Why F#?
(batsov.com)431 points by bozhidar 2 days ago
431 points by bozhidar 2 days ago
For reference, F# slack is already dead, community has moved to Discord https://discord.com/invite/fsharp-196693847965696000
Curious since you don't expand on it on the blog: in what way did Haskell's purity make it difficult to you?
Having used Haskell in production for a bit now, I don't even notice its purity. Most functions are in some kind of I/O context making it similar as other languages, except with the option of running without I/O capabilities for functions that shouldn't need it.
For me, Haskell's image and ideal of purity are what made it difficult when I started out. I tried learning the language by reimplementing a program I'd previously done imperatively, that was (in hindsight) obviously hard to do in a plain pure way, ended up learning about zippers and knot-tying to do something in a less efficient and more confusing way than just using something like STArray because I had this idea from reading about Haskell that this was not only a good way to do things, but would be magically fast because GHC. (It was not.)
These days I'd just do such a task more-or-less imperatively in Haskell, and I would be well guided by the types in doing so. But I also feel like you have to make a few such mistakes if you want to get a good intuition and taste for when it's good do things purely and when imperatively.
I honestly don't remember as it was +/-6 years ago. I had started learning Haskell and got to that conclusion. Maybe that I am now more versed in FP I would arrive at another conclusion? I don't know.
Another thing that was hard to grasp for me were the special operators like =<<, ., $, etc. I was using Xmonad, but those operators create a barrier to understanding exactly what happened in the file.
In the end, F# was in my (personal) experience much more approachable, and it let me learn the functional concepts along the way.
You gotta remember people are often picking languages based on what they can easily find out about it and extrapolating/guessing about what problems they'll run into with their expected use.
A few years ago on here I had an interesting conversation with someone who wasn't going to use rescript for something because they didn't like how it handled object types. I can't remember ever using an object type in rescript; we all just convert js objects to record type in the extern binding. But that's not information easily available to someone who has never used the language.
Same thing here I think. If you don't already have familiarity with this paradigm, it's hard to imagine what using an IO monad for side effects is like. It's not easy to tell how hard it'll be to learn it, how much it may affect the rest of your code, etc. It's easy to imagine someone (shit even me a few years ago) going "eh I'll take the language with the big easy escape hatches just in case."
> You gotta remember people are often picking languages based on what they can easily find out about it and extrapolating/guessing about what problems they'll run into with their expected use.
This is a good observation.
As someone who writes a lot of Lisp, I'm inclined to agree as the amount of people that have never written any Lisp yet immediately reject it over syntax over fears that it somehow hampers development is a (to me) surprisingly large number of people.
If I recall correctly, one of the motivating factors for Rescript was to reduce the perceived/real distance between Reason and JS in order to attract more JS devs, as Reason was so heavily associated with OCaml.
Haskell taught me the important differences between IO and parity and forever influenced the way I program.
Even so the complexity here and the sheer number of mind bending concepts makes me not want to use it.
There was a large group of folks that left Ruby on Rails for Elixir (even has a similar looking syntax), yet it wasn't on your list of languages to consider. Just curious, was there a particular reason?
I should have mentioned in the message, but I was looking for a strongly typed language. I was an avid-user of dynamically-typed languages, but that particular Ruby on Rails app became unmaintainable, and part of the culprit was due to the dynamic typing. I hoped that using a statically typed language would make it easier to maintain a complex app in the long term. And I must say that it totally materialised, to the point that I don't want to develop in dynamically typed languages anymore.
Here's an example: as I said in my original message, I was a complete stranger to the dotnet ecosystem, and I learned the F# language at the same time. And I decided to develop the app as a library project to be used by the web app. I completely missed the prevalence of the async approach in the dotnet, and all my code was synchronous. One day, about half-way in the project, I realised I needed to switch to async code. Had this happened in a dynamically typed project, it would have been hell for me. Maybe it's me that can't grasp a project well enough, but I need the type-guardrails to find my way in large refactorings. And with the strong types, large refactorings can be done confidently. They don't replace tests, but make the refactoring process much more smooth.
The app is open source and its code is at: https://gitlab.com/myowndb/myowndb It doesn't have a lot of users, not the least due to lack of marketing and polishing the user experience. But I am satisfied of what I learned developing it!
This is a really minor point, but "strongly typed" and "statically typed" are not interchangeable terms. In the context of your comments here, you are exclusively interested in the static nature of the type system, rather than anything about the "strength" of it (which is something totally different and inconsistently defined).
These days there's Gleam[0], as a strongly typed alternative for the BEAM virtual machine. Of all the languages I haven't used yet, it seems to hit the safe + minimalistic + productive sweet spot the best. (Yes the C-inspired syntax is slightly off-putting, but syntax is the least important aspect of a language.)
[0]: https://gleam.run/
Hey, thank you for sharing your app's source code. I'll definitely check it out, I was really looking for such apps on F# open source projects!
Apache Spark, Delta Lake are written Scala. Being JVM based, it has a large ecosystem. Scala seems like a better choice than F#.
I'm sure it can be the better choice, but for me it was not. It seems there was some incompatibility between me and Scala. I find it such a complex language and I never managed to wrap my head around it. As I said F# was my last choice at the start of my evaluation, and Scala was high on the list due to the Java ecosystem. But in the end it didn't work out for me.
F# on the JVM would be great though!
I think Clojure is the better option if you want to do FP using the JVM ecosystem. The problem (for me, anyway) I've run into with Scala is that it supports both functional programming and object-oriented programming. Every code base I've worked on in Scala has ended up being a hodgepodge of both, which I find annoying.
However, the best functional programming language is, of course, Elixir. :D
> Every code base I've worked on in Scala has ended up being a hodgepodge of both
Is there something about that that has bothered you? Working in Scala codebases, I've found the best ones to work in are the ones that embrace Scala's multiparadigm nature. When programmers try to solve every problem with OO, they end up adding more and more layers to get the job done. When programmers try to solve every problem with FP, they end up resorting to sophisticated techniques that are unapproachable for other engineers. I think the simple parts of OO and the simple parts of FP go much, much further together than simple OO or simple FP can go by themselves. Have you seen something different?
Elixir getting a strong type system is interesting, but watch out for gleam though
But they still miss the computation expressions, which open interesting possibilities like https://github.com/CaptnCodr/Fli and https://github.com/fsprojects/FsHttp
Evaluated F# vs Clojure. Speed of certain algorithms just lacked for me. Value types particularly in tail recursive stacks shines in F# compared to the JVM in general. As usual YMMV
Isn't Clojure similarly (or even moreso) multiparadigm?
i don't think it is. i would say it is functional + bridges to the jvm (which is why it has been ported to many other platforms... there is not that much stuff in the language itself).
it is functional (value) programming first. there are tools to hook in the object jvm stuff but this is not the natural grain of the language.
clojure is pretty much all values and functions (and some macroes).
+ some concurrency stuff
there is no class, there is no inheritance, you don't even have information hiding (no private etc.). you have protocols and multimethods.
(well technically there is private because java but it is not obvious to use and not what you expect, you will very rarely see that in clojure codebases)
honestly it is a nice small yet powerful language, with not too many kludges. my personal coding is either clojure or rust (which has way more kludges, but better than the other stuff in the typed fast compiled world at least for me).
