "meant to be read by humans/not meant to be read by humans". JSON is the king of meant to be written by computers read by humans, grpc and swift and protobuf and arrow can duke it out the written by computer/not read corner. We are missing good options in written by humans half.

Dhall and Cue come to mind as ones that _feel_ more designed

https://github.com/dhall-lang/dhall-lang

https://cuelang.org/docs/usecases/configuration/

Interesting...

Me, the programmer finds those kinda cool.

And the sysadmin in me developed a dislike of both within 1 minute of looking at them.

Honestly, I think a good configuration library should be more than a spec, it should come with a library that handles parsing/validation. See, there are two sides to configuration, the user and the program. Knowledge about the values, defaults and types should live on the program side and should be documented. Then the user side of configuration can be clean and easy to read/write and most important of all, allow the user to accomplish the most common configuration without having to learn a new config language on top of learning the application.

> Honestly, I think a good configuration library should be more than a spec, it should come with a library that handles parsing/validation

You just described CUELang.

The type system allows to define a schema as well as the data, in the same file, or in 2 separate ones. Then you can call either a cli tool (that works on linux, windows or mac) or use the Go lib (or bind to it).

For compat, cue can import and export to yaml, json and protobuf, as well as validate them.

Isn't Dhall basically the same (=have the same set of features)?

In the same way python and js are basically the same.

Exactly. So if I'm going to learn/use one of them, there's no clear winner, really. Both also seem to also have about the same amount of adoption (zero?).

Ok, you have convinced me to give it a serious look.

> Yeah, I think it's because nobody sat down and methodically created it.

I think it's the opposite. There isn't a single config file that suits all needs.

Especially when you realize config isn't a single thing.

http://mikehadlow.blogspot.com/2012/05/configuration-complex...

I guess that could also be the case.

I haven’t studied it, I am just generally feeling unhappy about most software configuration.

About Ansible, I think it gained it's success partially due to YAML.

Ansible is worse than Puppet and CFEngine in many ways, but it is superior in the user interface.

It managed to not only be a config management solution, but provide a universal config language that most apps could be configured with. So for a lot of use cases, if you know Anisible/YAML then you don't have to learn a new configuration language on top of learning a new application.

The problem with Ansible is it's not universal, because most app playbooks, are configured in the worst possible way. In my experience typically you get handed an Ansible script, something which you'd hoped was declarative but isn't (like a version that apt-get grabs isn't fixed, or even, gets patched) then suddenly a downstream templated command fucks up, and the person who wrote the script isn't around anymore (or you don't trust their chops because they are a blowhard that worked at Google/Facebook and had a coddling ops team behind them in the past) or worse it's from "community" and has a billion hidden settings that you can't be bothered to grok - and so you have to dig so many layers down that you are better off just fucking rewriting the Ansible script to do the one thing which probably should have been four lines.

In any case, I found Ansible scripts to have like a 3 month half life. If we were lucky. I'm not bitter.

haha, I can go on lengthy rants about every single configuration management system that I have used.

My dream configuration system should revert to default when the config is removed (keeping data). Have a simple/easy user interface. Have maintained modules with sane defaults for the 500 most common server software. I would rather there be no module than an abandoned one with unsafe defaults, that way it is clear that I would have to maintain my own if I want to use that particular piece of software. Performant, it really shouldn't take more than a few minutes to apply a config change. No more than 30 min for initial run.

Early on, Ansible was primarily agent-less from the start which made it ridiculously easy to sneak into existing infrastructure and manual workflows. I probably would not have been able to stand up Puppet or Salt or whatever but I could run Ansible all by myself with no one to stop me :).

I'm curious what your thoughts are on a config language I'm working on.

GitHub.com/vitiral/zoa

It has both binary and textual representation (with the first byte being able to distinguish them), and the syntax is clean enough I'm planning on extending it into a markup language as well.

Even if you have sensible defaults don't you still need to be able to parse configured changes?

