fix(backend): improve API version detection with retry logic for health probes

Signed-off-by: Noste <83548733+Noooste@users.noreply.github.com>

backend/internal/services/admin_factory.go

@@ -39,8 +39,14 @@ type AdminServiceResult struct {
 	APIVersion   string
 }
 
+// errProbeNotFound means the probed route returned 404. Garage v2.x also serves
+// /v1/health, so a 404 on /v2 is the only reliable "this is a v1 server" signal.
+var errProbeNotFound = errors.New("probe endpoint not found")
+
 func NewAdminService(cfg *config.GarageConfig, logLevel string) (*AdminServiceResult, error) {
-	if err := probeEndpoint(cfg, "/v2/GetClusterHealth"); err == nil {
+	// retry so a startup fails doesn't lock us to a v1 client.
+	err := probeEndpointWithRetry(cfg, "/v2/GetClusterHealth")
+	if err == nil {
 		logger.Info().Str("api_version", "v2").Msg("Detected Garage admin API v2")
 		svc := NewGarageV2AdminService(cfg, logLevel)
 		return &AdminServiceResult{
@@ -50,7 +56,17 @@ func NewAdminService(cfg *config.GarageConfig, logLevel string) (*AdminServiceRe
 		}, nil
 	}
 
-	if err := probeEndpoint(cfg, "/v1/health"); err == nil {
+	// only fall back to v1 on a real 404
+	// other errors mean the server is up but the probe failed transiently; picking v1 against
+	// a v2.x server breaks /v1/status with "v1/ endpoint is no longer supported" (issue #78).
+	if !errors.Is(err, errProbeNotFound) {
+		return nil, fmt.Errorf(
+			"could not detect Garage admin API version at %s: %w. Ensure Garage v1.1+ is running and the admin API is reachable",
+			cfg.AdminEndpoint, err,
+		)
+	}
+
+	if err := probeEndpointWithRetry(cfg, "/v1/health"); err == nil {
 		logger.Info().
 			Str("api_version", "v1").
 			Msg("Detected Garage admin API v1 — cluster statistics and per-node details will be unavailable")
@@ -68,6 +84,26 @@ func NewAdminService(cfg *config.GarageConfig, logLevel string) (*AdminServiceRe
 	)
 }
 
+const probeAttempts = 4
+
+// probeEndpointWithRetry retries transient failures with backoff, but returns a
+// 404 immediately, a missing route won't appear on a retry.
+func probeEndpointWithRetry(cfg *config.GarageConfig, path string) error {
+	var err error
+	backoff := 250 * time.Millisecond
+	for attempt := range probeAttempts {
+		err = probeEndpoint(cfg, path)
+		if err == nil || errors.Is(err, errProbeNotFound) {
+			return err
+		}
+		if attempt < probeAttempts-1 {
+			time.Sleep(backoff)
+			backoff *= 2
+		}
+	}
+	return err
+}
+
 func probeEndpoint(cfg *config.GarageConfig, path string) error {
 	session := azuretls.NewSession()
 	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
@@ -86,6 +122,9 @@ func probeEndpoint(cfg *config.GarageConfig, path string) error {
 	}
 	defer resp.RawBody.Close()
 
+	if resp.StatusCode == http.StatusNotFound {
+		return fmt.Errorf("probe %s returned status 404: %w", path, errProbeNotFound)
+	}
 	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
 		return fmt.Errorf("probe %s returned status %d", path, resp.StatusCode)
 	}

backend/internal/services/admin_factory_test.go

@@ -89,3 +89,61 @@ func TestDetectVersion_Unreachable(t *testing.T) {
 		t.Fatal("expected error for unreachable server")
 	}
 }
+
+// Garage v2.x serves /v1/health too, so a transient failure of the /v2 probe
+// must not cause a permanent downgrade to the (broken on v2.x) v1 client.
+// Regression test for https://github.com/Noooste/garage-ui/issues/78
+func TestDetectVersion_V2_TransientProbeFailure(t *testing.T) {
+	var v2Hits int
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/v2/GetClusterHealth":
+			v2Hits++
+			if v2Hits < 3 { // fail the first two attempts, then recover
+				w.WriteHeader(http.StatusServiceUnavailable)
+				return
+			}
+			w.Header().Set("Content-Type", "application/json")
+			w.Write([]byte(`{"status":"healthy"}`))
+		case "/v1/health": // v2.x still answers this
+			w.Header().Set("Content-Type", "application/json")
+			w.Write([]byte(`{"status":"healthy"}`))
+		default:
+			w.WriteHeader(http.StatusNotFound)
+		}
+	}))
+	t.Cleanup(srv.Close)
+
+	cfg := &config.GarageConfig{AdminEndpoint: srv.URL, AdminToken: "tok"}
+	result, err := NewAdminService(cfg, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.APIVersion != "v2" {
+		t.Fatalf("expected v2 after transient probe failure, got %s", result.APIVersion)
+	}
+}
+
+// A server that answers /v1/health but returns a server error (not 404) for
+// /v2/GetClusterHealth must NOT be detected as v1, because v2.x servers also
+// answer /v1/health. Falling through to v1 here is the issue #78 misdetection.
+func TestDetectVersion_DoesNotDowngradeOnV2ServerError(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/v2/GetClusterHealth":
+			w.WriteHeader(http.StatusServiceUnavailable)
+		case "/v1/health":
+			w.Header().Set("Content-Type", "application/json")
+			w.Write([]byte(`{"status":"healthy"}`))
+		default:
+			w.WriteHeader(http.StatusNotFound)
+		}
+	}))
+	t.Cleanup(srv.Close)
+
+	cfg := &config.GarageConfig{AdminEndpoint: srv.URL, AdminToken: "tok"}
+	result, err := NewAdminService(cfg, "")
+	if err == nil && result.APIVersion == "v1" {
+		t.Fatal("must not downgrade to v1 when /v2 returns a server error; v2.x also serves /v1/health")
+	}
+}