I did try F#, but I was new to .NET ecosystem. For 1 "hello world" I was quite surprised by how many project files and boilerplate was generated by .NET, which put me off.
I am all for FP, immutable, and modern languages. But then where are the jobs and which companies care if you write good code?
Now everyone wants languages which are easy to use with AI, while reducing workforce and "increased productivity". I have been programming for 20 years and know 4-5 languages, in India it was worse but in EU at-least I can make a sustainable living by writing Java / TypeScript. I cannot even find jobs with Kotlin + TypeScript which pay well, forget getting jobs in Elixir / Clojure / F# (there maybe a handful of opportunities if I will relocate for around 70K/year). That is why I have mostly given up on learning niche languages.
I understand your perspective. I like to view niche languages as a medium for learning. For instance, I enjoy using Rust in my personal projects—even if many of these projects may never be released—because the lessons on immutability, functional programming constructs, and trait-oriented programming significantly enhance my day-to-day work. Therefore, I believe that learning niche languages, even in the absence of a robust job market, is worthwhile.
I'm not sure I'd call Rust a "niche language" any more (perhaps in ~2018) - it's in common use across many big technology companies.
Just look at the job market. There are far more jobs for Go programmers and Go isn't particularly huge.
Compared with C/C++, Java, C#, Javascript, Python, Typescript, PHP, all the rest can be considered niche.
>" I was quite surprised by how many project files and boilerplate was generated by .NET, which put me off.
With which language are you comparing with?
Because there's afaik csproj and maybe .sln
and both of them are let's be frank - foundational for almost all projects that arent just hello world.
Otherwise you end up with some cmakes or something similar that want to achieve something similar
F# is quite usable with AI. All AI models are perfectly capable of generating idiomatic F# code. In fact, because it has a nice type system, if you ask the AI to model the problem well with types before implementing, hallucinated bugs are also easier caught.
I find F# easy to use with AI, mainly because it's statically typed (which results in compiler errors when the LLM generates non-working code) and it's very expressive, which allow me to more easily comprehend what the LLM is trying to do.
It is possible to start your project with the script possibilities offered by F# (as mentioned in the blog post). It is absolutely a viable approach and I even blogged about it a couple of months ago: https://www.asfaload.com/blog/fsharp-fsx-starting-point/
> But then where are the jobs and which companies care if you write good code?
Oh man, that is poignant :( They always say they do in the job description, but it always a different story once you get there.
Opportunities do exist, even when they’re few and far between. I learned Rust in my spare time because I was really interested in it. Then we stumbled across something that would have really benefitted from a cross platform library and lo and behold, I got to use my Rust knowledge, even though the vast majority of my day job doesn’t use it.
I like F#'s syntax when all you're doing is pure logic. But when you have to interface with any IO like a database or REST call or something, you have to abandon the elegance of ML syntax and use these ugly computation blocks. In C# you can do something like this:
var post = await _postService.getById(id);
let getPostById id = async {
let! post = blogPostService.getPostById id
return post
}
let post = getPostById 42 |> Async.RunSynchronously
1. Do all logic in ML-syntax, then pass data into a computation block and handle I/O operations as the last part of your function, then return unit OR
2. Return a C#-style Task<> and handle all I/O in C#
Either way, ML-style languages don't seem like they're designed for the kind of commercial CRUD-style applications that 90% of us find ourselves paid to do.
I find it personally better for CRUD applications than C# and I've written my share in both languages. Your syntax comparisons aren't exactly comparable in the sense that you haven't put in the wrapping/boilerplate around the C# code - you can't just await anywhere. You are also using an async which to run needs to know which context - this can be nice when you don't want to run the composed Task/Async on the current sync context. These days you stick to tasks if you want C# like behavior - and there's libraries to take away some SyncContext overload via custom F# CE's if you want.
The equivalent C# to your F# would be
task { return! _postService.getById(id) }
To make C# comparable to your F# code (tasks are not the same so not quite true) you would need to define a method around it, and find a way if you want to run the resulting Task synchronously to do that safely.
public async Task<Post> GetPostById(id) => await blogPostService.getPostById(id);
// This is not entirely eq - since tasks are hot
this.GetPostById(42).ResultF# is a big language, it is a ML multi paradigm language that interoperates with C# so there is a lot of necessary complexity and many ways to do the same thing. A strong benefit of this is the ability to create a working functional paradigm prototype that can iteratively be refined to a faster version of itself by hot spot optimizing the slower parts with equivalent highly mutable functions while staying within the same language. Similar how one would use python and C++ and over time replace the python code with C++ code where performance is important.
For the specific case of C# use of await it is unfortunate that C# didn't design this feature with F# interop in mind so it does require extra steps. F# did add the task builder to help with this so the 'await' is replaced with a 'let!' within a task builder block.
let getById(id:int) : Task<string> = failwith "never"
let doWork(post:string) : unit = failwith "never"
let doThing() = task {
let! post = getById(42);
doWork(post); }
let getPostById1(id:int) : Async<string> = async { return! getById(id) |> Async.AwaitTask }
let getPostById2(id:int) : Async<string> = getById(id) |> Async.AwaitTask
let getPostById3 : int -> Async<string> = getById >> Async.AwaitTaskIt is best to just use task CE full-time unless you need specific behavior of async CEs.
The author of the original comment, however, does not know this nor tried verifying whether F# actually works seamlessly with this nowadays (it does).
Writing asynchronous code in F# involves less syntax noise than in C#. None of that boilerplate is required, F# should not be written that way at all.
`var post = await _postService.getById(id);`
the F# equivalent is
`let! post = _postService.getById id`
In C#, you can't use the await keyword in a non async method, so I find the argument short sighted.
I don't see how that changes things. You'd have to async it all the way to the top but the syntax is still cleaner than F#. If you're using an Asp.Net controller you just declare the handler as async Task<IActionResult> and it's fine. Even program main methods can be async these days
As sibling comment pointed out, it's just .fsproj manifest and Program.fs file. What boilerplate do you speak of? It's on the opposite end boilerplate-wise to projects made in e.g. Java or TypeScript.
For F#, projects are needed to make full applications or libraries. Otherwise, you can simply write F# scripts with .fsx and execute them via 'dotnet fsi {SomeScript.fsx}'.
(obviously you can also specify dotnet fsi as shebang and integrate these scripts into general scripting on Unix systems - it's very productive)
I suspect they were either referring to pre-.NET Core days before the new project formats came out or they're creating projects in Visual Studio and checking all the optional boxes. There indeed did used to be a lot more required boilerplate to get some code running. Now you can run a .NET project quite nicely in VS Code with 2 total files.
Well also if you're using Visual Studio it will generate solution files as well, not just fsproj. I grew up doing C/C++ so boilerplate project/IDE/make files as well as build objects are something I expect to see. I think people who work in primarily JIT'd/interpreted languages are used to just having a directory tree full of source files and having some CLI tool manage everything for them. Maybe a dependency list file as well, but that's about it. Python is like this, and javascript CAN be like this
I'm completely convinced that F# (along with Scala, Haskell, and OCaml) adoption has stalled due to having ridiculously bad build systems. More significantly, they are being passed up in favor of Rust, which is a great language but nonetheless a bad fit for a lot of problem domains, simply because Rust has a superior build system. Hell, 80% of the reason I choose Rust over C++ for embedded work is because of the build system.
