Request: auto-connect snapd-control and system-files plugs for helio

Snap: helio (revision 1, pending manual review) Source: https://github.com/canonical/helio (snap/snapcraft.yaml)

What it is

Helio is a proof of concept of how Sunbeam could be simplified to have less surface of attack and potential for bugs. Helio deploys and operates OpenStack 2026.1 from one CLI, without Juju, Kubernetes or Terraform. The helio snap contains the CLI and helio-node, a root daemon on each cloud node that applies the desired state. It plays the role Juju’s machine agent plays for Sunbeam.

Helio’s other snaps (helio-services, helio-mysql, helio-rabbitmq, helio-ingress) are strictly confined, and so is helio itself.

snapd-control (installation, and auto-connection if possible)

helio-node’s job is managing snaps on its node. It installs, refreshes, holds and connects Helio’s service snaps and the upstream data-plane snaps (MicroOVN, MicroCeph, openstack-hypervisor). It also delivers configuration with snap set and starts oneshot task apps. All of this goes through the snapd REST API.

  • helio-node has no generic “run a command” or “install any snap” operation. It applies a node document from Helio’s control plane, validated against a fixed schema, and performs a fixed set of operations.
  • It listens only on a Unix socket in $SNAP_COMMON and checks peer credentials (root, or the dedicated helio system user that SSH pins to a forced command). It opens no network port.
  • Every change goes to an audit log.

Without auto-connection the administrator runs snap connect helio:snapd-control once per machine; the helio prepare script does this today. helio-node connects its other plugs itself through snapd-control.

system-files: one plug per purpose, narrowest paths I could find

microovn-control

  • Write: /var/snap/microovn/common/state/control.socket
  • Read: /var/snap/microovn/common/state/database

Bootstrap and join MicroOVN, issue join tokens, enable/disable its services, remove members. helio-node calls MicroOVN’s microcluster REST API, the same calls the microovn CLI makes; a strict snap can’t run another snap’s CLI. Reading state/database (existence only) stops us bootstrapping over a member whose daemon is down, which would fork the cluster.

microceph-control

  • Write: /var/snap/microceph/common/state/control.socket
  • Read: /var/snap/microceph/common/state/database

The same for MicroCeph: bootstrap, join, tokens, disk enrollment, pool replica counts, member removal.

host-trust

  • Write: /var/lib/snapd/hostfs/usr/local/share/ca-certificates, /etc/ssl/certs/ca-certificates.crt

openstack-hypervisor verifies RabbitMQ’s TLS certificate with the host trust store, so the cloud’s CA has to be in it. We write the CA file and append it to the bundle in a marked block (a later update-ca-certificates keeps it).

iscsi-initiator-name

  • Write: /etc/iscsi

Nodes attaching iSCSI volumes from storage appliances need an initiator name. open-iscsi only generates one when iscsid first starts, so helio-node writes one if none exists and never changes an existing one.

host-config

  • Read only: /etc/ssh/sshd_config, /etc/ssh/sshd_config.d, /etc/ssh/helio, /etc/ssh/ssh_host_ecdsa_key.pub, /etc/ufw/ufw.conf, /etc/modprobe.d, the host’s os-release and systemd unit directory under hostfs

Preflight checks before deployment: sshd AllowUsers/AllowGroups on hardened hosts, ufw state, a blacklisted squashfs module, the host OS, whether iscsid/multipathd are installed. The host key (public half only) is pinned so other control nodes can verify this one over SSH.

On reaching into MicroOVN and MicroCeph

I know microovn-control and microceph-control reach into other snaps I do not own or control. The cleaner design would be a slot from MicroOVN and MicroCeph that exposes their control socket; neither has one today, and since this is just a proof of concept, I don’t have any plan to ask for plugs in these snaps. I just want to make this proof of concept easy to install and test by uploading the snaps to snapstore.

Happy to answer questions or narrow anything further.

This request has been added to the queue for review by the @reviewers team.

Hi @Thinking-Dragon

I don’t feel particularity comfortable granting use these super privileged interfaces, specially snapd-control, in the global store just for a proof of concept. I believe distributing it via alternative methods, such as github releases and installing it with --dangerous would be more appropriate until it is agreed this is right direction. Let me know if this may work for you for the time being.

Thanks