Field notes from
the team.
Technical writing on GPU compute, CI/CD performance, and GitHub Actions infrastructure. Built and maintained by the engineers behind machine.dev.
Writing
archive.
6 posts indexed. Ordered by newest first.
Every job now records its own metrics
Every job now records CPU, memory, disk, network — and on GPU runners, GPU utilisation, memory, temperature, and power draw — drawn as sparklines on the job page. Generally available today, on by default, and free. Plus how to dial the sampling interval with one label.
READ_POST ›One runner, one job
When we added per-job disk controls, jobs started landing on the wrong runners. The subset/superset quirk in how GitHub matches self-hosted runners, and the small fix we shipped: one packed label with a unique id per job.
READ_POST ›Disk config, tunable and visible. Cost breakdown on every job.
Get started with GPU-powered GitHub Actions at a fraction of the cost of other solutions.
READ_POST ›We open sourced nat-zero: scale-to-zero NAT instances for AWS
We open sourced nat-zero — a Terraform module that replaces always-on NAT infrastructure with on-demand NAT instances that start when your workloads need internet access and shut down when they don't.
READ_POST ›Duct tape is cancelled. Pipelines only.
Everyone obsesses over the models. Nobody wants to talk about the plumbing. How proper ML engineering with GitHub Actions and machine.dev turns chaos into reproducible pipelines.
READ_POST ›Why We Built machine.dev
Get started with GPU-powered GitHub Actions at a fraction of the cost of other solutions.
READ_POST ›
Skip the plumbing.
Ship the model.
$10 free compute on signup. Two minutes to connect your GitHub org. torch.cuda.is_available() == True on the first run.