Cirrus CI continues to amaze
This writeup is not sponsored, nor am I affiliated with Cirrus Labs. I’m just a pleasantly delighted user of 1.5 years.
If you’re a FLOSS maintainer hosting your projects on GitHub, Cirrus CI can offer you three killer features.
Don’t snigger! Even if you don’t care about FreeBSD as a target platform, you can still take advantage of its affinity with macOS. Both operating systems are descendants of BSD, and some build failures affect the whole family. If you don’t have a Mac lying around, you can’t legally run OS X in a virtual machine, and so FreeBSD becomes a very compelling option. Not to mention that its source is open, and it has up-to-date documentation.
All your jobs can simultaneously use up to 16 CPUs for Linux, 12 for macOS, 8 for FreeBSD, and 8 for Windows. You are free to slice them however you want (except on macOS). For example, you can give a linter job a single core, while the compilation jobs get four each. Furthermore, the limits are per-user, so you won’t get starved of CI time no matter how many pull requests you get!
The only meaningful competition here is GitHub Actions: they provide an unlimited number of dual-core VMs. But if you have just a handful of heavy-weight jobs, you’re still better off with Cirrus.
You can probably get a competitive proposition from CircleCI and Travis, but they bill by a minute, so the dynamics will change every time you change your pipeline. Either way, you probably don’t want to wake up to an email notification saying that you’ve run out of your compute credits for the month.
You pay for concurrency by accepting some flakiness. Cirrus uses AWS Spot instances to make such high limits viable, and Amazon can take those machines back at any time. Cirrus papers over this by automatically restarting affected jobs, but GitHub would still display them as “failed” until the restarted ones succeed. From my experience, about once a week some job won’t be restarted, or won’t start at all.
Dockerfile as an environment
You specify a path to a Dockerfile; Cirrus builds and caches the image, then creates a container and runs all your tasks inside of it.
Contrast this with what other CI providers give you. At best, you are allowed to specify a Docker image to use. If no image fits your needs perfectly, you’ll have to build, publish, and maintain your own. At worst, your CI provider has a VM image with some popular stuff pre-installed, and you have to waste CI time installing the rest—every time your job runs.
This feature can break reproducibility. You can’t download a Docker image from Cirrus cache; you have to build one locally. As a result, your container might be slightly different from the one in CI, and you won’t reproduce the failure. This hasn’t bitten me yet, but I think it’s just a matter of time.
Too good to be true?
With Travis changing their billing just a few months ago, it’s an obvious question to ask: how long will the Cirrus-FLOSS paradise last? @email@example.com says they’re an “NYC-based startup”, which to me indicates that their future is less clear than that of GitHub (owned by Microsoft) or Travis (owned by Idera). Maybe this post will help Cirrus get a few more paying customers, and thus delay the exit a bit :)
If/when the time comes, I don’t think that migrating off Cirrus will be any more complicated than migrating off any other CI provider. The dockerfile feature means you’ll have your build environments in Docker already, which is widely supported (if not as conveniently as in Cirrus). The pipeline config you won’t keep anyway, as no two CI providers use the same format. If Cirrus limits the resources, you could keep just the FreeBSD jobs on there, as other platforms are well-represented with other providers.
Other than that, my biggest gripe with Cirrus CI is lack of support for any forges other than GitHub. When I migrate to a different one, it’d be sad to leave all this wealth of features behind. But until then, I think Cirrus is going to serve well.
Drop me a line! (wonder where’s the comments form?)