diff --git a/internal/locate/region_cache.go b/internal/locate/region_cache.go
index 84b9b40b6..18de47f70 100644
--- a/internal/locate/region_cache.go
+++ b/internal/locate/region_cache.go
@@ -549,6 +549,18 @@ func NewRegionCache(pdClient pd.Client) *RegionCache {
 	return c
 }
 
+// only used fot test.
+func newTestRegionCache() *RegionCache {
+	c := &RegionCache{}
+	c.storeMu.stores = make(map[uint64]*Store)
+	c.tiflashComputeStoreMu.needReload = true
+	c.tiflashComputeStoreMu.stores = make([]*Store, 0)
+	c.notifyCheckCh = make(chan struct{}, 1)
+	c.ctx, c.cancelFunc = context.WithCancel(context.Background())
+	c.mu = *newRegionIndexMu(nil)
+	return c
+}
+
 // clear clears all cached data in the RegionCache. It's only used in tests.
 func (c *RegionCache) clear() {
 	c.mu = *newRegionIndexMu(nil)
@@ -558,8 +570,8 @@ func (c *RegionCache) clear() {
 }
 
 // thread unsafe, should use with lock
-func (c *RegionCache) insertRegionToCache(cachedRegion *Region, invalidateOldRegion bool, shouldCount bool) {
-	c.mu.insertRegionToCache(cachedRegion, invalidateOldRegion, shouldCount)
+func (c *RegionCache) insertRegionToCache(cachedRegion *Region, invalidateOldRegion bool, shouldCount bool) bool {
+	return c.mu.insertRegionToCache(cachedRegion, invalidateOldRegion, shouldCount)
 }
 
 // Close releases region cache's resource.
@@ -1491,9 +1503,32 @@ func (mu *regionIndexMu) removeVersionFromCache(oldVer RegionVerID, regionID uin
 // It should be protected by c.mu.l.Lock().
 // if `invalidateOldRegion` is false, the old region cache should be still valid,
 // and it may still be used by some kv requests.
-func (mu *regionIndexMu) insertRegionToCache(cachedRegion *Region, invalidateOldRegion bool, shouldCount bool) {
-	oldRegion := mu.sorted.ReplaceOrInsert(cachedRegion)
-	if oldRegion != nil {
+// Moreover, it will return false if the region is stale.
+func (mu *regionIndexMu) insertRegionToCache(cachedRegion *Region, invalidateOldRegion bool, shouldCount bool) bool {
+	newVer := cachedRegion.VerID()
+	oldVer, ok := mu.latestVersions[newVer.id]
+	// There are two or more situations in which the region we got is stale.
+	// The first case is that the process of getting a region is concurrent.
+	// The stale region may be returned later due to network reasons.
+	// The second case is that the region may be obtained from the PD follower,
+	// and there is the synchronization time between the pd follower and the leader.
+	// So we should check the epoch.
+	if ok && (oldVer.GetVer() > newVer.GetVer() || oldVer.GetConfVer() > newVer.GetConfVer()) {
+		logutil.BgLogger().Debug("get stale region",
+			zap.Uint64("region", newVer.GetID()), zap.Uint64("new-ver", newVer.GetVer()), zap.Uint64("new-conf", newVer.GetConfVer()),
+			zap.Uint64("old-ver", oldVer.GetVer()), zap.Uint64("old-conf", oldVer.GetConfVer()))
+		return false
+	}
+	// Also check and remove the intersecting regions including the old region.
+	intersectedRegions, stale := mu.sorted.removeIntersecting(cachedRegion, newVer)
+	if stale {
+		return false
+	}
+	// Insert the region (won't replace because of above deletion).
+	mu.sorted.ReplaceOrInsert(cachedRegion)
+	// Inherit the workTiKVIdx, workTiFlashIdx and buckets from the first intersected region.
+	if len(intersectedRegions) > 0 {
+		oldRegion := intersectedRegions[0].cachedRegion
 		store := cachedRegion.getStore()
 		oldRegionStore := oldRegion.getStore()
 		// TODO(youjiali1995): remove this because the new retry logic can handle this issue.
@@ -1506,15 +1541,6 @@ func (mu *regionIndexMu) insertRegionToCache(cachedRegion *Region, invalidateOld
 		if InvalidReason(atomic.LoadInt32((*int32)(&oldRegion.invalidReason))) == NoLeader {
 			store.workTiKVIdx = (oldRegionStore.workTiKVIdx + 1) % AccessIndex(store.accessStoreNum(tiKVOnly))
 		}
-		// If the old region is still valid, do not invalidate it to avoid unnecessary backoff.
-		if invalidateOldRegion {
-			// Invalidate the old region in case it's not invalidated and some requests try with the stale region information.
-			if shouldCount {
-				oldRegion.invalidate(Other)
-			} else {
-				oldRegion.invalidateWithoutMetrics(Other)
-			}
-		}
 		// Don't refresh TiFlash work idx for region. Otherwise, it will always goto a invalid store which
 		// is under transferring regions.
 		store.workTiFlashIdx.Store(oldRegionStore.workTiFlashIdx.Load())
@@ -1523,21 +1549,27 @@ func (mu *regionIndexMu) insertRegionToCache(cachedRegion *Region, invalidateOld
 		if store.buckets == nil || (oldRegionStore.buckets != nil && store.buckets.GetVersion() < oldRegionStore.buckets.GetVersion()) {
 			store.buckets = oldRegionStore.buckets
 		}
-		mu.removeVersionFromCache(oldRegion.VerID(), cachedRegion.VerID().id)
-	}
-	mu.regions[cachedRegion.VerID()] = cachedRegion
-	newVer := cachedRegion.VerID()
-	latest, ok := mu.latestVersions[cachedRegion.VerID().id]
-	if !ok || latest.GetVer() < newVer.GetVer() || latest.GetConfVer() < newVer.GetConfVer() {
-		mu.latestVersions[cachedRegion.VerID().id] = newVer
 	}
 	// The intersecting regions in the cache are probably stale, clear them.
-	deleted := mu.sorted.removeIntersecting(cachedRegion)
-	for _, region := range deleted {
+	for _, region := range intersectedRegions {
 		mu.removeVersionFromCache(region.cachedRegion.VerID(), region.cachedRegion.GetID())
+		// If the old region is still valid, do not invalidate it to avoid unnecessary backoff.
+		if invalidateOldRegion {
+			// Invalidate the old region in case it's not invalidated and some requests try with the stale region information.
+			if shouldCount {
+				region.cachedRegion.invalidate(Other)
+			} else {
+				region.cachedRegion.invalidateWithoutMetrics(Other)
+			}
+		}
 	}
+	// update related vars.
+	mu.regions[newVer] = cachedRegion
+	mu.latestVersions[newVer.id] = newVer
+	return true
+}
 
