The RIO Monad

I'd really intended to write a blog post on a different topic this week. But given that I did some significant refactoring in the Stack codebase related to a number of recent posts, let's just knock this one out and get to other topics another time.

I've played with the idea of a RIO (Reader + IO) monad a number of times in the past, but never bit the bullet to do it. As I've been hashing out ideas with some people and working through cleanups on Stack, it became clear that the time was right to try this out. And after having explored a bunch of other options, the arguments in favor of this approach are clearer in my head (which hopefully means clearer in this post too).

I'm going to describe the kinds of problems I wanted to address in Stack, look at various alternative solutions, and point out where each fell short. I'm going to talk about Stack because

1. I've implemented multiple solutions in that codebase
2. it's a fairly large application, and
3. it makes particularly good usage of multiple environment types, as I'll be explaining shortly.

If you're terribly impatient, you can take a peek at the code change. In reality, the diff to the codebase is significantly larger than this: I had to make a number of other refactorings happen to make this one possible, in particular removing StackM. But this captures the core idea.

What's happening in Stack?

When you first run Stack, it knows nothing about your project, your configuration, or anything else. It has to parse environment variables, command line options, and config files to get basic information like "should I be doing a verbose and colored logging output." Before any real work happens (which may involve logging), we need to load up our logging function and other basic settings.

Next, we need to load up some general configuration values, like "do we use the system-wide GHC." This does not include project-specific configuration values for two reasons:

1. If we don't need project-specific information, the parsing is unneeded overhead
2. We'll need the general config in order to load up the project config

Once we have the project information, we need to resolve it into real information about packages in the global database, which actual compiler version is available, and so on.

This creates a hierarchy of loaded configuration information, which we're going to call:

• Runner: basic info on logging, whether we're using Docker or Nix, etc
• Config: general config
• BuildConfig: project specific config
• EnvConfig: information on the actual build environment

Each of these configuration values contains its parent, so that a Config also knows about logging, and EnvConfig also knows project specific config. Awesome. (In the actual codebase, there are a few other levels to this, but this is a good enough explanation to move ahead with for now.)

Some parts of our code base are pure. We'll ignore those for now, and focus only on the parts that perform some kind of IO.

Pass individual values

Let's take the (arguably) simplest approach first. I described some data types above collecting many values together. But what if we completely ignore such helper types, and simply deal directly with their constituent values. For example, consider if we had:

data Runner = Runner
  { runnerLog :: Text -> IO () -- print a log message
  , runnerInDocker :: Bool
  }

data Config = Config
  { configRunner :: Runner
  , configSystemGHC :: Bool
  }

someFunc
  :: (Text -> IO ())
  -> Bool -- ^ system GHC?
  -> IO ()

Then I could use someFunc like so:

someFunc (runnerLog (configRunner config)) (configSystemGHC config)

There are two nice advantages of this approach:

1. It's crystal clear exactly which pieces of information a function needs in order to run. For example, we can tell that someFunc doesn't care about whether we're running in Docker, but does care about how to log and whether we're using the system GHC.
2. There are no complicated concepts at all introduced; someone with the most basic knowledge of Haskell could follow this function signature most likely.

However, there are two major problems as well:

1. It's really, really tedious. Having to pull out all of the values we need each time we call a function sucks. And if you suddenly need some extra information, you'll need to add that, and potentially add that value to all of the calling functions up the tree.
2. I'd argue it's not type safe. Passing around random Bools like that is a recipe for disaster. If you saw someFunc (runnerLogFunc runner) (runnerInDocker runner), you probably wouldn't notice the bug, and GHC wouldn't help either. You can mitigate that with copious usage of newtypes, but that's yet more tedium.

Pass composite values

The solution to those two problems is pretty simple: pass around the Runner, Config, et al values instead. And since Runner is contained inside Config, Config inside BuildConfig, and BuildConfig inside EnvConfig, we only ever need to pass in one of these values (plus whatever other arguments our function needs). This looks like:

someFunc :: Config -> IO ()

I also like the fact that, within the function, you'll now be using configSystemGHC, which is very explicit about what the value represents. Not bad.

