GHC 7.8, transformers 0.3, and lenient lower bounds

12 May 2014 Michael Snoyman

In the Stackage maintainer's agreement, there's a section about keeping your package compatible with the newest versions of all dependencies. What the maintainer's agreement doesn't (yet) discuss is when it's important to be compatible with old versions of a package. The reasons for this are not immediately obvious, especially as it affects a smaller subset of the Hackage author population. This blog post will cover some of the reasons for this goal.

The original impetus for writing this was to get one specific message across: please continue supporting transformers-! For the explanation of why, please keep reading.

Non-upgradeable packages

The simplest case to discuss is packages like base and template-haskell. Not only are these packages shipped with GHC, but they cannot be upgraded. As a result, if you have a package that says base >= 4.7, it will only work with GHC 7.8 and later. Users who are still using 7.6 (or 7.4... or earlier... yes, those people do in fact exist) will have no means of using your package.

That of course brings up a question of how many versions of GHC you want to support. I'd highly recommend always supporting the most recent Haskell Platform release, as many users (especially Windows users) stick to that. Going back an extra version as well isn't a bad idea either, especially as some distributions (e.g., Ubuntu) tend to ship relatively old GHC versions.

Upgradeable, GHC-shipped packages

This issue is more subtle. In addition to non-upgradeable packages, GHC includes a number of packages which can be installed separately, resulting in one copy of the package in your global database, and one in your user database. (Yes, you can also install into the global database, but I'm covering the common case here.) Examples of these packages are bytestring, binary, and containers.

The first problem with this is that it can lead to end-user confusion. How many of you have tried working in GHCi, or just compiling code with ghc --make, and gotten a message along the lines of "Could not match type ByteString with ByteString"? That usually comes from two versions of a package being available.

Now that's just a bit of an annoyance, and building your code with cabal will almost always avoid it. But there's a second, more serious problem. Some of these upgradeable packages are in turn depended upon by non-upgradeable packages. For example, template-haskell depends on containers. As a result, imagine if you try to use containers 0.5 and template-haskell when on GHC 7.4. Since template-haskell depends on containers-, you'll run into issues.

Another problem is the ghc package (aka GHC-the-library). With GHC 7.8.2, I have the following dependencies for the installed ghc package:

depends: Cabal-

So if I try to use- for example- transformers and a package requiring ghc at the same time, I'll run into a conflict. And there are actually a large number of such packages; just doctest has over 100 dependencies.

Haskell Platform

The last reason is the one I hear the most pushback about from package authors. The Haskell Platform pegs users at specific versions of dependencies. For example, the most recent HP release pegs text at Now imagine that you write a package that depends on text >= 1.0. A user with the Haskell Platform installed will likely get warnings from cabal when installing your package about conflicting versions of text, and possibly breaking other packages that depend on it.

I can tell you what I've personally done about this situation. For my open source packages, I make sure to keep compatibility with the Haskell Platform released version of a package. Sometimes this does lead to some ugliness. Two examples are:

  • streaming-commons has to have a copy of some of the streaming text code, since it was not available before text 1.1. (And due to an issue with cabal, we can't even conditionally include the code.)
  • In chunked-data, I wasn't able to rely upon the hGetChunk function, and instead needed to use CPP to include a far less efficient backup approach when using older versions of text.

In the Stackage project, I run versions of the build both with and without Haskell Platform constraints. There are actually a whole slew of conditionals in the version selection which say "if you're using HP, then use this older version of a dependency." However, as time goes on, more and more packages are simply not supporting the HP-pegged versions of packages anymore.

Future changes

I'm not commenting here on the value of HP-pegged versions, but simply pointing out a reality: if you want your users to have a good experience, especially Windows users, it's probably a good idea to keep compatibility with the older HP-provided versions. I also think the ramifications of the HP approach really need to be discussed by the community, it seems like there's not much discussion going on about the impact of the HP.

Also, regarding the packages shipped with GHC: there have certainly been discussions about improving this situation. I know that removing the Cabal dependency from ghc has been discussed, and would certainly improve the situation somewhat. If others want to kick off a conversation on improving things, I'd be happy to participate, but I frankly don't have any concrete ideas on how to make things better right now.

comments powered by Disqus