This is interesting, because I generally feel like getting the types right actually IS most of the work that needs to be done, and once the types are correct the implementation code just follows from the available operations necessary. I've never used Rust, so perhaps it's different in this language (I understand memory management is quite onerous there?), but this is how things feel for me in C# or TypeScript.
In my world, anything the type system doesn’t do for me, I must do myself in unit tests.
It’s actually possible to learn a type system and get good/fast at it. But typing out all those goddamn test cases never gets any easier, and updating them when the code changes is a nightmare.
My example for something that's simpler than Haskell but useful would be OCaml.
- Written in Haskell
- Not a tool for manipulating Haskell
- With a substantial user base
- Other than Pandoc, xmonad, or shellcheck
Partial credit if you can name closed source software instead.
For Go, off the top of my head, Docker/K8s would qualify, but if we want to eliminate that, Hugo, Caddy, Gogs, etcd, fzf…
https://github.com/search?q=stars%3A%3E1000+language%3AGo&ty...
https://github.com/search?l=&o=desc&q=stars%3A%3E1000+langua...
If anything, you should have asked about the prospects of OCaml, shouldn't you? (Spoiler alert: they are grimmer than Haskell's.)
Mostly, Go is full of wasted opportunities. It would have perhaps been a worthwhile language had they come up with Go instead of C in the 1970s. https://github.com/ksimka/go-is-not-good is a good summary.
I would suggest having a look at D as an example of what an evolution of C could look. An evolution with fewer wasted opportunities.
We also used OCaml at Bloomberg quite a bit. (And they are still using it, but I'm not there anymore.)
But there's not that much more in the commercial OCaml world.
Haskell has actually a bit of a broader adoption. The open source library situation is also better.
(For the record, I like both OCaml and Haskell. And Janestreet are doing a great job.
And commercial and open source adoption are important. But they are not the end to every discussion.)
I think I've heard the name darcs, but have no idea what it is or does. I may have heard of postgrest but I think the whole category of a thin layer over the DB is dumb, so I refuse to learn about it. :-) (If you want to talk to your DB, use the native SQL drivers. Don't just invent the same thing but as REST or GraphQL for no reason. Backend system should connect SQL to SQL. If you want a browser to be able to talk to your DB, then you will need auth and want bundling of calls and object translation and suddenly the thin translation layer isn't thin anymore.)
Anyway, the point is that Haskell has been around for a long time and is very popular with HN/Reddit users, but unlike Go and Rust, it has produced a very small amount of OSS.
I tried using darcs for a bit. But the designers were a bit too ambitious: they had an elegant concept to solve all rebases automatically at least in principle. Alas, the early implementations sometimes ran into corner cases that had exponential runtime. Which in practice meant that it hang forever for all a user could tell.
As far as I can recall, that behaviour is fixed now. But they missed their window of opportunity, when other dvcs became really popular. Like git.
Interestingly, git had the opposite philosophy: they explicitly only resolve simple conflicts in a simple way, and bubble up anything slightly more complicated to the user.
About SQL: Haskell users wouldn't want to use SQL directly. They want their compiler to yell at them when they get their database interactions wrong.
So they don't want some much to have a layer on top of SQL that hides things, but they'd rather want some support to forbid nonsensical SQL queries.
Interestingly, at one of my previous jobs we had 'relational object mappers' in our Haskell code. What that means is that we used relations as data structures inside our Haskell code, but had to interact with a mostly object oriented world on the outside.
Relations make fabulous data structures for expressing business logic. Take the leap in expressivity coming from eg dumb C to Python's ad-hoc dicts and tuples, and then imagine making a similar step again, and you arrive at relations.
Especially the various kinds of joins are good at expressing many business concerns. But projections and maps etc as well. Basically, everything that makes SQL useful, but embedded in a better language and just as a datastructure.
Codd's paper on relational algebra (https://www.seas.upenn.edu/~zives/03f/cis550/codd.pdf) is just as relevant to in-memory programming as it is to data bases.
> Anyway, the point is that Haskell has been around for a long time and is very popular with HN/Reddit users, but unlike Go and Rust, it has produced a very small amount of OSS.
The ML family of languages that Haskell is a part of has been around for even longer, and many of the beloved features of Haskell stem from that older legacy.
I don't have high hopes of the ML family becoming more mainstream. But I am really glad to see many advances born in the land of functional programming, and ML or Haskell in particular, making it out into the wider world.
The poster child is garbage collection. Which was invented for Lisp in the first place.
(You can do functional programming without garbage collection, but it requires much more finesse and understanding than they had when Lisp was young. And even then, garbage collection is probably still the right trade-off when you programme is not resource constrained.)
Garbage collection is pretty much the default for new languages these days. People expect a good rationale for any deviation.
More recently we have seen first class functions, closures and lambdas make it out into the wider world. Even Java and C++ have picked up some of those.