It baffles me that there are languages with non-profit foundations and are financially backed by multiple corporations which still have bad build systems. It is the most important investment you can make into a programming language.
I don't really get this take, at least wrt Haskell--I know less about the others. I don't disagree that it's not ideal, but I still would compare the combination of nix+cabal in Haskell favorably with anything I've used in Python, JS, Ruby, Clojure...and other ecosystems I'm forgetting. Python and JS in particular I've always found absolutely miserable to work with when it comes to dealing with dependencies. So I don't believe that this is why folks aren't choosing Haskell (I think it has a lot more to do with how different it is from most people's programming language experience, and how so much of the documentation is aimed at non-beginners).
It’s always puzzled me that so many languages have their own build systems and package managers. Why aren’t programming language-agnostic build systems like Bazel and Buck more popular? It seems so strange that every new programming language essentially has to reinvent the wheel multiple times, inventing a new build system, package manager, formatter, linter, etc. I wonder if we’ll ever see something like LLVM for these technologies.
Firstly a strong desire to self-host: write the build system in the language itself.
Secondly, often very differently shaped requirements. The dotnet SDK tries to keep its build specification (.csproj) files editable by Visual Studio, which is why most of the stuff in them is just XML properties.
You probably could build C#/F# with Bazel but that's not what Microsoft chose, and you kind of need to stay aligned with them and the large amount of MSBuild files in the SDK.
Because there are differences that go beyond just the language syntax. The build processes are very different for C++ vs. C#, vs. F#, etc.
C++: invoke compiler for all compilation units, invoke linker to combine object files into executable
C#: invoke compiler once with all source files listed
F#: same as C# AFAIK, except file order matters!
C++ I have worked only a little with it, more so with C, but AFAIK order of files does kind-of matter, so at least declarations need to be present in each file upfront before use.
C# has a multi-pass compiler so that it can compile and link the components from multiple files, without need of placeholder declarations, regardless of the order the symbols appear in the files.
F# has a single pass compiler, which keeps the compiler implementation simpler, but the file, and symbol definition order does matter that way. This is totally intentional, this is supposed to make the codebase more straightforward, with which I personally agree with. This avoids the need for declarations and centralization of them, the includes all the baggage that comes with that approach, and all the complexity C# has. I have rarely found a limiting factor, though there are some cases when it can be a bit inconvenient, for me the application setup/composition (~DI, but I prefer more static approach in F#) needed some cumbersome refactoring in some cases (have only vague memories by now, and yes, I know co-recursive types exists)
I really like F#, but rarely have to opportunity to work in it.
(writing all the below while being aware you likely know much more about OCaml than I do...!)
Possibly `eval $(opam env)` is something that should just go in your ~/.zshrc
The OCaml folks have done some work recently to improve the onboarding documentation, which I think is going in a positive direction
e.g. https://ocaml.org/docs/installing-ocaml (the eval as a one-off post install command)
And then guiding people to use 'switches' https://ocaml.org/docs/opam-switch-introduction, which I totally missed when I started with the language.
> Local switches are automatically selected based on the current working directory.
The only issues I've had with OCaml's build system is using "ocamlopt", "ocamlbuild", "ocamlfind" manually, but this was solved by OASIS and now Dune. I don't need to think about it. It automatically compiles when I save the file in Emacs. Very easy to set it up (one time setup).
My two cents is that F# hasn't received the same care and attention as C# and working with it can be awkward.
At the same time, a lot of the cool features (list comprehension, pattern matching, immutable records) have slowly trickled into c#, giving even less incentive to switch
I personally find the way these features were shoehorned into the C# syntax an eyesore, I have quite some C# experience, and I think the language is getting more and more convoluted and noisy, with ever less coherent syntax.
On the other hand many of these features are really convenient and handy in F#. Adding many of the oh-my-gamedev-such-speed features from C# to F# also makes its syntax less pleasant to work with.
Personally I also think that the C# async model is terrible, and was a grave mistake. The F# async model with computation expressions, and explicit control over their execution context was a better approach, and I'm really sorry the hack-something-together to unblock the event loop WPF/frontend-dev usecase won over the more disciplined backend-focused approach.
I have the opposite experience. Being able to write stuff like this is refreshing:
public record Name(string First, string? Last = null);
public record Register(Name[] Members);
...
var register = new Register([new("John", "Doe"), new("Neo")])
For me the pain is twofold:
a) it poisons all interfaces it touches (common trait of async in other languages as well)
b) C# async Task -s typically are created in Running state without any easy control over when, where and how they will execute. Controlling these things is far from trivial, and and requires lot of extra effort.
In F# the traditional async block is a builder for an async workflow, and you could then submit this workflow to an executor that is easy to configure for the execute model best suites you, eg. thread pool, single thread with continuations, maximum number of "operations" in flight, etc. The fact that it is not started right away also makes it easy to create your own executors.
Having to deal with backpressure in C# style async is way harder IMO. On the other hand when writing a UI app, always having to submit to an executor might seem inconvenient, and you generally don't have to handle thousands of concurrent requests all reaching out to a (different) backend and avoid DOS-ing it. This is why I wrote that this way made with a frontend-centric approach in my opinion.
I use C++ with Meson and Conan and I can say that there is an initial learning curve if you need to adapt Conan recipes but once you get past, the results are quite good.
Lets settle on a finally good enough. I can give honest compliments on learning from the past problems and from the better examples in the indsutry.
You may not remember the early .net core times with yeoman and other then-current javascript ecosystem originated things applied in really cumbersome, half-assed ways, with lacking docs and always being in flux for years. The project.json era was terrible.
Also msbuild was way worse 10-15 years ago...
Mono with automake was special circle of hell IMO, I have very small exposure but it was really unproductive and painful.
> I'm completely convinced that F# (along with Scala, Haskell, and OCaml) adoption has stalled due to having ridiculously bad build systems.
> Hell, 80% of the reason I choose Rust over C++ for embedded work is because of the build system.
Perl is mature, too.
Maven is awful. SBT is awful. Gradle is awful. I've used them all professionally, and the best I can say about them is that you can get the job done with them.
Newer languages and newer build systems are much better experiences, because of decades more hindsight and because language designers think about tooling from the start. Java was designed with the assumption that all software projects were built with Make, and with no ambition to improve on that. There was no Java-specific build tool until Ant was released as a standalone tool circa 2000.
> What is wrong with CMake for C++?
Granted, most of what's wrong with CMake is the problem it solves. Probably there's no solution that wouldn't be at least close to as awful as CMake. But it is objectively a hideous experience compared to any language created in the last 15 years.
How long are Scala compilation times for a Hello World CLI, a Hello World basic webapp and for a reasonably sized production size code base?
My guess is that it's much longer than for the equivalent Java apps.
> What is wrong with CMake for C++?
