This is a fork of [begriffs/heroku-buildpack-ghc](https://github.com/begriffs/heroku-buildpack-ghc begriffs/heroku-buildpack-ghc).
Its been modified to pass over control to the app/repo itself to allow for more projects with less mainstream build processes.
My specific use case was dependencies managed by private submodules, which Heroku doesn't really handle, so my bin/build
script would pull in the private repos, add them as sources to the cabal sandbox and build the main project.
This also allowed me to separate the heroku build/runtime configuration and deployment from the project itself
So the main two modifications are as follows:
- Does not call
cabal
to build the project, callsbin/build
, meaning the working directory for thebin/build
script will be the root of the repository. - Clears the sandbox sources before updating the cached version of the sandbox.
What follows is the original documentation from [begriffs/heroku-buildpack-ghc](https://github.com/begriffs/heroku-buildpack-ghc begriffs/heroku-buildpack-ghc).
This buildpack supports frameworks like Yesod, Snap, and Happstack with the latest stable GHC binaries. Putting Haskell web applications online should be easy, and now it is. Try it for yourself.
**Note**, this buildpack does a basic build from scratch. For a faster build, try [mietek/haskell-on-heroku](https://github.com/mietek/haskell-on-heroku) which detects your project dependencies and fetches a custom pre-built sandbox.
Here's how to go from zero to "hello world" on Heroku. You'll need to install the Haskell Platform and the Heroku Toolbelt on your local machine, then do this:
# Generate a barebones snap app called snapdemo
mkdir snapdemo && cd $_
cabal sandbox init
cabal install snap
cabal exec snap init barebones
# Tell Heroku how to start the server
echo 'web: cabal run -- -p $PORT' > Procfile
# Create a git repo and deploy!
git init .
echo "dist\n.cabal-sandbox\ncabal.sandbox.config" > .gitignore
git add *
git commit -m 'Initial commit'
heroku create --stack=cedar-14 --buildpack https://github.com/begriffs/heroku-buildpack-ghc.git
git push heroku master
The first deploy is slowest as the environment downloads and bootstraps. Subsequent deploys use cached binaries and cached cabal packages to go faster.
The first time you try to deploy a big framework like Yesod the compilation can take so long that Heroku cuts it off. If this happens fear not, you can build your app with an Anvil server.
# Enable Anvil builds
heroku plugins:install https://github.com/ddollar/heroku-anvil
# Move big build artifacts out of the way or else the upload
# to Anvil will be very slow
mkdir -p /tmp/deploy-stash ; mv .cabal-sandbox /tmp/deploy-stash ; mv dist /tmp/deploy-stash
# Build your slug and cache without any time limits
heroku build -r -b https://github.com/begriffs/heroku-buildpack-ghc.git
# Use Anvil-generated cache next time we do a regular git push to Heroku
heroku config:set EXTERNAL_CACHE=$(cat .anvil/cache)
# Bring your sandbox etc back
mv /tmp/deploy-stash/.cabal-sandbox . ; mv /tmp/deploy-stash/dist .
After the first deploy using Anvil you can go back to the regular deploy process. This is because the cabal sandbox etc are cached by Anvil and will be retrieved, making future builds incremental and fast.
Remember to heroku config:unset EXTERNAL_CACHE
after your first
successful regular (post-Anvil) git push.
Cabal sometimes gets confused on Heroku and tries installing outdated packages. If you have your app working locally you can constrain the remote package versions to match your local environment. Just do this:
cabal freeze
git add cabal.config
# commit and push to fix remote build
You can change build settings through Heroku environment variables.