First class functions belong relate to a wider theme of 'first class' elements of your programming language. I remember that eg in Perl you had to jump through extra hoops to make an array of arrays. That's because arrays were not by default treated the same as any other value.
I think Python did a lot for the mainstream here: Python is pretty good at letting you assign many of its constructs (like classes or functions or ints or tuples etc) to variables and pass them around just like any other value. In the understanding of the Python community, they see that just as good OOP, of course.
Combinators like map and filter have become popular in mainstream languages.
Algebraic data types have made it to a few languages. With that comes structural pattern matching, and the compiler complaining when you miss a case.
Tuples are pretty much expected in any new language these days.
The very article we are commenting on talks about generics.
Immutable data types are something every new language is supposed to have thought about. Even C++ has const and Java has final. Go's designers were asked to justify their very limited support for immutability.
Many people go so far as suggesting immutability should be the default in new languages. (And you could very well imagine a dialect of C++ where 'const' is implied, and you need 'mutable' everywhere, if you want to override that.)
There's quite a few more examples. Of course, correlation is not causation, and so not everything that showed up in functional programming languages before it showed up in the mainstream means that the mainstream actually got it from FP.
---
In summary, you are right that Haskell has not been used in OSS or commercial products as much as eg Go, but I am just happy that the rest of the world is moving in the Right Direction.
Most of my favourite design choices are the ones that make OCaml a functional programming language. So mostly not apt for Go. But here's a list of what I think one could adapt while staying true to the idea of a C-like language:
C had unions. They had some obvious problems. Go's solution was to get rid of them. OCaml instead went for discriminated unions (https://www.drdobbs.com/cpp/discriminated-unions/240009296) and made the compiler check that you are handling all cases correctly. In OCaml parlance, they are called algebraic data types.
C had null pointers. Go inherited them. OCaml uses the general mechanism of discriminated unions to represent cases like 'end of a linked list' or 'no data' or 'unknown' instead. And the compiler will check for you that you either handled those cases (when declared) or that you don't create those cases (when not explicitly declared to be there).
Some of C's constructs are generic. Like accessing arrays or taking the address of a value or dereferencing a pointer. To give an example, p works for a pointer of any type and does the right thing. And the compiler will tell you off, if you mix up the types.
As a user of C you can not add your own generic operations. You can use copy-and-paste or cast to and fro void or sometimes use function pointers in a clever way.
Go goes the same route. For example the built-in datastructures are generic, like maps or arrays.
Similar to C, a user of Go can not add their own generic operations. Go's sorting https://golang.org/pkg/sort/#Interface demonstrates what I mean by clever use of function pointers. (Note, how even that sorting interface still has lots of copy-and-paste when implementing and how it only really works for in-place sorting algorithms. I'm not sure how much trouble it would give in a multi threaded environment.)
I suspect Go's aversion to user-creatable generics comes from C++. C++'s templates confuse ad-hoc polymorphism and parametric polymorphism.
If C++'s unholy mess were the only game in town, Go's stance of restricting generics to the grownups (ie language implementors) would be perfectly reasonable.
OCaml carefully separates these two kinds of polymorphism.
Parametric polymorphism basically means 'this function or data structure works identically with any type'. That's what we also call generics. Eg the length of an array doesn't depend on what's in the array. (See https://en.wikipedia.org/wiki/Parametric_polymorphism)
When compiling you only need to create code once regardless of type. (Modulo unboxed types.)
Examples of ad-hoc polymorphism (https://en.wikipedia.org/wiki/Ad_hoc_polymorphism) are things like function overloading to do different things for different data types.
An infamous example are the bitshift operators from C whose cute looks got them roped into IO duty in C++.
In OCaml parametric polymorphism is handled as the simple concept it is. Ad hoc polymorphism (and unboxing for parametric polymorphism) are more complicated, and rightly make up the more complicated bits of OCaml types system.
Go could have gone with parametric polymorphism and left out ad-hoc polymorphism to stay simple.
(Parametric polymorphism synergises well with algebraic data types. Eg you can model the possibility of a null-pointer via an 'option' type, and have generic functionality to handle it. Same for the slightly richer concept of a value that's 'either an error message or a proper value'.)
But discussions about generics are a bit of a mine-field in Go.
Type inference. Go's type inference only goes one way, and is very limited. OCaml type inference goes backwards and forwards. Rust's system is perhaps a good compromise: it has similar machinery to OCaml, but they intentionally restrict it to only consider variable usage information from inside the same function.
(Generics restricted to parametric-polymorphism-only don't make type inference any harder.)
Back to something uncontroversial: Go allows tuples and pattern matching against them only when returning from functions and has lots of specialised machinery like 'multiple return types'.
OCaml just allows you tuples anywhere any other value could occur. You can pattern match against tuples and arbitrary other types. The compiler ensures that your pattern matching considers all cases.
I am sure there are more things to learn. But those few examples shall suffice for now. I feel especially strong about tuples as a wasted opportunity that wouldn't impact how the language feels at all. (From a practical point of view discriminated unions would probably make a bigger impact.)