It doesn't manage dependencies.
Our shop converted 6 years ago, from C# to exclusively F#. I also author and maintain some packages (falco, donald, validus and others). The language is tough to learn if you're coming from a C-style language. But worth the effort and experience. It's extremely concise a true delight to build programs in that are fast, robust and durable.
There are a few drawbacks, depending on your perspective:
- compilation is slower than c# and hot reload isn't supported (it's in progress)
- there are very few opportunities to use it professionally
- hiring devs can be challenging
Hiring devs is perfectly fine if you don't look for F# skills - just hire generally smart people, and allow them 1-2 weeks to get comfortable with F#. Make them just solve problems from project euler or something.
For those who have already done functional programming, they wont take more than 2 days to start getting productive. For those who have written a lot of code, it will take them ~2 weeks to pick up functional thinking.
Anyone who is still uncomfortable with F# after 1 month - well that's a strong signal that the dev isn't a fast learner.
Additionally, I've never had anyone reject our job offer because we do F#. I'm sure a whole bunch of people might only be looking for python or javascript jobs, but that's fine because I'm not looking for them. I always have more people who I want to hire but I can't due to budget constraints.
Source: direct experience - I run a pure F# company with a team size of ~80.
> Anyone who is still uncomfortable with F# after 1 month - well that's a strong signal that the dev isn't a fast learner.
For myself, I feel that I have "full" command of C# as a programming language, but also how to structure projects, how to isolate modules, how to decouple code, how to set up a build system from scratch for C#, how do deploy and scale applications built with C#, what the strengths and weaknesses are, etc. My definition of "comfort" would entail a broader understanding of not just the syntax, but of the runtime ecosystem in which that code operates.
Hi Isaac ;) Of course you can train people. But in my experience they take a lot longer to learn than you suggest.
The problem is, many recruiters don't work with this mindset. If they're hiring a Java developer, and they get a CV from someone who has 1 year of Java experience and 5 years of C# experience, they see 1 year of experience, and immediately put it on the "unqualified" pile.
As someone who quite likes f#: It seems like a chicken and egg problem, not many companies doing f# because not many devs know it and not many devs learning it because not many companies are doing it.
I certainly wish I were doing f# professionally, but I only ever found 1 job listing for it, and that was in vienna while I am located like 200km away from it :(
Speaking of elm: I really like elmish for the frontend, when I need to make a dynamic page in the first place. Maybe that could be to your interest? (It transpiles to react under the hood via fable, which you can webpack into a drop in bundle. But I digress)
Jobs are kind of rare. I have to learn F# hoping maybe I will find a job in one or two years. And if I find it, they might want someone with F# work experience.
How does the typing system work for F#?
From the article, it looks like it's mostly dynamically typed. Or is it inferred? Or is it something else?
Like, if I write
let hello value =
print value
hello "world"
hello 2
To me, that'd be a point that might steer me away from the language. Deducible types seem vital to larger and long lived projects.
It's statically typed and inferred.
With regards to your example, the print/printfn (equivalent of Write/WriteLine) functions are a bit funny in F#. They don't actually take bound string values directly. You need to specify the type (which could be a string, a number, obj, etc)
https://learn.microsoft.com/en-us/dotnet/fsharp/language-ref...
F# is a statically typed language with gradual typing and full type inference.
Given
let hello value =
printfn "%A" value
hello "world"
hello 2
Thanks for your work on falco, it’s a really great library.
>- there are very few opportunities to use it professionally
- hiring devs can be challenging
Those drawbacks are huge, IMO.
I learned F# in 2013 and had a lot of fun with it, some of that code remains on Github (e.g. a 2D platformer game for windows https://github.com/lihaoyi/FSharpMetro/tree/master/Applicati...).
My experience was that it was a surprisingly nice language with a surprisingly warty user experience: papercuts ranging from naming conventions and function call styles (`|> List.map` vs `.Select`), basic syntax (`foo.[0]` to lookup arrays), type system features (F# doesn't have covariance/contravariance even though C# does), IDE support (back then was only Visual Studio, whose support for F# was inferior to C#).
Ended up settling on Scala after that, as a language with its own Warts, but one that somehow managed to feel a more cohesive than F# did despite having largely the same featureset and positioning.
F# was my first functional language and one that changed how I look at programming, but at the same time I'm happy to not actually have to use it for serious programming!
F# supports both functional and OO call styles. That's why you have both `|> List.map` and `.Select`. It can be a bit confusing at first, but the interoperability with C# is worth it.
Array lookup in modern F# is just `foo[0]`.
Subtyping is much less common in F# than in C#, so the need for covariance/contravariance is correspondingly lower. Personally, I've never needed it.
F# support in Visual Studio is now excellent. You can also develop F# in VS Code.
FWIW I think writing F# is a really cohesive experience in day-to-day work. While there are usually at least two ways to do things, due to .NET interoperability requirements, it’s usually pretty clear which way is the “right” way to do something.
F# feels kind of similar to Python in this regard, where there might be more than one way to do it, but there is community and ecosystem consensus on what is the right way.
I think a lot of credit should go to Don Syme for this; he seems to have a very clear vision of what F# should and should not be, and the combination of features ends up being very tasteful and well composed.
I think these days F# is probably a big more polished than what you remember, so perhaps it's worth giving it another shot.
Being a hosted language always requires certain compromises (something that was also apparent in Scala). I used to do Scala professionally in its early days, but for me it felt it added just as much complexity as it addressed. I focused on Clojure back then (on the FP side at least), and I do think that F# probably brings more to the table than Scala. (if one is not constrained to Java, that is)
The tooling story is not great, but I've almost never seen great tooling for a language that's not super popular. I'm guessing what you get today with Rider is more or less as good as what VS has to offer.
Interesting - I'm curious where your thoughts are now on using FP/Scala in 2025?
I've always looked at F# with envy as it is a hosted ML that will have extremely battle tested bindings to the important day to day stuff via C# (Darklang's stories of struggling with postgres and AWS when using OCaml was a good cautionary tale on the risks of using less common langs as a startup)
Never had a chance to try out Scala but am a seasoned Clojurian, as an outsider it seemed Scala suffered a little from being not opinionated enough so the ML family has been more appealing to tinker with even though Scala supports type classes out the box and will also have great ecosystem support via the JVM
It's never been a better time to try Scala if you're interested in the FP side. It's still very much "not opinionated" but the community and ecosystem have benefited from a certain convergence and given up on the "better Java" front which is served by Kotlin more adequately. Of course you can still consume any Java library when needed but it's best to avoid it if possible. Today you can pick between several ecosystems:
- Typelevel libraries: modern and more welcoming take on ScalaZ ideas, rich and mature, extensible libraries that are written in "tagless final" style.
- ZIO: concrete "super monad" with 3 type parameters, shuns the Haskell baggage and category theory lingo but pretty much the same concepts, compile-time autowiring dependency injection, a bit less mature.
- Kyo: new effects system on the block, pushing Scala 3's type system to the limit to stack effects using an "auto-flattening" monad (sorry if I butchered the description).
- Li Haoyi's own ecosystem that sticks to the standard library and JVM built-in mechanisms whenever possible, focused on Python style expressiveness, only more functional and with stronger types.
