Drupal 8 uses YAML* as its configuration language because JSON doesn't support comments. That simple. Thank you for YAML, it does deliver for us: it's human readable and it's easy to parse (see below).
* I mean, it uses an ill defined subset of YAML. The definition is "whatever the Symfony YAML parser supports".
You know what else is human readable, easy to parse if you're using PHP, and supports comments?
PHP.
I understand why some languages rely on common configuration file formats.
I don't understand why the popular dynamic script-y languages don't more commonly use the natively-expressable associative/list data structures that they're famous for making convenient.
Using includes/imports is not the greatest idea ever.
Your configuration file is one of your program interface. It's something that must be well define. If your configuration file is a programing language this interface is not that well defined.
Also you expose yourself to all kind of weird bugs because some (too smart for their own good) people will monkey patch your software using it.
It adds a lot of unnecessary stuff in the configuration file, things like ';' or '$' are not really useful.
Lastly, common configuration file format are good because there are... common. You can have 2 pieces of software in 2 different languages accessing the same configuration file. A common example of that is configuration management, There are a lot of modules/formula in salt/ansible/puppet/chef doing fine parsing of the configuration files and permits fine grain settings, and I'm not mentioning augeas. If your configuration is a php/python/perl/ruby file good luck with that.
I know it's really common for php applications to do configuration files in php, but frankly, it's a bit annoying.
> If your configuration file is a programing language this interface is not that well defined.
While I do agree with the rest of your comment I don't think they were advocating using the full language for configuration, just the maps/arrays/etc. (e.g. Python's `literal_eval`).
If a key objection/perceived threat is that this might give someone an insertion point they're not meant to use for code ... well, let's consider that we're talking about applications distributed as interpreted language source here. Disallowing code-as-config isn't even closing the door of this particular barn after the horse has left, it's putting two strands of police line tape across the bottom half of the gap where the door was never installed and hoping any equines thinking of passage politely consider the message in case it hadn't already occurred to them which side of the entrance they preferred to be on.
Consider this: Design and optimize for the common case.
Why do we have config files? Because developers actually want a place dedicated to simple or structured application configuration data, for which PHP assignments with arrays + primitives can function at least as effectively as JSON. Most developers would prefer that config data get loaded quickly so the application can get on to doing actual app-y things. Using the language for this means you're parsing at least as fast as you can interpret and you can also take advantage of any code caching that's part of your deployment (especially nice in the PHP-likely event that config settings would be reloaded with every request).
Abuse isn't likely to be the common case. The end users you invoked certainly aren't going to be the ones looking for opportunities to insert code over data. Developers have other places to put code and, as mentioned, probably actually want a place dedicated to data. You're still right that of course someone will do it, just like someone will inevitably create astronaut architecture hierarchy monstrosities in any language with classical inheritance or make potentially hidden/scary changes to language function using metaprogramming facilities.
But potential for abuse doesn't automatically mean a feature should be disallowed.
A lot of the time it's better to let people who can be circumspect have the benefits of a potential approach, and if somebody thinks they need to solve a problem by using a technique that's arguably abuse, well, let them either find out why it's a bad idea or enjoy having solved their problem in an unusual way. Not the end of the world. Possibly even legit.
You can use arbitrary tools to programmatically generate YAML (or JSON, or XML, any of the other "data only" formats.) This allows for tools to drive other tools by generating a spec file and feeding it in. See e.g. Kubernetes for a good example of that.
There's no language that I'm aware of that can natively generate PHP syntax, and there's no common multi-language-platform library for generating PHP syntax. I think that's most of the reason.
To contradict myself, though: Ruby encodes Gemfiles and Rakefiles as Ruby syntax. And Elixir encodes Mixfiles, Mix.Config files, Distillery release-config files, and a bunch of other common data formats as Elixir syntax.
And, of course, pretty much every Lisp just serializes the sexpr representation of the live config for its config format (which means that, frequently, a lot of Lisp runs code at VM-bootstrap time, because people write Turing-complete config files.)
> There's no language that I'm aware of that can natively generate PHP syntax
This is a solid argument against using PHP (or any such language) as a cross-language data interchange format. There are others :) And I totally agree you want a language independent format for anything you might have to feed across an ecosystem of tools.
For a PHP-system generating/altering its own config files... PHP's `var_export` generates a PHP-parseable string representation of a variable (though it sadly doesn't use the short array syntax).
Turing-complete config files probably have some hazards, like Lisp itself does. YMMV regarding whether those hazards can be avoided by circumspect developers or need to be fenced off.
