Small Scale GitOps
While working on a bunch of tools for WaffleHacks, a hackathon I helped organize that recently finished, I kept running into issues deciding when to deploy something onto our server. No matter whether it was a small or large change, I would still need to go through the same arduous process of:
- SSHing into the server
- pulling the changes from GitHub for each service
- restarting the service(s)
- waiting for the application to start, and hoping it worked
Depending on how long I waited between deployments, this could take anywhere from 5 minutes to an hour. Time I could be spending being more productive, like working on new features or maybe doing homework.
This same process continued for around 2 months when I decided to attempt to automate this workflow. I began working on a quick and dirty tool I sensibly named AutoDeploy. Using webhooks from GitHub and a configuration file located at the root of the repository, it would pull and run a set of predefined commands on every push. It worked well enough for a while, but it was very brittle. If any dependencies changed or a new application needed to be added, AutoDeploy would blindly run through its commands and the deployment would most likely fail. Not to mention that there was no way to know whether the deployment was successful because AutoDeploy never reported its status anywhere.
Shortly after setting up AutoDeploy, I stopped making and deploying as many changes since it was becoming more of a burden to deploy than before. I quickly realized I needed a new solution.
A system development/management pattern where
- git is the SINGLE source of truth for a system
- git is the SINGLE place where we operate ALL environments
- ALL changes are observable and verifiable
Using something like ArgoCD or Flux seemed like the perfect solution to my deployments problem. Just containerize all the applications and use a single repository to store all the Kubernetes manifests.
All done, right? Unfortunately not.
Everything seemed great until I took a look at the cost of running this. At the minimum it would cost double, maybe even triple what we were paying for our server, granted that was only $20/month. However, for a bunch of broke college students trying to run a hackthon with, at the time, no sponsors, this was a no-go. So I went searching for a smaller solution that could be used on a single server, or could scale down to a single server.
After searching for about a week, I couldn’t find anything. Either I was using the wrong terms, nothing like it exists, or they are all environment specific solutions that are not public. So, Like any overly enthusiastic programmer, I decided to write one myself.
Around 3 months later, after weeks of on-off work interspersed throughout my internship, WaffleMaker was born. Why the name? One of my inspirations for the tool was Beekeeper made by HackGT, so WaffleMaker seemed fitting given the hackathon’s name was WaffleHacks.
Similarly to Flux and ArgoCD, WaffleMaker allows using a single git repository as the single source of truth for deployments. However, since it does not run on Kuberentes, it can’t benefit from having a pre-defined manifest format for any resources, or a nice API to hook into for deploying them. As such, WaffleMaker is what you get when you smash Kubernetes, ArgoCD, and Vault together into one application. It does some light container monitoring to prevent an application from being inaccessible, receives webhook events whenever a push is received on the source repository, and injects secrets into containers from Vault.
I wrote a custom manifest format to define each service (example manifest). It allows specifying the base image, any configuration through environment variables, secrets to pull from Vault, and the image to deploy from. The benefit of using Vault for secret management, is that you can generate credentials for things on the fly. In the case of WaffleMaker, I use it to generate credentials for AWS and a PostgreSQL database. This means less credential management, and it still stays secure.
In addition to WaffleMaker, we are using GitHub Actions at WaffleHacks to automatically build container images on each push to any branch. This allows us to automatically update the running container ensuring the latest version is deployed. Our tagging scheme is such that each image is tagged with the commit hash and the branch it was pushed to. When configuring the service, we can specify which tags are allowed to be updated from, preventing development versions from accidentally being deployed.
Currently, WaffleMaker is deployed and being used (our source of truth). Looking ahead, there are a handful of changes and process improvements I would like to make:
- make the subdomain name not based on the file path
- allow explicit communication between containers
- utilize container health checks for restarting a service
That’s all for this post, thanks for reading! If you know of any existing solutions or see anywhere WaffleMaker could be improved, feel free to comment below.