- I'd skip Akka/Pekko libraries but it's still an interesting piece of software if you need actor based, stateful cluster sharding.
Martin Odersky and the LAMP are focused on "capabilities" and we should eventually see something like direct-style algebraic effects, or like Kyo but without monads.
Also we have much better build systems than before (Scala-CLI, Mill, sbt has improved a lot too), binary backwards compatibility since Scala 3, and a very capable LSP backend if you don't like IntelliJ IDEA.
So, there is a gazillion ways to do things and the community didn't settle on one? I don't find that very attractive when trying to learn a new language.
As far as I can tell F# is one of those things where every single user is extremely happy. This happens rarely and I really am curious about the thing but never had time to get into it. I'm also pretty well versed in the .net ecosystem so it's probably gonna be easy.
Any tips? What kind of workflows might benefit the most if I were to incorporate it (to learn..)?
The funny thing is that you can write very similar code in C#, so maybe you don't need to switch which language you're using as a CLR frontend.
using System.Linq;
using System;
var names = new string[] {"Peter", "Julia", "Xi" };
names.Select(name => $"Hello, {name}").ToList().ForEach(greeting => Console.WriteLine($"{greeting}! Enjoy your C#"));
As far as fluency goes, that’s not very impressive.
%w{Peter Julia Xi}.map{"Hello, #{it}"}.each{puts "#{it}! Enjoy your Ruby"}
That is, each language will have its strong and weak points, but so far on "fluency" I’m not aware of anything that really beat Ruby far beyond as Ruby does compared to other mainstream programming languages.
I think LINQ is inspired by SQL. You can do whatever you can with SQL, it's just that the data source might differ IEnumerable with some in memory data, IQueryable with some DB. Or you can use async enumerable and your data source can be whatever web API or protocol.
You can write a vast majority of your C# codebase in a functional style if you prefer to.
All the good stuff has been pirated from F# land by now: First-class functions, pattern matching, expression-bodied members, async functional composition, records, immutable collections, optional types, etc.
I don't know if there's a name for it but essentially F# is where the language designers can push the boundaries and try extremely new things that 99% of users will not want or need, but eventually some of them are such good ideas that they feed back into C#.
Maybe that's just research, and I'm glad that Microsoft hasn't killed F# (I do work there, but I don't write F# at work.)
> F# is where the language designers can push the boundaries
It really isn't, not anymore. F# now evolves conservatively, just trying to remove warts and keep up with C# interop.
And even then some C# features were considered too complex/powerful to implement (e.g. variance, scoped refs) or implemented in weaker, incompatible ways when C#'s design is considered messy (e.g. F#'s non-nullable constraints disallow value-types, which breaks for some generic methods written in C#, sadly even part of the System libs).
> There’s no reason to do ToList there
In this case, I would move it to the very end if we are concerned about the underlying data shifting when the collection is actually enumerated.
Forgetting to materialize LINQ results can cause a lot of trouble, oftentimes in ways that happily evade detection while a debugger is attached.
Yes, ForEach isn't idiomatic but he could use Select instead.
Modern C# collection expressions make the definition of names closer to F#:
string[] names = ["Peter", "Julia", "Xi"];
var names = ["Peter", "Julia", "Xi"];
["Peter", "Julia", "Xi"].Select(name => $"Hello, {name}").ToList().ForEach(greeting => Console.WriteLine($"{greeting}! Enjoy your C#"));I did try that initially and got
<source>(5,1): error CS9176: There is no target type for the collection expression.
It seems reasonable to have it default to Array if it's ambiguous, maybe there's a downside I'm not aware of.
Maybe you can submit a proposal/issue to the C# language team? I'd vote for it!
I love LINQ, maybe a little too much. I can end up writing monster oneliners to manipulate data in just the right way. I love list comprehensions in python too, since they can work in similar ways
That could be shortened to
names.ForEach(name=>Console.WriteLine($"Hello, {name}! Enjoy your C#"));
F# shines on the back end, where its functional-first style is very adept at crunching data. Think about data flows in your system: Any place where you use LINQ in C# today to select/filter/transform data might be even better in F#. Parsing is also a great F# use case (e.g. parser combinators), although a fairly narrow niche.
Personally I think F# is excellent for writing ye olde CRUD applications, especially as the business logic becomes more complex. F# is really good at domain modeling, as creating types comes with minimal overhead. C# has improved a lot in this area (eg record types) but it’s still got a long way to go.
I wrote a tutorial about how to get up and running with web dev in F# that might be of interest: https://functionalsoftware.se/posts/building-a-rest-api-in-g...
Thank you, for someone interested in using F#, that is great.
I see you use Giraffe but I wonder how hard would it be to use Web API or to mix F# projects with C# projects in the same solution.
"As far as I can tell F# is one of those things where every single user is extremely happy" Isn't it because language has rather small community of passionate people, who are devoted to their language of choice?
F# popularity is somewhere between CHILL, Clipper and Raku langs, that are probably as obscure as F# for typical software dev.
I know Raku from Perl fame, and F# because it’s Microsoft, but CHILL and Clipper are totally new to me, so in my own humble experience these two latter look far more obscure. :D
If we consider number of jobs, it's probably on par with Rust.
Here's a Saas that actually makes money written fully in F#
Here's a rust ray tracer compiled to web assembly written in F#
https://ncave.github.io/fable-raytracer/
source: https://github.com/ncave/fable-raytracer?tab=readme-ov-file
The navbar takes up half my screen on an iPhone 13 mini :/
F# is great
Even if you never write a single line, it’s a fantastic illustrative language.
For example I refer to https://fsharpforfunandprofit.com/ all the time for functional programming ideas.
F# is beautiful, but I could never crack the nut and get fluent in it. I think the big problem is I only know a little C#, so it is difficult to figure out the object oriented methods that F# depends on. It was the same thing with Clojure and Scala for the JVM. I have zero interest in first learning C# or Java, just to use those platforms.
> It was the same thing with Clojure and Scala for the JVM. I have zero interest in first learning C# or Java, just to use those platforms.
FWIW, I managed to learn Clojure without knowing anything from Java/JVM before diving into it. I did have plenty of JS experience which made ClojureScript easier to pick up I suppose, but Java works more or less the same as every other Algol-like and/or OOP language out there so if you have any previous experience with something like that, basic Java is trivial to pick up even accidentally.
I'm aware, but you need to understand the .NET ecosystem to get anything practical done (at least when I was using it in 2017). All the books written on it (I own 3) are also the same way and assume you're a skilled C# dev.
100% this, I spent many months going through the most recent books on F# including one which the latest version was only released last year I think.
They all seem to try and shield you from the fact that you are much better placed if coming from C# (which everyone seems to refer to as .net these days) and have a solid understanding of the .net class library.
All the main web frameworks sit on top of asp.net and pretty much all official documentation for that is in c#
Such a shame because I learnt so much about types from trying to crack f# for real world application. fsharpforfunandprofit taught me heaps which I apply to other languages, but I don't want to become a c# developer which comes with all the years of changing best practices to be able to really be productive in f#.
