Skip to content

Latest commit

 

History

History
133 lines (95 loc) · 5.89 KB

README.md

File metadata and controls

133 lines (95 loc) · 5.89 KB
Raw

CircleCI

Utility Warehouse template Go monorepo

This repo is an abbreviated copy of one used by one of the teams inside Utility Warehouse. It's been built for tight CI integration and developer productivity.

Making it your own

If you're creating a new repo from this template, you'll want to do a search-replace on github.com/uw-labs/go-mono and uwlabs (replacing yourorg, yourrepo and yourscm with your own):

$ find . -type d -name "uwlabs" -exec sh -c 'mv {} $(dirname {})/yourorg' \;
$ find . -type f -exec sed -i 's;github.com/uw-labs/go-mono;yourscm.com/yourorg/yourrepo;g' {} +
$ find . -type f -exec sed -i 's;uwlabs;yourorg;g' {} +

Configure DOCKER_USER and DOCKER_PASSWORD in your CircleCI settings for release builds to work. The registry path to use is configured in the release CI job.

CI Setup

The CI setup uses Circle CI, partly because of the powerful caching features, and partly because Circle CI machine users have access to a local docker socket, which allows us to run integration tests using go test which work both locally and in CI. The CI jobs rely on Go build and test caching to be efficient even with a larger number of services. The same CI setup is used to efficiently test and publish over 130 applications within UW.

Linters

Automatic publishing

The release CI job automatically figures out what needs building and publishes a docker image to the configured registry. It requires the setting of DOCKER_USER and DOCKER_PASSWORD in the Circle CI configuration environment variables.

The Dockerfile used to build the images is here. It can be edited as necessary, just make sure to run make generate after changing it.

For an example of this, the user-api is published automatically to the local GitHub docker registry whenever it requires rebuilding.

Repository Layout

  • cmd - Utilities and service applications.
  • pkg - Shared packages.
  • proto - Protobuf definitions & generated code.
  • vendor - Vendored third-party dependencies.

Makefile

A top level Makefile exists to help you perform common actions within the monorepo. Recipes include:

  • format - Formats your .go source code
  • install-generators - Installs all necessary generators for the generate step
  • generate - Runs all generators required within the monorepo
  • lint-imports - Runs the import linter.

How do I add a new application?

  1. Create a new folder in cmd for your service E.g. cmd/my-new-service.
  2. Create a main.go
  3. If you want to automatically build a docker container, add a deploy.yml file.

Adding protofiles

Add your own protofiles under proto/uwlabs/. Follow the folder structure laid out in the buf documentation.

The calculate-releases script

calculate-releases is the magic that calculates exactly what applications need to be rebuilt based on file changes to the application directly, or any of its transative dependencies. It is called in CI on every branch push, to calculate which applications to build docker images for.

It relies on go list -json ./... for dependency information.

The deploy script

deploy is responsible for building and publishing a docker image to the specified registry. It's usually run on the output of the calculate-releases script. It defines a custom format for build configuration and metadata. This can be extended to include things such as kubernetes deployment targets, extra application metadata and more. It is currently run automatically against every branch push in CI.

The deploy.yml file

Use a deploy.yml together with any main packages that you want to deploy to configure automatic docker container building and publishing. The deploy.yml file allows a single configuration parameter:

  • name

    Used to configure the name of the docker image pushed to the registry.

Why a vendor directory?

When evaluating solutions to two problems, the vendor directory became the primary candidate:

  1. How do we keep CI builds as fast as possible?

    We first implemented this using module caching, where the first job would download all the modules and cache them for future jobs. It meant a lot of extra boilerplate in the CircleCI configuration files, and it never worked well for the Docker builds. Vendoring means we have all the source code available at all time, and completely removes the need for caching. This sped up builds by roughly 50% in testing.

  2. How do we ensure we only release applications that have changed when we perform dependency updates?

    This could be done with some custom tooling that can discover file changes between module updates, but this is nontrivial, and we already had a solution that worked with the vendor directory.

The CI pipeline will fail on your pull request if you add a new dependency that is not within the /vendor directory. If you've added a new dependency, make sure you run go mod tidy and go mod vendor to ensure your dependencies are up-to-date and vendored.