Runpool
Runpool is a Docker-native control plane for autoscaling ephemeral CI
runners on a single, capacity-bounded host. It translates provider demand
into durable assignments, isolated per-job execution capsules, and optional
repository-scoped cache lanes — without requiring Kubernetes.
Architecture
flowchart LR
provider["GitHub Actions scale sets"] -->|"durable deliveries"| controller["Runpool controller"]
controller -->|"Moby API"| host["Docker Engine 28+"]
host --> capsule["Per-job capsule"]
host --> gateway["Per-job egress gateway"]
capsule -->|"DNS and HTTP proxy"| gateway
capsule --> runtime["Runner + dockerd + job containers"]
state[("SQLite state")] --- controller
cache[("Repository cache lane")] -. "optional" .- capsule
The controller has no inbound service dependency. It maintains outbound
provider sessions, reconciles durable state with Docker, and admits work only
while scheduling and disk budgets are healthy. See the
architecture guide for package boundaries, identity,
recovery, and capacity semantics.
Why Runpool
- Clean execution: every workload gets a fresh runner, workspace, Docker
daemon data root, and control filesystem.
- Bounded capacity: runner, inner daemon, child containers, and egress
gateway share an aggregate cgroup budget.
- Durable delivery: assignments are persisted before broker
acknowledgement; redelivery is idempotent and ambiguous execution is held
for operator review.
- Restricted egress: the default profile provides kernel-enforced
no-route isolation plus a policy-enforcing DNS and HTTP relay.
- Safe ownership: cleanup acts only on resources carrying the expected
instance labels; Runpool never performs daemon-wide pruning.
- Provider-neutral lifecycle core: delivery, attempt, lease, cache, and
cleanup state use opaque provider keys. GitHub-specific configuration and
metadata remain at the composition and adapter boundary; GitHub Actions is
the only implemented provider.
Security boundary
Runpool is a resource-management, environment-hygiene, and network-policy
boundary for CI workloads approved by the operator.
[!WARNING]
Runpool is not a hostile code sandbox. It bounds and cleans up what an
approved workload does; it does not contain code you would not have run.
Where that boundary ends:
- the controller holds the Docker socket, which is host-root authority;
- capsules run a privileged inner Docker daemon;
- GitHub's required
--jitconfig interface makes the one-run JIT bundle
observable to the assigned workload, although it is kept out of controller
persistence, Docker metadata, logs, and later capsules;
shared-daemon deliberately shares the Engine's compromise domain with the
platform and its services; dedicated-daemon reduces that blast radius but
is not a VM boundary;- public fork pull requests are outside the supported model.
Under public-internet-only, direct egress is denied. Proxy-aware HTTP
clients can use absolute-form requests or CONNECT to allowed addresses on
ports 80 and 443. CONNECT is an opaque tunnel; Runpool applies destination
and port policy, not application-layer inspection. git+ssh, direct TCP,
non-DNS UDP, IPv6, and other ports fail closed. Read the complete
threat model before deployment.
Evaluate locally
Prerequisites are Go 1.26.0 or newer and a Linux host with rootful Docker
Engine 28.
The selected release-qualification target is documented in the
support matrix.
git clone https://github.com/rhobuild/runpool.git
cd runpool
go build -trimpath -o runpool ./cmd/runpool
./runpool config validate --file internal/config/testdata/example.yaml
go test ./...
That builds the controller and checks a configuration. Reaching a job
takes three more things: the two images, a target on GitHub, and a
workflow that names the tier.
docker build -f build/capsule/Dockerfile -t runpool-capsule:dev .
docker build -f build/controller/Dockerfile -t runpool:dev .
[!NOTE]
runpool-capsule:dev is the exact name a development build looks for. A
release binary carries its capsule by digest and refuses to be pointed
elsewhere; from source, this tag is the pairing.
Then a configuration, a credential, and the state directory:
RUNPOOL_GITHUB_URL=https://github.com/<owner>/<repo> \
RUNPOOL_GITHUB_TOKEN_FILE=/run/secrets/runpool/token \
RUNPOOL_HOST_TOPOLOGY=dedicated-daemon \
RUNPOOL_STATE_DIR="$PWD/state" \
./runpool doctor
RUNPOOL_STATE_DIR is $PWD/state here rather than the default
/var/lib/runpool/state, which the deployed controller uses from inside
its container and which a shell on the host cannot create without root.
doctor proves the host and the credential, and prints the runs-on
label for every tier it would serve. Then ./runpool serve, and a
workflow in that repository:
jobs:
build:
runs-on: runpool-standard
The complete example workflow shows what
a job gets: its own Docker daemon, and egress through the relay.
Deployment covers the GitHub side — which target
shapes exist, what a runner group has to grant, and which credential to
use — and the runbook covers operating it.
The reference Compose deployment asks for a released digest and refuses to
start with no image at all; that it is a digest rather than a moving tag is
yours to hold, and deployment covers verifying one. The
images above are how a source checkout runs.
[!CAUTION]
Do not use a production credential for local experimentation. A token that
can reach real repositories reaches them from a host you are still testing.
Deployment
Runpool provides a canonical Docker Compose manifest for its headless
controller: no domain, reverse-proxy route, or public port is required.
[!IMPORTANT]
A version is release-qualified only after a no-skip qualification run on the
exact platform frozen before its candidate in
build/platform.lock.json. Each release carries
the record that run produced, bound to its commit and to the image digests it
publishes; release readiness lists the gates and
the evidence each one leaves.
| Platform |
Entry point |
| Docker Compose |
deploy/compose/compose.yaml |
| A Compose platform |
Deploy the canonical Compose file; a platform catalog entry, where the platform has one, derives its template from that same contract |
Read the deployment guide before installing. The
Compose contract reads an operator-managed
configuration file; the example selects
shared-daemon for a host shared with a platform, and illustrates an explicit reserve for
colocated services. Measure that reserve before deployment. Choose
dedicated-daemon only on an exclusive CI host.
Project documentation
| Guide |
Purpose |
| Deployment |
Compose, platforms, credentials, and lifecycle |
| Product contract |
Guarantees, exclusions, and compatibility surfaces |
| Architecture |
Components, dependency direction, state, and recovery |
| Configuration |
Supported settings and validation rules |
| CLI reference |
Generated commands, flags, and exit-code behaviour |
| Support matrix |
Engine compatibility, release reference, and provider scope |
| Runbook |
Operations, recovery, backup, GC, and uninstall |
| Threat model |
Trust boundary, defences, one known weakness, and accepted exposure |
| ADRs |
Architectural decisions and measured constraints |
| Release readiness |
The objective gates a release must satisfy |
Contributing and security
Read CONTRIBUTING.md before opening a pull request. Report
security issues privately as described in SECURITY.md; never
open a public issue for a suspected vulnerability.
Runpool is maintained by Rhobuild and licensed under
the Apache License 2.0.
Плоскость управления в духе Docker, которая автоматически масштабирует временные CI-раннеры на одном хосте с ограниченной ёмкостью и без Kubernetes: потребность провайдера превращается в устойчивые назначения задач и изолированные капсулы выполнения для каждой задачи.