Move nonDatagramSized below the transportFallback methods

* [infrastructure] allow docker image overrides for getting started

Make dashboard and server image configurations overrideable via environment variables

* [infrastructure] update Traefik gRPC rule to include ProxyService PathPrefix

* make Traefik and CrowdSec images configurable via environment variables

.github/workflows/golang-test-darwin.yml

@@ -45,4 +45,11 @@ jobs:
         run: git --no-pager diff --exit-code
 
       - name: Test
-        run: NETBIRD_STORE_ENGINE=${{ matrix.store }} CI=true go test -tags=devcert -exec 'sudo --preserve-env=CI,NETBIRD_STORE_ENGINE' -timeout 5m -p 1 $(go list ./... | grep -v -e /management -e /signal -e /relay -e /proxy -e /combined)
+        run: NETBIRD_STORE_ENGINE=${{ matrix.store }} CI=true go test -coverprofile=coverage.txt -tags=devcert -exec 'sudo --preserve-env=CI,NETBIRD_STORE_ENGINE' -timeout 5m -p 1 $(go list ./... | grep -v -e /management -e /signal -e /relay -e /proxy -e /combined)
+
+      - name: Upload coverage reports to Codecov
+        uses: codecov/codecov-action@e79a6962e0d4c0c17b229090214935d2e33f8354 #v6.0.1
+        with:
+          token: ${{ secrets.CODECOV_TOKEN }}
+          slug: netbirdio/netbird
+          flags: unit,client

.github/workflows/golang-test-linux.yml

@@ -158,7 +158,16 @@ jobs:
         run: git --no-pager diff --exit-code
 
       - name: Test
-        run: CGO_ENABLED=1 GOARCH=${{ matrix.arch }} CI=true go test -tags devcert -exec 'sudo' -timeout 10m -p 1 $(go list ./... | grep -v -e /management -e /signal -e /relay -e /proxy -e /combined)
+        run: CGO_ENABLED=1 GOARCH=${{ matrix.arch }} CI=true go test -coverprofile=coverage.txt -tags devcert -exec 'sudo' -timeout 10m -p 1 $(go list ./... | grep -v -e /management -e /signal -e /relay -e /proxy -e /combined)
+
+      - name: Upload coverage reports to Codecov
+        if: matrix.arch == 'amd64'
+        uses: codecov/codecov-action@e79a6962e0d4c0c17b229090214935d2e33f8354 #v6.0.1
+        with:
+          token: ${{ secrets.CODECOV_TOKEN }}
+          slug: netbirdio/netbird
+          flags: unit,client
+
 
   test_client_on_docker:
     name: "Client (Docker) / Unit"
@@ -276,9 +285,17 @@ jobs:
         run: |
           CGO_ENABLED=1 GOARCH=${{ matrix.arch }} \
           go test ${{ matrix.raceFlag }} \
-            -exec 'sudo' \
+            -exec 'sudo' -coverprofile=coverage.txt \
             -timeout 10m -p 1 ./relay/... ./shared/relay/...
 
+      - name: Upload coverage reports to Codecov
+        if: matrix.arch == 'amd64'
+        uses: codecov/codecov-action@e79a6962e0d4c0c17b229090214935d2e33f8354 #v6.0.1
+        with:
+          token: ${{ secrets.CODECOV_TOKEN }}
+          slug: netbirdio/netbird
+          flags: unit,relay
+
   test_proxy:
     name: "Proxy / Unit"
     needs: [build-cache]
@@ -326,7 +343,15 @@ jobs:
       - name: Test
         run: |
           CGO_ENABLED=1 GOARCH=${{ matrix.arch }} \
-          go test -timeout 10m -p 1 ./proxy/...
+          go test -timeout 10m -p 1 -coverprofile=coverage.txt ./proxy/...
+
+      - name: Upload coverage reports to Codecov
+        if: matrix.arch == 'amd64'
+        uses: codecov/codecov-action@e79a6962e0d4c0c17b229090214935d2e33f8354 #v6.0.1
+        with:
+          token: ${{ secrets.CODECOV_TOKEN }}
+          slug: netbirdio/netbird
+          flags: unit,proxy
 
   test_signal:
     name: "Signal / Unit"
@@ -377,9 +402,17 @@ jobs:
         run: |
           CGO_ENABLED=1 GOARCH=${{ matrix.arch }} \
           go test \
-            -exec 'sudo' \
+            -exec 'sudo' -coverprofile=coverage.txt \
             -timeout 10m ./signal/... ./shared/signal/...
 
+      - name: Upload coverage reports to Codecov
+        if: matrix.arch == 'amd64'
+        uses: codecov/codecov-action@e79a6962e0d4c0c17b229090214935d2e33f8354 #v6.0.1
+        with:
+          token: ${{ secrets.CODECOV_TOKEN }}
+          slug: netbirdio/netbird
+          flags: unit,signal
+
   test_management:
     name: "Management / Unit"
     needs: [build-cache]
@@ -445,10 +478,18 @@ jobs:
           CGO_ENABLED=1 GOARCH=${{ matrix.arch }} \
           NETBIRD_STORE_ENGINE=${{ matrix.store }} \
             CI=true \
-          go test -tags=devcert \
+          go test -tags=devcert -coverprofile=coverage.txt \
             -exec "sudo --preserve-env=CI,NETBIRD_STORE_ENGINE" \
             -timeout 20m ./management/... ./shared/management/...
 
+      - name: Upload coverage reports to Codecov
+        if: matrix.arch == 'amd64'
+        uses: codecov/codecov-action@e79a6962e0d4c0c17b229090214935d2e33f8354 #v6.0.1
+        with:
+          token: ${{ secrets.CODECOV_TOKEN }}
+          slug: netbirdio/netbird
+          flags: unit,management
+
   benchmark:
     name: "Management / Benchmark"
     needs: [build-cache]
@@ -687,6 +728,14 @@ jobs:
           CGO_ENABLED=1 GOARCH=${{ matrix.arch }} \
           NETBIRD_STORE_ENGINE=${{ matrix.store }} \
           CI=true \
-          go test -tags=integration \
+          go test -tags=integration -coverprofile=coverage.txt \
           -exec 'sudo --preserve-env=CI,NETBIRD_STORE_ENGINE' \
           -timeout 20m ./management/server/http/...
+
+      - name: Upload coverage reports to Codecov
+        if: matrix.arch == 'amd64'
+        uses: codecov/codecov-action@e79a6962e0d4c0c17b229090214935d2e33f8354 #v6.0.1
+        with:
+          token: ${{ secrets.CODECOV_TOKEN }}
+          slug: netbirdio/netbird
+          flags: integration,management

.github/workflows/wasm-build-validation.yml

@@ -65,7 +65,7 @@ jobs:
 
           echo "Size: ${SIZE} bytes (${SIZE_MB} MB)"
 
