backendcluster: route PD and backend traffic through cluster DNS

Author: YangKeao

go.mod

@@ -38,6 +38,7 @@ require (
 	go.uber.org/mock v0.5.2
 	go.uber.org/ratelimit v0.2.0
 	go.uber.org/zap v1.27.0
+	golang.org/x/net v0.48.0
 	google.golang.org/grpc v1.63.2
 )
 
@@ -272,7 +273,6 @@ require (
 	golang.org/x/crypto v0.47.0 // indirect
 	golang.org/x/exp v0.0.0-20240909161429-701f63a606c0 // indirect
 	golang.org/x/mod v0.31.0 // indirect
-	golang.org/x/net v0.48.0 // indirect
 	golang.org/x/oauth2 v0.30.0 // indirect
 	golang.org/x/sync v0.19.0 // indirect
 	golang.org/x/sys v0.40.0 // indirect

pkg/balance/observer/health_check.go

@@ -19,6 +19,11 @@ import (
 	"go.uber.org/zap"
 )
 
+type BackendNetwork interface {
+	HTTPClient(clusterName string) *http.Client
+	DialContext(ctx context.Context, network, addr, clusterName string) (net.Conn, error)
+}
+
 // HealthCheck is used to check the backends of one backend. One can pass a customized health check function to the observer.
 type HealthCheck interface {
 	Check(ctx context.Context, backendID string, info *BackendInfo, lastHealth *BackendHealth) *BackendHealth
@@ -48,20 +53,44 @@ type security struct {
 type DefaultHealthCheck struct {
 	cfg     *config.HealthCheck
 	logger  *zap.Logger
-	httpCli *http.Client
+	network BackendNetwork
 }
 
 func NewDefaultHealthCheck(httpCli *http.Client, cfg *config.HealthCheck, logger *zap.Logger) *DefaultHealthCheck {
-	if httpCli == nil {
-		httpCli = http.NewHTTPClient(func() *tls.Config { return nil })
+	return NewDefaultHealthCheckWithNetwork(newDefaultBackendNetwork(httpCli), cfg, logger)
+}
+
+func NewDefaultHealthCheckWithNetwork(network BackendNetwork, cfg *config.HealthCheck, logger *zap.Logger) *DefaultHealthCheck {
+	if network == nil {
+		network = newDefaultBackendNetwork(nil)
 	}
 	return &DefaultHealthCheck{
-		httpCli: httpCli,
+		network: network,
 		cfg:     cfg,
 		logger:  logger,
 	}
 }
 
+type defaultBackendNetwork struct {
+	httpCli *http.Client
+}
+
+func newDefaultBackendNetwork(httpCli *http.Client) *defaultBackendNetwork {
+	if httpCli == nil {
+		httpCli = http.NewHTTPClient(func() *tls.Config { return nil })
+	}
+	return &defaultBackendNetwork{httpCli: httpCli}
+}
+
+func (n *defaultBackendNetwork) HTTPClient(string) *http.Client {
+	return n.httpCli
+}
+
+func (n *defaultBackendNetwork) DialContext(ctx context.Context, network, addr, _ string) (net.Conn, error) {
+	var dialer net.Dialer
+	return dialer.DialContext(ctx, network, addr)
+}
+
 func (dhc *DefaultHealthCheck) Check(ctx context.Context, _ string, info *BackendInfo, lastBh *BackendHealth) *BackendHealth {
 	bh := &BackendHealth{
 		BackendInfo: *info,
@@ -96,10 +125,13 @@ func (dhc *DefaultHealthCheck) checkSqlPort(ctx context.Context, info *BackendIn
 		return
 	}
 	addr := info.Addr
+	clusterName := info.ClusterName
 	b := backoff.WithContext(backoff.WithMaxRetries(backoff.NewConstantBackOff(dhc.cfg.RetryInterval), uint64(dhc.cfg.MaxRetries)), ctx)
 	err := http.ConnectWithRetry(func() error {
 		startTime := time.Now()
-		conn, err := net.DialTimeout("tcp", addr, dhc.cfg.DialTimeout)
+		dialCtx, cancel := context.WithTimeout(ctx, dhc.cfg.DialTimeout)
+		conn, err := dhc.network.DialContext(dialCtx, "tcp", addr, clusterName)
+		cancel()
 		setPingBackendMetrics(addr, startTime)
 		if err != nil {
 			return err
@@ -134,7 +166,8 @@ func (dhc *DefaultHealthCheck) checkStatusPort(ctx context.Context, info *Backen
 
