pagurus — one copy of a service, failing over across an Incus cluster

A hermit crab lives in one shell at a time and moves when it has to. pagurus keeps a container workload (an incus-compose stack, or plain Incus containers) running on at most one member of an Incus cluster at any instant, and on some member whenever a quorum is up. When the member running it dies, hangs or is cut off, the copy there is fenced before another member starts its own.

It is for clusters without shared storage, where Incus' own cluster healing cannot move an instance off a dead member. Each member keeps its own copy of the stack in its own project; pagurus decides which one may run.

                 etcd (3 nodes)      the office: a conch election on a lease
                 ▲        ▲        ▲
   ┌─────────────┴─┐ ┌────┴──────┐ ┌┴──────────────┐
   │ pagurus@node1 │ │ @node2    │ │ @node3        │   one agent per member, under systemd's watchdog
   │  leads: runs  │ │ standby   │ │ standby       │
   └──────┬────────┘ └───────────┘ └───────────────┘
          │ incus-compose up / incus stop / cgroup.kill
   project whoami-node1   whoami-node2   whoami-node3   (each pinned to its own member)

The leader holds a lease in etcd (conch runs the election). Before its lease can expire it stops its copy; if its agent is gone, peers stop that copy through the Incus API, and an optional watchdog reboots a host whose agent wedges. The safety argument, and the gap it leaves, are in docs/00_design.md.

What it assumes — read these before relying on it:

• at most one member impaired at a time, and a working etcd quorum;
• containers only for the self-fence's last resort (cgroup.kill); VMs only with the instances starter and an Incus-API fence;
• stateless workloads, or the opt-in volume replication in docs/02_replication.md, which can lose the writes since the last sync;
• a residual gap: with no reboot-armed watchdog and fence.unverified=proceed, a leader cut off from both etcd and Incus whose agent is also dead, on a host that stays up, can overlap a new leader. docs/00_design.md § The residual gap states it exactly.

Requirements

• An Incus cluster of ≥ 3 members (or 2 members plus an etcd member elsewhere as a witness), cgroup2, systemd.
• etcd v3, three nodes.
• conch on PATH (the version in go.mod), and incus-compose for the compose starter.

Try it

lab/ builds a throwaway three-member cluster on one Incus host and runs the failure drills against it, refereed by pagurus audit:

lab/up.sh && lab/drill.sh    # ~3 min up, ~4 min of drills; lab/down.sh removes it

The same lab runs in CI on every push (.github/workflows/lab.yml).

Docs

00_design.md	process shape, the safety argument, the residual gap, timing, architecture
01_operations.md	provisioning, reaching Incus and etcd, starters, updates, metrics, the CLI
02_replication.md	stateful services: replicated volumes and their limits
03_fencing.md	fence policies, VM and peer fences, watchdog modes
04_comparison.md	vs Incus cluster healing, Proxmox HA, Pacemaker, keepalived, Nomad
05_results.md	drill results from the author's cluster
06_lab.md	the lab

Status

In production on the author's three-member cluster since October 2026 (stateless and replicated services). Releases: GitHub releases (static linux amd64/arm64). Interfaces may still change before 1.0; see CHANGELOG.md.

License: see LICENSE. Security reports: SECURITY.md.