quickstart is the Ansible playbook that stands up an Eru cluster on Ubuntu hosts: an etcd cluster
for metadata, eru-core as the scheduler, its resource plugins, and any number of nodes registered
with core and running eru-agent. One inventory file describes the cluster, one command deploys it,
and quickstart.sh collapses the whole thing onto a single machine.
Documentation: projecteru2.github.io/quickstart (source in docs/).
One node, on a fresh Ubuntu host, as root:
curl -fsSL https://raw.githubusercontent.com/projecteru2/quickstart/master/quickstart.sh | bashSeveral nodes, from any machine with ansible-core installed:
git clone https://github.com/projecteru2/quickstart.git &&cd quickstart
cp inventory.yml.example inventory.yml
$EDITOR inventory.yml # your hosts, and which of them run etcd, core and workloads
make upThe control machine needs root SSH access to every host in the inventory — keys authorized and
host keys accepted — before make up; that is plain Ansible, nothing Eru-specific.
eru-cli lands on the core hosts, already pointed at core through ERU. A deploy needs a spec
file: it carries the app name and the entrypoints, while the flags carry the placement and the
resources.
cat > /tmp/nginx.yaml <<'EOF'appname: nginxentrypoints: web: commands: - nginx - -g - "daemon off;" publish: - "80"EOF
eru-cli workload deploy \
--pod eru --entry web --image nginx:alpine \
--network eru --count 1 --cpu 1 --memory 256M --storage 1G \
/tmp/nginx.yaml--pod is the node group to deploy on, --entry names an entrypoint from the spec, and --network
is the CNI network the playbook configured on every containerd node — pass host instead to share
the node's network namespace. --storage is accounted by the resource-storage plugin, which the
playbook installs alongside core.
eru-cli pod nodes eru shows what the cluster holds, and eru-cli workload list nginx what is
running. verify.sh runs the whole loop — cache the image, deploy, get, exec,
logs, stop, start, remove — against a freshly deployed cluster:
./verify.sh # or ./verify.sh <nodename>| Group | Host gets |
|---|---|
etcd | etcd, from the upstream release tarball, as a systemd unit |
core | eru-core, the resource-storage and resource-gpu plugins, and eru-cli |
node_containerd | containerd, runc, the CNI plugins, eru-agent, and a node registration |
node_process | oras and eru-agent; workloads run as transient systemd units |
node_cocoon | eru-agent and a node registration over an existing cocoon install — cocoon itself, per its own docs, and /dev/kvm come first |
Core reaches every node over SSH with a key pair the playbook generates for it, so nodes expose no
daemon API of their own. A host may belong to etcd, core and one node group at once — the
single-node bootstrap does exactly that — but not to both node groups, because one host is one Eru
node with one engine.
- core — stateless gRPC resource scheduler
- agent — per-node daemon reporting node and workload status
- cli — command line client for the core API
- resource-extend — external resource plugins (gpu, storage)
- footstone — container images the project builds with
make check # ansible-playbook --syntax-check against $(INVENTORY)
make lint # ansible-lint over every playbook, role and template
make up # run the playbook; TAGS=etcd,core limits it to those roles
make all # check, then lintmake help lists every target. Both checks run in CI on every push and pull request.
This project is licensed under the MIT License. See LICENSE.