Sorry if I am coming across as bitter but I just can't see learning f# in isolation from c# which is an absolute shame.
One way to look at it, but consuming OOP libraries doesn't turn code into OOP.
Also, FSharp.Core (which most F# code leans heavily on) is not OOP at all.
F# promotes object programming, doesn't proscribe mutability, encourages function and data approach.
It offers simple access to the different paradigms, with some opinionated choices (e.g. preventing leaning on OOP beyond an arbitrary stretch, like no "protected", only explicit interface implementation, etc.).
Why not? It does it very well, better than C# in my opinion. At least that was the case ten years ago when I last used C# and played with F#.
oop is just another tool in your toolbox; if f# provides it and it's the best way to express a given algorithm you should definitely go ahead and use it. oop got a bad name due to people trying to shoehorn it into places it was not the best way to express something, and it seems like you're making the inverse mistake here.
I've found that as C# gains much of the features that F# has and will soon gain more (pattern matching, functions as first class data types, great fp libraries, etc) the "moat" that F# has over C# has gotten smaller. I write most of my c# code in a primarily functional style, but I still have the advantage of using the libraries in their own native ways that follow the examples given by microsoft and other vendors.
Watching c# eat f# features as someone who has dabbled in f# lightly for over a decade has been wild. And supposedly DU's are in the works but multiple years out.
Though one thing I doubt c# ever gets that I love when I'm writing f# is pipeline operators. I love the way they read, from object/collection being worked on and then a list of operations being run on it in order from left to right (or you can do right to left if you need to for some particular reason).
The killer feature for me is type providers. I need to read a lot of CSV files of varying formats, and the CSV Type Provider lets me make quick work of them in a type-safe manner.
https://fsprojects.github.io/FSharp.Data/library/CsvProvider...
Agreed. Type providers bring metaprogramming to F# in a way that’s both powerful and innovative. The concept is amazing, enabling a new frontier for dynamically extending type safety. My only surprise is how little fanfare it has received since its introduction.
For a more pragmatic take on static metaprogramming, the manifold project[1] for Java is worth a look. Unlike F#, which leans toward expansive schemas, Manifold focuses on contained, compile-time integrations—handling JSON, XML, SQL, GraphQL, and even other languages in a seamless, type-safe way.
Why would Lispers feel at home with its (whitespace delimited) syntax? Quite the strange claim.
I know this isn't a common rant, but I hate so-called functional language still bowing to the "infix mathematical operator special case" dogma, when those are just binary (variadic in Lisp) functions.
Always found it pretty appealing, otherwise. And no ";;"!
I put together a quick-start guide to F# Computation Expressions — showing how you can go from C# async/await all the way to Result<> workflows with let!... and!... expressions, and even a custom validation {} CE. [0]
This is a practical side of F# that doesn’t get enough spotlight — but one I’m using daily.
I work in an industry where F# is a non-starter but I still love to use it outside of work. Also what I learn from F# and functional programming still benefits me in my day job (mostly c++). I think I tend to write a bit safer code after understanding the benefits of a functional language.
> Trivia: F# is the language that made the pipeline operator (|>) popular.
I’m not a dabbler in exotic languages, so this definition of “popular” was puzzling. I’ve literally never seen that operator before. Maybe I need to get out more.
I'm familiar with it via Julia, and I believe recent versions of R also have it.
It made it to Stage 2, which is some traction: https://github.com/tc39/proposal-pipeline-operator
It has been "stuck" at Stage 2 for a while, though.
I've used it in Racket, which is not exactly "popular" but common in universities as a teaching language. I actually built a fairly popular Tor hidden search engine in Racket nearly a decade ago, but I quickly shut it down when I ran some stats on the most common queries.
F# is a wonderful language, one that I write as my daily for the second time during my carrier. It _baffles_ me how it isn't more popular than it is because it truly is very very good. And I say this as an experienced and avid functional programmer.
We even do the frontend in it using Fable and Elmish, which is to say: we basically write our frontends in Elm, but the platform is .NET.
Because it’s great, and people that think otherwise are dead to me.
Curried functions combined with that magnificent pipe operator, overlaid on the .NET runtime. Don Syme et al knocked it out of the park.
It's the one programming language that changed how I think about programming.
I'm only talking about the version before type providers. Then it got messy.
Before that, we could (and I did) recompile fsi.exe to do some custom prompt manipulation. It was a slog, but it worked, but then Microsoft faded from my life. Still, that early version (I believe 2.0) F# is just magnificent.
F# is up to version 9 now, and has only improved over time, IMHO. Type providers are a very small part of the story and can be avoided entirely if you want.
I tried F# when it was first released and was not a fan, but it sounds like that impression is a little outdated. C# has come so far in that time it’s almost a new language. I’ll have to take another look.
Why would type providers be avoided? It seemed to me like a nice metaprogramming feature, akin to what Zig does with comptime types (except runtime?)
I don't doubt it, but I don't run Microsoft software any more. I've seen enough embrace, extend, and extinguish in my lifetime to not depend on them for my code's execution environment.
My current work needs nothing the .NET environment provides that I can't use python's standard libraries to get done, or bash and C if I need to.
But I'm lucky to no longer be in a corporate environment, so I don't need to consume commercial services, which was much easier using WCF within .NET. Back in my previous life, constructing n-tiered services on top of SqlServer using WCF was slick, indeed.
To any who are interested in how to construct such n-tiered applications simply but securely and precisely, I highly suggest Juval Lowy's IDesign system. He had three specific videos that I watched three or four times each until I understood his distillation of his vast expertise. Of course, Mr. Lowy is one of the co-designers of WCF, which was an excellent bit of tech.
It has great ideas but because of all these conveniences it is very bad for performance based programming making it slower than C#. I like the ideas in Roc language to make functional programming as fast as imperative by controlling allocations in things like closures
It might be great, but IME, MSFT docs and tooling are subpar, and anything dotnet related is/was a disaster.
As a 20+ year C# developer I’ve tried several times to learn/use F#. Despite being interested in FP, my brain is having trouble figuring out how to structure my code in F#. In C# I’d either build a service (or use the mediator pattern) for the domain and a repository for data access. With F# it’s functions all the way down and it feels unnatural (as silly as that may sound).
along with what jcmontx said: F# is structured bottom from top. As in you can't reference something that is later defined earlier. I find that naturally leads to getting a decent enough structure "for free" because it forces you to have your basic functionality early on and build on that later.
That also, IMO, makes untangling/splitting up parts of the codebase easier as well.
By the way, I consider https://www.bartoszsypytkowski.com/dealing-with-complex-depe... to be the canonically correct way to do DI when you want to inject more than like two dependencies.
Have you tried using a book or a tutorial to see how things can be done using idiomatic F#?
That makes sense when one is used to the Visual Studio organization of solutions and projects, with some main method somewhere being the entry point, unless it's a WCF service or somesuch that gets run via a service manager.
I only used F# at its command line, fsi.exe, to give me commandline access to .NET for exploration, testing, and munging data. Over time, I built up quite a library of usable functions that I'd have the fsi.exe program pre-load when I kicked it off, leaving me at the prompt with all .NET namespaces and my code ready and accessible.
