seankelleher.ie

Docker for the Build Process

| 7 min read

This article covers the use of Docker as part of the build process, for both local development and continuous integration builds. In particular, it addresses the practice of building projects using docker build, and the alternative approach of using docker run to perform builds.

Many articles already espouse the benefits of using Docker for local builds, as part of the "dev loop", but to recap a few:

  • Running your local builds in Docker gives you better parity with the central build pipeline when the latter is also using Docker.

  • Your local builds can run in a minimal environment, reducing dependencies and the risk of depending on tools and resources that aren't available in the central build pipeline. This also helps with the issue of using mismatched versions of tools.

  • It's possible to work on a project without having any of the tools/languages/frameworks installed locally. This is particularly useful when working on projects for a short term, or for working on different projects that depend on conflicting versions of the same tools.

Building in Docker Images

When using Docker as part of the build process, some projects build the project artefacts as part of an image build. A typical Dockerfile to define a service may look like the following:

FROM golang:1.14.3-stretch

RUN \
    apt-get update \
    && apt-get install -y \
        fortune-mod \
    && ln -s /usr/games/fortune /bin/fortune

ENV GO111MODULE=on

COPY go.mod go.sum /go/src/github.com/ezanmoto/hello/

WORKDIR /go/src/github.com/ezanmoto/hello

RUN go mod download

COPY . /go/src/github.com/ezanmoto/hello

RUN go build ./...

ENTRYPOINT ["./hello"]

One of the most obvious problems with this approach is that this image defines both the build environment and the run environment. This leads to a number of issues such as increased image size, an increased number of attack vectors (because of all of the extra packages required for the build), and the more subtle issue of mixing contexts (for example, is a particular dependency required for the build time or the run time?). Thankfully, Docker added multi-stage Docker builds, which allows us to separate the definition of the build image and the run image, and allows us to make the run image as minimal as possible.

However, another problem exists in the form of the COPY commands:

  • COPY . /go/src/github.com/ezanmoto/hello means that a new image needs to be built every time we want to use Docker to test a code change.

  • COPY . /go/src/github.com/ezanmoto/hello actually results in the entire codebase being copied every time that Docker is used to build the project (assuming something in the codebase has changed). This is more of an issue in bigger projects, where it can take a few seconds to copy all files, adding delays to the development loop.

  • docker build has to be run a lot in this setup for debugging purposes; every time you change the Dockerfile to try something different you'll need to rebuild the image, further compounding any delays encountered by re-downloading project dependencies and copying the codebase into the image. It's not very interactive.

  • COPY go.mod go.sum /go/src/github.com/ezanmoto/hello/ can be a pain-point when working with Docker images. With dependency managers such as npm install and go get that can handle automatically fetching packages, it's useful to be able to quickly try out different combinations. However, because the COPY step will cause all packages to be re-downloaded instead of just the updates, this adds friction, and a resistance to using Docker in the development loop, when it comes to testing and updating different dependencies.

Building in Docker Containers

A straightforward alternative is to build artefacts in containers instead of in images, utilising bind-mounts instead of COPY. For example, instead of the image-based build in the previous section, the following can be used:

build.Dockerfile:

FROM golang:1.14.3-stretch

RUN \
    apt-get update \
    && apt-get install -y \
        fortune-mod \
    && ln -s /usr/games/fortune /bin/fortune

ENV GO111MODULE=on

WORKDIR /go/src/github.com/ezanmoto/hello

build_img.sh:

proj='ezanmoto/hello'
build_img="$proj.build"
run_img="$proj"

bash scripts/docker_rbuild.sh \
    "$build_img" \
    "latest" \
    --file='build.Dockerfile' \
    .

mkdir -p target
docker run \
    --rm \
    --mount="type=bind,src=$(pwd),dst=/go/src/github.com/ezanmoto/hello" \
    "$build_img:latest" \
    bash -c '
        set -o errexit

        go mod download
        CGO_ENABLED=0 go build -o target ./...
    '

bash scripts/docker_rbuild.sh \
    "$run_img" \
    'latest' \
    .

Dockerfile:

FROM alpine:3.11.3

RUN apk add fortune

COPY target/hello /bin

ENTRYPOINT ["/bin/hello"]

There is a notable increase in the amount of code that's present here, but there are also some immediate benefits:

  • Fewer image builds: The build image only needs to be built once, until its actual definition changes, as opposed to being built every time there's a code change. The run image only needs to be built in the CI environment unless it's being tested locally. When debugging the run image, rebuilding the image multiple times has less friction because changing files in the project won't break the cache.

  • No redundant copying: The build is being run using the host directory so the delay incurred from copying the build context over and over is avoided, without needing to play around with .dockerignore.

  • Project dependencies that are located within the project directory (e.g. .node_modules) can be kept from previous runs, even if the dependency/lock files change, and even if the definition of the build environment changes. Volumes can be used to cache dependencies that exist outside the project directory.