But there are downsides here too:

1. We're giving more information than is strictly needed to these functions (in our example: the runnerInDocker value). I'm honestly not too terribly concerned about this. What is slightly more concerning is how easy it is to accidentally depend on a larger value than you need, like using BuildConfig in the function signature instead of Config. This makes it harder to test functions, harder to understand what they do, and harder to reuse them. But this can be addressed with discipline, code review, and (in theory at some future date) static analysis tools. (That last bit makes more sense the the typeclass approach mentioned later.)
2. Let's say we've got someFunc2 :: BuildConfig -> IO () that wants to use someFunc. We'll need to explicit extract the Config value from the BuildConfig to pass it in. This is tedious and boilerplate-y, but honestly not that bad.
3. With some functions—especially the logging functions—having to pass in the Runner value each time feels more tedious. Compare logInfo "Calling ghc-pkg" with logInfo (configRunner config) "Calling ghc-pkg". Again, not terrible, but certainly an overhead and line noise.

I want to be clear: this approach isn't bad, and has a lot of simplicity benefits. I know people who advocate for this approach in production software, and I wouldn't argue strongly against it. But I do think we can do better aesthetically and ergonomically. So let's keep going.

ReaderT IO

There's a pretty well-known approach to passing around an environment like Config: the ReaderT monad transformer. We can whip that out here easily enough:

someFunc :: ReaderT Config IO ()
someFunc2 :: ReaderT BuildConfig IO ()
logInfo :: Text -> ReaderT Runner IO ()

This solves the problem of explicitly passing around these environments. But we've still got to somehow convert our environments appropriately when calling functions. This may look like:

someFunc :: ReaderT Config IO ()
someFunc = do
  config <- ask
  liftIO $ runReaderT (logInfo "Inside someFunc") (configRunner config)

Some of this could get extracted to helper functions, but let's face it: this is just ugly.

MonadLogger

We're using the monad-logger library in Stack, which has a typeclass that looks roughly like this:

class MonadLogger m where
  logInfo :: Text -> m ()

We can change around our functions to look like this:

someFunc :: MonadLogger m => ReaderT Config m ()
someFunc = logInfo "Inside someFunc"

This works because monad-logger defines an instance for ReaderT like so:

instance MonadLogger m => MonadLogger (ReaderT env m) where
  logInfo = lift . logInfo

This reads much more nicely, but it's weird that we've got our logging function defined in Config and in the m type variable. Also: which concrete representation of m are we going to use at the end of the day? We could use LoggingT like so:

newtype LoggingT m a = LoggingT ((Text -> IO ()) -> m a) -- defined in monad-logger

runMyStack :: Config -> ReaderT Config (LoggingT IO) a -> IO a
runMyStack config (ReaderT f) = do
  let LoggingT g = f config
   in g (runnerLogFunc (configRunner config))

But this is starting to feel clunky. And imagine if we added a bunch of other functionality like logging: the many different layers of transformers would get much worse.

Custom ReaderT

These are solvable problems:

newtype MyReaderT env m a = MyReaderT (ReaderT env m a)

instance MonadReader env (MyReaderT env m) where
  ask = MyReaderT ask
instance MonadIO m => MonadLogger (MyReaderT Runner m) where
  logInfo msg = do
    runner <- ask
    liftIO $ runnerLogFunc runner msg

And then we'll need a bunch of other instances for MonadLogger MyReaderT depending on which concrete environment is used. Ehh.... OK, fair enough. We'll deal with that wrinkle shortly. Our functions now look like:

someFunc :: MonadIO m => MyReaderT Config m ()
someFunc = do
  logInfo "Inside someFunc"
  config <- ask
  liftIO $ someFunc3 config
someFunc3 :: Config -> IO () -- does something else, who cares

Not too terrible I guess.

Be more general!

It turns out we can generalize our signature even more. After all, why do we need to say anything about MyReaderT? We aren't using its implementation at all in someFunc. Here's the type that GHC will be able to derive for us:

someFunc :: (MonadReader Config m, MonadIO m, MonadLogger m) => m ()

Nifty, right? Now we can be blissfully unaware of our concrete implementation and state explicitly what functionality we need via typeclass constraints.

But how exactly are we going to call someFunc from someFunc2? If we were still using MyReaderT, this would look like:

someFunc2 :: MonadIO m => MyReaderT BuildConfig m ()
someFunc2 = do
  buildConfig <- ask
  let config = bcConfig buildConfig :: Config
  runMyReaderT someFunc config

But that's relying on knowing the concrete representation. We're trying to avoid that now.