-} // searchCachedRegion finds a region from cache by key. Like `getCachedRegion`,
+// searchCachedRegion finds a region from cache by key. Like `getCachedRegion`,
 // it should be called with c.mu.RLock(), and the returned Region should not be
 // used after c.mu is RUnlock().
 // If the given key is the end key of the region that you want, you may set the second argument to true. This is useful
diff --git a/internal/locate/region_cache_test.go b/internal/locate/region_cache_test.go
index 91a7603d1..7c06b8b59 100644
--- a/internal/locate/region_cache_test.go
+++ b/internal/locate/region_cache_test.go
@@ -55,6 +55,7 @@ import (
 	"github.com/tikv/client-go/v2/internal/mockstore/mocktikv"
 	"github.com/tikv/client-go/v2/kv"
 	pd "github.com/tikv/pd/client"
+	uatomic "go.uber.org/atomic"
 )
 
 func TestRegionCache(t *testing.T) {
@@ -1567,6 +1568,8 @@ func (s *testRegionCacheSuite) TestBuckets() {
 
 	// update buckets if it's nil.
 	cachedRegion.getStore().buckets = nil
+	// we should replace the version of `cacheRegion` because of stale.
+	s.cluster.PutRegion(r.GetId(), newMeta.RegionEpoch.ConfVer, newMeta.RegionEpoch.Version, []uint64{s.store1, s.store2}, []uint64{s.peer1, s.peer2}, s.peer1)
 	s.cluster.SplitRegionBuckets(cachedRegion.GetID(), defaultBuckets.Keys, defaultBuckets.Version)
 	s.cache.UpdateBucketsIfNeeded(cachedRegion.VerID(), defaultBuckets.GetVersion())
 	waitUpdateBuckets(defaultBuckets, []byte("a"))
@@ -1833,3 +1836,192 @@ func (s *testRegionRequestToSingleStoreSuite) TestRefreshCacheConcurrency() {
 
