I'll never get how merely having function names that reflect the use case, plus a stripped down or absent standard library, qualifies as a DSL. I know some people have long used "DSL" in this way, especially among LISP fans, but... I just don't get it. If I want a DSL it's because I want something that gives me, e.g., novel control flow constructs a la AWK, or highly specialized semantics a la regular expressions, that directly suit the problem domain. If I'm not getting that kind of power, why tie myself to some esoteric dependency? Either way you're adopting a tremendous maintenance burden; it better be worth your while.
I'm a huge fan of Lua and have used it for many projects in different roles, but never once thought of any particular case as having created a DSL, even when stripping the environment to just a few, well-named, self-describing functions.
I don't mean to criticize this particular project. Good code is good code. It's just the particular conceptualization (one shared by many others, to be fair) of what a "DSL" means that bugs me.
generally this would be called an "embedded domain specific language". Some languages are relatively flexible to change the syntax. For example Common Lisp has reader macros to change the token syntax and macros to change the Lisp syntax. With that one can create all kinds of embedded languages, incl. domain specific languages (languages which are specific to a special domain). Examples would be embedded logic languages, query languages, rule based languages, languages to describe user interfaces, etc. The Common Lisp standard has a notorious example for that, a complex LOOP construct, which uses a very different syntax: https://www.lispworks.com/documentation/HyperSpec/Body/m_loo...
There are other real-world examples out there, for example an embedded domain specific language to describe 3d objects in the domain of parametric CAD, for description of technical things like turbines or other parts of an aircraft.
This is basically a lib with some extra syntax to parse CSV Files.
A 'proper' DSL would require a very specific domain where it is applied to. Like document creation, or solving a certain problem and only that but not much else. Turing completeness is usually not required as well.
For example Matlab or LaTex are domain specific as well as SQL. Those are used to do math, create documents or do mangle tables.
Imho just renaming forEach to Map to parse CSV files does not fit the bill so the linked example is not that great.
This project basically a DSL builder thingy with a text processor demo.
As an aside:
I am missing the most important thing when it comes to CSV which is configuration of the input.
Might be because this is more of an example but it is usually a sign of a lets say more academic project.
Working with CSV, is usually a source of a lot of "good fun" where many hours can be spent.
Because your average CSV is often a SSV or TSV in some ISO that is everything but not UTF8. It usually contains line breaks which have been renamed to funny icons by some combo of tool and operating system. Also there are weird escape character in orders which are not consistent on every row. Also sometimes you have titles sometimes not. Dates make no sense, are language depended and in a weird format the intern made up 10 years ago. And even numbers are weird too, like 12e^-25 or '0.00' or '10.000,0'. Then you get columns which really should be 2 or more, or there are lines which span multiple rows.
Ihmo it is way better for robust CSV parsing to have a really low level approach where you rigorously check for everything (and return everything to sender that does not fit).
1. Given that their stated use-case for this tool is generation of data for testing purposes, why would input configuration be relevant?
2. It doesn't have any CSV-associated features to begin with. If for some reason you wanted to use this for that, you would interact with your own bound Go functions to facilitate working with CSV, just as demonstrated in the article. Within those functions, you can setup any configuration your heart desires.
Any naming of types, functions or variables that you do while solving a problem in software is creating a "language" of terms that are specific to the domain.
A well-constructed fluent API can read a lot like what you would call a DSL. A configuration language in YAML is both YAML and a DSL for config.
Without those function names that you talk about, you wouldn't really recognize a language. Those standard library function names give the "batteries included" kind of feeling. The size of the community libraries furthers that accessibility and productivity feeling of the language. Furthermore, you can certainly create control flows via libraries--with callbacks specifically--and create any kind of novel branching structure your domain had need of.
A good API is a sort of DSL. If it well reflects how you think about the domain and helps you express instructions within that domain. The language can be very different. We experience that in our own language when we hear people talking with heavy use of jargon we don't recognize: they might as well be speaking another language.
But overall I agree with you that I would probably save the use of the jargon term DSL for novel syntax targeted at a specific domain.
I think that's too strict. For example, JAX is (imo) an eDSL, but it doesn't have a specialized syntax, grammar or constructs - on the contrary, it's meant to feel just like Numpy. The thing that makes it special is the interpretation of those constructs.
Mojo just discovered it ain't happening: https://news.ycombinator.com/item?id=41927000
If I create this kind of mapped functions DSLs I can assure that things will be done a a certain way vs the borderline infinite possibilities of code.
and by DSL they mean "extension language". like vimscript or emacs lisp. or guile or python...
so you can easily add primitives to your application specific language, in go, specifically.
https://gitlab.com/gitlab-org/vulnerability-research/foss/li...