 	addr := net.JoinHostPort(info.IP, strconv.Itoa(int(info.StatusPort)))
 	b := backoff.WithContext(backoff.WithMaxRetries(backoff.NewConstantBackOff(dhc.cfg.RetryInterval), uint64(dhc.cfg.MaxRetries)), ctx)
-	resp, err := dhc.httpCli.Get(addr, statusPathSuffix, b, dhc.cfg.DialTimeout)
+	clusterName := info.ClusterName
+	resp, err := dhc.network.HTTPClient(clusterName).Get(addr, statusPathSuffix, b, dhc.cfg.DialTimeout)
 	if err == nil {
 		var respBody backendHttpStatusRespBody
 		err = json.Unmarshal(resp, &respBody)
@@ -176,7 +209,8 @@ func (dhc *DefaultHealthCheck) queryConfig(ctx context.Context, info *BackendInf
 
 	b := backoff.WithContext(backoff.WithMaxRetries(backoff.NewConstantBackOff(dhc.cfg.RetryInterval), uint64(dhc.cfg.MaxRetries)), ctx)
 	var resp []byte
-	if resp, err = dhc.httpCli.Get(addr, configPathSuffix, b, dhc.cfg.DialTimeout); err != nil {
+	clusterName := info.ClusterName
+	if resp, err = dhc.network.HTTPClient(clusterName).Get(addr, configPathSuffix, b, dhc.cfg.DialTimeout); err != nil {
 		return
 	}
 	var respBody backendHttpConfigRespBody

pkg/balance/observer/health_check_test.go

@@ -5,10 +5,12 @@ package observer
 
 import (
 	"context"
+	"crypto/tls"
 	"encoding/json"
 	"net"
 	"net/http"
 	"strings"
+	"sync"
 	"sync/atomic"
 	"testing"
 	"time"
@@ -17,6 +19,7 @@ import (
 	"github.com/pingcap/tiproxy/lib/util/logger"
 	"github.com/pingcap/tiproxy/lib/util/waitgroup"
 	"github.com/pingcap/tiproxy/pkg/testkit"
+	httputil "github.com/pingcap/tiproxy/pkg/util/http"
 	"github.com/stretchr/testify/require"
 )
 
@@ -120,6 +123,59 @@ func TestSupportRedirection(t *testing.T) {
 	require.False(t, health.SupportRedirection)
 }
 
+func TestHealthCheckUsesClusterNetwork(t *testing.T) {
+	lg, _ := logger.CreateLoggerForTest(t)
+	cfg := newHealthCheckConfigForTest()
+	backend, info := newBackendServer(t)
+	defer backend.close()
+	backend.setServerVersion("1.0")
+	backend.setHasSigningCert(true)
+	info.ClusterName = "cluster-a"
+
+	network := &mockBackendNetwork{
+		httpCli: httputil.NewHTTPClient(func() *tls.Config { return nil }),
+	}
+	hc := NewDefaultHealthCheckWithNetwork(network, cfg, lg)
+	health := hc.Check(context.Background(), backend.sqlAddr, info, nil)
+	require.True(t, health.Healthy)
+	require.Contains(t, network.httpClusters(), "cluster-a")
+	require.Contains(t, network.dialClusters(), "cluster-a")
+}
+
+type mockBackendNetwork struct {
+	httpCli *httputil.Client
+	mu      sync.Mutex
+	https   []string
+	dials   []string
+}
+
+func (n *mockBackendNetwork) HTTPClient(clusterName string) *httputil.Client {
+	n.mu.Lock()
+	n.https = append(n.https, clusterName)
+	n.mu.Unlock()
+	return n.httpCli
+}
+
+func (n *mockBackendNetwork) DialContext(ctx context.Context, network, addr, clusterName string) (net.Conn, error) {
+	n.mu.Lock()
+	n.dials = append(n.dials, clusterName)
+	n.mu.Unlock()
+	var dialer net.Dialer
+	return dialer.DialContext(ctx, network, addr)
+}
+
+func (n *mockBackendNetwork) httpClusters() []string {
+	n.mu.Lock()
+	defer n.mu.Unlock()
+	return append([]string(nil), n.https...)
+}
+
+func (n *mockBackendNetwork) dialClusters() []string {
+	n.mu.Lock()
+	defer n.mu.Unlock()
+	return append([]string(nil), n.dials...)
+}
+
 type backendServer struct {
 	t            *testing.T
 	sqlListener  net.Listener

pkg/balance/router/router.go

@@ -80,6 +80,7 @@ type BackendInst interface {
 	Healthy() bool
 	Local() bool
 	Keyspace() string
+	ClusterName() string
 }
 
 // backendWrapper contains the connections on the backend.

pkg/balance/router/router_static.go

@@ -82,6 +82,7 @@ func (r *StaticRouter) OnConnClosed(backendID, redirectingBackendID string, conn
 type StaticBackend struct {
 	addr     string
 	keyspace string
+	cluster  string
 	healthy  atomic.Bool
 }
 
