How to Use Ephemeral Runners to Control Cost

An ephemeral runner exits after one job, so it never lingers idle and never carries state that inflates the next run.

Pass --ephemeral at registration. Paired with an autoscaler, this gives clean, cost-predictable capacity that appears for a job and disappears after.

Steps

• Register the runner with --ephemeral.
• Let it pick up exactly one job, then exit.
• Have the autoscaler replace it only when the queue needs it.

Register an ephemeral runner

./config.sh --url https://github.com/my-org \
  --token "$RUNNER_TOKEN" \
  --ephemeral \
  --name "runner-$(hostname)" \
  --labels linux,ephemeral
./run.sh   # process exits cleanly after one job

Gotchas

• Ephemeral runners lose their cache between jobs; use a shared cache backend to keep hit rates up.
• Each cold start reinstalls tools, so bake a custom image to avoid paying setup time every job.
• Non-ephemeral runners can leak state and secrets across jobs, a security cost as well as a compute one.

Related guides