 	cancel()
 }
+
+func TestRegionCacheWithDelay(t *testing.T) {
+	suite.Run(t, new(testRegionCacheWithDelaySuite))
+}
+
+type testRegionCacheWithDelaySuite struct {
+	suite.Suite
+	mvccStore mocktikv.MVCCStore
+	cluster   *mocktikv.Cluster
+	store     uint64 // store1 is leader
+	region1   uint64
+	bo        *retry.Backoffer
+
+	delay      uatomic.Bool
+	delayCache *RegionCache
+	cache      *RegionCache
+}
+
+func (s *testRegionCacheWithDelaySuite) SetupTest() {
+	s.mvccStore = mocktikv.MustNewMVCCStore()
+	s.cluster = mocktikv.NewCluster(s.mvccStore)
+	storeIDs, _, regionID, _ := mocktikv.BootstrapWithMultiStores(s.cluster, 1)
+	s.region1 = regionID
+	s.store = storeIDs[0]
+	pdCli := &CodecPDClient{mocktikv.NewPDClient(s.cluster), apicodec.NewCodecV1(apicodec.ModeTxn)}
+	s.cache = NewRegionCache(pdCli)
+	pdCli2 := &CodecPDClient{mocktikv.NewPDClient(s.cluster, mocktikv.WithDelay(&s.delay)), apicodec.NewCodecV1(apicodec.ModeTxn)}
+	s.delayCache = NewRegionCache(pdCli2)
+	s.bo = retry.NewBackofferWithVars(context.Background(), 5000, nil)
+}
+
+func (s *testRegionCacheWithDelaySuite) TearDownTest() {
+	s.cache.Close()
+	s.delayCache.Close()
+	s.mvccStore.Close()
+}
+
+func (s *testRegionCacheWithDelaySuite) TestInsertStaleRegion() {
+	r, err := s.cache.findRegionByKey(s.bo, []byte("a"), false)
+	s.NoError(err)
+	fakeRegion := &Region{
+		meta:          r.meta,
+		syncFlag:      r.syncFlag,
+		lastAccess:    r.lastAccess,
+		invalidReason: r.invalidReason,
+	}
+	fakeRegion.setStore(r.getStore().clone())
+
+	newPeersIDs := s.cluster.AllocIDs(1)
+	s.cluster.Split(r.GetID(), s.cluster.AllocID(), []byte("c"), newPeersIDs, newPeersIDs[0])
+	newPeersIDs = s.cluster.AllocIDs(1)
+	s.cluster.Split(r.GetID(), s.cluster.AllocID(), []byte("b"), newPeersIDs, newPeersIDs[0])
+
+	r.invalidate(Other)
+	r2, err := s.cache.findRegionByKey(s.bo, []byte("c"), false)
+	s.NoError(err)
+	s.Equal([]byte("c"), r2.StartKey())
+	r2, err = s.cache.findRegionByKey(s.bo, []byte("b"), false)
+	s.NoError(err)
+	s.Equal([]byte("b"), r2.StartKey())
+
+	stale := !s.cache.insertRegionToCache(fakeRegion, true, true)
+	s.True(stale)
+
+	rs, err := s.cache.scanRegionsFromCache(s.bo, []byte(""), []byte(""), 100)
+	s.NoError(err)
+	s.Greater(len(rs), 1)
+	s.NotEqual(rs[0].EndKey(), "")
+
+	r3, err := s.cache.findRegionByKey(s.bo, []byte("a"), false)
+	s.NoError(err)
+	s.Equal([]byte("b"), r3.EndKey())
+}
+
+func (s *testRegionCacheWithDelaySuite) TestStaleGetRegion() {
+	r1, err := s.cache.findRegionByKey(s.bo, []byte("a"), false)
+	s.NoError(err)
+	r2, err := s.delayCache.findRegionByKey(s.bo, []byte("a"), false)
+	s.NoError(err)
+	s.Equal(r1.meta, r2.meta)
+
+	// simulates network delay
+	s.delay.Store(true)
+	var wg sync.WaitGroup
+	wg.Add(1)
+	go func() {
+		r2.invalidate(Other)
+		_, err := s.delayCache.findRegionByKey(s.bo, []byte("b"), false)
+		s.NoError(err)
+		wg.Done()
+	}()
+	time.Sleep(30 * time.Millisecond)
+	newPeersIDs := s.cluster.AllocIDs(1)
+	s.cluster.Split(r1.GetID(), s.cluster.AllocID(), []byte("b"), newPeersIDs, newPeersIDs[0])
+	r1.invalidate(Other)
+	r, err := s.cache.findRegionByKey(s.bo, []byte("b"), false)
+	s.NoError(err)
+	s.Equal([]byte("b"), r.meta.StartKey)
+	r, err = s.cache.findRegionByKey(s.bo, []byte("c"), false)
+	s.NoError(err)
+	s.Equal([]byte("b"), r.meta.StartKey)
+
+	s.delay.Store(false)
+	r, err = s.delayCache.findRegionByKey(s.bo, []byte("b"), false)
+	s.NoError(err)
+	s.Equal([]byte("b"), r.meta.StartKey)
+	wg.Wait()
+	// the delay response is received, but insert failed.
+	r, err = s.delayCache.findRegionByKey(s.bo, []byte("b"), false)
+	s.NoError(err)
+	s.Equal([]byte("b"), r.meta.StartKey)
+	r, err = s.delayCache.findRegionByKey(s.bo, []byte("a"), false)
+	s.NoError(err)
+	s.Equal([]byte("b"), r.meta.EndKey)
+}
+
+func generateKeyForSimulator(id int, keyLen int) []byte {
+	k := make([]byte, keyLen)
+	copy(k, fmt.Sprintf("%010d", id))
+	return k
+}
+
+func BenchmarkInsertRegionToCache(b *testing.B) {
+	b.StopTimer()
+	cache := newTestRegionCache()
+	r := &Region{
+		meta: &metapb.Region{
+			Id:          1,
+			RegionEpoch: &metapb.RegionEpoch{},
+		},
+	}
+	rs := &regionStore{
+		workTiKVIdx:              0,
+		proxyTiKVIdx:             -1,
+		stores:                   make([]*Store, 0, len(r.meta.Peers)),
+		pendingTiFlashPeerStores: map[uint64]uint64{},
+		storeEpochs:              make([]uint32, 0, len(r.meta.Peers)),
+	}
+	r.setStore(rs)
+	b.StartTimer()
+	for i := 0; i < b.N; i++ {
+		newMeta := proto.Clone(r.meta).(*metapb.Region)
+		newMeta.Id = uint64(i + 1)
+		newMeta.RegionEpoch.ConfVer = uint64(i+1) - uint64(rand.Intn(i+1))
+		newMeta.RegionEpoch.Version = uint64(i+1) - uint64(rand.Intn(i+1))
+		if i%2 == 0 {
+			newMeta.StartKey = generateKeyForSimulator(rand.Intn(i+1), 56)
+			newMeta.EndKey = []byte("")
+		} else {
+			newMeta.EndKey = generateKeyForSimulator(rand.Intn(i+1), 56)
+			newMeta.StartKey = []byte("")
+		}
+		region := &Region{
+			meta: newMeta,
+		}
+		region.setStore(r.getStore())
+		cache.insertRegionToCache(region, true, true)
+	}
+}
+
+func BenchmarkInsertRegionToCache2(b *testing.B) {
+	b.StopTimer()
+	cache := newTestRegionCache()
+	r := &Region{
+		meta: &metapb.Region{
+			Id:          1,
+			RegionEpoch: &metapb.RegionEpoch{},
+		},
+	}
+	rs := &regionStore{
+		workTiKVIdx:              0,
+		proxyTiKVIdx:             -1,
+		stores:                   make([]*Store, 0, len(r.meta.Peers)),
+		pendingTiFlashPeerStores: map[uint64]uint64{},
+		storeEpochs:              make([]uint32, 0, len(r.meta.Peers)),
+	}
+	r.setStore(rs)
+	b.StartTimer()
+	for i := 0; i < b.N; i++ {
+		newMeta := proto.Clone(r.meta).(*metapb.Region)
+		newMeta.RegionEpoch.ConfVer = uint64(i + 1)
+		newMeta.RegionEpoch.Version = uint64(i + 1)
+		region := &Region{
+			meta: newMeta,
+		}
+		region.setStore(r.getStore())
+		cache.insertRegionToCache(region, true, true)
+	}
+}
diff --git a/internal/locate/sorted_btree.go b/internal/locate/sorted_btree.go
index 6165f7614..f718aba3d 100644
--- a/internal/locate/sorted_btree.go
+++ b/internal/locate/sorted_btree.go
@@ -38,6 +38,8 @@ import (
 	"bytes"
 