This, and the security problems of executable code as configuration, are why the OpenBSD people mandate that /etc/rc.conf is not general-purpose shell script, and why the systemd people mandate that /etc/os-release is similarly not. People want to be able to parse configuration files like this with something other than fully-fledged shell language interpreters; and they want these things to not be vectors for command injections.
Settings.py is uniquely bad, though, IMO because it tries to be a badly defined dict(), instead of exposing proper configuration interfaces. Ruby config files are common and usually fairly great, see for example the Vagrantfiles.
And you won't have to generate your config files (parsing, maaaaaybe), because those needs are covered by the fact that the files are programs. They are _already_ generating a configuration.
> And you won't have to generate your config files (parsing, maaaaaybe), because those needs are covered by the fact that the files are programs. They are _already_ generating a configuration.
Yes, theoretically, if settings.py was a "generator" format that you ran as a pre-step (like you do to get parser-generators like Bison to spit out source files for you to work with), and this generator actually spat out something like a settings.json, and all the rest of the infrastructure actually dealt with the settings.json rather than the generator, then, yes, it wouldn't matter. Tools in other languages could just generate the settings.json directly.
As it stands, none of those things are true, so tools in other languages actually need to do something that outputs settings.py files.
Galaxy brain: if your config is programmable, it can read whatever terrible configuration format you want. That means my settings.py (yes, I'm forced to use Django) is configured via environment, which is populated by k8s from - gasp - JSON files.
That means that if I wanted to configure Vagrant with JSON, there is no force in the universe that could stop me.
If the config file is actually a normal program, then it can do normal program things, then any benefit from using JSON instead is nullified by the fact that you can still use JSON. In turn, if your tools primary configuration is via a more limited settings, you're stuck with it. Not even "generators in other languages" allow comparable runtime flexibility.
Yup, totally agree with you, settings.py has always been a pain in the ass. Not really an acute one but the kind that is uncomfortable but not enough to make you do something about it.
> There's no language that I'm aware of that can natively generate PHP syntax.
Actually, I've had to use PHP to output a PHP configuration array for a project that required config in PHP.
`var_export($foo)` will output valid PHP code for creating the array $foo. In my case I was doing horrible things to create the array in my pseudo-makefile, then using `var_export()` to output the result. Note that you can run php from the Bash CLI with the `-r` flag, which helps.
Tcl works well for configuration files. You can strip away the extraneous commands in a sub-interpreter to prevent Turing completeness and add infix assignment to remove the monotony of the set command and what you get is a nice config format. If you need more power in the future you just relax some of the restrictions and use it as a script without breaking existing files.
People get really upset when they have to type "array(" instead of "[" or "{" (pre-PHP 5.something) and quotes instead of no quotes (and punting the character escape problem to something else) I guess.
Using code-as-data works really well in Lisp-like languages. Reading a Clojure project's project.clj file or a Lisp project's project.asdf file is pretty pleasant. A programming language's choice in how it decides to handle library config info for building and specifying dependencies (XML, makefiles, JSON, YAML, INI, nothing, etc...) will be a good indicator for the culture of the language around config files in general. Composer for PHP only came out in 2012.
Interestingly, the Lua programming language actually evolved from configuration files: https://www.lua.org/history.html (and is still officially deemed useful for writing them)
I use Lua for configuration files for both personal and work related projects [1]. You get comments and the ability to construct strings piecemeal (DRY and all that). It's easy to sandbox the environment, and while you can't protect against everything (basically, a configuration script can go into an infinite loop), if someone unauthorized does have access to the script, you have bigger things to worry about.
That was also one of the rationales behind TCL's design.
John Ousterhout explained in one of his early TCL papers that, as a "Tool Command Language" like the shell but unlike Lisp, arguments were treated as quoted literals by default (presuming that to be the common case), so you don't have to put quotes around most strings, and you have to use punctuation like ${}[] to evaluate expressions.
TCL's syntax is optimized for calling functions with literal parameters to create and configure objects, like a declarative configuration file. And it's often used that way with Tk to create and configure a bunch of user interface widgets.
Oliver Steel has written some interesting stuff about "Instance-First Development" and how it applies to the XML/JavaScript based OpenLaszlo programming language, and other prototype based languages.
>The equivalence between the two programs above supports a development strategy I call instance-first development. In instance-first development, one implements functionality for a single instance, and then refactors the instance into a class that supports multiple instances.
>[...] In defining the semantics of LZX class definitions, I found the following principle useful:
>Instance substitution principal: An instance of a class can be replaced by the definition of the instance, without changing the program semantics.