Once you get access to your database's data, it's easy to write queries against it and then play with the data. I could then port the F# processing bits that worked into my C# projects as necessary, but it was far easier to do it that way than to write the logic deep within complex multi-project solution files, where the various classes are spread throughout the projects' files.
I also just really enjoyed using F#.
F# is getting some traction in Norway. I know for a fact that places such as NRK (BBC eqvivalent), Resoptima, Frende Forsikring and my current employer, REN are all using F#.
I just gave a talk about how we use F# at REN: https://vimeo.com/1070647821
Two nice things about F# are that you can introduce it into an organisation using the dotnet ecosystem and that you can use all libraries in dotnet, which is a huge advantage over OCaml.
Otherwise, I am happy with OCaml, but F# has also a place in this world.
F# was my favorite language, but - You have to chose the language of the Domain that you are working in, e.g, Swift for native iOS Development. Supabase Backend requires TypeScript, etc. - LLMs don't care about F#.
I've been using F# professionally for the past seven years across different contexts. First in a small software shop and now while bootstrapping a SaaS company. Some observations:
* It’s easier to attract smart developers to an F# project than to a [mainstream language] project. This was one of my driving beliefs when I introduced F# seven years ago. https://www.paulgraham.com/pypar.html. This is probably just as true for languages like Elixir, Clojure, ... But F# is what we went with.
Small Software Shop Context
* We operated in a small market where customers eventually dictated our tech stack (.NET & React). In that market, F# was a major advantage—it allowed junior developers to build apps that "just worked" with minimal regressions. Even with mediocre code quality, I felt confident that we could refactor safely at any time.
* I constantly had to justify F# to clients, which was exhausting. We always delivered decent results, so it worked out, but my partners were never as confident in defending F#.
Bootstrapping a SaaS Company
* F# has been invaluable for shipping features quickly and taking shortcuts when needed.
* Three years in, our codebase is large and contains its fair share of messy parts. But we can still develop new features at high speed with minimal regressions. Refactoring is relatively safe and straightforward.
* Compilation speed is the Achilles’ heel. If you don’t monitor it, the compiler slows down to the point where it impacts productivity. Earlier this year, waiting over a minute for feedback after a small change became unbearable. A lot of our "clean-up" work focuses on optimizing compilation times. We're still learning, but we’re optimistic that we can restructure the project to significantly improve build performance.
EDIT: maybe one more point. I see a lot of C# vs F# popping up here. Yes, C# has all the features that F# has. But do not underestimate how well designed F# is. It is an extremely simple language to learn compared to C#. There is a very limited amount of keywords to learn. And they compose extremely well. If you learned F# 7 years ago, took a break, and came back today, you'd simply write the same boring code that you would have written 7 years ago. And along the way you'd find out that some things have gotten a bit nicer over time.
I’ll add that features isn’t really a constructive way to think about the differences between C# and F#. C# has more ‘features’ than most other programming languages. One of the core features of F# is less features and churn. Also, I really appreciate the way F# maintainers agonize over how to maintain the language and keep it coherent. You don’t get the feel they’re chasing features at all.
As a 20+ year C# dev...where do I learn how to structure apps in F#? In C# my ASP.NET Controller might use a service (or a mediator) to execute some domain logic and that in turn will use a repository pattern (or EF DbContext) to update/query a database. How are dependencies injected in? It seems like there are multiple ways of going about it, but I don't have enough knowledge of F# to know 'the proper way' to do it.
The situation is a bit more complex in F# than C#, as there are multiple ways to do it. Scott Wlaschin has a good overview post about it here:
https://fsharpforfunandprofit.com/posts/dependencies/
FWIW you can do it exactly the same way you do it in C#; it’s not “wrong”, it might just feel a bit out of place.
I wrote an article about it. https://medium.com/@lanayx/dependency-injection-in-f-the-mis...
Just start. Use whatever style you are used to. Use controllers. Adapt your style as F# pulls you deeper inside the pit of success. You'll struggle the first couple of features, but you'll reach a sweet spot between your current style and functions relatively fast.
The problem with F#, Clojure and Elixir (hosted languages)
For F# , you need some basic C# knowledge For Clojure, you need some basic Java knowledge For Elixir, you need some basic Erlang knowledge
I like all 3 languages but usually each vm have a primary language, and each hosted language eventually become hosted on that primary language not the vm
I understand that for many task simple, to medium complexity, you might not need that, but it seem as you try to be more advanced you hit the wall of having to learn you host vm primary language
> For Elixir, you need some basic Erlang knowledge
As an Elixir programmer, this does not resonate. Basically the only thing I've ever felt I needed to understand Erlang for was ets, but let's be honest, that's not really proper Erlang but just the terrible ets query syntax. And all this requires is "ability to read enough erlang term syntax to be able to understand the ets manual". I don't think I could write a single line of correct Erlang by heart.
I feel like that's different in F#, where you still need to know lots of .NET internals which are all documented in C#y terms with C# examples etc. Elixir wraps pretty much all good Erlang/OTP internals in nice Elixiry modules, which solves that quite nicely.
Elixir has its warts but this really isn't one of them.
> For F# , you need some basic C# knowledge For Clojure, you need some basic Java knowledge For Elixir, you need some basic Erlang
Honestly, none of this really rings a bell. Having used all three options, I never felt that. Well, I already knew C# before getting into F#, but honestly, it felt like my C# knowledge at the time was more of a distraction. Been using Clojure for nine years, never done any serious Java and have not felt any need for it. Not knowing Erlang wasn't a problem with Elixir, like at all.
That’s a reassuring thing to hear as a new clojure learner who has little interest in java.
What bits of java have you ended up needing? Like do you often use java libraries that don’t have clojure wrappers?
I feel like I’m often running up against little things. I’ll google “how to do xyz in clojure” and the top SO answer is to use a java library that apparently everybody already knows about, cause so many clojurists came from java first!
> What bits of java have you ended up needing?
The same with Clojurescript and JS (and probably with Clojure-Dart) - you have nice interop with the hosting platform. The need for learning anything about Java (while writing Clojure) basically boils down to finding API documentation for a specific class and simply using it. That's all. That's all you'd ever need.
I worked with Elixir for over five years without knowing anything about Erlang. I know Erlang now, but only because I was interested in learning it, not because I needed to do so to write Elixir code.
I'm waiting for the excellent F# Article about railroad-coding to be posted here. Required F#e reading :) No not trolling. Am on mobile and traveling else I would have searched-posted
I assume you're talking about Scott Wlaschin's "Railway Oriented Programming": https://fsharpforfunandprofit.com/rop/
I worked a lot in F# and loved it. I love that it has a lot of great functional ideas without being too pedantic about being 100% functional all the time. (You can have mutating state or just call arbitrary C#.) I took a lot of its insights into my daily python code too. I especially love match.
Last time i tried F# i got bit by the weird concurrency story. There was async/task and somwhow they did not play well together. Also the dev tooling (for vim) was subpar, compared to ocaml (lsp). Compile times also was on the slower side.
`async` is F#'s original implementation of async programming. It is the precursor to C#'s await/async.