When I was much younger, I was told by my elders that to a first approximation a function doesn't need a name because it's defined by its range and domain. If it's hard to discern what the function does just by the range and domain then your type system isn't powerful enough. This insight is as powerful as it is superficially wrong.
1) Consider easing functions in animation (see easings.net for examples). They all have domain and range "real numbers between 0 and 1", yet they are all meaningfully different.
2) Consider filters in audio processing. They all have domain and range "streams of real numbers", yet they are all meaningfully different. The fact that they have the same type is important - it allows composition of filters, which I'd hope would sound exciting to a typed programmer!
3) Consider sorting algorithms. Even under a very advanced type system, the domain would be "arrays of numbers" and the range would be "sorted arrays of numbers". Yet there are tons of sorting algorithms that are meaningfully different. The fact that they have the same type is important - it allows a sorting library to evolve without breaking applications.
See the common thread? Types enable mixing and matching functionality. If a type allows only one functionality, it's useless for that purpose.
I was thinking a similar thought: sine vs cosine. But I think that falls under 'superficially wrong'.
Can confirm, this happened to me in Java constantly. I once spent an hour writing a big utility class using generics, even though I needed it for only one type in the code. "What if I need it for something else? Copying the code and replacing the types is very inefficient!"
Turned out that Java, being Java, was unable to instantiate an empty array of a generic type, so the very elegant code in the utility class now meant having to handle a bunch of special null cases in the application. I actually left it in for a while, but thankfully changed my mind before merging it as I'm sure the team wouldn't have appreciated having to deal with even more null in unexpected places because I got attached to my shiny generic code.
In general, I agree with your sentiment. Though in some cases, an abstraction can be justified even when used only once.
Take the simple case of a function. Generally a function only communication with its surroundings via parameters and returned values. A function that respects those conventions makes reading code much easier, and can be worth it even if used only once.
In contrast, a code block inside a much larger stretch of code doesn't give the reader any explicit guidance one how it intends to interact with the rest of the system. Even if it's exactly the same code as in the function.
Similarly, a for-loop can do almost anything. But using map or filter immediately tells you that only a few things are possible here. Even a fold is more restricted than a general for-loop.
'Theorems for free' makes a very similar argument for generics. https://people.mpi-sws.org/~dreyer/tor/papers/wadler.pdf
This is especially true with Rust where you also need to consider who owns what data. With C# it’s much too to model things that make sense as a type zoo but where the more important problems like ownership and access patterns is lost in the cracks.
The typical sign in Rust is when you need a lifetime on a struct so you realize it has someone else’s data. That usually means I stop to think before progressing. In C# it’s very easy to miss that big red flag.
Obviously it’s always worth revisiting and simplifying a design, but I still don’t think the initial design or designs were a waste of time if it gave me an understanding of the problem space.
I spent 2 days implementing a (terrible) discriminated union struct type in C# (for types I don’t control, so I can’t just add an interface) rather than just return `Object` from a couple of methods.
Type systems need to be more expressive if we’re going to be able to increase developer productivity without costing program safety.
The .NET CLI itself places limits on what .NET-based languages can actually do: for example, structural-typing at method boundaries in CIL is only possible by having the compiler generate build-time adapter classes and interfaces - which means that methods accepting structural-types can’t be easily called by non-strucutural-typing-aware languages. The CLI also imposes single-inheritance - so a .NET language cannot implement multiple inheritance at all - and because non-virtual calls cannot be intercepted at all (without hacks like abusing .NET Remoting’s special-cases in the CLR) it’s not possible to have a “every call is virtual” system like Java’s in the CLR (note to self: find out how J# did it). Implementing Go-style composition is hamstrung by the fact we can’t patch the vtable at runtime either (reified vtables would be nice...).
I dare say it - but I think 2020 marks the decline of C# and the CLR because we’re held-back by decisions made over 15 years ago. .NET 5 isn’t showing any sign of significant improvements to the CLR’s type system or fundamental concepts.
A type system without native or at least low-impedance support for discriminated unions is, well, not a very good type system.
> Type systems need to be more expressive if we’re going to be able to increase developer productivity without costing program safety.
Lots of type systems support discriminated unions directly, or at least make it trivial to implement them. C# perhaps doesn’t, but that's a problem with C#, not typed programming in general.
I write a lot of Rust nowadays, so I often need to keep myself in check and make sure I'm not getting sidetracked. When I'm writing internal code, the priority is to just get it done but still avoiding shortcuts. This is mainly avoiding .unwrap() / .clone() / taking care to handle Result correctly, as well as regularly running clippy to pick up the obviously silly stuff.
I find the most painful part of this strategy is when you want to revisit completed code to reduce copying as that generally requires a lot of type changes. At least when you get to this point though and haven't taken shortcuts, you've got functioning software and it should be fairly fault tolerant.