Queueing and capacity
What happens when the machine is full, the order in which jobs start and how quickly a job is picked up.
When the machine is full
A job that cannot start yet stays queued in GitHub. It is not failed and it is not sent anywhere else. It starts when a running job finishes and frees enough of the machine for its size.
If jobs wait too long, use smaller sizes for jobs that do not need the memory, or a larger machine. Usage shows how to read the machine’s history before you decide, and Machines lists the classes.
Waiting adds nothing to the price. There is no spending limit to raise, because nothing is charged by the minute.
GitHub cancels a job that has been queued for more than 24 hours. That limit is GitHub’s.
The order in which jobs start
Waiting jobs are considered oldest first, across all pools on the machine. A job starts when all of these hold:
- its pool is below its runner limit;
- its size fits the CPU, memory and disk the machine has left;
- its pool is not in the 60-second pause that follows a shortage reported by the machine.
A waiting job that does not fit reserves the capacity it needs. Smaller, younger jobs cannot keep starting ahead of it and starve a larger size.
Why a job is waiting
Battleship records one reason for each waiting job. In GitHub the job shows as queued. Whoever operates the machine sees these reasons on its Action Runners page. In your own organisation’s dashboard, Usage shows the work placed on your machines and the refusals that named them.
| Reason | Meaning |
|---|---|
capacity | Running or older waiting jobs hold the machine’s budget, or the machine reported a shortage and is retried after 60 seconds. |
runner_limit | The pool is at its runner limit. |
no_backend | No pool serves the size the job asked for. No runner is started. |
standby | The pool serves jobs through pre-connected runners; the job counts toward how many the pool keeps ready. |
How quickly a job starts
By default every job gets a new virtual machine that is cloned, booted and connected after the job is queued. In live runs on our own pool, that took 33–37 seconds from queued to started.
A pool can keep one to four pre-connected standby runners. A job that found an idle one started 2–3 seconds after it was queued in our startup benchmark, which ran jobs one after another against a single standby runner. Jobs beyond the idle standby runners still wait for a new virtual machine.
These are measurements of one setup, not a promised start time. Each standby runner is a running virtual machine that holds its memory while it waits, so it takes a slot from the machine’s budget. A standby pool requires a runner group restricted to exactly one private repository.
Cancellation, timeouts and missed events
Cancelling a workflow in GitHub cancels the job on its runner, as on any runner. Each pool also has a time limit, set for the pool between 1 minute and 6 hours. A job that passes it is asked to stop and is ended 30 seconds later. Pools of the battleship-slim size are limited to 15 minutes.
Battleship learns about new jobs from GitHub’s workflow_job events.
A job is never replayed after an uncertain result. If Battleship cannot prove what happened to a runner, it keeps that slot for inspection instead of starting the job again.
The example day on the home page
The home page plays one working day twice: on a metered runner with a daily budget, and on one Frigate. It is an example with fixed inputs, not a measurement. These are the inputs:
| Input | Value |
|---|---|
| Work | 46 pushes between 08:00 and 18:00, most of them in the afternoon. Each push starts 3 jobs, so 138 jobs in the day. |
| Length of a job | 8 minutes, every job |
| Battleship | One Frigate running 4 jobs at once. A free slot picks up a waiting job in 2.5 seconds, the middle of the 2 to 3 seconds we measured with a standby runner. $144 a month. |
| Metered runner | $0.042 per job minute, the GitHub-hosted 16 vCPU list price read on 17 Nov 2026. No limit on jobs at once. |
| Budget | $15 a day. When it is reached nothing starts. We assume it takes 25 minutes for someone to notice and raise it by $15. |
| Month | 22 working days, each like this one |
Arrival times come from a fixed seed, so the day plays the same for every reader. The same day, every working day: about $1,020 a month on a meter, $144 on Battleship.
The example assumes every working day is this busy. A quieter day costs less on a meter, and a team that runs few jobs pays less on a metered runner than for a machine. Benchmarks gives the list prices behind the comparison.