Porting to Polysemy

Many years ago, when I first started using free monads in anger, I was tasked with porting our giant codebase to something that used an effect system. While it was a noble goal, my efforts slowly imploded upon their own weight. I didn’t know how to go about doing such a dramatic refactoring on a live codebase, and unwisely tried to do the whole thing in a single PR. A month later, as you might expect, it became overwhelming obvious that we were never going to merge the thing, and it died there.

Several years older (and wiser), I’ve recently been contracted to port another codebase to Polysemy. Today we hit our first big milestone, and the experience has gone swimmingly. I wanted to spend some time today discussing how to actually go about Polysemizing a codebase. It’s not too onerous if you proceed cautiously. The trick is to do several passes over the codebase, each time introducing a few more effects, but at no point ever actually changing any code paths.

Getting Your Foot in the Door🔗

The first step is to introduce Polysemy into your codebase. Your program is almost certainly structured around a main application monad, and that’s the right place to start. As a first step, we will swap out IO for Sem. For example, if your main monad were:

newtype App a = App
  { unApp :: ReaderT Env (ExceptT AppError IO) a
  }

we will change it to:

newtype App r a = App
  { unApp :: Member (Final IO) r => ReaderT Env (ExceptT AppError (Sem r)) a
  }

This change exposes the effect row (the r type variable,) and asserts that we always have a Final IO member in that row. Exposing r means we can gradually introduce Member constraints in application code as we begin teasing apart effects, and Final IO gives us a way to implement MonadIO for App. Let’s start with that:

instance MonadIO (App r) where
  liftIO a = App $ lift $ lift $ embedFinal a

Due to some quirks of how Haskell deals with impredicativity, this function can’t be written point-free.

This change of App to App r isn’t the end-goal; it’s just enough that we can get Polysemy into the project without it being a huge change. In the medium term, our goal is to eliminate the App newtype altogether, leaving a bare Sem in its place. But one step at a time.

You’ll need to rewrite any instances on App that you were previously newtype deriving. This sucks, but the answer is always just to lift. You might find that some instances used to be derived via IO, and thus now cannot be implemented via lift. In these cases, don’t be afraid to give an orphan instance for Sem r; orphans are bad, but we’ll be cleaning this all up very soon.

Take some time to get everything compiling. It’s a lot of drudgery, but all you need to do is to add the r type variable to every type signature in your codebase that mentions App.

You will also need an introduction function, to lift Polysemy actions into App:

liftSem :: Sem r a -> App r a
liftSem a = App $ lift $ lift a

as well as an elimination function which will evolve as you add effects. At some point in your (existing) program, you will need to actually run App down to IO. It probably looks something like this:

runApp :: Env -> App a -> IO (Either AppError a)
runApp env = runExceptT . flip runReaderT env . unApp

instead we are going to create the canonical interpretation down to IO:

type CanonicalEffects =
  '[ Final IO
   ]

canonicalAppToIO :: Env -> App CanonicalEffects a -> IO (Either AppError a)
canonicalAppToIO env
  = runFinal
  . runExceptT
  . flip runReaderT env
  . unApp

As we pull effects out of the program, we will add them to CanonicalEffects, and their interpreters to canonicalAppToIO. But for now, this function is very boring.

Once everything is up and compiling, all of the old tests should still pass. We haven’t changed anything, just installed some new machinery. But importantly, all of code paths are still exactly the same. Remember, this is a refactoring task! The goal is to do lots of little refactors, each one pulling out some effect machinery, but not changing any code paths. The entire porting project should be a series of no-op PRs that slowly carve your codebase into one with explicitly described effects.

First Effects🔗

Your medium term goal is to eliminate the Final IO constraint inside of App, which exists only to provide a MonadIO instance. So, our real goal is to systematically eliminate raw IO from App.

The usual culprits here are database access, HTTP requests, and logging. If your team has been disciplined, database access and HTTP requests should already be relatively isolated from the rest of the codebase. Isolated here means “database calls are in their own functions,” rather than being inlined directly in the application code whenever it wants to talk to the database. If your database accesses are not isolated, take some time to uninline them before continuing.

Our next step is to identify CRUD groups on the database. We generously interpret the “read” in CRUD to be any queries that exist against the logical datastructure that you’re serializing in the database. These CRUD groups might be organized by table, but they don’t necessarily need to be; by table is good enough for now if it corresponds to how the queries exist today.

For each CRUD group, we want to make a new Polysemy effect, and thread it through the application, replacing each direct call to the database with a call to the effect action. Finish working on each effect before starting on the next; each group makes for a good PR.