-          if [ ${SIZE} -gt 58720256 ]; then
-            echo "Wasm binary size (${SIZE_MB}MB) exceeds 56MB limit!"
+          if [ ${SIZE} -gt 62914560 ]; then
+            echo "Wasm binary size (${SIZE_MB}MB) exceeds 60MB limit!"
             exit 1
           fi

client/internal/connect.go

@@ -118,7 +118,6 @@ func (c *ConnectClient) RunOniOS(
 	networkChangeListener listener.NetworkChangeListener,
 	dnsManager dns.IosDnsManager,
 	stateFilePath string,
-	cacheDir string,
 ) error {
 	// Set GC percent to 5% to reduce memory usage as iOS only allows 50MB of memory for the extension.
 	debug.SetGCPercent(5)
@@ -128,7 +127,6 @@ func (c *ConnectClient) RunOniOS(
 		NetworkChangeListener: networkChangeListener,
 		DnsManager:            dnsManager,
 		StateFilePath:         stateFilePath,
-		TempDir:               cacheDir,
 	}
 	return c.run(mobileDependency, nil, "")
 }

client/internal/debug/debug.go

@@ -250,7 +250,6 @@ type BundleGenerator struct {
 	syncResponse   *mgmProto.SyncResponse
 	logPath        string
 	tempDir        string
-	statePath      string
 	cpuProfile     []byte
 	capturePath    string
 	refreshStatus  func() // Optional callback to refresh status before bundle generation
@@ -275,7 +274,6 @@ type GeneratorDependencies struct {
 	SyncResponse   *mgmProto.SyncResponse
 	LogPath        string
 	TempDir        string // Directory for temporary bundle zip files. If empty, os.TempDir() is used.
-	StatePath      string // Path to the state file. If empty, the ServiceManager default path is used.
 	CPUProfile     []byte
 	CapturePath    string
 	RefreshStatus  func()
@@ -297,7 +295,6 @@ func NewBundleGenerator(deps GeneratorDependencies, cfg BundleConfig) *BundleGen
 		syncResponse:   deps.SyncResponse,
 		logPath:        deps.LogPath,
 		tempDir:        deps.TempDir,
-		statePath:      deps.StatePath,
 		cpuProfile:     deps.CPUProfile,
 		capturePath:    deps.CapturePath,
 		refreshStatus:  deps.RefreshStatus,
@@ -814,11 +811,8 @@ func (g *BundleGenerator) addSyncResponse() error {
 }
 
 func (g *BundleGenerator) addStateFile() error {
-	path := g.statePath
-	if path == "" {
-		sm := profilemanager.NewServiceManager("")
-		path = sm.GetStatePath()
-	}
+	sm := profilemanager.NewServiceManager("")
+	path := sm.GetStatePath()
 	if path == "" {
 		return nil
 	}

client/internal/debug/debug_ios.go

@@ -1,36 +0,0 @@
-//go:build ios
-
-package debug
-
-import (
-	"path/filepath"
-
-	log "github.com/sirupsen/logrus"
-)
-
-// swiftLogFile is the Swift app log written by the iOS app into the same log
-// directory as the Go client log, so it can be collected into the bundle.
-const swiftLogFile = "swift-log.log"
-
-// addPlatformLog collects logs for the iOS debug bundle. iOS has no logcat or
-// systemd journal, so we rely on file-based logs. addLogfile handles the Go
-// client log (logPath) with rotation, the stderr/stdout companions and
-// anonymization. The iOS app writes its own Swift log into the same directory,
-// so we add it alongside the Go log.
-func (g *BundleGenerator) addPlatformLog() error {
-	if err := g.addLogfile(); err != nil {
-		return err
-	}
-
-	if g.logPath == "" {
-		return nil
-	}
-
-	swiftLogPath := filepath.Join(filepath.Dir(g.logPath), swiftLogFile)
-	if err := g.addSingleLogfile(swiftLogPath, swiftLogFile); err != nil {
-		// The Swift log is best-effort: the app may not have written it yet.
-		log.Warnf("failed to add %s to debug bundle: %v", swiftLogFile, err)
-	}
-
-	return nil
-}

client/internal/debug/debug_nonandroid.go

@@ -1,4 +1,4 @@
-//go:build !android && !ios
+//go:build !android
 
 package debug
 

client/ios/NetBirdSDK/client.go

@@ -17,7 +17,6 @@ import (
 
 	"github.com/netbirdio/netbird/client/internal"
 	"github.com/netbirdio/netbird/client/internal/auth"
-	"github.com/netbirdio/netbird/client/internal/debug"
 	"github.com/netbirdio/netbird/client/internal/dns"
 	"github.com/netbirdio/netbird/client/internal/listener"
 	"github.com/netbirdio/netbird/client/internal/peer"
@@ -26,7 +25,6 @@ import (
 	"github.com/netbirdio/netbird/formatter"
 	"github.com/netbirdio/netbird/route"
 	"github.com/netbirdio/netbird/shared/management/domain"
-	types "github.com/netbirdio/netbird/upload-server/types"
 )
 
 // ConnectionListener export internal Listener for mobile
@@ -67,7 +65,6 @@ func init() {
 type Client struct {
 	cfgFile               string
 	stateFile             string
-	cacheDir              string
 	recorder              *peer.Status
 	ctxCancel             context.CancelFunc
 	ctxCancelLock         *sync.Mutex
@@ -78,20 +75,16 @@ type Client struct {
 	onHostDnsFn           func([]string)
 	dnsManager            dns.IosDnsManager
 	loginComplete         bool
+	connectClient         *internal.ConnectClient
 	// preloadedConfig holds config loaded from JSON (used on tvOS where file writes are blocked)
 	preloadedConfig *profilemanager.Config
-
-	stateMu       sync.RWMutex
-	connectClient *internal.ConnectClient
-	config        *profilemanager.Config
 }
 
 // NewClient instantiate a new Client
-func NewClient(cfgFile, stateFile, cacheDir, deviceName string, osVersion string, osName string, networkChangeListener NetworkChangeListener, dnsManager DnsManager) *Client {
+func NewClient(cfgFile, stateFile, deviceName string, osVersion string, osName string, networkChangeListener NetworkChangeListener, dnsManager DnsManager) *Client {
 	return &Client{
 		cfgFile:               cfgFile,
 		stateFile:             stateFile,
-		cacheDir:              cacheDir,
 		deviceName:            deviceName,
 		osName:                osName,
 		osVersion:             osVersion,
@@ -168,13 +161,8 @@ func (c *Client) Run(fd int32, interfaceName string, envList *EnvList) error {
 	c.onHostDnsFn = func([]string) {}
 	cfg.WgIface = interfaceName
 
-	connectClient := internal.NewConnectClient(ctx, cfg, c.recorder)
-	c.setState(cfg, connectClient)
-	// Persist the latest sync response so DebugBundle can include the network
-	// map. On iOS this is backed by disk to keep it out of the constrained
-	// process memory (see the syncstore package).
-	connectClient.SetSyncResponsePersistence(true)
-	return connectClient.RunOniOS(fd, c.networkChangeListener, c.dnsManager, c.stateFile, c.cacheDir)
+	c.connectClient = internal.NewConnectClient(ctx, cfg, c.recorder)
+	return c.connectClient.RunOniOS(fd, c.networkChangeListener, c.dnsManager, c.stateFile)
 }
 
 // Stop the internal client and free the resources
@@ -186,83 +174,6 @@ func (c *Client) Stop() {
 	}
 
