tl;dr: F# has a long learning curve (it takes a long time to master) but productivity happens quickly.
I recently wrote my first real application in F#. Although I've been using F# for many of my scripting needs for about a year and use F# to explore iOS APIs regularly, this application is the tool in which I'm going to live for the next several months. As such, I wanted to use a modicum of decent practices: separating UI from behavior and test-driven (or at least test-supported) development. On the other hand, every day I spent writing the app was a day I wasn't doing the task it was meant to aid, so I needed my programming time to be primarily about implementation, not Functional Purity or pushing the corners of the envelope of my understanding of F#.
The actual project is proprietary, so I can't be particularly specific nor share code, but it involves a lot of fiddly details surrounding the part that actually requires expertise. Without automation, the fiddly details end up not only taking a significant amount of time in themselves but they prevent me from achieving flow, which I think is an even greater problem for my overall productivity. (Note to self: Write blogpost titled "Work harder, not smarter," about the importance of achieving flow.)
One of the major aspects of the task involves writing XML. As an example of the type of trade-off that I found easy to make, I mutate the XElements rather than using a Zipper or something like that.
On the other hand, it's amazing how quickly and constantly mutability bites you in the ass. This is one aspect of functional programming that I do think is a legitimate surprise. Although I believe "prefer immutability to mutability" is just good programming and not a revelation brought down from the mountain by Haskell, when you write code in which most of your identifiers are assignment-only, it highlights how often your troubles rotate around those with a mutable modifier or values that depend on the order in which things are initialized.
On the other other hand, the stringly-typed nature of the world is not to be dismissed. My biggest bug, which lay there for 2 days being passed by my test suite, was an XPath mistake. I also have some regular expressions in the program; these happen to be pretty straightforward, but obviously regexes are another place where the advantages of strong typing are often undermined by convenience.
As Peter Norvig told me during a conversation about type systems and the Mars Climate Orbiter (the one lost due to using the wrong units of measure):
I don't know of any language, no matter how type-strict, that forces you to tag the string "123.45" in a file with the units of force (newtons vs foot-pounds), nor do I know of any language, no matter how type-loose, in which you could not impose such a convention if you wanted to.
I very much wanted to separate the domain logic from the interface. Since I'm going to be living in this tool for months, there is a definite advantage if I can run it on my iPhone and Nexus 7 as well as my desktop Mac (did I mention that I was writing this using mono?). Since one can write an F# iOS app in 39 lines of code and since cross-platform applications are easy to write with Xamarin, this is straightforward. On the other hand, I needed to make sure that I was correct that the, you know, functionality was worthwhile in the first place.
So I wanted my initial version to be a command-line app. It's been awhile since I've written an input loop and I don't know if I've ever done one with a Model-View-Controller architecture before!
Although I took the Coursera course on Reactive Programming, staying event-driven was another example of letting pragmatics override a learning opportunity (no actors, no reactive extensions... Erik Meijer would be so ashamed of me...). This is one of the few places where F# disappointed me a little. My events look like this:
type EventType = Foo | Bar | Bat
type ParamArgs< 'T>(value : 'T, change : EventType) =
inherit System.EventArgs
member this.Change = change
member this.Value = value
type ChangeDelegate< 'T> = delegate of obj * ParamArgs< 'T> -> unit
type MyType =
let fooEvent = new Event<changedelegate <string>, ParamArgs<string>>()
[<clievent>]
member this.FooEvent = fooEvent.Publish
And on the receiving side:
do myItem.FooEvent.AddHandler(
new ChangeDelegate<string>(
fun sender args -> //handle the event
)
)
Which, now that I type it out, doesn't look all that bad. But given how spoiled you get about not spending your time finger-typing in F#, it feels crufty.
As an example of how F# generally spoils you, I offer the "forward pipe operator":
let baz = foo |> bar |> bat
Which should bring a nod from anyone familiar with the UNIX shell. It's the equivalent of, say:
let baz = bat(bar(foo()))
//or
let fooResult = foo()
let barResult = bar(fooResult)
let baz = bat(barResult)
But it's superior to both those forms. It's just easier to read than the in-to-out, right-to-left nested function call (putting aside the parentheses, about which more later...). But it's superior to the second in a more fundamental way. Phil Karlton tells us
There are only two hard things in Computer Science: cache invalidation and naming things. [And off-by-one errors.]
The really nice thing about piping is the way it ties together the expression without intermediate names. The pipes indicate
Here,
foo,bar, andbatare logically connected: you can't replacefooandbarwithout considering the effect onbat, the calculation ofbatis not only tied up with the calculation ofbar, it's also tied up with the calculation offoo. BUT the result offooand the result ofbararen't important outside of the calculation ofbat! The result ofbatis not just used immediately: it has a name and therefore has some importance.
That's a lot of context to get from a language operator! It's not just a convenience or a way to avoid a keystroke or two. It's a good thing that you miss as soon as you work in another language.
