prepare_snap: establish connection to all stores before pausing admin (#51449)

br/pkg/backup/prepare_snap/BUILD.bazel

@@ -35,7 +35,7 @@ go_test(
     timeout = "short",
     srcs = ["prepare_test.go"],
     flaky = True,
-    shard_count = 7,
+    shard_count = 8,
     deps = [
         ":prepare_snap",
         "//br/pkg/utils",

br/pkg/backup/prepare_snap/prepare.go

@@ -155,7 +155,7 @@
 		zap.Int("retry_limit", p.RetryLimit),
 		zap.Duration("lease_duration", p.LeaseDuration))
 	p.retryTime = 0
-	if err := p.prepareConnections(ctx); err != nil {
+	if err := p.PrepareConnections(ctx); err != nil {
 		log.Error("failed to prepare connections", logutil.ShortError(err))
 		return errors.Annotate(err, "failed to prepare connections")
 	}
@@ -385,23 +385,31 @@
 }
 
 func (p *Preparer) streamOf(ctx context.Context, storeID uint64) (*prepareStream, error) {
-	s, ok := p.clients[storeID]
+	_, ok := p.clients[storeID]
 	if !ok {
+		log.Warn("stream of store found a store not established connection", zap.Uint64("store", storeID))
 		cli, err := p.env.ConnectToStore(ctx, storeID)
 		if err != nil {
 			return nil, errors.Annotatef(err, "failed to dial store %d", storeID)
 		}
-		s = new(prepareStream)
-		s.storeID = storeID
-		s.output = p.eventChan
-		s.leaseDuration = p.LeaseDuration
-		err = s.InitConn(ctx, cli)
-		if err != nil {
-			return nil, err
+		if err := p.createAndCacheStream(ctx, cli, storeID); err != nil {
+			return nil, errors.Annotatef(err, "failed to create and cache stream for store %d", storeID)
 		}
-		p.clients[storeID] = s
 	}
-	return s, nil
+	return p.clients[storeID], nil
+}
+
+func (p *Preparer) createAndCacheStream(ctx context.Context, cli PrepareClient, storeID uint64) error {
+	s := new(prepareStream)
+	s.storeID = storeID
+	s.output = p.eventChan
+	s.leaseDuration = p.LeaseDuration
+	err := s.InitConn(ctx, cli)
+	if err != nil {
+		return err
+	}
+	p.clients[storeID] = s
+	return nil
 }
 
 func (p *Preparer) pushWaitApply(reqs pendingRequests, region Region) {
@@ -414,17 +422,31 @@
 	p.inflightReqs[region.GetMeta().Id] = *region.GetMeta()
 }
 
-func (p *Preparer) prepareConnections(ctx context.Context) error {
+// PrepareConnections prepares the connections for each store.
+// This will pause the admin commands for each store.
+func (p *Preparer) PrepareConnections(ctx context.Context) error {
 	log.Info("Preparing connections to stores.")
 	stores, err := p.env.GetAllLiveStores(ctx)
 	if err != nil {
 		return errors.Annotate(err, "failed to get all live stores")
 	}
+
+	log.Info("Start to initialize the connections.", zap.Int("stores", len(stores)))
+	clients := map[uint64]PrepareClient{}
 	for _, store := range stores {
-		_, err := p.streamOf(ctx, store.Id)
+		cli, err := p.env.ConnectToStore(ctx, store.Id)
 		if err != nil {
-			return errors.Annotatef(err, "failed to prepare connection to store %d", store.Id)
+			return errors.Annotatef(err, "failed to dial the store %d", store.Id)
+		}
+		clients[store.Id] = cli
+	}
+
+	for id, cli := range clients {
+		log.Info("Start to pause the admin commands.", zap.Uint64("store", id))
+		if err := p.createAndCacheStream(ctx, cli, id); err != nil {
+			return errors.Annotatef(err, "failed to create and cache stream for store %d", id)
 		}
 	}
+
 	return nil
 }