 	c.ctxCancel()
-	c.setState(nil, nil)
-}
-
-// DebugBundle generates a debug bundle, uploads it and returns the upload key.
-// It works with or without a running engine: when the engine is up it reuses
-// the live config, sync response and client metrics; otherwise it loads the
-// config from disk (or the preloaded tvOS config).
-func (c *Client) DebugBundle(anonymize bool) (string, error) {
-	cfg, cc := c.stateSnapshot()
-
-	// If the engine hasn't been started, load config so we can reach management.
-	if cfg == nil {
-		if c.preloadedConfig != nil {
-			cfg = c.preloadedConfig
-		} else {
-			var err error
-			// Use DirectUpdateOrCreateConfig to avoid atomic file operations
-			// (temp file + rename) blocked by the tvOS sandbox.
-			cfg, err = profilemanager.DirectUpdateOrCreateConfig(profilemanager.ConfigInput{
-				ConfigPath:    c.cfgFile,
-				StateFilePath: c.stateFile,
-			})
-			if err != nil {
-				return "", fmt.Errorf("load config: %w", err)
-			}
-		}
-	}
-
-	deps := debug.GeneratorDependencies{
-		InternalConfig: cfg,
-		StatusRecorder: c.recorder,
-		TempDir:        c.cacheDir,
-		StatePath:      c.stateFile,
-	}
-
-	if cc != nil {
-		resp, err := cc.GetLatestSyncResponse()
-		if err != nil {
-			log.Warnf("get latest sync response: %v", err)
-		}
-		deps.SyncResponse = resp
-
-		if e := cc.Engine(); e != nil {
-			if cm := e.GetClientMetrics(); cm != nil {
-				deps.ClientMetrics = cm
-			}
-		}
-	}
-
-	bundleGenerator := debug.NewBundleGenerator(
-		deps,
-		debug.BundleConfig{
-			Anonymize:         anonymize,
-			IncludeSystemInfo: true,
-		},
-	)
-
-	path, err := bundleGenerator.Generate()
-	if err != nil {
-		return "", fmt.Errorf("generate debug bundle: %w", err)
-	}
-	defer func() {
-		if err := os.Remove(path); err != nil {
-			log.Errorf("failed to remove debug bundle file: %v", err)
-		}
-	}()
-
-	uploadCtx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
-	defer cancel()
-
-	key, err := debug.UploadDebugBundle(uploadCtx, types.DefaultBundleURL, cfg.ManagementURL.String(), path)
-	if err != nil {
-		return "", fmt.Errorf("upload debug bundle: %w", err)
-	}
-
-	log.Infof("debug bundle uploaded with key %s", key)
-	return key, nil
 }
 
 // SetTraceLogLevel configure the logger to trace level
@@ -443,12 +354,11 @@ func (c *Client) ClearLoginComplete() {
 }
 
 func (c *Client) GetRoutesSelectionDetails() (*RoutesSelectionDetails, error) {
-	_, connectClient := c.stateSnapshot()
-	if connectClient == nil {
+	if c.connectClient == nil {
 		return nil, fmt.Errorf("not connected")
 	}
 
-	engine := connectClient.Engine()
+	engine := c.connectClient.Engine()
 	if engine == nil {
 		return nil, fmt.Errorf("not connected")
 	}
@@ -560,12 +470,11 @@ func prepareRouteSelectionDetails(routes []*selectRoute, resolvedDomains map[dom
 }
 
 func (c *Client) SelectRoute(id string) error {
-	_, connectClient := c.stateSnapshot()
-	if connectClient == nil {
+	if c.connectClient == nil {
 		return fmt.Errorf("not connected")
 	}
 
-	engine := connectClient.Engine()
+	engine := c.connectClient.Engine()
 	if engine == nil {
 		return fmt.Errorf("not connected")
 	}
@@ -591,11 +500,10 @@ func (c *Client) SelectRoute(id string) error {
 }
 
 func (c *Client) DeselectRoute(id string) error {
-	_, connectClient := c.stateSnapshot()
-	if connectClient == nil {
+	if c.connectClient == nil {
 		return fmt.Errorf("not connected")
 	}
-	engine := connectClient.Engine()
+	engine := c.connectClient.Engine()
 	if engine == nil {
 		return fmt.Errorf("not connected")
 	}
@@ -619,22 +527,6 @@ func (c *Client) DeselectRoute(id string) error {
 	return nil
 }
 
-// setState stores the running engine state so DebugBundle can reuse the live
-// config and ConnectClient. It is cleared on Stop.
-func (c *Client) setState(cfg *profilemanager.Config, cc *internal.ConnectClient) {
-	c.stateMu.Lock()
-	defer c.stateMu.Unlock()
-	c.config = cfg
-	c.connectClient = cc
-}
-
-// stateSnapshot returns the current config and ConnectClient under the lock.
-func (c *Client) stateSnapshot() (*profilemanager.Config, *internal.ConnectClient) {
-	c.stateMu.RLock()
-	defer c.stateMu.RUnlock()
-	return c.config, c.connectClient
-}
-
 func formatDuration(d time.Duration) string {
 	ds := d.String()
 	dotIndex := strings.Index(ds, ".")

infrastructure_files/getting-started.sh

@@ -311,11 +311,12 @@ initialize_default_values() {
   NETBIRD_STUN_PORT=3478
 
   # Docker images
-  DASHBOARD_IMAGE="netbirdio/dashboard:latest"
+  DASHBOARD_IMAGE=${DASHBOARD_IMAGE:-"netbirdio/dashboard:latest"}
   # Combined server replaces separate signal, relay, and management containers
-  NETBIRD_SERVER_IMAGE="netbirdio/netbird-server:latest"
-  NETBIRD_PROXY_IMAGE="netbirdio/reverse-proxy:latest"
-
+  NETBIRD_SERVER_IMAGE=${NETBIRD_SERVER_IMAGE:-"netbirdio/netbird-server:latest"}
+  NETBIRD_PROXY_IMAGE=${NETBIRD_PROXY_IMAGE:-"netbirdio/reverse-proxy:latest"}
+  TRAEFIK_IMAGE=${TRAEFIK_IMAGE:-"traefik:v3.6"}
+  CROWDSEC_IMAGE=${CROWDSEC_IMAGE:-"crowdsecurity/crowdsec:v1.7.7"}
   # Reverse proxy configuration
   REVERSE_PROXY_TYPE="0"
   TRAEFIK_EXTERNAL_NETWORK=""
@@ -656,7 +657,7 @@ render_docker_compose_traefik_builtin() {
     if [[ "$ENABLE_CROWDSEC" == "true" ]]; then
       crowdsec_service="
   crowdsec:
-    image: crowdsecurity/crowdsec:v1.7.7
+    image: $CROWDSEC_IMAGE
     container_name: netbird-crowdsec
     restart: unless-stopped
     networks: [netbird]
@@ -687,7 +688,7 @@ render_docker_compose_traefik_builtin() {
 services:
   # Traefik reverse proxy (automatic TLS via Let's Encrypt)
   traefik:
-    image: traefik:v3.6
+    image: $TRAEFIK_IMAGE
     container_name: netbird-traefik
     restart: unless-stopped
     networks:
@@ -771,7 +772,7 @@ $traefik_dynamic_volume
     labels:
       - traefik.enable=true
       # gRPC router (needs h2c backend for HTTP/2 cleartext)
-      - traefik.http.routers.netbird-grpc.rule=Host(\`$NETBIRD_DOMAIN\`) && (PathPrefix(\`/signalexchange.SignalExchange/\`) || PathPrefix(\`/management.ManagementService/\`))
+      - traefik.http.routers.netbird-grpc.rule=Host(\`$NETBIRD_DOMAIN\`) && (PathPrefix(\`/signalexchange.SignalExchange/\`) || PathPrefix(\`/management.ManagementService/\`) || PathPrefix(\`/management.ProxyService/\`))
       - traefik.http.routers.netbird-grpc.entrypoints=websecure
       - traefik.http.routers.netbird-grpc.tls=true
       - traefik.http.routers.netbird-grpc.tls.certresolver=letsencrypt

management/internals/server/server.go

@@ -122,7 +122,7 @@ func (s *BaseServer) Start(ctx context.Context) error {
 	s.errCh = make(chan error, 4)
 
 	if s.autoResolveDomains {
-		s.resolveDomains(srvCtx)
+		s.ResolveDomains(srvCtx)
 	}
 
 	s.PeersManager()
@@ -398,10 +398,10 @@ func (s *BaseServer) serveGRPCWithHTTP(ctx context.Context, listener net.Listene
 	}()
 }
 
-// resolveDomains determines dnsDomain and mgmtSingleAccModeDomain based on store state.
+// ResolveDomains determines dnsDomain and mgmtSingleAccModeDomain based on store state.
 // Fresh installs use the default self-hosted domain, while existing installs reuse the
 // persisted account domain to keep addressing stable across config changes.
-func (s *BaseServer) resolveDomains(ctx context.Context) {
+func (s *BaseServer) ResolveDomains(ctx context.Context) {
 	st := s.Store()
 
 	setDefault := func(logMsg string, args ...any) {

management/internals/server/server_resolve_domains_test.go

@@ -22,7 +22,7 @@ func TestResolveDomains_FreshInstallUsesDefault(t *testing.T) {
 	srv := NewServer(&Config{NbConfig: &nbconfig.Config{}})
 	Inject[store.Store](srv, mockStore)
 