Regrettably, I don't feel as positively about F#'s indentation rules. I'm not going to go into too much detail on this, because I think it's just a matter of personal preference (I take the increasing ubiquity of Python in the science community as evidence that I'm in the minority...). I can imagine that my irritation might go away with better tooling and some more-understandable compiler errors.
In F#, as with several other functional languages, enclosing function arguments with parentheses is optional: one can say Foo a as opposed to Foo(a). I'm pretty resistant to that: I don't find Foo a easier to read at all. But the pipe operators (there's also a backwards pipe) make the "point-free style" of programming more palatable. You don't say |> fun a -> Foo (a) |> fun b -> Bar (b), you just say Foo |> Bar and that's that. Admittedly, this occasionally leads to code that strikes me as somewhat cryptic: optionList |> List.collect Option.toList converts a list of Options into a list of the values of the Somes (i.e., flatMap id).
Another structural thing that slowed me down a bit is F#'s many kinds of types (I would say that F# has a plethora of kinds of types). I'm a believer in Functional in the Small, Object-Oriented in the Large, so I like to hang both data and responsibilities on a type. But with F#, I have to decide between classes (class), records (type), and discriminated unions (type |) (and 14 other kinds). The good news is that in my limited experience you can start with a record and then, if you need to refactor it into a class, it's trivial (I may be insanely wrong about that, having only needed a class rather than a record once).
This proliferation of somewhat-overlapping concepts is about ensuring compatibility the Common Language Runtime and 95% of the time, it's not something that you have to think about.
With my events in place and having decided on records, programming went very well. It's trivial to use NUnit with F# and I quickly had 3 projects up and iterating: the model project, the CLI view & controller project, and my test suite.
I also used Xamarin Studio's F# Interactive console regularly.
.
All of which was very pleasant, but I wonder if one downside of the “small functions doing one precise thing” of the functional approach is that it becomes easy to just keep composing function calls on the results of previous function calls rather than trying to create a consistent and sufficient data structure / object model. The drive towards refactoring a concise, task-driven API doesn’t seem as dominant since you’re simultaneously valuing the “easy to compose” tiny functions.
With small, easy-to-compose functions and a lot of flexibility about scoping you immediately start circling the issues of coupling and cohesion. General-purpose functions are, by definition, not very cohesive: they happen to be composed together to deliver value. Hmmm....
Even in a small program such as this, the lack of robust refactoring tools jumps out. When you’re learning a language, you just plough away at small programs, but as soon as you start writing even a few-module program, organizational and refactoring instantly becomes concerns.
Speaking of "small programs," this program (which is focused, but not entirely trivial) is under 1KLoC including its test suite. I use craploads of higher-order functions, but with LINQ, I could structure a C# program very similarly. I also use pattern matching in several places and that would take more lines of C# code. I use the Option type, so that eliminates lines of C# dedicated to null-checking.
But in general I think I could (and would) write a very similarly-structured C# program to solve the problem. I would have written it faster, but not that much faster (maybe 10-15% faster). And the C# program would have, I think, about 20% or maybe even 33% more LOC; not that many more expressions but some more and then quite a few lines dedicated to closing parentheses and blocks. (Regarding my previous snarky comments about parentheses and indenting, I quote Walt Whitman:
Do I contradict myself? Very well, then I contradict myself, I am large, I contain multitudes.
)
One thing that became clear is that I haven't come even close to mastering F#, despite having used it semi-regularly for a year. It's learning curve is longer and higher than most languages. For instance, it now seems to me that I could have used a custom computation expression (dangerously close to the M-word) in a way that would have allowed me a lot of flexibility in evolving the app. Even more, I think that type providers might be an enormously important facility, but it's an area where I haven't even explored.
But as I said, I don't think my productivity suffered all that much relative to C#, a language I've worked with since it's very first beta. In other words, while the road to F# mastery is long, productivity comes much faster. I paid some penalty relative to C# on this project, I'll pay less penalty on the next, less on the next after that... And I can see in a few smallish projects such as this, getting to the point where my productivity is in my standard range.
Scala was my day-to-day language for 2 years and I'd choose Scala over Java without hesitation (and no, Java 8 doesn't change my mind). Mastering Scala involves understanding a very sophisticated and complex type system and there are certainly advantages to that, but (for me) the mental effort to know that type system has limited practical benefits relative to a less-sophisticated system. But I see in some of these F# facilities (such as type providers) the potential for techniques that are not just O(n) relative to my C# approaches. The F# learning curve is high, but I think that the long road to mastery might have more of a transformative change in my approach to programming.
Nonetheless, programming is programming and programming is hard. As soon as I thought I was ready to go, I discovered that some important data came in a variety of shapes. So I had to spend time discovering those patterns and working with them. That type of work is much easier in a language with a good set of higher-order list processing functions and built-in pattern matching. And, I think, the smaller functions characteristic of FP made it easier to isolate the change, but nothing defeats this essential characteristic of programming, which is that it’s far more about exceptions and business rules and capricious “that’s just the way it is” aspects than it is about mathematical elegance.
As for the application, in the first three days of this week, I averaged 5x my "doing it manually" productivity. So take that, Randall.