Method UpgradeBackend

core/services/nodes/unloader.go:202–227

UpgradeBackend sends a backend.upgrade request-reply to a worker node. The worker stops every live process for this backend, force-reinstalls from the gallery (overwriting the on-disk artifact), and replies. The next routine InstallBackend call spawns a fresh process with the new binary - upgrade it

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

// (default 15m). Real-world worst case observed: 8-10 minutes for large
// CUDA-l4t backend images on Jetson over WiFi.
func (a *RemoteUnloaderAdapter) UpgradeBackend(nodeID, backendType, galleriesJSON, uri, name, alias string, replicaIndex int, opID string, onProgress func(messaging.BackendInstallProgressEvent)) (*messaging.BackendUpgradeReply, error) {
	subject := messaging.SubjectNodeBackendUpgrade(nodeID)
	xlog.Info("Sending NATS backend.upgrade", "nodeID", nodeID, "backend", backendType, "replica", replicaIndex, "opID", opID)

	sub := a.subscribeProgress(nodeID, opID, onProgress)

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

	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
}

// installWithForceFallback is the rolling-update fallback used by
// DistributedBackendManager.UpgradeBackend when backend.upgrade returns