I have recently discovered Docker's new multi-stage build feature. This has been a great help in answering my question of how to reduce the size of my haskell images for deploying, as the main haskell image on Docker Hub is over 1GB before you start adding things. This also means you don't have to maintain two Dockerfiles.
With multi stage builds we can use the haskell docker image to build our distributable executable, then copy this to a smaller image (in this case phusion/baseimage).
Here's an example Dockerfile for a project called haskell-example:
FROM haskell as build-env
WORKDIR /opt/server
RUN cabal update
COPY haskell-example.cabal .
RUN cabal install --only-dependencies -j4
ADD . .
RUN cabal install
FROM phusion/baseimage
WORKDIR /opt/server
COPY --from=build-env /opt/server/dist .
CMD ./build/haskell-example/haskell-example
Before, using just the haskell image: 1.87GB.
After, using multi-stage with baseimage: 249MB.
Image caching
There's just one problem with this method - the intermediate images are all marked as dangling. This means that if you do a docker system prune and rebuild, you will have install all your dependencies again, which takes a lot of time in the haskell world.
The workaround for this is to build an image using the build-env stage. This is done by specifying the target when building.
docker build -t haskell-example-dev --target build-env .
This keeps the image in the cache so it won't be prune'd.
Next steps
So we've shrunk our deployable image using phusion/baseimage, can we go further? The next step would be to use something like alpine as the production image, a mere 4.14MB. That's what I'll talk about next time :)
This post was cross-posted to my personal blog.