-	srv.resolveDomains(context.Background())
+	srv.ResolveDomains(context.Background())
 
 	require.Equal(t, DefaultSelfHostedDomain, srv.dnsDomain)
 	require.Equal(t, DefaultSelfHostedDomain, srv.mgmtSingleAccModeDomain)
@@ -40,7 +40,7 @@ func TestResolveDomains_ExistingInstallUsesPersistedDomain(t *testing.T) {
 	srv := NewServer(&Config{NbConfig: &nbconfig.Config{}})
 	Inject[store.Store](srv, mockStore)
 
-	srv.resolveDomains(context.Background())
+	srv.ResolveDomains(context.Background())
 
 	require.Equal(t, "vpn.mycompany.com", srv.dnsDomain)
 	require.Equal(t, "vpn.mycompany.com", srv.mgmtSingleAccModeDomain)
@@ -56,7 +56,7 @@ func TestResolveDomains_StoreErrorFallsBackToDefault(t *testing.T) {
 	srv := NewServer(&Config{NbConfig: &nbconfig.Config{}})
 	Inject[store.Store](srv, mockStore)
 
-	srv.resolveDomains(context.Background())
+	srv.ResolveDomains(context.Background())
 
 	require.Equal(t, DefaultSelfHostedDomain, srv.dnsDomain)
 	require.Equal(t, DefaultSelfHostedDomain, srv.mgmtSingleAccModeDomain)

shared/relay/client/client.go

@@ -9,12 +9,14 @@ import (
 	"net/url"
 	"strings"
 	"sync"
+	"sync/atomic"
 	"time"
 
 	log "github.com/sirupsen/logrus"
 
 	auth "github.com/netbirdio/netbird/shared/relay/auth/hmac"
 	"github.com/netbirdio/netbird/shared/relay/client/dialer"
+	netErr "github.com/netbirdio/netbird/shared/relay/client/dialer/net"
 	"github.com/netbirdio/netbird/shared/relay/healthcheck"
 	"github.com/netbirdio/netbird/shared/relay/messages"
 )
@@ -172,6 +174,19 @@ type Client struct {
 	stateSubscription *PeersStateSubscription
 
 	mtu uint16
+
+	// transportFallback, when set, records datagram-too-large failures so a
+	// datagram-sized transport is avoided on subsequent connects. Shared via
+	// the manager.
+	transportFallback *transportFallback
+	// datagramFallbackTriggered guards a single fallback per connection so a
+	// burst of oversized datagrams triggers one reconnect, not many.
+	datagramFallbackTriggered atomic.Bool
+}
+
+// SetTransportFallback wires the shared datagram-transport fallback tracker.
+func (c *Client) SetTransportFallback(tf *transportFallback) {
+	c.transportFallback = tf
 }
 
 // NewClient creates a new client for the relay server. The client is not connected to the server until the Connect
@@ -361,12 +376,13 @@ func (c *Client) Close() error {
 }
 
 func (c *Client) connect(ctx context.Context) (*RelayAddr, error) {
-	dialers := c.getDialers()
+	mode := transportModeFromEnv()
+	dialers := c.getDialers(mode)
 
 	var conn net.Conn
 	if c.serverIP.IsValid() {
 		var err error
-		conn, err = c.dialRaceDirect(ctx, dialers)
+		conn, err = c.dialRaceDirect(ctx, mode, dialers)
 		if err != nil {
 			c.log.Infof("dial via server IP %s failed, falling back to FQDN: %v", c.serverIP, err)
 			conn = nil
@@ -375,6 +391,9 @@ func (c *Client) connect(ctx context.Context) (*RelayAddr, error) {
 
 	if conn == nil {
 		rd := dialer.NewRaceDial(c.log, dialer.DefaultConnectionTimeout, c.connectionURL, dialers...)
+		if mode.sequential() {
+			rd.WithSequential()
+		}
 		var err error
 		conn, err = rd.Dial(ctx)
 		if err != nil {
@@ -382,6 +401,7 @@ func (c *Client) connect(ctx context.Context) (*RelayAddr, error) {
 		}
 	}
 	c.relayConn = conn
+	c.datagramFallbackTriggered.Store(false)
 
 	instanceURL, err := c.handShake(ctx)
 	if err != nil {
@@ -396,7 +416,7 @@ func (c *Client) connect(ctx context.Context) (*RelayAddr, error) {
 }
 
 // dialRaceDirect dials c.serverIP, preserving the original FQDN as the TLS ServerName for SNI.
-func (c *Client) dialRaceDirect(ctx context.Context, dialers []dialer.DialeFn) (net.Conn, error) {
+func (c *Client) dialRaceDirect(ctx context.Context, mode TransportMode, dialers []dialer.DialeFn) (net.Conn, error) {
 	directURL, serverName, err := substituteHost(c.connectionURL, c.serverIP)
 	if err != nil {
 		return nil, fmt.Errorf("substitute host: %w", err)
@@ -406,6 +426,9 @@ func (c *Client) dialRaceDirect(ctx context.Context, dialers []dialer.DialeFn) (
 
 	rd := dialer.NewRaceDial(c.log, dialer.DefaultConnectionTimeout, directURL, dialers...).
 		WithServerName(serverName)
+	if mode.sequential() {
+		rd.WithSequential()
+	}
 	return rd.Dial(ctx)
 }
 
@@ -631,13 +654,53 @@ func (c *Client) writeTo(containerRef *connContainer, dstID messages.PeerID, pay
 	}
 
 	// the write always return with 0 length because the underling does not support the size feedback.
-	_, err = c.relayConn.Write(msg)
+	conn := c.relayConn
+	_, err = conn.Write(msg)
 	if err != nil {
-		c.log.Errorf("failed to write transport message: %s", err)
+		if errors.Is(err, netErr.ErrDatagramTooLarge) {
+			c.onDatagramTooLarge(conn, err)
+		} else {
+			c.log.Errorf("failed to write transport message: %s", err)
+		}
 	}
 	return len(payload), err
 }
 
+// onDatagramTooLarge reacts to a datagram rejected as too large for the path.
+// When a non-datagram transport is available, it records a fallback for this
+// server and closes the connection so the reconnect avoids datagram-sized
+// transports. A single fallback is triggered per connection regardless of how
+// many oversized datagrams arrive. cause carries the datagram size and budget.
+func (c *Client) onDatagramTooLarge(conn net.Conn, cause error) {
+	// Handle one oversized datagram per connection; a burst triggers a single
+	// fallback (and a single log line), not many.
+	if !c.datagramFallbackTriggered.CompareAndSwap(false, true) {
+		return
+	}
+
+	// If the selected mode offers no non-datagram transport (e.g. pinned to a
+	// datagram-sized transport), reconnecting would just re-fail, so leave the
+	// connection up rather than loop.
+	if len(nonDatagramSized(c.baseDialers(transportModeFromEnv()))) == 0 {
+		c.log.Warnf("%s, but no non-datagram transport is available, not falling back", cause)
+		return
+	}
+
+	// Without the shared tracker a reconnect would just select the same
+	// transport again and re-fail, so leave the connection up rather than loop.
+	if c.transportFallback == nil {
+		c.log.Debugf("%s, but no transport fallback configured, leaving connection up", cause)
+		return
+	}
+
+	window := c.transportFallback.recordFailure(c.connectionURL)
+	c.log.Warnf("%s, avoiding datagram-sized transport for %s", cause, window)
+
+	if err := conn.Close(); err != nil {
+		c.log.Debugf("close relay connection for transport fallback: %s", err)
+	}
+}
+
 func (c *Client) listenForStopEvents(ctx context.Context, hc *healthcheck.Receiver, conn net.Conn, internalStopFlag *internalStopFlag) {
 	for {
 		select {