br/pkg/backup/prepare_snap/prepare_test.go

@@ -110,15 +110,24 @@ type mockStores struct {
 	mu               sync.Mutex
 	stores           map[uint64]*mockStore
 	onCreateStore    func(*mockStore)
+	connectDelay     func(uint64) <-chan struct{}
 	onConnectToStore func(uint64) error
 
 	pdc *tikv.RegionCache
 }
 
 func newTestEnv(pdc pd.Client) *mockStores {
 	r := tikv.NewRegionCache(pdc)
+	stores, err := pdc.GetAllStores(context.Background())
+	if err != nil {
+		panic(err)
+	}
+	ss := map[uint64]*mockStore{}
+	for _, store := range stores {
+		ss[store.Id] = nil
+	}
 	ms := &mockStores{
-		stores:        map[uint64]*mockStore{},
+		stores:        ss,
 		pdc:           r,
 		onCreateStore: func(ms *mockStore) {},
 	}
@@ -138,7 +147,14 @@ func (m *mockStores) GetAllLiveStores(ctx context.Context) ([]*metapb.Store, err
 
 func (m *mockStores) ConnectToStore(ctx context.Context, storeID uint64) (PrepareClient, error) {
 	m.mu.Lock()
-	defer m.mu.Unlock()
+	defer func() {
+		m.mu.Unlock()
+		if m.connectDelay != nil {
+			if ch := m.connectDelay(storeID); ch != nil {
+				<-ch
+			}
+		}
+	}()
 
 	if m.onConnectToStore != nil {
 		err := m.onConnectToStore(storeID)
@@ -147,8 +163,8 @@ func (m *mockStores) ConnectToStore(ctx context.Context, storeID uint64) (Prepar
 		}
 	}
 
-	_, ok := m.stores[storeID]
-	if !ok {
+	s, ok := m.stores[storeID]
+	if !ok || s == nil {
 		m.stores[storeID] = &mockStore{
 			output:         make(chan brpb.PrepareSnapshotBackupResponse, 16),
 			successRegions: []metapb.Region{},
@@ -456,3 +472,41 @@ func TestSplitEnv(t *testing.T) {
 	require.Equal(t, cc.PrepareClient.(*counterClient).send, 1)
 	require.ElementsMatch(t, cc.PrepareClient.(*counterClient).regions, tinyRequest.Regions)
 }
+
+func TestConnectionDelay(t *testing.T) {
+	log.SetLevel(zapcore.DebugLevel)
+	req := require.New(t)
+	pdc := fakeCluster(t, 3, dummyRegions(100)...)
+	ms := newTestEnv(pdc)
+	called := 0
+	delayConn := make(chan struct{})
+	blocked := make(chan struct{}, 64)
+	ms.connectDelay = func(i uint64) <-chan struct{} {
+		called += 1
+		if called == 2 {
+			blocked <- struct{}{}
+			return delayConn
+		}
+		return nil
+	}
+	ctx := context.Background()
+	prep := New(ms)
+	connectionPrepareResult := make(chan error)
+	go func() {
+		connectionPrepareResult <- prep.PrepareConnections(ctx)
+	}()
+	<-blocked
+	ms.mu.Lock()
+	nonNilStore := 0
+	for id, store := range ms.stores {
+		// We must not create and lease (i.e. reject admin command from any tikv) here.
+		if store != nil {
+			req.True(store.leaseUntil.Before(time.Now()), "%d->%s", id, store.leaseUntil)
+			nonNilStore += 1
+		}
+	}
+	req.GreaterOrEqual(nonNilStore, 2)
+	ms.mu.Unlock()
+	delayConn <- struct{}{}
+	req.NoError(<-connectionPrepareResult)
+}

br/pkg/backup/prepare_snap/stream.go

@@ -70,6 +70,7 @@ func (p *prepareStream) InitConn(ctx context.Context, cli PrepareClient) error {
 	p.cli = cli
 	p.clientLoopHandle, ctx = errgroup.WithContext(ctx)
 	ctx, p.stopBgTasks = context.WithCancel(ctx)
+	log.Info("initializing", zap.Uint64("store", p.storeID))
 	return p.GoLeaseLoop(ctx, p.leaseDuration)
 }