Running Build Bots On Premise

Ryan James Spencer

Late November I did a video series discussing continuous integration and automation strategies for projects. I used GitHub Actions as they aided me in demonstrating configuration of pipelines without setting up supporting infrastructure. If you are a developer making things having a fast response time for feedback is crucial and continuous integration helps drastically.

When you use a SaaS offering for CI you are stuck using whatever tiers and upgrades are on offer. For the last six months, however, I've not used a SaaS offering for my infrastructure and instead have chosen to run computers in my home. I use buildkite to pick my own infrastructure. I did the numbers for renting my ideal EC2 instances on AWS and figured I could pay the same amount of money to buy a machine or two to do my bidding that would still be relevant four to five years later. Buildkite has an offering for an elastic stack build agent that can scale to zero when idle but I the stack configuration was too bloated to my liking. Nonetheless, having the ability to opt into whatever infrastructure you please has some cool consequences and I doubt this will be my last post on the subject!

Regardless of running compute locally or in the cloud, one can choose if they want to pay the overhead of virtualization or let things build on bare metal, which makes sense for slower machines. In the case of local compute this is a "true" bare metal unless you are paying a cloud provider the money to not rent in a co-tenant server. It does help to build things in a virtualized environment with the excuse that it is a "clean room" but if you want to do a compile check or run some tests you probably don't care about dirt.

Initially I bought three raspberry pis; two B+'s and one Zero. The intent was to run docker on them but I had forgot the host needs to be the same architecture as the image you intend to build on and I often use x86_64 images. There was nothing stopping me from converting these little boxes to running the languages directly for tests and basic checks. I have yet to see any architecture specific failures with the languages I tend to build for. These agents don't produce artifacts as I don't need ARM releases.

Later I took a box I use for streaming video, a 2011 Mac mini I upgraded to 16GB of memory, and equipped it with a few buildkite agents and docker. I have since reduced the box to running a sole agent. This build bot has been the most frustrating because the box isn't only doing CI related things. It's easy for docker to randomly die or get OOMed as the browsers can take up a fair portion of CPU and memory when streaming video, often ~20-30% CPU load, and it makes no sense for me to constrain the docker daemon unnecessarily. Eventually I split configurations into two types: those that use docker or those that don't.

I also own a rather beefy Intel ATX tower that sometimes participates as a build bot. I recently dissected an older tower to contribute whatever parts I could collect to build another ATX box for full-time builds to ease pressure off the mac mini on release builds. No matter what the machine is, I try to use it do any kind of automation, CI related or not. I was curious if anyone else was crazed enough like me to do this. I poised a question on twitter a bit indirectly:

My intent was to find out how often people buy various machines and actually put them to use for purposes beyond what they use directly. I got several awesome responses, primarily that often people buy new machines every 4-7yrs and, no, this isn't really a common thing, at least for the people following me and keen enough to have respond.

That said, a few cool machines were mentioned:

  • NUC is a small form-factor device sold by Intel. There are equivalent ones such as from System76 and if you care about open firmware it's well worth supporting them!
  • Mintbox3 -pro and -basic both seem really cool and are competitively priced in comparison to building your own machine. They are also fanless if noise is a concern!
  • You wouldn't run this as a node in the cluster itself, but I had never heard of the GDP portable ultrabooks that seem like they'd be useful for quickly SSH'ing into a box without having to carry around a full-sized laptop.

Inevitably there are other computers I'd love to own "just cus":

You don't need a super computer to build and test software. Unless you are looking for a laptop or need really bespoke hardware you can generally build a desktop machine that'll run laps around most of its earlier variants. Building your own also gives you a degree of customisation although, to be fair, it is partly on par with cloud compute options: if you want to upgrade your file system storage to a higher IOPS device, for example, you can abstractly do that with a cloud provider, although you get a much finer degree of resolution with your own computers.

Now that renting compute from cloud providers is commonplace, I suspect most people would favor the ease of cattle-based systems administration and simply slay any misbehaving servers. I find doing local systems administrations to be a bit educational and cathartic in the sense of being a master of what you have and working within limitations.

In terms of availability it's ok! There is always the risk that my children flip a power switch or something goes wrong with my internet connection or power. This sort of thing is already solved if you have, say, an AWS EC2 spot instance in an autoscaling group where the terminating box will swiftly be replaced. Because of this risk I occasionally supplement with cloud compute temporarily.

There are a few lingering things such as:

  1. External access into the specific network that has the bots. Something like and wireguard could supplant this if configured correctly. I remember seeing Brad Fitzpatrick asking about ways to do a TLS based termination into his home network awhile back.

  2. Doing smarter things with scaling agents in and out as ewll as scaling to zero, regardless of location. I had one crazy thought which was to have agents scale out if local bots have pending work but haven't picked anything up in X amount of time, say an hour tops. The newly spun up agent could sit around for an hour attempting to pick up work and then kill itself when idle, too.

If you or someone you know runs local compute at home I'd love to get in touch and see what usages are in practice out there. I'm always curious to see how people are using on premise computing rather than switching entirely over to cloud compute.