And the other problem: that MonadLogger instance was annoying. We'd really rather say "anything MyReaderT that has a Runner in its environment is a MonadLogger." Can we do this?

Has type classes

Yes we can!

class HasRunner env where
  getRunner :: env -> Runner
instance HasRunner Runner where
  getRunner = id
instance HasRunner Config where
  getRunner = configRunner
class HasRunner env => HasConfig env where
  getConfig :: env -> Config
instance HasConfig Config where
  getConfig = id
-- And so on with BuildConfig and EnvConfig

A bit repetitive, but we only need to define these typeclasses and instances once. Let's see what MonadLogger looks like:

instance (MonadIO m, HasRunner env) => MonadLogger (MyReaderT env m) where
  logInfo msg = do
    runner <- asks getRunner
    liftIO $ runnerLogFunc runner msg

Much better, just one instance.

And finally, how does this affect our someFunc2 problem above?

someFunc
  :: (MonadReader env m, HasConfig env, MonadIO m, MonadLogger m)
  => m ()

someFunc2
  :: (MonadReader env m, HasBuildConfig env, MonadIO m, MonadLogger m)
  => m ()
someFunc2 = someFunc

This works because the HasBuildConfig constraint implies the HasConfig constraint. This is the general approach that the Stack codebase took until last week, so clearly it works. Our code is fully general, we don't need to do any explicit parameter passing, there's no need to explicitly convert from BuildConfig to Config or from Config to Runner. Besides some pretty verbose type signatures, this solves a lot of our problems.

Or does it?

Exception handling

I just realized that I need to do some exception handling in someFunc2. Cool, no big deal:

someFunc2 :: (MonadReader env m, HasBuildConfig env, MonadIO m, MonadLogger m) => m ()
someFunc2 =
  someFunc `catch`
  \e -> logInfo (T.pack (show (e :: IOException)))

Who sees the problem? Well, which catch function are we using? If we're using Control.Exception, then it's specialized to IO and this won't typecheck. If we're using Control.Monad.Catch (the exceptions package), then we need to add a MonadCatch constraint. And if we're using Control.Exception.Lifted (the lifted-base package), we need a MonadBaseControl IO m.

Alright, let's assume that we're using one of the latter two. Now our type signature looks like this:

someFunc2
  :: ( MonadReader env m, HasBuildConfig env, MonadIO m, MonadLogger m
     , MonadCatch m)
  => m ()

That's getting a bit long in the tooth, but OK. However, I've now got a bigger problem. As previously discussed, MonadCatch can play a bit fast-and-loose with monadic state (and MonadBaseControl even more so). Sure, in our concrete monad there isn't any monadic state. But this type signature doesn't tell us that, at all. So I'm left worried that someone may call this function with a StateT on top and we'll introduce a subtle bug into the code.

Sound far-fetched? In fact, it's worse than you realize. Stack does a lot of code that involves concurrency. It also does a number of things where it defers IO actions to be run the first time a value is demanded (check out the RunOnce module). In previous versions of the codebase, we had a number of places where we had to discard intermediate monadic state. Again, this shouldn't be a problem in practice, since our concrete monad was isomorphic to ReaderT. But the types don't prove it!

Introducing MonadUnliftIO

Last week's blog post gets us closer to the mark. MonadUnliftIO is like MonadCatch and MonadBaseControl, except it witnesses that there is no monadic state. Changing our imports around and then setting our type signature to the following restores sanity:

someFunc2
  :: ( MonadReader env m, HasBuildConfig env, MonadLogger m, MonadUnliftIO m
     )
  => m ()

With that change, Stack was able to get rid of its state-dropping code wrapping around the monad-control library, which is very nice. But it's not all perfect yet:

• Those type signatures are a mouthful
• Should we add in extra constraints, like MonadThrow or MonadBaseControl? It makes it more convenient to use existing library functions, but it makes the type signatures even longer
• Let's say I want to modify the env value a bit (such as temporarily turning off logging). I could use local, but (1) it's not even clear from the type signature that using local will affect the MonadLogger instance, and (2) local won't allow us to change the datatype itself (e.g., switch from BuildConfig to Config)

More concrete

This is peak generality. We're using typeclasses to constrain both the monad we're using, and the reader environment. Here's the insight I had last week*: why do both? There are really solid arguments for using the HasRunner-style typeclasses on the environment: it avoids explicit conversion and lets a function automatically work in a larger context (i.e., it provides us with subtyping, yay Object Oriented programming).

