The woes of multiple package versions

September 23, 2014

GravatarMichael Snoyman

When I've answered the same question more than three times, it's usually time to write up a blog post explaining the situation in more detail, and just link people to that in the future. This is such a blog post.

Many people working with Haskell end up with one of these two classes of inexplicable GHC errors:

  1. Cannot match ByteString with ByteString, with a whole bunch of package name and version junk as well.

  2. SomeTransformerT is not an instance of MonadTrans, when the documentation clearly indicates that SomeTransformerT does define such an instance.

How can a ByteString not be a ByteString? Well, there are two ways I can think of. The first is that you're accidentally trying to mix up a strict ByteString with a lazy ByteString, whose types clearly don't unify. While that problem does pop up, in my experience most people figure that one out pretty quickly. By the time someone's asking a question on Stack Overflow/Reddit/haskell-cafe, it's usually much more insidious: there are two copies of bytestring on their system.

Imagine this situation: you install GHC 7.6, which ships with bytestring-0.10.0.2. You then install text via cabal install text. A few days later, someone mentions that bytestring-0.10.4.0 is out, and it's all new and shiny, so you go ahead and install it with cabal install bytestring. Everything works wonderfully, and life is good. Then you decide to experiment with text a bit, so you write the following program:

{-# LANGUAGE OverloadedStrings #-}
import Data.Text.Encoding (encodeUtf8)
import qualified Data.ByteString.Char8 as S8

main :: IO ()
main = S8.putStrLn $ encodeUtf8 "Hello World!"

Woe unto you! GHC rejects your program with:

foo.hs:6:22:
    Couldn't match expected type `S8.ByteString'
                with actual type `bytestring-0.10.0.2:Data.ByteString.Internal.ByteString'
    In the return type of a call of `encodeUtf8'
    In the second argument of `($)', namely `encodeUtf8 "Hello World!"'
    In the expression: S8.putStrLn $ encodeUtf8 "Hello World!"

When is a ByteString not a ByteString? Here, apprently. Now it turns out the GHC is actually giving you quite of bit of useful information, you just need to know what to look for. It's expecting the type S8.ByteString, which expands to Data.ByteString.Char8.ByteString, which in reality is just a type synonym for Data.ByteString.Internal.ByteString. So what GHC really means is that it can't unify the following two types:

expected:                     Data.ByteString.Internal.ByteString
actual:   bytestring-0.10.0.2:Data.ByteString.Internal.ByteString

Now the difference just jumps out at you: the actual type comes from the bytestring-0.10.0.2 package, whereas the first comes from... well, somewhere else. As I'm sure you're guessing right now, that "somewhere else" is bytestring-0.10.4.0, but GHC doesn't bother telling us that, since including that level of information in every error message would be overwhelming. To step through why this came up exactly:

  • text is installed against bytestring-0.10.0.2 (it was the only version of bytestring available at the time you installed text).
  • Therefore, encodeUtf8 will generate a ByteString value from version 0.10.0.2.
  • Your program imports Data.ByteString.Char8, which is provided by both bytestring-0.10.0.2 and bytestring-0.10.4.0.
  • GHC's default dependency resolution is: take the latest version of each package, in this case 0.10.4.0.
  • Now we have a S8.putStrLn function expecting a 0.10.4.0 ByteString, but an encodeUtf8 function returning a 0.10.0.2 ByteString.

So how do we work around this problem? I can think of three ways:

  1. Explicitly tell GHC which version of the bytestring package you want to use to force consistency, e.g. runghc -package=bytestring-0.10.0.2 foo.hs.
  2. Never use GHCi, runghc, or ghc directly from the command line. Instead, always create a cabal file first. cabal's default dependency resolution will force consistent package loading.
  3. Don't wind up in the situation in the first place, by ensuring you only have one version of each package installed.

That last point is what I strongly recommend to all users. And this is exactly the design goal around Stackage, so it will hopefully not come as a surprise that that's exactly what I recommend most Haskell users use to get their packages installed.

Let's demonstrate that second case of MonadTrans. This time, let's try it with GHC 7.8.3. GHC ships with transformers-0.3.0.0. Next, we'll install the either package with cabal install either. Once again, someone comes along and tells us about a shiny new package, transformers-0.4.1.0. Dutifully, we upgrade with cabal install transformers-0.4.1.0. And then we try to run the following simple program:

import Control.Monad.Trans.Class
import Control.Monad.Trans.Either

main :: IO ()
main = do
    x <- runEitherT $ lift $ putStrLn "hey there!"
    print (x :: Either () ())

GHC mocks you with:

foo.hs:6:23:
    No instance for (MonadTrans (EitherT ()))
      arising from a use of ‘lift’
    In the expression: lift
    In the second argument of ‘($)’, namely
      ‘lift $ putStrLn "hey there!"’
    In a stmt of a 'do' block:
      x <- runEitherT $ lift $ putStrLn "hey there!"

"But EitherT is an instance of MonadTrans!" you insist. That may be true, but it's an instance of the wrong MonadTrans. The either package is built against transformers-0.3.0.0, whereas you've imported lift from transformers-0.4.1.0. This can be worked around as above, with runghc -package=transformers-0.3.0.0 foo.hs. And yet again, my strong recommendation is: use Stackage.

There's one more particularly painful thing I need to point out. Some packages are bundled with GHC, and are depended on by the ghc package. The special thing about the ghc package is that it cannot be reinstalled without installing a new version of GHC itself. Any packages depended on by the ghc package cannot be unregistered without breaking ghc, which would in turn break libraries like doctest and hint. If you follow these points to conclusion, this means that you should never upgrade GHC-bundled libraries. I wrote a blog post on this topic, and the takeaway is: please, always support older versions of packages like bytestring, transformers, and- of course- base.

There's one final case I want to mention. Try running cabal install data-default-0.5.1 http-client, and then run the following program:

import Data.Default
import Network.HTTP.Client

main :: IO ()
main = withManager defaultManagerSettings $ \man -> do
    res <- httpLbs def man
    print res

You'll get the irritating error message:

foo.hs:6:20:
    No instance for (Default Request) arising from a use of ‘def’
    In the first argument of ‘httpLbs’, namely ‘def’
    In a stmt of a 'do' block: res <- httpLbs def man
    In the expression:
      do { res <- httpLbs def man;
           print res }

But if you look at http-client, Request is in fact an instance of Default. "Alright, I know what's going on here" you say. Certainly there are two versions of data-default installed, right? Actually, no, that's not the case. Have a look at the following:

$ ghc-pkg list | grep data-default
    data-default-0.5.1
    data-default-class-0.0.1

There's just a single version of each of these packages available. So why are we getting our mysterious error message? Once again, it's because we have two versions of the Default class. After data-default version 0.5.1, data-default split into a number of packages, and the Default class migrated into data-default-class. http-client defines an instance for Default from data-default-class. And if you use data-default version 0.5.2 or higher, it will simply re-export that same class, and everything will work.

However, our cabal install command forced the installation of the older data-default (0.5.1) which defines its own Default typeclass. Therefore, we end up with two separate Default classes that don't unify. This is a problem that exists whenever packages are split or joined, which is why you should embark on such refactorings with great care.

As it happens, this is yet another problem that is solved by using Stackage, since it forces a consistent set of versions for data-default and data-default-class.

Comments

comments powered by Disqus

Archives