shared/relay/client/dialer/capability.go

@@ -0,0 +1,18 @@
+package dialer
+
+// DatagramSized is implemented by dialers whose connections carry each write in
+// a single datagram, so a write can be rejected when it exceeds the path's
+// datagram budget (e.g. QUIC). Transports without this capability (e.g.
+// WebSocket over TCP) impose no per-write size limit, so the relay client can
+// fall back to them when a datagram-sized transport rejects a write as too
+// large. The capability is advertised per dialer rather than hardcoded, so a
+// new transport only needs to declare whether it is datagram-sized.
+type DatagramSized interface {
+	DatagramSized()
+}
+
+// IsDatagramSized reports whether d produces datagram-sized connections.
+func IsDatagramSized(d DialeFn) bool {
+	_, ok := d.(DatagramSized)
+	return ok
+}

shared/relay/client/dialer/net/err.go

@@ -4,4 +4,9 @@ import "errors"
 
 var (
 	ErrClosedByServer = errors.New("closed by server")
+
+	// ErrDatagramTooLarge is returned when a transport message exceeds the
+	// QUIC datagram size the path to the relay can carry. The relay client
+	// treats it as a signal to fall back to a non-datagram transport.
+	ErrDatagramTooLarge = errors.New("datagram frame too large")
 )

shared/relay/client/dialer/quic/conn.go

@@ -8,7 +8,6 @@ import (
 	"time"
 
 	"github.com/quic-go/quic-go"
-	log "github.com/sirupsen/logrus"
 
 	netErr "github.com/netbirdio/netbird/shared/relay/client/dialer/net"
 )
@@ -52,11 +51,8 @@ func (c *Conn) Read(b []byte) (n int, err error) {
 }
 
 func (c *Conn) Write(b []byte) (int, error) {
-	err := c.session.SendDatagram(b)
-	if err != nil {
-		err = c.remoteCloseErrHandling(err)
-		log.Errorf("failed to write to QUIC stream: %v", err)
-		return 0, err
+	if err := c.session.SendDatagram(b); err != nil {
+		return 0, c.writeErrHandling(err, len(b))
 	}
 	return len(b), nil
 }
@@ -95,3 +91,15 @@ func (c *Conn) remoteCloseErrHandling(err error) error {
 	}
 	return err
 }
+
+// writeErrHandling normalizes SendDatagram errors. A datagram that exceeds the
+// path's QUIC packet budget is mapped to ErrDatagramTooLarge (annotated with the
+// datagram size and path budget) so the relay client can fall back to a
+// non-datagram transport.
+func (c *Conn) writeErrHandling(err error, size int) error {
+	var tooLarge *quic.DatagramTooLargeError
+	if errors.As(err, &tooLarge) {
+		return fmt.Errorf("%w: %d byte datagram over path budget %d", netErr.ErrDatagramTooLarge, size, tooLarge.MaxDatagramPayloadSize)
+	}
+	return c.remoteCloseErrHandling(err)
+}

shared/relay/client/dialer/quic/quic.go

@@ -9,6 +9,7 @@ import (
 	"time"
 
 	"github.com/quic-go/quic-go"
+	"github.com/quic-go/quic-go/logging"
 	log "github.com/sirupsen/logrus"
 
 	nbnet "github.com/netbirdio/netbird/client/net"
@@ -23,6 +24,12 @@ func (d Dialer) Protocol() string {
 	return Network
 }
 
+// DatagramSized marks QUIC as a datagram-sized transport: relay traffic is
+// carried in QUIC DATAGRAM frames, which must fit a single packet.
+func (d Dialer) DatagramSized() {
+	// Intentional marker method; presence is the capability signal.
+}
+
 func (d Dialer) Dial(ctx context.Context, address, serverName string) (net.Conn, error) {
 	quicURL, err := prepareURL(address)
 	if err != nil {
@@ -47,6 +54,7 @@ func (d Dialer) Dial(ctx context.Context, address, serverName string) (net.Conn,
 		MaxIdleTimeout:    4 * time.Minute,
 		EnableDatagrams:   true,
 		InitialPacketSize: nbRelay.QUICInitialPacketSize,
+		Tracer:            connectionTracer(quicURL),
 	}
 
 	udpConn, err := nbnet.ListenUDP("udp", &net.UDPAddr{Port: 0})
@@ -74,6 +82,28 @@ func (d Dialer) Dial(ctx context.Context, address, serverName string) (net.Conn,
 	return conn, nil
 }
 
+// connectionTracer returns a QUIC tracer that logs the DPLPMTUD result and the
+// reason a relay connection closed, so the path MTU settled on and teardown
+// cause are visible in logs. Lines carry the relay address as a structured
+// field, matching the rest of the relay client logging.
+func connectionTracer(addr string) func(context.Context, logging.Perspective, quic.ConnectionID) *logging.ConnectionTracer {
+	relayLog := log.WithField("relay", addr)
+	return func(context.Context, logging.Perspective, quic.ConnectionID) *logging.ConnectionTracer {
+		return &logging.ConnectionTracer{
+			UpdatedMTU: func(mtu logging.ByteCount, done bool) {
+				if done {
+					relayLog.Infof("QUIC path MTU settled at %d", mtu)
+					return
+				}
+				relayLog.Debugf("QUIC path MTU probing at %d", mtu)
+			},
+			ClosedConnection: func(err error) {
+				relayLog.Debugf("QUIC connection closed: %v", err)
+			},
+		}
+	}
+}
+
 func prepareURL(address string) (string, error) {
 	var host string
 	var defaultPort string

shared/relay/client/dialer/race_dialer.go

@@ -32,6 +32,7 @@ type RaceDial struct {
 	serverName        string
 	dialerFns         []DialeFn
 	connectionTimeout time.Duration
+	sequential        bool
 }
 
 func NewRaceDial(log *log.Entry, connectionTimeout time.Duration, serverURL string, dialerFns ...DialeFn) *RaceDial {
@@ -53,7 +54,21 @@ func (r *RaceDial) WithServerName(serverName string) *RaceDial {
 	return r
 }
 
+// WithSequential makes Dial try the dialers in order, falling back to the next
+// only when one fails to connect, instead of racing them concurrently.
+//
+// Mutates the receiver and is not safe for concurrent reconfiguration; a
+// RaceDial is intended to be constructed per dial and discarded.
+func (r *RaceDial) WithSequential() *RaceDial {
+	r.sequential = true
+	return r
+}
+
 func (r *RaceDial) Dial(ctx context.Context) (net.Conn, error) {
+	if r.sequential {
+		return r.dialSequential(ctx)
+	}
+
 	connChan := make(chan dialResult, len(r.dialerFns))
 	winnerConn := make(chan net.Conn, 1)
 	abortCtx, abort := context.WithCancel(ctx)
@@ -72,6 +87,30 @@ func (r *RaceDial) Dial(ctx context.Context) (net.Conn, error) {
 	return conn, nil
 }
 
