Skip to content

GitHub Actions runners on GCP

backmesh runs each GitHub Actions job on its own Compute Engine VM in your Google Cloud project, spot unless the label says on demand or the job reruns after a spot reclaim. The VM is deleted when the job ends, and Google bills you for it.

A job’s VM is the first type in its label’s list that has capacity:

Label vCPU Memory Types, in order
backmesh-2 2 8 GiB n2-standard-2, n2d-standard-2
backmesh-4 4 16 GiB n2-standard-4, n2d-standard-4
backmesh-8 8 32 GiB n2-standard-8, n2d-standard-8
backmesh-16 16 64 GiB n2-standard-16, n2d-standard-16

The KVM labels use n2 only, since n2d doesn’t support nested virtualization. GCP has no GPU labels yet: anycloud can’t launch a T4 VM there. When GCP is the control plane’s cloud, its labels are plain, as above; otherwise they name it, such as backmesh-gcp-4.

Runners use a fixed list of US regions, not discovered through your cloud CLI or chosen on the command line: every one whose catalog offers every type above, which leaves out us-central2 (almost no general-purpose types), narrowed to the ones your account allows. Each job tries the cheapest region first, then another when Google refuses for capacity or quota. More.

Install works through your gcloud CLI, signed in with a project. It turns on the Compute, IAM, IAM Credentials and Security Token Service APIs, and creates:

  • A role connection: a custom role, a service account, and a workload identity pool and provider with their IAM bindings. They trust only your control plane, so no service account keys are stored.
  • Cache buckets, through that connection: one per runner region, bm-cache-<id>-<region>, for the cache.
  • The control plane, when GCP is its cloud: an e2-standard-2 VM and an e2-small VM for the listener, with their disks and public addresses, and a checkpoint bucket for its backups.

Uninstall deletes the VMs and buckets. It asks before deleting the checkpoint bucket, and leaves the role connection’s setup for you to delete in the console.

Install can’t read GCP’s quotas, so it names the ones to raise instead. The one that bites first is CPUS_ALL_REGIONS, 32 vCPUs by default across all regions: 8 backmesh-4 jobs at once. Raise it before big bursts. Each runner region also has N2D_CPUS and N2_CPUS, and PREEMPTIBLE_CPUS for spot or CPUS for on demand. Runners past a quota wait, and backmesh status warns when they do, as in this example with illustrative numbers:

⚠ 9 runners waited for GCP quota in us-east4, up to 16.7 min.
Raise CPUS_ALL_REGIONS above its default 32 to start them right away.

These are catalog prices in us-central1, spot in the lowest zone, for a backmesh-4 VM while its job runs:

Type Spot $/h On demand $/h
n2-standard-4 0.1165 0.1942
n2d-standard-4 0.0899 0.169

Other sizes scale with their vCPUs. Each job also pays for its VM’s boot and its 100 GB disk. The control plane costs about $74 a month. At those prices, 150,000 minutes a month of backmesh-4 jobs cost at least $299–$365 with the control plane, before disks, boots and runners started ahead of jobs; see the costs. Spot prices change by the hour.