feat: support smooth upgrade with checking csi sock (#10)

When the model CSI driver is deployed as a DaemonSet pod, newly created
pod may delete the sock file that the old pod was listening on. If the
new pod hasn't started properly yet, this causes a period of service
unavailability.

This change adds a check to verify if the sock file exists and is
listening, if it is, the new pod will exit with an error.

Signed-off-by: imeoer <[email protected]>

Author: imeoer

pkg/server/server.go

@@ -56,6 +56,20 @@ func ensureSockNotExists(ctx context.Context, sockPath string) error {
 	return nil
 }
 
+func isSockListening(sockPath string) bool {
+	if _, err := os.Stat(sockPath); err != nil {
+		return false
+	}
+
+	conn, err := net.DialTimeout("unix", sockPath, time.Second*5)
+	if err != nil {
+		return false
+	}
+	defer func() { _ = conn.Close() }()
+
+	return true
+}
+
 type Server struct {
 	cfg *config.Config
 	svc *service.Service
@@ -110,6 +124,15 @@ func (server *Server) Run(ctx context.Context) error {
 		}
 	}
 
+	if server.cfg.Get().IsNodeMode() {
+		if endpoint, err := url.Parse(server.cfg.Get().CSIEndpoint); err == nil {
+			if endpoint.Path != "" && isSockListening(endpoint.Path) {
+				err := errors.Errorf("CSI socket %s is already listening", endpoint.Path)
+				return err
+			}
+		}
+	}
+
 	if server.cfg.Get().PprofAddr != "" {
 		eg.Go(withFatalError(func() error {
 			endpoint, err := url.Parse(server.cfg.Get().PprofAddr)