+// dialSequential tries each dialer in order, returning the first connection and
+// falling back to the next on failure.
+func (r *RaceDial) dialSequential(ctx context.Context) (net.Conn, error) {
+	for _, dfn := range r.dialerFns {
+		if err := ctx.Err(); err != nil {
+			return nil, err
+		}
+		attemptCtx, cancel := context.WithTimeout(ctx, r.connectionTimeout)
+		r.log.Infof("dialing Relay server via %s", dfn.Protocol())
+		conn, err := dfn.Dial(attemptCtx, r.serverURL, r.serverName)
+		cancel()
+		if err != nil {
+			if errors.Is(err, context.Canceled) {
+				return nil, err
+			}
+			r.log.Errorf("failed to dial via %s: %s", dfn.Protocol(), err)
+			continue
+		}
+		r.log.Infof("successfully dialed via: %s", dfn.Protocol())
+		return conn, nil
+	}
+	return nil, errors.New("failed to dial to Relay server on any protocol")
+}
+
 func (r *RaceDial) dial(dfn DialeFn, abortCtx context.Context, connChan chan dialResult) {
 	ctx, cancel := context.WithTimeout(abortCtx, r.connectionTimeout)
 	defer cancel()

shared/relay/client/dialer/race_dialer_test.go

@@ -250,3 +250,66 @@ func TestRaceDialFirstSuccessfulDialerWins(t *testing.T) {
 		}
 	}
 }
+
+func TestRaceDialSequentialFallback(t *testing.T) {
+	logger := logrus.NewEntry(logrus.New())
+	serverURL := "test.server.com"
+
+	var firstDialed, secondDialed bool
+	preferred := &MockDialer{
+		protocolStr: "quic",
+		dialFunc: func(ctx context.Context, address string) (net.Conn, error) {
+			firstDialed = true
+			return nil, errors.New("quic unreachable")
+		},
+	}
+	fallbackConn := &MockConn{remoteAddr: &MockAddr{network: "ws"}}
+	fallback := &MockDialer{
+		protocolStr: "ws",
+		dialFunc: func(ctx context.Context, address string) (net.Conn, error) {
+			secondDialed = true
+			return fallbackConn, nil
+		},
+	}
+
+	rd := NewRaceDial(logger, DefaultConnectionTimeout, serverURL, preferred, fallback).WithSequential()
+	conn, err := rd.Dial(context.Background())
+	if err != nil {
+		t.Fatalf("expected fallback to succeed, got %v", err)
+	}
+	if conn != fallbackConn {
+		t.Errorf("expected fallback connection, got %v", conn)
+	}
+	if !firstDialed || !secondDialed {
+		t.Errorf("expected both dialers attempted in order, first=%v second=%v", firstDialed, secondDialed)
+	}
+}
+
+func TestRaceDialSequentialPreferredWins(t *testing.T) {
+	logger := logrus.NewEntry(logrus.New())
+	serverURL := "test.server.com"
+
+	preferredConn := &MockConn{remoteAddr: &MockAddr{network: "quic"}}
+	preferred := &MockDialer{
+		protocolStr: "quic",
+		dialFunc: func(ctx context.Context, address string) (net.Conn, error) {
+			return preferredConn, nil
+		},
+	}
+	fallback := &MockDialer{
+		protocolStr: "ws",
+		dialFunc: func(ctx context.Context, address string) (net.Conn, error) {
+			t.Errorf("fallback dialer must not be tried when preferred succeeds")
+			return nil, errors.New("should not happen")
+		},
+	}
+
+	rd := NewRaceDial(logger, DefaultConnectionTimeout, serverURL, preferred, fallback).WithSequential()
+	conn, err := rd.Dial(context.Background())
+	if err != nil {
+		t.Fatalf("expected preferred to succeed, got %v", err)
+	}
+	if conn != preferredConn {
+		t.Errorf("expected preferred connection, got %v", conn)
+	}
+}

shared/relay/client/dialers_generic.go

@@ -9,11 +9,42 @@ import (
 	"github.com/netbirdio/netbird/shared/relay/client/dialer/ws"
 )
 
-// getDialers returns the list of dialers to use for connecting to the relay server.
-func (c *Client) getDialers() []dialer.DialeFn {
-	if c.mtu > 0 && c.mtu > iface.DefaultMTU {
-		c.log.Infof("MTU %d exceeds default (%d), forcing WebSocket transport to avoid DATAGRAM frame size issues", c.mtu, iface.DefaultMTU)
-		return []dialer.DialeFn{ws.Dialer{}}
+// getDialers returns the ordered dialers for connecting to the relay server. It
+// applies the datagram fallback generically: if this server recently rejected a
+// datagram-sized transport, those dialers are dropped, leaving the rest.
+func (c *Client) getDialers(mode TransportMode) []dialer.DialeFn {
+	dialers := c.baseDialers(mode)
+
+	if c.transportFallback != nil && c.transportFallback.avoidDatagramSized(c.connectionURL) {
+		if filtered := nonDatagramSized(dialers); len(filtered) > 0 {
+			c.log.Infof("relay recently rejected a datagram-sized transport, avoiding it")
+			return filtered
+		}
 	}
-	return []dialer.DialeFn{quic.Dialer{}, ws.Dialer{}}
+	return dialers
+}
+
+// baseDialers returns the ordered dialers for the mode, before any datagram
+// fallback filtering. For racing modes (auto) the order is irrelevant; for
+// prefer modes the first entry is tried before falling back to the second.
+func (c *Client) baseDialers(mode TransportMode) []dialer.DialeFn {
+	switch mode {
+	case TransportModeWS:
+		c.log.Infof("%s=ws, using WebSocket transport", EnvRelayTransport)
+		return []dialer.DialeFn{ws.Dialer{}}
+	case TransportModeQUIC:
+		c.log.Infof("%s=quic, using QUIC transport", EnvRelayTransport)
+		return []dialer.DialeFn{quic.Dialer{}}
+	}
+
+	all := []dialer.DialeFn{quic.Dialer{}, ws.Dialer{}}
+	if mode == TransportModePreferWS {
+		all = []dialer.DialeFn{ws.Dialer{}, quic.Dialer{}}
+	}
+
+	if c.mtu > 0 && c.mtu > iface.DefaultMTU {
+		c.log.Infof("MTU %d exceeds default (%d), avoiding datagram-sized transports", c.mtu, iface.DefaultMTU)
+		return nonDatagramSized(all)
+	}
+	return all
 }

