Module compute 

Compute-backend assembly (moved from the binary — node-library N2).

Builds this node’s compute BackendRegistry and scheduler Node inventory from the optional [compute] config, capability-detecting Docker, native-container, and embedded-VMM backends. Lives here (not in the backend-agnostic boatramp-server) because it depends on the concrete backend crates; the binary and future assemble() call build_compute.

Structs

DeployDnsSource

The internal-DNS control-plane source (Linux): snapshots the co-located fleet from the DeployStore replica state into the two maps the resolver needs — IP → (project, workload) (the source-IP reverse map, from every replica so an unknown source is recognised as such) and (project, workload) → healthy replica IPs (the name → IP forward map, mirroring DeployEndpointResolver’s healthy-running filter). A workload resolves ONLY to its own project’s healthy replicas; the isolation scoping itself lives in the pure resolver, which keys its answer by the source IP’s project.

NodeComputeControl

The node’s ComputeControl: restart a replica (stop it + drop its observed state so the reconcile loop relaunches it). Backs POST /api/compute/maintenance/restart (admin-scoped). Resolves the target replica’s backend from its persisted state and calls ComputeBackend::stop; the server nudges the reconcile loop afterwards so relaunch is prompt.

NodeComputeExec

The node’s ComputeExec: resolve a workload’s running replica from the control-plane state, pick its backend, and run the command inside it. Backs POST /api/compute/{name}/exec; the API gates it behind the allow_compute_exec posture. Only the shared-kernel backends (native container, remote docker) actually implement ComputeBackend::exec; the rest surface as ExecError::Unsupported.

NodeComputeVolumes

The node’s ComputeVolumes: list + reclaim persistent volumes. Backs GET /api/compute/volumes + DELETE /api/compute/volumes/{name} (admin-scoped). Lists every volume-capable backend’s on-node volumes, flags which are still referenced by a registered workload’s active spec (in use vs orphaned), and refuses to remove an in-use volume unless forced — so compute rm <workload> (which unregisters it, then the reconcile loop stops the replica) is the safe precondition for reclaiming its volume.

Functions

adopt_running_replica_ips

Adopt the IPs of already-running replicas into each compute backend’s fresh-on-boot IP pool, so the backend reserves live addresses before the reconcile loop allocates any new one. Reads every persisted replica across all projects, parses each endpoint’s IPv4 address (from the routable endpoint host, falling back to the <ip>:<port> backend_ref), groups them by backend, and hands each backend the (workload, replica, ip) tuples for its own replicas.

build_compute

Build this node’s compute [BackendRegistry] + scheduler [Node] inventory from the optional [compute] config. Backends are capability-detected: a reachable Docker daemon ⇒ docker; Linux ⇒ the native container backend; Linux + /dev/kvm ⇒ the in-process vmm-embedded microVM backend (strongest isolation). Absent config ⇒ an empty registry + a node advertising nothing, so the reconcile loop stays a no-op.

spawn_internal_dns

Start the per-project internal DNS resolver task when it is enabled and the container backend + bridge are active. Binds UDP gateway:53 (the bridge gateway the container backend assigns), forwarding non-internal queries to compute.dns_upstream. Returns the detached task handle, or None when the resolver is off, the container backend isn’t present (nothing to serve names for), or the subnet/upstream is malformed. Linux-only: it binds a socket and is meaningful only where the container/bridge code runs; a non-Linux node returns None (the seam is compiled out).