* To be fair, I've discussed this idea with others many times in the past, so it's neither a new insight nor my insight. But it helps move the story along if I say it that way :)

But can you make the same argument about allowing the monad to be general? I'm going to argue that, no, you can't. We know the application is always using the same concrete monad under the surface, just with different environments. You may argue that this limits consumers of our API: what if they wanted to use an RWST or ExceptT monad when calling these functions? Well, two responses:

1. We're not writing a generally-consumable library. We're writing an application. Our consumers are all within the application (or people crazy enough to use Stack's explicitly unstable API). And within our application, we know this is never happening.
2. We've already added a MonadUnliftIO constraint on our functions, and stated that we need that constraint for sanity purposes. Guess what: that limits our API to things which are isomorphic to ReaderT anyway. So we may as well just force usage of our own variant of ReaderT that handles the MonadLogger instance.

Here's one version of this idea:

someFunc :: (HasConfig env, MonadUnliftIO m) => MyReaderT env m ()
someFunc2 :: (HasBuildConfig env, MonadUnliftIO m) => MyReaderT env m ()

We get our MonadReader and MonadLogger instances for free from specifying MyReaderT. We still get subtyping by using the Has* typeclasses on the environment. And by specifying MonadUnliftIO on the base monad, we get MonadUnliftIO and MonadIO too.

Should we be that general?

We kept an m type variable in our signature. Do we need it? In reality, not at all. Concretely, that m is always going to turn out to be IO. Also, if we had some reason we needed to run in some other monad, we can just do this:

runMyReaderT :: env -> MyReaderT env m a -> m a

helper :: MonadIO m => env -> MyReaderT env IO a -> m a
helper env = liftIO . runMyReaderT env

Meaning we can always get back our more general signature. But interestingly, we're doing something even more surprising: the MonadUnliftIO constraint has been replaced with MonadIO. That's because, inside our MyReaderT stack, we're just relying on IO itself, and leveraging its lack of monadic state. Then we can liftIO that IO action into any transformer on top of IO, even one that doesn't provide MonadUnliftIO.

Do we need a transformer?

Alright, what value is the transformer providing? I'm going to argue none at all. Every usage of our MyReaderT has m specified as IO. So let's finally knock out the final simplification and introduce the RIO (Reader+IO) monad:

newtype RIO env a = RIO (env -> IO a)
runRIO :: MonadIO m => env -> RIO env a -> m a
runRIO env (RIO f) = liftIO (f env)

This has instances for Monad, MonadReader, MonadUnliftIO, and can even support MonadThrow, MonadCatch, or MonadBaseControl. When using those latter functions, though, we can look at our type signatures and realize that, since RIO by definition has no monadic state, they're perfectly safe to use.

Our type signatures look like:

someFunc :: HasConfig env => RIO env ()
someFunc2 :: HasBuildConfig env => RIO env ()

The only typeclass constraints we're left with are on the environment, which is exactly what we said (or at least I said) we wanted, in order to allow the useful subtyping going on in our application.

I'll argue that we've lost no generality: using runRIO makes it trivial to use all of our functions in a different transformer stack. And as opposed to using constraints like MonadLogger, it's trivial to fully unwrap our transformer, play with the contents of the environment, and do new actions, e.g.:

env <- ask
let runner = getRunner env
    modRunner = turnOffLogging runner
runRIO modRunner someActionThatShouldBeQuiet

This approach is currently on the master branch of Stack. I'm biased, but I think it greatly helps readable and approachability for the codebase.

I'm strongly considering spinning off this newtype RIO into its own package (or adding it to unliftio, where it's a good fit too). I'm also considering extracting the HasLogFunc typeclass to the monad-logger library.

The m*n instance problem

The MonadLogger typeclass suffers from what's known as the m*n instance problem. When I wrote MonadLogger, I had to define instances for all of the concrete transformers in the transformers package.