In OpenLaszlo, you can create trees of nested instances with XML tags, and when you define a class, its name becomes an XML tag you can use to create instances of that class.
That lets you create your own domain specific declarative XML languages for creating and configuring objects (using constraint expressions and XML data binding, which makes it very powerful).
The syntax for creating a bunch of objects is parallel to the syntax of declaring a class that creates the same objects.
So you can start by just creating a bunch of stuff in "instance space", then later on as you see the need, easily and incrementally convert only the parts of it you want to reuse and abstract into classes.
>I don't understand why the popular dynamic script-y languages don't more commonly use the natively-expressable associative/list data structures that they're famous for making convenient.
You picked the wrong language... PHP comes with its own JSON parser. And INI and XML and even CSV.
But, the reason is that, generally, you want config files to describe data or state only. Yes, you could just make your config native code, but then the temptation to add functions and methods and logic to that becomes irresistible and soon your config is an application that needs its own config.
Config formats need to be simple, and preferably not Turing complete.
Because it's just in general incredibly short sighted to think that your config file is never going to be read by code written in another language.
There's also an argument about whether making configuration files able to execute arbitrary code is a good idea. You get straight into the JavaScript 'eval' problems which we've spent a decade escaping.
I think some of it is PLOP (Principle of Least Power).
$CFG = random() > 0.5 ? "yes" : "no";
...is likely "too powerful". It'd be nice if there were ways in certain programming languages to do something like "drop privileges" to avoid loops, function calls, external access, etc.
The makers of Drush, the cli for Drupal, subscribed to your line of thinking in the early versions and inventory items were defined in PHP files. Migrating from that will be interesting.
Because that forces the end user, who might not know anything about the programming language one’s application is written in to wrestle with the low level implementation details. In the words of Keith Wesolowski, the programmer assumes that the end user is a “Linux superengineer”, which is almost always a wrong assumption to make.
I totally agree that in the ideal world, JSON should support comments. I yearn for them, and none of the in-band work-arounds or post-processing tools are acceptable substitutes.
But to play the devil's advocate, how would JSON be able to support round-tripping comments like XML can, since <!-- comments --> are part of the DOM model that you can read and write, while JSON // and /* comments */ are invisible to JavaScript programs. There's nowhere to store the comments in the JSON object model, which you would need to be able to write them back out later!
On important feature of JSON is being able to read and write JSON files with full fidelity and not lose any information like comments. XML can do that, but JSON can't. To fix that you'd have to go back and redesign (and vastly complicate) fundamental JavaScript objects and arrays and values, to be as complex and byzantine as the DOM API.
The less-than-ideal situation we're in isn't JSON's fault or JavaScript's fault, because JSON is just a post-hoc formalization of something that was designed for a different purpose. But JSON is rightly more popular than XML, because it's extremely simple, and nicely impedance matched with many popular languages.
YAML suffers from the same problem as JSON that it can't round-trip comments like XML can, but it fails to be as simple as JSON, is almost as complex as XML, and doesn't even map directly to many popular languages (as the article points out, you can't use a list as a dict key in Python, PHP, JavaScript, or Go, etc).
You can sidestep some of JSON's problems by representing JSON as outlines and tables in spreadsheets, without any need for syntax and sigils like brackets, braces, commas, no commas, quoting, escaping, tabs, spaces, etc, but in a way that supports rich formatted comments and content (you can even paste pictures and live charts into most spreadsheets if you like), and even dynamic transformations with spreadsheet expressions and JavaScript.
In fact YAML is probably more complex than XML; the specification of YAML, when I print it into PDF, is about three times as long as that of XML 1.0. (And XML 1.0 also describes DTD, which is kind of a simple type validation for XML and thus includes much more than just serialization syntax.)
> But to play the devil's advocate, how would JSON be able to support round-tripping comments like XML can, since <!-- comments --> are part of the DOM model that you can read and write, while JSON // and /* comments */ are invisible to JavaScript programs.
It doesn't support it for whitespace in general (if you deserialize into JS object model or equivalent), so why would it be any different for comments specifically? It's just not a design goal of the format.
Although, of course, it's quite possible to have a JSON parser that preserves representation. It'll just have a non-obvious mapping to the host language because of all the comment and whitespace nodes etc.
I'm kinda sad that JSON has been struggling for like 15 years to get comments. Is there like some kind of gestapo that's saying no or something? All it takes is for the maintainers of probably 15 popular libraries to start handling comments.