# allow the buildpack to see environment vars
heroku labs:enable buildpack-env-arg
# then set the variable of your choice
heroku config:set VARIABLE=value
Here are the options
name | description | default |
---|---|---|
CLEAR_CACHE | Force everything to reinstall from scratch by setting to 1. | 0 |
EXTERNAL_CACHE | Url of replacement buildpack cache tarball. Useful for Anvil. | |
GHC_VER | GHC version to download or build | 7.8.2 |
CABAL_VER | Version of cabal-install | 1.20.0.0 |
PREBUILT | Base url for the prebuilt binary cache | https://s3.amazonaws.com/heroku-ghc |
PRE_SCRIPT | Path to an executable, relative to checkout root. If present, this will be chmodded (+x) and ran before anything else. |
heroku run bash # shell access
heroku run cabal repl # Haskell repl with all app modules loaded
- Latest binaries: GHC 7.8.2, cabal-install 1.20.0.0
- Uses cabal >=1.20 features to run the app and repl
- Exposes Haskell platform binaries to your app and scripts
- Uses prebuilt binaries for speed but...
- ...can fall back to building the standard GHC distribution
There are a number of ways to improve this buildpack. Please see the Github issues for ideas.
In order to contribute to the build script it will help to understand how Heroku's deployment process works, and how that affects GHC. Heroku provides three areas for storing files during build: a cache directory, a working directory, and a build directory.
The cache, called $CACHE_DIR
in the script, persists between
deployments. We use it to avoid building binaries more than once. The
working directory, called $WORKING_HOME
, does not persist between
builds, or even after the build script is done. Seems we should avoid
this area, right? Well GHC has some idiosyncracies that make this area
quite useful as you will see. Finally the build destination directory,
$BUILD_DIR
, holds the git repo the user pushes and gets copied into
what will be /app
in the deployed application. During build it lives
in a weird nonce filename.
We want to use GHC binaries at build time to compile the app, and we
would also like those binaries to be available in the app environment
after deployment (so people can use runhaskell
or cabal repl
). Seems
like we should install directly to $BUILD_DIR
. There's one problem:
GHC breaks if you move it to a new path after installation because many
of its binaries are just scripts with hard-coded full paths in them to
other GHC files. And as you remember, Heroku is going to move things in
$BUILD_DIR
to /app
. So the trick will be to install to /app
in the
working directory, use GHC there, then copy that installation to the
build directory which will be renamed in the deployed application and
not notice it has been moved.
GHC is also sensitive to having libgmp
named just right. We don't
have privileges to adjust /usr/lib
in the deployed app so we create a
symbolic link in a place we are permitted and set linker variables in
the shell so that everything can build.
Some packages have external dependencies (i.e. non-Haskell dependencies which cannot be satisfied by cabal
). If you come across such a package, check in contribs
to see if someone has already created patches for the dependencies. If they have, you should be able to
patch -p0 < contribs/DEP.patch
to patch the buildpack and proceed. If not, you'll need to modify bin/compile
yourself. Contributing these changes back as patches is appreciated.
As new versions of GHC and Cabal are released we should build them for Heroku and put them on S3 to speed up future deploys for everyone. Luckily the buildpack can do the building too.
Adjust the GHC_VER
and CABAL_VER
environment vars and then
deploy. It will build the new binaries from the standard GHC
distribution. Then copy the results to S3 like this:
heroku run bash
# now SSH'd into the server
cd /app/vendor
curl -L http://softlayer-ams.dl.sourceforge.net/project/s3tools/s3cmd/1.5.0-alpha1/s3cmd-1.5.0-alpha1.tar.gz | tar zx
s3cmd-1.5.0-alpha1/s3cmd --configure
# ^^^ answer the configuration questions
tar zcf heroku-ghc-[VERSION].tar.gz ghc-[VERSION]/
tar zcf heroku-cabal-install-[VERSION].tar.gz cabal-install-[VERSION]/
s3cmd-1.5.0-alpha1/s3cmd put heroku-ghc-[VERSION].tar.gz s3://[BUCKET]
s3cmd-1.5.0-alpha1/s3cmd put heroku-cabal-install-[VERSION].tar.gz s3://[BUCKET]
Thanks to Brian McKenna and others for their work on heroku-buildpack-haskell which inspired and informed this buildpack. For a history of that project's contributions and ideas see [this article] (http://blog.begriffs.com/2013/08/haskell-on-heroku-omg-lets-get-this.html).