I’ve been supporting the CI needs of a JS project and would like to assemble the lessons learned from the app here so that I can more clearly advise future JS projects on how to make their efforts more “CI friendly”.
The main things I’ve noticed are summarised below and covered in the section that follow. I include a “why” and a section on possible approaches to accomplishing this.
- Do More Processing Ahead of Time
- Single command to start App or Tests
- Incremental Processing
- Avoid Unmergable Custom Forks of Upstream Dependencies
- Use Packages Instead of Cloning Repos
- Have an Explicit Version
Do More Processing Ahead of Time
Why? Speed. Allow compilation-like tasks (CSS transforms, TS to JS transpilation, image processing, etc) to be done completely ahead of time.
In a containerised environment create a “cache” image that has all the prep work mentioned above applied and then use it as a base image for your deployments (whether for running tests or the service proper). In a similar fashion a VM image (VMware or an AWS AMI) can also be created ahead of time in a non-containerised environment.
To facilitate re-use of the ready-to-run code it could also be packaged into an OS package such as Debian’s .deb, RPM or Chocolatey. This can allow people to install particular version locally in VMs.
Single command to start App or Tests
Why? Simplicity. Simple startup makes CI easier and less error prone. It also makes it much easier for people to set up local development environments.
Consolidate tasks into, for example, a single grunt task so that CI folks can just use that to get things going. Even better would be for the app to know how to query environment variables to find its dependent services rather than relying on config files - this is very useful in containerised environments like Mesos with Marathon or Kubernetes.
Why? Speed. Unnecessarily processing unchanged files greatly adds to the time it takes to run/test your app and also adds compute expenses.
Only transform/compile/process what has changed rather than everything. For example, if one typescript file has changed then only transform that single file to JS rather than everything. Likewise for binary processing of items such as images. I am still research this so can hopefully post more on this topic or point to the work of others.
Avoid Unmergable Custom Forks of Upstream Dependencies
Why? Avoid lock-in. By using a custom dependency only useful for your team you are locked into that dependency and cannot easily upgrade to get new features and stability improvements.
If a framework or library you are using does not meet your needs but can be easily modified then by all means do so. But then don’t sit on the changes and neglect to submit a PR upstream. If you make changes that look like upstream won’t accept them then question the direction you’re taking the code. Likely there is a more elegant way to accomplish what you want to do.
Use Packages Instead of Cloning Repos
Why? Speed. Simplicity. It is quick and easy to install a package.
By using packages you can more simply pull down dependencies and also make your build more repeatable by depending on particular versions rather than what happens to be in a repo’s branch. Also, in CI environments, using a package repo like NPM is easier than arranging access to (often private) Git repositories which require juggling SSH keys or other credentials.
Have an Explicit Version
Why? Simplicity. Having a single version for your whole project makes it easier to package and deploy.
It may seem obvious but sometimes projects just organically grow and are built from their SCM repo without an explicit version. Put in place something so that each build has some kind of unique version even if it is just the Git commit hash or simply an incrementing number. Or you can do something more sophisticated like semantic versioning.