shared/relay/client/dialers_generic_test.go

@@ -0,0 +1,101 @@
+//go:build !js
+
+package client
+
+import (
+	"os"
+	"testing"
+
+	log "github.com/sirupsen/logrus"
+	"github.com/stretchr/testify/assert"
+
+	"github.com/netbirdio/netbird/client/iface"
+	"github.com/netbirdio/netbird/shared/relay/client/dialer"
+	netErr "github.com/netbirdio/netbird/shared/relay/client/dialer/net"
+	"github.com/netbirdio/netbird/shared/relay/client/dialer/quic"
+	"github.com/netbirdio/netbird/shared/relay/client/dialer/ws"
+)
+
+// TestDatagramSizedCapability locks the capability the generic fallback relies
+// on: QUIC is datagram-sized, WebSocket is not.
+func TestDatagramSizedCapability(t *testing.T) {
+	assert.True(t, dialer.IsDatagramSized(quic.Dialer{}), "QUIC must advertise datagram-sized")
+	assert.False(t, dialer.IsDatagramSized(ws.Dialer{}), "WebSocket must not advertise datagram-sized")
+}
+
+func protocols(dialers []dialer.DialeFn) []string {
+	out := make([]string, len(dialers))
+	for i, d := range dialers {
+		out[i] = d.Protocol()
+	}
+	return out
+}
+
+func TestGetDialers(t *testing.T) {
+	const url = "rels://relay.example:443"
+
+	tests := []struct {
+		name     string
+		mode     string
+		mtu      uint16
+		preferWS bool
+		want     []string
+	}{
+		{name: "auto races quic and ws", mode: "auto", mtu: iface.DefaultMTU, want: []string{"quic", "WS"}},
+		{name: "ws pinned", mode: "ws", mtu: iface.DefaultMTU, want: []string{"WS"}},
+		{name: "quic pinned", mode: "quic", mtu: iface.DefaultMTU, want: []string{"quic"}},
+		{name: "prefer-quic orders quic first", mode: "prefer-quic", mtu: iface.DefaultMTU, want: []string{"quic", "WS"}},
+		{name: "prefer-ws orders ws first", mode: "prefer-ws", mtu: iface.DefaultMTU, want: []string{"WS", "quic"}},
+		{name: "mtu above default forces ws", mode: "auto", mtu: iface.DefaultMTU + 100, want: []string{"WS"}},
+		{name: "sticky fallback forces ws in auto", mode: "auto", mtu: iface.DefaultMTU, preferWS: true, want: []string{"WS"}},
+		{name: "sticky fallback forces ws in prefer-quic", mode: "prefer-quic", mtu: iface.DefaultMTU, preferWS: true, want: []string{"WS"}},
+		{name: "quic pin overrides sticky fallback", mode: "quic", mtu: iface.DefaultMTU, preferWS: true, want: []string{"quic"}},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Setenv(EnvRelayTransport, tc.mode)
+			if tc.mode == "" {
+				os.Unsetenv(EnvRelayTransport)
+			}
+
+			tf := newTransportFallback()
+			if tc.preferWS {
+				tf.recordFailure(url)
+			}
+
+			c := &Client{
+				log:               log.WithField("test", t.Name()),
+				connectionURL:     url,
+				mtu:               tc.mtu,
+				transportFallback: tf,
+			}
+
+			assert.Equal(t, tc.want, protocols(c.getDialers(transportModeFromEnv())))
+		})
+	}
+}
+
+// TestStickyFallbackAfterDatagramTooLarge verifies the full chain: an oversized
+// datagram records a fallback that makes the next dial pick WebSocket, the way a
+// reconnect would after the connection is closed.
+func TestStickyFallbackAfterDatagramTooLarge(t *testing.T) {
+	const url = "rels://relay.example:443"
+	t.Setenv(EnvRelayTransport, string(TransportModeAuto))
+
+	c := &Client{
+		log:               log.WithField("test", t.Name()),
+		connectionURL:     url,
+		mtu:               iface.DefaultMTU,
+		transportFallback: newTransportFallback(),
+	}
+
+	// First dial races both transports.
+	assert.Equal(t, []string{"quic", "WS"}, protocols(c.getDialers(transportModeFromEnv())))
+
+	// An oversized datagram records the fallback for this server.
+	c.onDatagramTooLarge(&closeTrackingConn{}, netErr.ErrDatagramTooLarge)
+
+	// The reconnect now sticks to WebSocket.
+	assert.Equal(t, []string{"WS"}, protocols(c.getDialers(transportModeFromEnv())))
+}

shared/relay/client/dialers_js.go

@@ -7,7 +7,11 @@ import (
 	"github.com/netbirdio/netbird/shared/relay/client/dialer/ws"
 )
 
-func (c *Client) getDialers() []dialer.DialeFn {
+func (c *Client) getDialers(_ TransportMode) []dialer.DialeFn {
 	// JS/WASM build only uses WebSocket transport
 	return []dialer.DialeFn{ws.Dialer{}}
 }
+
+func (c *Client) baseDialers(_ TransportMode) []dialer.DialeFn {
+	return []dialer.DialeFn{ws.Dialer{}}
+}

shared/relay/client/manager.go

@@ -79,23 +79,30 @@ type Manager struct {
 
 	cleanupInterval      time.Duration
 	keepUnusedServerTime time.Duration
+
+	// transportFallback is shared across home and foreign relay clients so a
+	// datagram-too-large failure makes that server avoid datagram-sized transports across reconnects.
+	transportFallback *transportFallback
 }
 
 // NewManager creates a new manager instance.
 // The serverURL address can be empty. In this case, the manager will not serve.
 func NewManager(ctx context.Context, serverURLs []string, peerID string, mtu uint16, opts ...ManagerOption) *Manager {
 	tokenStore := &relayAuth.TokenStore{}
+	tf := newTransportFallback()
 
 	m := &Manager{
-		ctx:        ctx,
-		peerID:     peerID,
-		tokenStore: tokenStore,
-		mtu:        mtu,
+		ctx:               ctx,
+		peerID:            peerID,
+		tokenStore:        tokenStore,
+		mtu:               mtu,
+		transportFallback: tf,
 		serverPicker: &ServerPicker{
 			TokenStore:        tokenStore,
 			PeerID:            peerID,
 			MTU:               mtu,
 			ConnectionTimeout: defaultConnectionTimeout,
+			TransportFallback: tf,
 		},
 		relayClients:            make(map[string]*RelayTrack),
 		onDisconnectedListeners: make(map[string]*list.List),
@@ -287,6 +294,7 @@ func (m *Manager) openConnVia(ctx context.Context, serverAddress, peerKey string
 	m.relayClientsMutex.Unlock()
 
 	relayClient := NewClientWithServerIP(serverAddress, serverIP, m.tokenStore, m.peerID, m.mtu)
+	relayClient.SetTransportFallback(m.transportFallback)
 	err := relayClient.Connect(m.ctx)
 	if err != nil {
 		rt.err = err

shared/relay/client/picker.go

@@ -29,6 +29,7 @@ type ServerPicker struct {
 	PeerID            string
 	MTU               uint16
 	ConnectionTimeout time.Duration
+	TransportFallback *transportFallback
 }
 
 func (sp *ServerPicker) PickServer(parentCtx context.Context) (*Client, error) {
@@ -70,6 +71,7 @@ func (sp *ServerPicker) PickServer(parentCtx context.Context) (*Client, error) {
 func (sp *ServerPicker) startConnection(ctx context.Context, resultChan chan connResult, url string) {
 	log.Infof("try to connecting to relay server: %s", url)
 	relayClient := NewClient(url, sp.TokenStore, sp.PeerID, sp.MTU)
+	relayClient.SetTransportFallback(sp.TransportFallback)
 	err := relayClient.Connect(ctx)
 	resultChan <- connResult{
 		RelayClient: relayClient,