`task` targets the .NET TPL instead, which is also what C#'s await/async and all of .NET *Async methods use.
While the `async` implementation still offers some benefits over `task` (cold vs. hot starts [0]), my advice is - if you're doing backend code on .NET, you should use task. The tigher integration with the .NET ecosystem & runtime results in better exception stack traces, easier debugging and faster performance.
[0] https://github.com/TheAngryByrd/IcedTasks?tab=readme-ov-file...
If you use task { } CE's you will get really good UX (it is recommended to use them over async CE's). They were introduced in F# 6.0 to address outstanding interoperability issues. As of now, writing asynchronous code is more convenient than C# (which is more convenient than Go or other languages which are less expressive w.r.t. writing highly concurrent code).
If you've had C# and F# co-exist in the same codebase, how do they co-exist? Is it like C# and VB.Net where a project (dll) is either C# or VB.Net, and they can reference each other?
Or: Is it more like the Swift / Objective C ecosystem where Swift, Objective C, and even straight C can co-exist in the same library?
In a mixed C# and F# codebase, generally when do you favor C# versus F#?
Coming from a C# background, what are the areas where F# is a better language?
Any success stories for F#, especially if it co-exists with C#? Any horror stories?
Yes mixing is just like with VB.NET.
When mixing, you often write business logic as self-contained F# libraries with a C#-friendly API; and use C# to integrate them with whatever imperative, reflection-heavy and DI-heavy frameworks .NET is promoting the current year, since those are filled with interop edge cases for F# anyway.
You really want to avoid a language sandwich though (e.g. C#/F#/C#), because you'll keep wishing to remove the middle layer. Sadly, the addition of source generators make this mistake even more appealing.
You have likely heard "functional core, imperative shell". This refers to having IO-heavy code that favors imperative patterns be written in C# and then have the actual domain logic core written in F# which is much better at expressing it. Because both languages are hosted on .NET, you simply achieve it by having two projects and having one reference another. It is very seamless F# and C# types are visible to each other if marked to be so.
The biggest advantage of F# is its gradual typing and full type inference which allows to massively reduce the amount of text required to describe application or domain logic. It is also extremely composable and I find doing async in F# somewhat nicer than in C# too. F# also has better nullability (or, rather, lack of thereof) assurances and, in my opinion, better UX for records.
That's just like how C# and VB.Net can co-exist in the same project. Would you pick the pattern of:
1: C# Library with interfaces and/or abstract base classes
2: F# library with implementations of those interfaces and base classes
3: C# program (console, web service, GUI, ect) that specifies the implementations in Dependency Injection
Or is there a simpler way for C# and F# to co-exist in the same project (dll or exe)?
You don't really need to split 1 & 2, since F# can define .NET interfaces and abstract classes just fine.
For that matter, you don't even need the interfaces if you wouldn't have had them in a C#-only solution. Just define the class in F# and use it directly from C#.
You still need a separate assembly for F#, but that doesn't imply dependency injection - again, just reference it and use it.
In the case of F#, the use cases are diminishing with every new C# release, since C# is getting better and better at the things F# is supposed to be strong at (record types, pattern-matching, etc.). Better to write the thing in C# using modern features of the more popular and capable language.
Not sure that C# is the more capable language - its still that F# is mostly a superset of C# although C# is catching up. I think they are on par w.r.t capability (i.e. CLR compatible). In terms of conciseness, and expression F# still wins IMO; and given some of the extra features can be more performant at times (e.g code templating/inlining vs just JIT attributes). Custom CE's (Async/TaskSeq), nested recursive seq's great for algorithm dev, DU's, etc. There's a lot of little features C# doesn't have (or don't quite fit) that when I write C# I'm forced to implement still more cruft code around.
IMO its generally more readable as well to non-dev than C# and to dev's outside the Java/C# ecosystem (e.g. Node, Go, etc). I've shown F# code back in the day to non-tech stakeholders and they typically understand it (e.g. data modelling).
Unions remain the killer F# feature missing from C#.
Also, basic object initialization in C# has turned into a nightmare with recent versions. You need a flowchart to select among the 18 syntax options which suite your current needs.
With F# (and other newer languages), record fields are either `T` or `T option`. No need to worry about whether the value needs to be computed in a constructor and then remain immutable, whether it needs to be initialized by an object initializer and/or a constructor or not, whether it needs to remain interior-ly mutable throughout the life of the record, and so on. (Although as I recall you do still need to consider null values assigned to non-nullable references in your F# code that consumes C#.)
There's a draft spec out there for discrete unions in C# fwiw. I wouldn't be surprised to see it in another version or two.
And while I agree that I don't love some of the object initialization patterns C# allows--I respect that other people might have different style than me and don't mind ignoring that those styles exist when writing my own stuff :)
My general rule goes something like:
1. Use record types for any simple data structure
2. Avoid using primary constructors (even on record types).
3. Use { get; init; } properties for everything unless there's a good reason not to.
4. For things that need to carry internal state, have different methods for mutations, emit events, etc., use a class with regular old constructors and either { get; } (for immutable) or { get; private set; } (mutable) properties as needed.
You may start to get a point when C# gets a two-directional type inference system. As it's now, any functional-looking code requires so much boiler plate that it's shorter and less bug-prone to copy your functions code everywhere you want to use them.
Just try to make any generic high order function in C#. Any one you can think of.
Using OneOf library or something similar instead of discriminated unions / sum types.
Trying to use a functional pipeline instead of DI.
F# can be nicer to use for a functional programming style.
It's not always about the features such as keywords, built-in functionality and types. It's also how language features work together.
C# is more fit for an imperative or OOP style when F# is more fit for a functional style.
F# was for me the best functional language when I looked at rewriting a Ruby on Rails app. I wanted to go with a functional language, as it seems to better fit my thinking and reasoning, and I looked at Haskell, Ocaml, Scala, F#.
Being a stranger to Microsoft technologies, F# was the least likely to be chosen, but easily became the first choice. Haskell's purity made it hard to adopt (for me), Ocaml's ecosystem is subpar (there wasn't even a clear choice for a library to interact with postgresql, I couldn't install the latest version due to its reliance on an obscure tool whose name I forgot and didn't get help on the forum), and Scala is seems complex....
F# was surprisingly easy to get started with. The community is mis-managed by a corporate-minded approach (requiring people to become member of the F# software foundation to get access to the official slack!), but its members are friendly, smart and ready to help. The ecosystem is great with access to all the dotnet libraries (some complain there's a mismatch as most of those are developed for use with C#, but I rarely got in trouble for using them).
There are also great libs and frameworks available. Like https://github.com/SchlenkR/FsHttp to easily interact with http servers, to the point that I find it easier to use than a dedicated library. Or https://github.com/CaptnCodr/Fli , to run commands. And last but not least, https://www.websharper.com/ is the best web framework I have encountered across all ecosystems. Their reactive approach to web ui really allows me to develop complex interfaces in a maintainable way.
This became a longer message than I thought, probably due to my enthousiasm for the language. For complete transparency, the situation is not perfect, and in my experience the tooling is not the best.
If you want more info, I blogged about it a couple of months ago: https://www.asfaload.com/blog/consider-fsharp/