At the end of the day I'm sure the reason we don't have JSON comments is somewhere listed in this page: xkcd.com/927/
I'm aware of at least three JSON libraries that at least can accept comments (Gson in lenient mode, Json.NET, and json-cpp are the ones I've used personally that do)-- it's hard to convince everyone that JSON needs comments, though, and comments are of limited utility if it's not guaranteed that they'll parse everywhere.
But you really only need comments in JSON if you're doing stuff like storing configuration in JSON, and JSON's too fiddly in general to be a great config file format (too easy to do something like forget a comma; no support for types beyond object, array, (floating point) number, and string). Something more like YAML without the wonky type inference would be better, IMO.
I believe Douglas Crockforf used to make the argument that JSON is not meant for human consumption and thus shouldn't be changed to better serve humans. I personally wish hjson (https://hjson.org) were to get more traction. I prefer it over both JSON and YAML.
> I removed comments from JSON because I saw people were using them to hold parsing directives, a practice which would have destroyed interoperability. I know that the lack of comments makes some people sad, but it shouldn't.
Well, then why not to allow a trailing comma in lists and objects? Computers don't care and they would even be happier, because they can then just pour array and object members with a trailing comma without concerning themselves whether this is the last member or not. (Dijkstra's train toilet problem comes to mind.) Also compare with XML, where each element is self-contained.
And why to model JSON syntax closely after JavaScript literal object syntax (which is actually more convenient, by the way) which, being taken from mainstream programming languages, naturally evolved to be written by humans in small amounts not by computers in large dumps? :)
VS Code uses JSON with comments for config files. [1]
Technically, this is not JSON. You won't be able to use a standard JSON parser without stripping comments first. But you can use a simple, JSON-like language with comments for config.
> simple
YAML is much, much more complicated than JSON.
Quoting a single word from the parent’s sentence is misleading. The sentence "YAML can be employed as a simple JSON-like language with comments." is true because JSON is YAML, so you can parse a JSON file with #-comments using a YAML parser.
Most YAML users don't need to look at the source for a YAML parser. I appreciate elegant simplicity, but I don't think parser complexity is the most important metric by which to judge a data interchange format.
If you use a YAML parser to parse JSON-with-comments, it will accept many inputs that don't correspond to JSON-with-comments, and furthermore is likely to report syntax errors that don't make sense to a user who only knows JSON.
So, this unnecessary parser complexity is a usability issue. You should use a parser for the config language you actually intend to support.
Jsonnet is awesome. We use it to generate our yaml files for kubernetes. YAML isn’t easy to parse, nor is it very flexible as a templating language. It gets cumbersome very quickly.
Jsonnet is a relief. Kubernetes should have been a dumb json config from the get go. JSON is ridiculously simple to parse and emit. It has huge interoperability as well with lots of programming languages.
I don't like it because it uses the = symbol which seems imperative rather than declarative. (Same with HCL, it might be a nitpick but these are languages I'm going to be using all the time.)
HOCON is interesting but at first glance it seems it might be too ambiguous for my tastes, because like YAML, because it supports both js-style ("//") and shell-style ("#") comments.
JSON plus comments is beautiful because it adds minimally to an unambiguous language which lends itself to automatic formatting (stringificiation).
I'd argue that = only feels imperative if you're used to imperative languages. Prolog and Haskell, both of which focus on being declarative, also both use the equals sign.
The trouble there is that your comments come in-band. What if you're trying to serialise something and you don't have the power to insist that it's not a dictionary with "comment" as a key?
Sigh. All I wanted to do is to say thanks for the YAML standard -- comments are important but not the only problem with JSON. And truly I can't be expected to remember all of this discussion from like six plus years ago. One thing I remember though, it the trailing comma problem -- we upstreamed a grammar change to Doctrine annotation so "foo, bar," is OK because PHP arrays accept that and it's bonkers trying to code a mostly PHP system without trailing comma support. Also, JSON is no fun to write , you need to have [] {} all correct where YAML is much easier. The less sigils the better and most of Drupal YAMLs only use the dash, the colon and the quote. This is the grave mistake Doctrine committed as well, instead of simple arithmetic (>=1.0) they used mysterious sigils in version specification (~1.0). Drupal is in the business of constantly accepting new contributors and (~R∊R∘.×R)/R←1↓ιR is not newbie friendly, no matter how you slice and dice it. There are certainly advantages of sigil heavy languages like APL and Perl but the scare factor is too high.
That's just ugly and you're mixing your comments with the data structure, which is potentially confusing. Also, Jason requires a lot more typing. I don't want to have to manually add in all the brackets, quotes and commas when editing config a file.
* I mean, it uses an ill defined subset of YAML. The definition is "whatever the Symfony YAML parser supports".