For example, maybe we’ve identified the following database accesses for table users:

insertUser       :: MonadDB m => UserName -> User -> m ()
lookupUser       :: MonadDB m => UserName -> m (Maybe User)
getUsersByRegion :: MonadDB m => Region -> m [User]
setUserLapsed    :: MonadDB m => UserName -> m ()
unsetUserLapsed  :: MonadDB m => UserName -> m ()
purgeUser        :: MonadDB m => UserNamr -> m ()

This CRUD group corresponds to an effect:

module App.Sem.UserStore where

data UserStore m a where
  Insert      :: UserName -> User -> UserStore m ()
  Lookup      :: UserName -> UserStore m (Maybe User)
  GetByRegion :: Region -> UserStore m [User]
  SetLapsed   :: UserName -> UserStore m ()
  UnsetLapsed :: UserName -> UserStore m ()
  Purge       :: UserName -> UserStore m ()

makeSem ''UserStore

We can now replace all calls across the codebase to insertUser a b with liftSem $ UserStore.insert a b. Doing so will require you to propagate a Member UserStore r constraint throughout the callstack. I really like this process. It’s a bit annoying to push constraints upwards, but it really gives you a good sense for the hidden complexity in your program. As it turns out, MonadIO is hiding a metric ton of spaghetti code!

All of this replacing and constraint propagating has given you dependency injection. But remember, at this step we’d like all of our changes to be no-ops, so we still need to inject the old codepath. For this we will make an interpreter of the UserStore effect:

module App.Sem.UserStore.IO where

import qualified TheDatabase as DB
import App.Sem.UserStore

userStoreToDB
    :: forall m r a
     . (Member (Embed m) r, MonadDB m)
    => Sem (UserStore ': r) a
    -> Sem r a
userStoreToDB = interpret $ embed @m . \case
  Insert un u    -> DB.insertUser un u
  Lookup un      -> DB.lookupUser un
  GetByRegion r  -> DB.getUsersByRegion r
  SetLapsed un   -> DB.setUserLapsed un
  UnsetLapsed un -> DB.unsetUserLapsed un
  Purge un       -> DB.purgeUser un

Make sure to add UserStore (and its dependency, Embed DB) to the head of CanonicalEffects:

type CanonicalEffects =
  '[ UserStore
   , Embed DB  -- dependency of UserStore
   , Embed IO  -- dependency of Embed DB
   , Final IO
   ]

and then we can update the canonical interpreter:

canonicalAppToIO :: Env -> App CanonicalEffects a -> IO (Either AppError a)
canonicalAppToIO env
  = runFinal
  . embedToFinal
  . runEmbedded @DB @IO (however you run the DB in IO)
  . userStoreToDB @DB
  . runExceptT
  . flip runReaderT env
  . unApp

The general principle here is that you add the new effect somewhere near the top of the CanonicalEffects stack, making sure to add any effects that your intended interpreter requires lower in the stack. Then, add the new interpreter to canonicalAppToIO, in the same order (but perhaps presented “backwards”, since function application is right to left.) Make sure to add interpreters for the depended-upon effects too!

As you pull more and more effects out, you’ll find that often you’ll already have the depended-upon effects in CanonicalEffects. This is a good thing — we will probably have several effects that can all be interpreted via Embed DB.

The benefit here is that we have now separated our application code from the particular choice of database implementation. While we want to use userStoreToDB in production, it might make less sense to use in local testing environments, where we don’t want to spin up a database. Instead, we could just write a little interpreter that emulates the UserStore interface purely in memory! Once you’ve fully exorcised IO from your codebase, this approach gets extremely powerful.

Choosing Effects🔗

Carving out your effects is probably the hardest thing to do here. What’s difficult is that you need to forget your instincts! Things that would make a good MTL-style typeclass are often terrible choices for effects.

Why’s that? There’s this extremely common pattern in the Haskell ecosystem for libraries that want to expose themselves to arbitrary applications’ monad stacks. To continue with the MonadDB example, it’s likely something like:

class (MonadIO m, MonadThrow m) => MonadDB m where
  liftDB :: DB a -> m a

While this works fine for a single underlying implementation, it’s an awful effect for the same reason: there’s only one interpretation! Any meaningful interpreter for MonadDB is equivalent to writing your own implementation of the database! It’s the same reason we don’t like IOIO is so big that every possible interpretation of it would necessarily need to be able to talk to the file system, to the network, to threads, and everything else that we can do in IO.

Instead, when you’re looking for effects to pull out, you need to forget entirely about the implementation, and just look at the abstract interface. Don’t use an HTTP effect to talk to a REST API — it’s too big, and would require you to implement an entire HTTP protocol. Instead, just define an effect that talks to exactly the pieces of the API that you need to talk to. Forget that it’s REST entirely! That’s an implementation detail, and implementation details are the domain of the interpreter, not the effect.

Furthermore, if you’re just using the standard Polysemy effects, pick the smallest effect that you can get away with. You’ll probably reach for Reader more often than you should. You don’t need to use Reader unless you need local — otherwise, prefer Input.

Summing Up🔗

That’s all I have for today, but I have a few more posts in mind for this series. One on how to actually go about testing all of this stuff, and another on how to follow up the refactoring of your new Polysemy codebase now that all of the IO has been removed.