Not always, sometimes all other options are just wrong. Or you can auto detect the correct setting.

I understand the pragmatic reasons for it being the way it is, but I still wish TOML didn't require all strings to be quoted.

This is why I like INI. It doesn't have these problems, because it doesn't try to wrangle the notion of nested objects (or lists) in the first place. The lack of a formal spec is a problem, sure, but it such a basic format that it's kind of self-explanatory.

When the problem is TOML not supporting easy nesting, a solution of "Don't nest." works just as well in TOML as it does in ini. It's not really an advantage of ini. Especially when a big factor in TOML not making it easy is that TOML uses the same kind of [section]\nkey=value formatting that ini does!

You can use toml as a better ini by limiting yourself to the key / value schema. It still superior because:

- it has a spec

- it has other types than strings

- you can always decide you actually need nested data, and add them later

I wrote an INI parser that has numerical, boolean, timestamp, MAC address, and IP address types ;) "advantages" of not having a spec!

Seriously: for application-specific config files, the lack of a formal spec can be kind of a nice thing. You can design your parser to the exact needs of your program, with data types that makes sense for your use case. Throw together a formal grammar for use in regression testing, and you're all set.

Obviously a formal spec is essential for data interchange, but that's why JSON exists. To me, YAML is in a gray area that doesn't need to exist. The same thing goes for TOML, but to a far lesser extent.

> it has other types than strings

But isn't the config file just a string?

Everything gets serialized to a string of bytes. The point is that you can fail at parsing when the value doesn't make sense, rather than failing at some point in the future when you decide to use the value and it doesn't make sense. And if you have a defined schema, you can have your editor validate it against the schema when saving, so you don't accidentally have "FILENOTFOUND" in a Boolean.

... lack of a hex float representation ...

We do, it’s called TOML. The future is here it’s just not equally distributed.

TOML sucks for list of tables simply because they intentionally crippled inline tables to only be able to occupy one line. For ideology reasons ("we don't need to add a pseudo-JSON"). Unless your table is small, it's going to look absolutely terrible being all crammed into one line.

https://github.com/toml-lang/toml/issues/516

The official way to do list of tables is (look at how much duplication there is)

  [[main_app.general_settings.logging.handlers]]
    name = "default"
    output = "stdout"
    level = "info"

  [[main_app.general_settings.logging.handlers]]
    name = "stderr"
    output = "stderr"
    level = "error"

  [[main_app.general_settings.logging.handlers]]
    name = "access"
    output = "/var/log/access.log"
    level = "info"

vs

  handlers = [
    {
      name = "default",
      output =  "stdout",
      level = "info",
    }, {
      name = "stderr",
      output =  "stderr",
      level = "error",
    }, {
      name = "access",
      output =  "/var/log/access.log",
      level = "info",
    },
  ]

I would still reach for TOML first if I only needed simple key-value configuration (never YAML), but for anything requiring list-of-tables I would seriously consider JSON with trailing commas instead.

I see the point and this is certainly a drawback of TOML but for me this is something of a boundary case between configuration and data.

When configuration gets so complicated that the configuration starts to resemble structured data I tend to prefer to switch to a real scripting language and generate JSON instead.

Expression languages like Nix, Jsonnet, Dhall, and Cue are really nice in these situations.

for this reason I couldn't see a CI platform ever seriously consider TOML

(someone may point out to me a CI platform that relies on TOML—which I welcome)

Rust is built on TOML. For better or worse.

Do you mean Cargo? Because Cargo is not a CI system. You never embed shell commands in a Cargo.toml.

If you need to program complex logic to build a crate, you don’t write TOML. You write a build.rs file in actual Rust.

If embedding shell commands in a configuration language is considered a CI system I think we are doomed.

> JSON with trailing commas

JSON5?

It's perfect until you do a lot of nesting..

...or any nesting. TOML sucks for anything non-trivial.

It makes me sad every time I see a newly announced tool that went for YAML instead of TOML.

XML is still good.