I think this fits nearly with the common understanding "DSL" doesn't it?
DSLs are such a trap for most projects that think they need them. Use lua or something off the shelf for scripting.
CEL exists for Go and (safely) solves many of the problems you might also want a DSL for.
The case for DSLs is often hard to justify in a project that has to be maintained for years.
As a DSL CEL is kinda pointless, since it does not create any additional convenience beyond the usual mainstream programming language syntax. It therefore leaves potential on the table, and as a tradeoff appeals to familiarity of syntax. As a configuration language it is usable, probably with reduced risk, compared to using Golang itself (no turing-completeness!).
I don't think it actually appeals to anyone, who considers creating a DSL for a good reason.
[1] https://github.com/google/cel-spec/blob/master/doc/intro.md
CEL-go:
At my last job, we spent 3 years building an electrical systems product with the same name as an eye scanning company. We had invested in marketing and sales that was practically useless once the executives finally agreed they needed to rebrand. It was the sort of thing that was pointed out on day zero, but the man in charge was too proud of the name. It wasn't even a very good name for our product.
And that was a bet-the-company type of choice. Lingo here is just some little library of no import to Gitlab's business.
Or probably grule? [1] https://github.com/hyperjumptech/grule-rule-engine
This is just another scripting language available. It only differs from something like Lua or Javascript in that it claims to have less complexity. Complexity which, while not fully elaborated on, apparently bogged down their efforts when they originally tried using Lua and an embeddable Go engine.
> Some popular DSLs most software developers use on a regular basis include Regular Expressions for pattern matching, AWK for text transformation or Standard Query Language for interacting with databases.
"Structured Query Language" is a so-called backronym. "Standard Query Language" fits the pattern as well. As does "Super Quick Lookup", if you want to have some fun with it.
https://en.wikipedia.org/wiki/Lingo_(programming_language)
#!/usr/bin/env ruby
require "csv"
puts CSV.foreach(ARGV[0], headers: true).sum { _1[1].to_f }.round(2)And where it should be used, I can't imagine you can't find a pre existing language (Cuelang maybe) instead.
I was expecting a section at the end where they demonstrate which services need a new language written just for it's configuration, but nope, just general examples.
Also, this should have a (2022) in the title.
Seems like it would be difficult to use for what they are trying to achieve. Lua would be a better fit, however it is noted that they tried it first, but ran into some kind of issue with it. So now Limbo is among the "off the shelf" options.
You have a specific DSL to generate test data and store code written using your DSL with your test cases.
You then have a security model that separates the responsibilities of the test team from that of other developers. The team can generate many test cases in a secure environment. You could then seek community input into your test processes with having to worry about rogue code.
But when you hit a problem that you need something like this for... you need something like this. The attempts to get around it or avoid it or do some unbelievably hacky thing leads to piles and piles of terrible, terrible code.
In 2024, though, I do try very hard to embed my DSLs in an existing serialization. It doesn't always work out, but, the case they show of directly embedding an AST into YAML is a worst-case scenario. In real life I've done things like specify a particular field carries an expr[1] expression to do that sort of thing, and then the structure of the rest of the file just follows normal serialization format.
[1]: https://github.com/expr-lang/expr , but I'm sure many static languages have something like this. If you don't know one, it's a good tool to put in the belt in case you ever need it.
At what point do you sit back and ask, is there a way to reduce complexity here? Is it not until you're joining a team where the last guy made dozens of DSL's for all his tools?
That was the main reason I was expecting more than a few real world use cases rather than generalized examples and no examples from their workplace.
However, a DSL could be made to do anything. Generate particular kinds of PDFs, Excel files, intentionally incorrect CSV/TSV files, zipbombs, and etc.
All that is required is to wire it up in the Go side of things and then write a script to make use of it.
Is there something about the current state of world that has invalidated some or all of the content in the article?
Heh that was my first instinct as to why they built this as well but more for providing ergonomic ways to generate k8s objects for complex gitlab specific CRDs.
Yes, you can generate your test data in the host language. That dramatically reduces upfront complexity.
But that doesn't come free. Now it is not easily transferrable to other languages. If you recall rsc's somewhat recent talk on testing, they benefited greatly by being able to use similar DSLs from other projects in tests to ensure feature parity. It can be a huge boon to have your test data scripts in an "agnostic" language for 30 years down the line when you are no longer working in the same environment.
Tradeoffs, as always.
> I was expecting more than a few real world use cases
Maybe that's the only use-case they know of? The project was clearly built for a specific purpose.
That's not to say that it cannot be found useful elsewhere, but that doesn't mean they have found it to be. If someone else finds it useful in other applications, I'm sure they'll write their own document about it. That's what the web is for, after all.