logservice: avoid dispatcher registration failure during schema store initialization

pkg/eventservice/event_broker.go

@@ -1032,6 +1032,9 @@ func (c *eventBroker) addDispatcher(info DispatcherInfo) error {
 			zap.Uint64("startTs", info.GetStartTs()), zap.String("span", common.FormatTableSpan(span)),
 			zap.Error(err),
 		)
+		// Mark removed to avoid processing notifications before unregister completes.
+		dispatcher.isRemoved.Store(true)
+		c.eventStore.UnregisterDispatcher(changefeedID, id)
 		status.removeDispatcher(id)
 		if status.isEmpty() {
 			c.changefeedMap.Delete(changefeedID)

pkg/eventservice/event_broker_test.go

@@ -172,6 +172,21 @@ func TestOnNotify(t *testing.T) {
 	log.Info("Pass case 6")
 }
 
+func TestAddDispatcherUnregisterOnSchemaStoreError(t *testing.T) {
+	broker, es, ss, _ := newEventBrokerForTest()
+	defer broker.close()
+
+	ss.registerTableError = errors.New("register schema store failed")
+
+	info := newMockDispatcherInfoForTest(t)
+	err := broker.addDispatcher(info)
+	require.Error(t, err)
+
+	_, ok := es.spansMap.Load(info.GetTableSpan())
+	require.False(t, ok)
+	require.Equal(t, uint64(1), es.unregisterCount.Load())
+}
+
 func TestScanRangeCappedByScanWindow(t *testing.T) {
 	broker, _, _, _ := newEventBrokerForTest()
 	// Close the broker, so we can catch all message in the test.

pkg/eventservice/event_service_test.go

@@ -17,6 +17,7 @@ import (
 	"context"
 	"fmt"
 	"sync"
+	"sync/atomic"
 	"testing"
 	"time"
 
@@ -163,6 +164,7 @@ type mockEventStore struct {
 	resolvedTsUpdateInterval time.Duration
 	dispatcherMap            sync.Map // key is common.DispatcherID, value is span
 	spansMap                 sync.Map // key is *heartbeatpb.TableSpan
+	unregisterCount          atomic.Uint64
 }
 
 func newMockEventStore(resolvedTsUpdateInterval int) *mockEventStore {
@@ -245,7 +247,9 @@ func (m *mockEventStore) UnregisterDispatcher(changefeedID common.ChangeFeedID,
 	span, ok := m.dispatcherMap.Load(dispatcherID)
 	if ok {
 		m.spansMap.Delete(span)
+		m.dispatcherMap.Delete(dispatcherID)
 	}
+	m.unregisterCount.Add(1)
 }
 
 func (m *mockEventStore) GetIterator(dispatcherID common.DispatcherID, dataRange common.DataRange) eventstore.EventIterator {

pkg/upstream/upstream.go

@@ -129,18 +129,24 @@ func CreateTiStore(ctx context.Context, urls string, credential *security.Creden
 		retry.WithBackoffBaseDelay(200),
 		retry.WithBackoffMaxDelay(4000),
 		retry.WithIsRetryableErr(func(err error) bool {
-			switch errors.Cause(err) {
-			case context.Canceled:
-				return false
-			}
-			return true
+			return isCreateTiStoreRetryable(ctx, err)
 		}))
 	if err != nil {
 		return nil, errors.WrapError(errors.ErrNewStore, err)
 	}
 	return tiStore, nil
 }
 
+func isCreateTiStoreRetryable(ctx context.Context, err error) bool {
+	switch errors.Cause(err) {
+	case context.Canceled:
+		// Only stop retrying if the caller's context is canceled.
+		// Otherwise treat it as transient (e.g. internal client cancellation).
+		return ctx.Err() == nil
+	}
+	return true
+}
+
 // init initializes the upstream
 func initUpstream(ctx context.Context, up *Upstream, cfg *NodeTopologyCfg) error {
 	ctx, up.cancel = context.WithCancel(ctx)

pkg/upstream/upstream_test.go

@@ -136,3 +136,16 @@ func TestRegisterTopo(t *testing.T) {
 		return len(resp.Kvs) == 0
 	}, time.Second*5, time.Millisecond*100)
 }
+
+func TestIsCreateTiStoreRetryable(t *testing.T) {
+	t.Parallel()
+
+	ctx := context.Background()
+	require.True(t, isCreateTiStoreRetryable(ctx, context.Canceled))
+	require.True(t, isCreateTiStoreRetryable(ctx, errors.Trace(context.Canceled)))
+
+	canceledCtx, cancel := context.WithCancel(context.Background())
+	cancel()
+	require.False(t, isCreateTiStoreRetryable(canceledCtx, context.Canceled))
+	require.False(t, isCreateTiStoreRetryable(canceledCtx, errors.Trace(context.Canceled)))
+}