Method InstallBackend

core/services/nodes/unloader.go:120–153

InstallBackend sends a backend.install request-reply to a worker node. Idempotent on the worker: if the (modelID, replica) process is already running, the worker short-circuits and returns its address; if the binary is on disk, the worker just spawns a process; only a missing binary triggers a full

(
	nodeID, backendType, modelID, galleriesJSON, uri, name, alias string,
	replicaIndex int,
	opID string,
	onProgress func(messaging.BackendInstallProgressEvent),
)

// it lives on a different NATS subject so it cannot head-of-line-block
// routine load traffic on the same worker.
func (a *RemoteUnloaderAdapter) InstallBackend(
	nodeID, backendType, modelID, galleriesJSON, uri, name, alias string,
	replicaIndex int,
	opID string,
	onProgress func(messaging.BackendInstallProgressEvent),
) (*messaging.BackendInstallReply, error) {
	subject := messaging.SubjectNodeBackendInstall(nodeID)
	xlog.Info("Sending NATS backend.install", "nodeID", nodeID, "backend", backendType, "modelID", modelID, "replica", replicaIndex, "opID", opID)

	// Subscribe to the per-op progress subject BEFORE publishing the install
	// request so we don't miss early events.
	sub := a.subscribeProgress(nodeID, opID, onProgress)

	reply, err := messaging.RequestJSON[messaging.BackendInstallRequest, messaging.BackendInstallReply](a.nats, subject, messaging.BackendInstallRequest{
		Backend:          backendType,
		ModelID:          modelID,
		BackendGalleries: galleriesJSON,
		URI:              uri,
		Name:             name,
		Alias:            alias,
		ReplicaIndex:     int32(replicaIndex),
		OpID:             opID,
	}, a.installTimeout)

	if sub != nil {
		_ = sub.Unsubscribe()
	}

	if err != nil && isNATSTimeout(err) {
		return nil, fmt.Errorf("%w (subject=%s nodeID=%s backend=%s): %v",
			galleryop.ErrWorkerStillInstalling, subject, nodeID, backendType, err)
	}
	return reply, err
}

// subscribeProgress subscribes to the per-op backend-install progress subject
// so the master can stream per-node download ticks while a worker installs or