When I wrote the MonadResource typeclass in resourcet, I had to do the same thing. And I had to provide instances for these for each new transformer I wrote, like HandlerT in yesod-core and ConduitM in conduit. That's a lot of typeclass instances. It gets even worse when you realize that dependency confusion problems and orphan instances that usually result.

By contrast, if we define additional functionality via typeclasses on the environment, this explosion of instances doesn't occur. Every transformer needs to define an instance of MonadReader, and then we can replace:

class MonadLogger m where
  logInfo :: Text -> m ()

with

class HasLogFunc env where
  getLogFunc :: env -> Text -> IO ()

logInfo :: (MonadReader env m, HasLogFunc env, MonadIO m) => Text -> m ()
logInfo msg = do
  logFunc <- asks getLogFunc
  liftIO (logFunc msg)

There is a downside to this approach: it assumes that all transformers have IO at their base. I'd argue that for something like MonadLogger, this is a fair assumption. But if you wanted to make it more general, you can in fact do so with a little more type trickery.

The principle I'm beginning to form around this is: don't define effects with a typeclass on the monad, define it with a typeclass on the environment.

Why not ReaderT?

I alluded to it, but let me say it explicitly: we need a new type like RIO instead of ReaderT to make all of this work. That's because typeclasses like MonadLogger define their instances on ReaderT to defer to the underlying monad. We need a new monad (or transformer) which explicitly takes responsibility for defining its own instances instead of deferring to the underlying monad's behavior.

Some notes on the Has typeclasses

The Has* typeclasses above work best when they properly define superclasses. It would have been much more irritating to write this code if I'd had to say:

someFunc2 :: (HasRunner env, HasConfig env, HasBuildConfig env) => RIO env ()

The superclasses on HasBuildConfig and HasConfig allow me to just state HasBuildConfig, which is great.

Also, I demonstrated the typeclasses above with accessor functions:

getRunner :: env -> Runner

In reality, Stack uses lenses for this (from the microlens package):

runnerL :: Lens' env Runner

This makes it much easier to make modifications to the environment, such as the "silence log messages" example above. (Though I think that example is totally made up and doesn't actually occur in the codebase.)

How about pure code?

I started off by saying this discussion applies only to IO code and doesn't apply to pure code. But what about pure code? One really, really bad option is to just say "don't write pure code." We're Haskellers (at least, I assume only a Haskeller would be enough of a masochist to get this far in the blog post). We know the value of partitioning off effects.

Another option is to rely on mtl-style typeclasses, such as MonadReader and MonadThrow. E.g.:

cannotFail :: (MonadReader env m, HasConfig env) => m Foo
canFail :: (MonadReader env m, HasConfig env, MonadThrow m) => m Bar

This is basically what we do in Stack right now, and has the benefit of unifying with RIO without explicitly lifting. Another approach would be to make these more concrete, e.g.:

cannotFail :: HasConfig env => Reader env Foo
canFail :: HasConfig env => ReaderT env (Either MyException) Bar

Or even ditching the transformers:

cannotFail :: HasConfig env => env -> Foo
canFail :: HasConfig env => env -> Either MyException Bar

Or even ditching the typeclass constraints:

cannotFail :: Config -> Foo
canFail :: Config -> Either MyException Bar

I'm frankly undecided on the right way forward. I like the current approach in that it unifies without explicit conversion, while still preserving the ability to use the code purely, test it purely, and see from its type that it has no side effects. Some people (they can speak up for themselves if they want) disagree with the concept of MonadThrow, and don't think it should be used. Another advantage of ditching MonadThrow is that it can allow more explicit exception types (notice the MyException above).

Regardless, take this message: don't use RIO as an excuse to dispense with purity. Strive for purity, be disciplined about making pure code pure, and then consider the RIO approach when you have to be effectful.

Using RIO in your application

I think this is a pattern I'd already recommend. It deserves more real world experience, and if I add RIO to a library on Hackage it will be easier to get started. There will always be a little bit of upfront cost with this (defining your environment data type and relevant typeclasses/instances), but I think that cost is well worth it.

If you're going to start using RIO yourself, please add a comment about it. It would be great to get some standardized effectful typeclasses defined to grow out the ecosystem.