feat: add check_tolerance to make node selection stable

component/outbound/dialer/alive_dialer_set.go

@@ -26,6 +26,7 @@ type AliveDialerSet struct {
 	dialerGroupName string
 	l4proto         consts.L4ProtoStr
 	ipversion       consts.IpVersionStr
+	tolerance       time.Duration
 
 	aliveChangeCallback func(alive bool)
 
@@ -43,6 +44,7 @@ func NewAliveDialerSet(
 	dialerGroupName string,
 	l4proto consts.L4ProtoStr,
 	ipversion consts.IpVersionStr,
+	tolerance time.Duration,
 	selectionPolicy consts.DialerSelectionPolicy,
 	dialers []*Dialer,
 	aliveChangeCallback func(alive bool),
@@ -53,6 +55,7 @@ func NewAliveDialerSet(
 		dialerGroupName:         dialerGroupName,
 		l4proto:                 l4proto,
 		ipversion:               ipversion,
+		tolerance:               tolerance,
 		aliveChangeCallback:     aliveChangeCallback,
 		dialerToIndex:           make(map[*Dialer]int),
 		dialerToLatency:         make(map[*Dialer]time.Duration),
@@ -146,14 +149,19 @@ func (a *AliveDialerSet) NotifyLatencyChange(dialer *Dialer, alive bool) {
 		bakOldBestDialer := a.minLatency.dialer
 		// Calc minLatency.
 		a.dialerToLatency[dialer] = latency
-		if alive && latency < a.minLatency.latency {
+		if alive && latency <= a.minLatency.latency-a.tolerance {
 			a.minLatency.latency = latency
 			a.minLatency.dialer = dialer
 		} else if a.minLatency.dialer == dialer {
+			if latency > a.minLatency.latency {
+				// Latency increases.
 				a.minLatency.latency = time.Hour
 				a.minLatency.dialer = nil
 				a.calcMinLatency()
 				// Now `a.minLatency.dialer` will be nil if there is no alive dialer.
+			} else {
+				a.minLatency.latency = latency
+			}
 		}
 		currentAlive := a.minLatency.dialer != nil
 		// If best dialer changed.
@@ -169,7 +177,8 @@ func (a *AliveDialerSet) NotifyLatencyChange(dialer *Dialer, alive bool) {
 					string(a.selectionPolicy): a.minLatency.latency,
 					"group":                   a.dialerGroupName,
 					"network":                 string(a.l4proto) + string(a.ipversion),
-					"dialer":                  a.minLatency.dialer.Name(),
+					"new dialer":              a.minLatency.dialer.Name(),
+					"old dialer":              bakOldBestDialer.Name(),
 				}).Infof("Group %vselects dialer", re)
 			} else {
 				// Alive -> not alive
@@ -199,7 +208,7 @@ func (a *AliveDialerSet) calcMinLatency() {
 		if !ok {
 			continue
 		}
-		if latency < a.minLatency.latency {
+		if latency <= a.minLatency.latency-a.tolerance {
 			a.minLatency.latency = latency
 			a.minLatency.dialer = d
 		}

component/outbound/dialer/dialer.go

@@ -37,6 +37,7 @@ type GlobalOption struct {
 	TcpCheckOptionRaw TcpCheckOptionRaw // Lazy parse
 	UdpCheckOptionRaw UdpCheckOptionRaw // Lazy parse
 	CheckInterval     time.Duration
+	CheckTolerance    time.Duration
 }
 
