Pure IO monad and Try Haskell

tl;dr: Try Haskell now has simple I/O built on a pure IO monad.

I wrote Try Haskell in 2010. I didn’t really update the codebase since. It was rather poor, but had generally been stable enough. Last week I finally set some time aside to rewrite it from scratch (barring the tutorial part). In doing so I gave it a fresh design, rewrote the backend and stopped using my patched version of mueval (faster, but less stable).

Aside from reproducing the functionality, I afterwards added a new piece of functionality: typing in a function name (or composition of functions) will simply show the generalized type of that expression. This removes the need for an explicit :t and is more friendly to newbies who would rather see something useful than an error.1

After that, Bob Ippolito requested that some simple IO operations be made available, like getLine, putStrLn, etc. People have requested this in the past, but it seemed more complicated back then. This time, it seemed rather easy to support simple input/output. The console library already supports continued prompts, so the question became simply how to do safe IO.

As Haskellers worth their salt know, the IO monad is not special. Examples of a monad ADT for newbies are available23, and we know that pre-monad Haskell IO was modelled as a request/response system4 before it got updated in Haskell 1.35 Fuelled by the nerdiness factor of this, and that it was the weekend, I rose to the challenge. I knew it wouldn’t be hard, but it would be fun with a real use-case (Try Haskell’s REPL).

My constraints were that it shouldn’t be a continuation-based library, because I cannot have any state in Try Haskell. The server evaluates an expression and returns the result. No other context is kept, no process kept open, and it should return immediately. Given that it’s rather hard to serialize closures, but rather easy to serialize a list of inputs and outputs (aka responses/requests), I thought I’d go that route.

In the end I settled on an ErrorT monad over a State monad containing Input and an Output. The inputs would be stdin lines as [String]. The outputs would be stdout lines and either a final value, or an interrupt.

runIO :: Input -> IO a -> (Either Interrupt a, Output)

Whenever the expression being evaluated runs getLine, it reads from the Input state and pops that line of text off the stack. When getLine tries to read something and the stack is empty, it throws an error (of the ErrorT monad), returning the interrupt InterruptStdin. For example, here is a return value:

λ> runIO (Input mempty mempty) (return ())
(Right ()
,Output {outputStdout = [], outputFiles = fromList []})

Here is an interrupt:

λ> runIO (Input mempty mempty) (putStrLn "Name:" >> getLine)
(Left InterruptStdin
,Output {outputStdout = ["Name:\n"],outputFiles = fromList []})

As a user of the library it is now my part of the dance to get some input from the user and then re-call6 the same function with more stdin input:

λ> runIO (Input ["Chris"] mempty) (putStrLn "Name:" >> getLine)
(Right "Chris"
,Output {outputStdout = ["Name:\n"],outputFiles = fromList []})

I also implemented trivial exceptions:

λ> runIO (Input ["Chris"] mempty) (throw (UserError "Woo!"))
(Left (InterruptException (UserError "Woo!"))
,Output {outputStdout = [],outputFiles = fromList []})

After that, it was only a matter of time before I implemented some simple file IO operations:

λ> runIO (Input ["Chris"]
                (M.fromList [("/foo.txt","Hello, World!")]))
         (readFile "/foo.txt")
(Right "Hello, World!"
,Output {outputStdout = []
        ,outputFiles = fromList [("/foo.txt","Hello, World!")]})

That resulted in the library pure-io which some thought was a joke. It supports enough of the subset of IO operations for, I think, a newbie to at least feel like they’re doing some realistic I/O. So I added it to Try Haskell! You can now run interactive commands and write/save/list files. Any file system operations you do will be saved in your browser’s local storage.

It’s really a rather nice referentially transparent IO service. Even if you run forever (getLine >>= putStrLn) :: IO (), it will run forever, but the server can be restarted inbetween. No state is stored on the server at all, it’s all in the client. All the client has to do is pass it back and forth when it communicates with the server.

I’d recommend Haskell intermediates (perhaps not newbies) to implement their own IO monad as a free monad, or as an mtl transformer, partly for the geeky fun of it, and partly for the insights.


  1. Like “No instance for (Show (a0 -> a0)) arising from a use of …” which is frankly a useless message to print in a REPL and it’s strange that this is GHCi’s default behaviour.↩︎

  2. Johan Kiviniemi demonstrates an alternative monad implementation as an ADT that you interpret.↩︎

  3. Russell O’Connor also talks about implementing IO as a free monad here.↩︎

  4. Request/response example API on StackOverflow. All shuffling between an actual interface and the user is left to someone else to deal with.↩︎

  5. “Monadic I/O has already become the de-facto standard in the various Haskell systems. We have chosen a fairly conservative, but extensible basic design (an IO monad with error handling),” in the changes list.↩︎

  6. Yes, that means running the same computation every time from scratch, like a transaction.↩︎