shared/relay/client/transport.go

@@ -0,0 +1,129 @@
+package client
+
+import (
+	"os"
+	"strings"
+	"sync"
+	"time"
+
+	log "github.com/sirupsen/logrus"
+
+	"github.com/netbirdio/netbird/shared/relay/client/dialer"
+)
+
+// EnvRelayTransport pins the relay transport. Valid values: "auto" (default,
+// race QUIC and WebSocket), "quic" (QUIC only), "ws" (WebSocket only),
+// "prefer-quic" / "prefer-ws" (try the preferred transport first, fall back to
+// the other only if it fails to connect; no race). The prefer modes trade a
+// slower connect when the preferred transport is blackholed for deterministic
+// transport selection.
+const EnvRelayTransport = "NB_RELAY_TRANSPORT"
+
+const (
+	// transportFallbackBase is the initial window a relay server avoids
+	// datagram-sized transports after a datagram is rejected as too large.
+	transportFallbackBase = 10 * time.Minute
+	// transportFallbackMax caps the pinned window when failures repeat.
+	transportFallbackMax = 60 * time.Minute
+)
+
+// TransportMode selects which relay dialers are used.
+type TransportMode string
+
+const (
+	TransportModeAuto       TransportMode = "auto"
+	TransportModeQUIC       TransportMode = "quic"
+	TransportModeWS         TransportMode = "ws"
+	TransportModePreferQUIC TransportMode = "prefer-quic"
+	TransportModePreferWS   TransportMode = "prefer-ws"
+)
+
+// transportModeFromEnv reads EnvRelayTransport, defaulting to auto for an empty
+// or unrecognized value.
+func transportModeFromEnv() TransportMode {
+	switch TransportMode(strings.ToLower(strings.TrimSpace(os.Getenv(EnvRelayTransport)))) {
+	case "", TransportModeAuto:
+		return TransportModeAuto
+	case TransportModeQUIC:
+		return TransportModeQUIC
+	case TransportModeWS:
+		return TransportModeWS
+	case TransportModePreferQUIC:
+		return TransportModePreferQUIC
+	case TransportModePreferWS:
+		return TransportModePreferWS
+	default:
+		log.Warnf("invalid %s value %q, using %q", EnvRelayTransport, os.Getenv(EnvRelayTransport), TransportModeAuto)
+		return TransportModeAuto
+	}
+}
+
+// sequential reports whether the mode tries dialers in order with fallback
+// instead of racing them concurrently.
+func (m TransportMode) sequential() bool {
+	return m == TransportModePreferQUIC || m == TransportModePreferWS
+}
+
+// transportFallback tracks relay servers that have rejected a datagram-sized
+// transport (a write too large for the path) and should temporarily avoid such
+// transports. It is shared across the relay manager so the preference survives
+// client recreation (foreign relay clients are evicted and rebuilt on
+// disconnect). Entries are keyed by server URL and expire after a window that
+// grows on repeated failures.
+type transportFallback struct {
+	mu      sync.Mutex
+	entries map[string]*fallbackEntry
+}
+
+type fallbackEntry struct {
+	until    time.Time
+	duration time.Duration
+}
+
+func newTransportFallback() *transportFallback {
+	return &transportFallback{entries: make(map[string]*fallbackEntry)}
+}
+
+// avoidDatagramSized reports whether serverURL is currently within a window
+// where datagram-sized transports should be avoided.
+func (f *transportFallback) avoidDatagramSized(serverURL string) bool {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	e := f.entries[serverURL]
+	return e != nil && time.Now().Before(e.until)
+}
+
+// recordFailure makes serverURL avoid datagram-sized transports for a window:
+// transportFallbackBase on the first failure, doubling up to transportFallbackMax
+// when a datagram transport fails again after a previous window expired. It
+// returns the active window duration.
+func (f *transportFallback) recordFailure(serverURL string) time.Duration {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+
+	now := time.Now()
+	e := f.entries[serverURL]
+	switch {
+	case e == nil:
+		e = &fallbackEntry{duration: transportFallbackBase}
+		f.entries[serverURL] = e
+	case now.Before(e.until):
+		return time.Until(e.until)
+	default:
+		e.duration = min(e.duration*2, transportFallbackMax)
+	}
+	e.until = now.Add(e.duration)
+	return e.duration
+}
+
+// nonDatagramSized returns the dialers from in that are not datagram-sized,
+// preserving order.
+func nonDatagramSized(in []dialer.DialeFn) []dialer.DialeFn {
+	out := make([]dialer.DialeFn, 0, len(in))
+	for _, d := range in {
+		if !dialer.IsDatagramSized(d) {
+			out = append(out, d)
+		}
+	}
+	return out
+}

shared/relay/client/transport_test.go

@@ -0,0 +1,140 @@
+package client
+
+import (
+	"net"
+	"os"
+	"testing"
+	"time"
+
+	log "github.com/sirupsen/logrus"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	netErr "github.com/netbirdio/netbird/shared/relay/client/dialer/net"
+)
+
+// closeTrackingConn records whether Close was called; only Close is exercised.
+type closeTrackingConn struct {
+	net.Conn
+	closed bool
+}
+
+func (c *closeTrackingConn) Close() error {
+	c.closed = true
+	return nil
+}
+
+func TestTransportModeFromEnv(t *testing.T) {
+	tests := []struct {
+		value string
+		want  TransportMode
+	}{
+		{"", TransportModeAuto},
+		{"auto", TransportModeAuto},
+		{"quic", TransportModeQUIC},
+		{"QUIC", TransportModeQUIC},
+		{"ws", TransportModeWS},
+		{" Ws ", TransportModeWS},
+		{"prefer-quic", TransportModePreferQUIC},
+		{"prefer-ws", TransportModePreferWS},
+		{"garbage", TransportModeAuto},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.value, func(t *testing.T) {
+			t.Setenv(EnvRelayTransport, tc.value)
+			if tc.value == "" {
+				os.Unsetenv(EnvRelayTransport)
+			}
+			assert.Equal(t, tc.want, transportModeFromEnv())
+		})
+	}
+}
+
+func TestTransportFallbackRecordAndExpiry(t *testing.T) {
+	const url = "rels://relay.example:443"
+	f := newTransportFallback()
+
+	assert.False(t, f.avoidDatagramSized(url), "no fallback recorded yet")
+
+	d := f.recordFailure(url)
+	assert.Equal(t, transportFallbackBase, d, "first failure pins for the base window")
+	assert.True(t, f.avoidDatagramSized(url), "datagram-sized transport avoided within the window")
+
+	// A second failure while still inside the window must not grow the window.
+	d = f.recordFailure(url)
+	assert.LessOrEqual(t, d, transportFallbackBase, "still within the active window")
+	require.NotNil(t, f.entries[url])
+	assert.Equal(t, transportFallbackBase, f.entries[url].duration, "duration unchanged inside window")
+
+	// Expire the window: datagram-sized transport allowed again.
+	f.entries[url].until = time.Now().Add(-time.Second)
+	assert.False(t, f.avoidDatagramSized(url), "window expired, datagram-sized transport allowed")
+}
+
+func TestTransportFallbackGrowsOnRepeat(t *testing.T) {
+	const url = "rels://relay.example:443"
+	f := newTransportFallback()
+
+	want := transportFallbackBase
+	for i := range 6 {
+		d := f.recordFailure(url)
+		assert.Equal(t, want, d, "window after %d expiries", i)
+
+		// expire the window so the next failure is treated as a repeat
+		f.entries[url].until = time.Now().Add(-time.Second)
+
+		want = min(want*2, transportFallbackMax)
+	}
+
+	assert.Equal(t, transportFallbackMax, f.entries[url].duration, "window caps at the max")
+}
+
+func TestOnDatagramTooLargeAuto(t *testing.T) {
+	const url = "rels://relay.example:443"
+	t.Setenv(EnvRelayTransport, string(TransportModeAuto))
+
+	tf := newTransportFallback()
+	c := &Client{
+		log:               log.WithField("test", t.Name()),
+		connectionURL:     url,
+		transportFallback: tf,
+	}
+	conn := &closeTrackingConn{}
+
+	c.onDatagramTooLarge(conn, netErr.ErrDatagramTooLarge)
+
+	assert.True(t, conn.closed, "connection closed to force reconnect")
+	assert.True(t, tf.avoidDatagramSized(url), "fallback recorded for the server")
+
+	// A second oversized datagram on the same connection must not re-close.
+	conn.closed = false
+	c.onDatagramTooLarge(conn, netErr.ErrDatagramTooLarge)
+	assert.False(t, conn.closed, "single fallback per connection")
+}
+
+func TestOnDatagramTooLargeQUICPinned(t *testing.T) {
+	const url = "rels://relay.example:443"
+	t.Setenv(EnvRelayTransport, string(TransportModeQUIC))
+
+	tf := newTransportFallback()
+	c := &Client{
+		log:               log.WithField("test", t.Name()),
+		connectionURL:     url,
+		transportFallback: tf,
+	}
+	conn := &closeTrackingConn{}
+
+	c.onDatagramTooLarge(conn, netErr.ErrDatagramTooLarge)
+
+	assert.False(t, conn.closed, "QUIC pin keeps the connection, no fallback redial")
+	assert.False(t, tf.avoidDatagramSized(url), "QUIC pin records no fallback")
+}
+
+func TestTransportFallbackPerServer(t *testing.T) {
+	f := newTransportFallback()
+	f.recordFailure("rels://a.example:443")
+
+	assert.True(t, f.avoidDatagramSized("rels://a.example:443"))
+	assert.False(t, f.avoidDatagramSized("rels://b.example:443"), "fallback is scoped to one server")
+}