Here are some other small benefits that I like:

  • Succinct definition of the build environment: This makes it easier for developers that prefer to develop locally to actually mirror the exact build environment that'll be used in the build pipeline, by following the steps outlined in build.Dockerfile.

  • The definition of the build environment, run environment and build instructions are all cleanly separated.

However, the biggest benefit that I get from this approach lies in the fact that I can use the build environment interactively, as outlined in the following section.

Interactive Build Environment

Now that the definition of the build environment has been separated from the build instructions and the definition of the run environment, it's possible to work interactively within the build environment with the local project directory mounted:

docker run \
    --interactive \
    --tty \
    --rm \
    --volume="$(pwd):/go/src/github.com/ezanmoto/hello" \
    'ezanmoto/hello' \
    bash

This starts a Bash session "in" the build environment, but with the project directory bind-mounted within the container. This means that any changes made inside the container are reflected locally, and vice-versa. This setup has numerous advantages when used as part of the development loop:

  • Assuming that the build pipeline uses the same build image, there is now almost total parity between the development environment and the build pipeline environment, meaning that there's less variability after a developer pushes changes. A developer can easily build locally, without creating new Docker images, in almost the same conditions as the build will occur in the build pipeline.

  • A subtle benefit of the previous point is that developers will be using the same version of dependencies as the build pipeline. For example, a new team member may use Go 1.14 locally but the build pipeline might still be on Go 1.12 for various reasons. New Go features will work locally but will break the build. Being able to run with the same versions of tools locally means that there is less chance of this occurring in practice.

  • Again following on from the previous point, updates are effortless, and safe across projects. For example, the image for the build pipeline may get updated from Go 1.12 to 1.15. It can often be a daunting task to update local installs, especially if there isn't a simple mechanism for removing old versions. With a specially-defined build image, local software doesn't need to be updated, but developers can instead simply work in the new build environment without installing the build tools locally. This also means that issues won't arise when a developer is working on two projects at the same time that require different versions of the same dependencies.

  • Developers don't have to install programs locally at all! As a somewhat extreme example, even though I work primarily in Go and Rust, I don't have them installed on my host. Instead, they're installed in my build images that I work in interactively. This also means that environments can be cleaned effortlessly after finishing a project - removing the build images for that project removes all the programming languages and tools that were being used for that project.

  • Setting up the development environment for a new developer is now handled automatically by the docker build process. Developers that want to replicate the setup locally can follow the instructions defined in build.Dockerfile manually.

  • Outside of aspects like bound directories, the build environment is decoupled from the host. This means that there's less risk of accidentally depending on things that are present in the host environment that aren't going to be present in the build pipeline. A simple example of this could be depending on Linux tools that are native to the Ubuntu development host when Debian is being used as the build environment.

Disadvantages

While the approach outlined above is my personal approach and preference for managing Docker images in a project, there are some notable caveats to it:

  • The presented approach depends on the use of bind-mounting volumes for the build environment. This can work well in practice when using Linux images on Linux hosts, but may be less practical on other platforms. Furthermore, build environments that nest Docker containers may encounter extra complexity with working with bind-mounted volumes, as paths are referenced relative to the host's filesystem.

  • When working interactively in the build environment you may quickly realise that users that you have defined locally don't exist in the build environment. Furthermore, trying to map users between these environments isn't the most straightforward - for example, do you do it at build time or at run time? This is perhaps one of the trickier aspects of running builds in containers instead of images, and could be a big argument in favour of the image approach. This is because any required users are usually better-defined in the image approach, typically being set up using the likes of adduser/useradd at the start of the image definition.

  • Some people may consider the bind-mount approach to be less "pure" because the container is exposed to the host, and may worry about the reproducibility of the setup, since reproducibility is one of the main benefits of Docker. However, the approach is no less reproducible than the approach that performs COPY . /src, as the entire host context is copied into the image. With both approaches, it is the responsibility of the build pipeline to ensure that the environment is clean and set up for reproducible results.

  • As mentioned in the previous section, this approach achieves almost total parity between the development environment and the build pipeline environment, but that doesn't mean that subtle differences can't emerge between the two.

    For example, I encountered an issue with this setup when using Samson for CI/CD. By default, Samson doesn't actually do a full clone of a repository when running a new build/deployment, but instead creates a new Git worktree using symbolic links. This meant that the Git repository mounted in the build environment was referencing a file on the host, which couldn't be resolved in the build environment. I wasn't using worktrees locally, so this issue wasn't occurring in my local environment.

    The resolution was straightforward, but less than ideal: to force a full checkout for projects that needed it. Still, it highlights the fact that differences between the local environment and the build pipeline can still manifest with this approach, and special attention should be made when working with symbolic links in particular.

Conclusion

Building code in Docker containers using Docker run is generally faster, more space-conserving and more amenable to debugging than building code as part of a docker build. Separating the build environment definition from build instructions also allows for greater parity between the development environment and the build pipeline, and allows for easier management of project dependencies.