@@ -120,3 +121,7 @@ func (b *StaticBackend) Keyspace() string {
 func (b *StaticBackend) SetKeyspace(k string) {
 	b.keyspace = k
 }
+
+func (b *StaticBackend) ClusterName() string {
+	return b.cluster
+}

pkg/manager/backendcluster/manager.go

@@ -7,6 +7,7 @@ import (
 	"context"
 	"crypto/tls"
 	"maps"
+	"net"
 	"slices"
 	"strings"
 	"sync"
@@ -16,7 +17,8 @@ import (
 	"github.com/pingcap/tiproxy/pkg/balance/metricsreader"
 	"github.com/pingcap/tiproxy/pkg/manager/infosync"
 	"github.com/pingcap/tiproxy/pkg/util/etcd"
-	"github.com/pingcap/tiproxy/pkg/util/http"
+	httputil "github.com/pingcap/tiproxy/pkg/util/http"
+	"github.com/pingcap/tiproxy/pkg/util/netutil"
 	"github.com/pingcap/tiproxy/pkg/util/waitgroup"
 	clientv3 "go.etcd.io/etcd/client/v3"
 	"go.uber.org/zap"
@@ -28,6 +30,8 @@ type Cluster struct {
 	etcdCli    *clientv3.Client
 	infoSyncer *infosync.InfoSyncer
 	metrics    *metricsreader.ClusterReader
+	httpCli    *httputil.Client
+	dialer     *netutil.DNSDialer
 }
 
 func (c *Cluster) Config() config.BackendCluster {
@@ -46,6 +50,14 @@ func (c *Cluster) GetPromInfo(ctx context.Context) (*infosync.PrometheusInfo, er
 	return c.infoSyncer.GetPromInfo(ctx)
 }
 
+func (c *Cluster) HTTPClient() *httputil.Client {
+	return c.httpCli
+}
+
+func (c *Cluster) DialContext(ctx context.Context, network, addr string) (net.Conn, error) {
+	return c.dialer.DialContext(ctx, network, addr)
+}
+
 type Manager struct {
 	lg         *zap.Logger
 	clusterTLS func() *tls.Config
@@ -54,6 +66,7 @@ type Manager struct {
 	wg      waitgroup.WaitGroup
 	cancel  context.CancelFunc
 	metrics *MetricsQuerier
+	network *NetworkRouter
 
 	mu struct {
 		sync.RWMutex
@@ -68,6 +81,7 @@ func NewManager(lg *zap.Logger, clusterTLS func() *tls.Config) *Manager {
 	}
 	mgr.mu.clusters = make(map[string]*Cluster)
 	mgr.metrics = NewMetricsQuerier(mgr)
+	mgr.network = NewNetworkRouter(mgr, clusterTLS)
 	return mgr
 }
 
@@ -195,10 +209,18 @@ func clusterReusable(cluster *Cluster, cfg config.BackendCluster) bool {
 
 func (m *Manager) buildCluster(ctx context.Context, cfg *config.Config, clusterCfg config.BackendCluster) (*Cluster, error) {
 	clusterCfg = normalizeCluster(clusterCfg)
-	etcdCli, err := etcd.InitEtcdClientWithAddrs(
+	nameServers, err := config.ParseNSServers(clusterCfg.NSServers)
+	if err != nil {
+		return nil, err
+	}
+	dialer := netutil.NewDNSDialer(nameServers)
+	httpCli := httputil.NewHTTPClientWithDialContext(m.clusterTLS, dialer.DialContext)
+
+	etcdCli, err := etcd.InitEtcdClientWithAddrsAndDialer(
 		m.lg.With(zap.String("cluster", clusterCfg.Name)).Named("etcd"),
 		clusterCfg.PDAddrs,
 		m.clusterTLS(),
+		dialer,
 	)
 	if err != nil {
 		return nil, err
@@ -217,13 +239,15 @@ func (m *Manager) buildCluster(ctx context.Context, cfg *config.Config, clusterC
 		cfg:        clusterCfg,
 		etcdCli:    etcdCli,
 		infoSyncer: infoSyncer,
+		httpCli:    httpCli,
+		dialer:     dialer,
 	}
 	cluster.metrics = metricsreader.NewClusterReader(
 		m.lg.With(zap.String("cluster", clusterCfg.Name)).Named("metrics"),
 		clusterCfg.Name,
 		cluster,
 		cluster,
-		http.NewHTTPClient(m.clusterTLS),
+		httpCli,
 		etcdCli,
 		config.NewDefaultHealthCheckConfig(),
 		m.cfgGetter,
@@ -277,6 +301,11 @@ func (m *Manager) HasBackendClusters() bool {
 func (m *Manager) MetricsQuerier() *MetricsQuerier {
 	return m.metrics
 }
+
+func (m *Manager) NetworkRouter() *NetworkRouter {
+	return m.network
+}
+
 // PrimaryCluster returns the only configured cluster when the cluster count is exactly one.
 // It exists for features that are only well-defined in the single-cluster case, such as VIP.
 func (m *Manager) PrimaryCluster() *Cluster {