What is aws-side-cluster?

This project is a proof of concept for stretching a home Kubernetes cluster into the cloud. A local workstation with four GPUs runs the main control plane and serves language models. When more capacity is needed, extra worker machines are started on AWS and connected back to the local network through WireGuard, a lightweight VPN tool. Those cloud workers can include plain CPU machines or a GPU machine running its own language models, and they join the same cluster as if they were sitting next to the local hardware. The idea driving the project is that cloud GPU workers should be disposable. They can be shut down after a few minutes of being idle to save money, then woken back up automatically when a request needs them. The author's companion article, linked in the README, frames the core tension the repo explores: even if a GPU worker turns itself off quickly, any decision it made while it was running still needs to be traceable afterward. So the repo also builds tooling to record and later reconstruct exactly what happened during a run, even after the machine that did the work is gone. Inside the repo you get the reproducible pieces of this setup: Terraform scripts to create the AWS network and machines, Helm charts to deploy everything onto Kubernetes, shell and Python scripts to manage the WireGuard connection and start or stop workers, a Grafana dashboard for monitoring, and a small terminal interface for replaying past decisions. There is also a component called a choreographer that runs multi model ceremonies, meaning coordinated tasks that involve several language models split between the local machines and the AWS worker, along with real logs of past runs. The documentation describes validated milestones dated between late June and early July 2026, including networking, GPU worker setup on an AWS L4 GPU instance, running two models on that single GPU at once, and automatic shutdown when idle. This is very much an experimental, hands on infrastructure project rather than a polished product, and it assumes familiarity with Kubernetes, Terraform, and cloud networking concepts.