 type InstanceOption struct {

component/outbound/dialer_group.go

@@ -39,10 +39,10 @@ type DialerGroup struct {
 func NewDialerGroup(option *dialer.GlobalOption, name string, dialers []*dialer.Dialer, p DialerSelectionPolicy, aliveChangeCallback func(alive bool, l4proto uint8, ipversion uint8)) *DialerGroup {
 	log := option.Log
 	var registeredAliveDialerSet bool
-	aliveTcp4DialerSet := dialer.NewAliveDialerSet(log, name, consts.L4ProtoStr_TCP, consts.IpVersionStr_4, p.Policy, dialers, func(alive bool) { aliveChangeCallback(alive, unix.IPPROTO_TCP, 4) }, true)
+	aliveTcp4DialerSet := dialer.NewAliveDialerSet(log, name, consts.L4ProtoStr_TCP, consts.IpVersionStr_4, option.CheckTolerance, p.Policy, dialers, func(alive bool) { aliveChangeCallback(alive, unix.IPPROTO_TCP, 4) }, true)
-	aliveTcp6DialerSet := dialer.NewAliveDialerSet(log, name, consts.L4ProtoStr_TCP, consts.IpVersionStr_6, p.Policy, dialers, func(alive bool) { aliveChangeCallback(alive, unix.IPPROTO_TCP, 6) }, true)
+	aliveTcp6DialerSet := dialer.NewAliveDialerSet(log, name, consts.L4ProtoStr_TCP, consts.IpVersionStr_6, option.CheckTolerance, p.Policy, dialers, func(alive bool) { aliveChangeCallback(alive, unix.IPPROTO_TCP, 6) }, true)
-	aliveUdp4DialerSet := dialer.NewAliveDialerSet(log, name, consts.L4ProtoStr_UDP, consts.IpVersionStr_4, p.Policy, dialers, func(alive bool) { aliveChangeCallback(alive, unix.IPPROTO_UDP, 4) }, true)
+	aliveUdp4DialerSet := dialer.NewAliveDialerSet(log, name, consts.L4ProtoStr_UDP, consts.IpVersionStr_4, option.CheckTolerance, p.Policy, dialers, func(alive bool) { aliveChangeCallback(alive, unix.IPPROTO_UDP, 4) }, true)
-	aliveUdp6DialerSet := dialer.NewAliveDialerSet(log, name, consts.L4ProtoStr_UDP, consts.IpVersionStr_6, p.Policy, dialers, func(alive bool) { aliveChangeCallback(alive, unix.IPPROTO_UDP, 6) }, true)
+	aliveUdp6DialerSet := dialer.NewAliveDialerSet(log, name, consts.L4ProtoStr_UDP, consts.IpVersionStr_6, option.CheckTolerance, p.Policy, dialers, func(alive bool) { aliveChangeCallback(alive, unix.IPPROTO_UDP, 6) }, true)
 
 	switch p.Policy {
 	case consts.DialerSelectionPolicy_Random,

config/config.go

@@ -19,6 +19,7 @@ type Global struct {
 	TcpCheckUrl    string            `mapstructure:"tcp_check_url" default:"http://cp.cloudflare.com"`
 	UdpCheckDns    string            `mapstructure:"udp_check_dns" default:"cloudflare-dns.com:53"`
 	CheckInterval  time.Duration     `mapstructure:"check_interval" default:"30s"`
+	CheckTolerance time.Duration     `mapstructure:"check_tolerance" default:"0"`
 	DnsUpstream    common.UrlOrEmpty `mapstructure:"dns_upstream" require:""`
 	LanInterface   []string          `mapstructure:"lan_interface"`
 	WanInterface   []string          `mapstructure:"wan_interface"`

control/control_plane.go

@@ -196,6 +196,7 @@ func NewControlPlane(
 		TcpCheckOptionRaw: dialer.TcpCheckOptionRaw{Raw: global.TcpCheckUrl},
 		UdpCheckOptionRaw: dialer.UdpCheckOptionRaw{Raw: global.UdpCheckDns},
 		CheckInterval:     global.CheckInterval,
+		CheckTolerance:    global.CheckTolerance,
 	}
 	outbounds := []*outbound.DialerGroup{
 		outbound.NewDialerGroup(option, consts.OutboundDirect.String(),

example.dae

@@ -10,6 +10,8 @@ global {
     tcp_check_url: 'http://cp.cloudflare.com'
     udp_check_dns: 'dns.google:53'
     check_interval: 30s
+    # Group will switch node only when new_latency <= old_latency - tolerance
+    check_tolerance: 50ms
 
     # Value can be scheme://host:port or empty string ''.
     # The scheme can be tcp/udp/tcp+udp. Empty string '' indicates as-is.