Skip to content

PitlaneEphemeral GitHub Actions runners with caches that stay warm

Each job gets a fresh runner on your own Docker host. Docker layers, package caches and toolchains stay on disk, so a private monorepo's CI went from ~6 min 30 on GitHub-hosted runners to ~3 min. One binary, a web UI, no Kubernetes.

Pitlane logo

Quick start ​

On a Linux host with systemd (Docker is installed if missing):

sh
curl -fsSL https://github.com/TuroYT/pitlane/releases/latest/download/install.sh | sudo sh

Open the printed URL, enter the setup token (pitlane setup-token), click Create GitHub App, install it, enable a repository, and use the label:

yaml
jobs:
  build:
    runs-on: pitlane

Docker Compose and Proxmox installs are in Getting started.

Faster than GitHub-hosted runners ​

Measured on a real-world private monorepo CI (pnpm workspace, Nuxt + Prisma, Postgres/Redis service containers, three Docker image builds), all jobs running in parallel, warm caches. Pitlane ran in an unprivileged Proxmox LXC container with 12 vCPU and a modest RAID disk.

GitHub-hosted (ubuntu-latest)Pitlane
Whole CI (longest job)~6 min 30~3 min
Job pickupqueue + VM boot6 s
Docker images & smoke test393 s127 s (3.1×)
Dependency install (setup step)30–40 s21–26 s
Prisma migrations77 s59 s
Lint & format40 s38 s
Unit tests161 s159 s

CPU-bound steps depend on your hardware: a job on a saturated host is not faster than on a 4-vCPU GitHub runner. The dashboard compares, for each of your jobs, the median duration on Pitlane with the same job on GitHub-hosted runners.

The dashboard ​

Pitlane dashboard: running jobs with their repository, workflow, branch or pull request and links to GitHub, the warm slots, and charts of jobs, slot utilization, queue wait and job durationPitlane dashboard: running jobs with their repository, workflow, branch or pull request and links to GitHub, the warm slots, and charts of jobs, slot utilization, queue wait and job duration

Screenshot with demo data.

How it works ​

GitHub Actions assigns jobs to pitlane over a long-poll; pitlane picks a warm Docker-in-Docker slot and starts a fresh runner that runs one job and is deleted; images, build cache, toolcache, package caches and the Docker Hub mirror are kept between jobsGitHub Actions assigns jobs to pitlane over a long-poll; pitlane picks a warm Docker-in-Docker slot and starts a fresh runner that runs one job and is deleted; images, build cache, toolcache, package caches and the Docker Hub mirror are kept between jobs

A slot is a long-lived Docker-in-Docker daemon. When GitHub assigns a job to your pitlane label, Pitlane picks a free slot (preferring one that already ran this repository, so its cache is hottest), starts a fresh runner container in the slot's network namespace, and lets it run exactly one job. Afterwards, the job's containers, networks, volumes and workspace are wiped, while images and build cache stay. Read more in How it works.

Limits ​

Pitlane is young, and has limits worth knowing before you install it:

  • Linux runners only (amd64 and arm64), on GitHub.com.
  • Each slot is a privileged Docker-in-Docker container: a job can take over the host. Run code you trust, on a host dedicated to CI (Security).
  • Public repositories run in strict mode (throwaway caches, wiped slot), which keeps jobs apart but is not a sandbox.

Why Pitlane compares it with actions-runner-controller, runscaler, GARM, terraform-aws-github-runner and hosted runners. Issues and pull requests are welcome.

Released under the MIT License.