 	"github.com/google/btree"
+	"github.com/tikv/client-go/v2/internal/logutil"
+	"go.uber.org/zap"
 )
 
 // SortedRegions is a sorted btree.
@@ -93,12 +95,16 @@ func (s *SortedRegions) AscendGreaterOrEqual(startKey, endKey []byte, limit int)
 
 // removeIntersecting removes all items that have intersection with the key range of given region.
 // If the region itself is in the cache, it's not removed.
-func (s *SortedRegions) removeIntersecting(r *Region) []*btreeItem {
+func (s *SortedRegions) removeIntersecting(r *Region, verID RegionVerID) ([]*btreeItem, bool) {
 	var deleted []*btreeItem
+	var stale bool
 	s.b.AscendGreaterOrEqual(newBtreeSearchItem(r.StartKey()), func(item *btreeItem) bool {
-		// Skip the item that is equal to the given region.
-		if item.cachedRegion.VerID() == r.VerID() {
-			return true
+		if item.cachedRegion.meta.GetRegionEpoch().GetVersion() > verID.ver {
+			logutil.BgLogger().Debug("get stale region",
+				zap.Uint64("region", verID.GetID()), zap.Uint64("ver", verID.GetVer()), zap.Uint64("conf", verID.GetConfVer()),
+				zap.Uint64("intersecting-ver", item.cachedRegion.meta.GetRegionEpoch().GetVersion()))
+			stale = true
+			return false
 		}
 		if len(r.EndKey()) > 0 && bytes.Compare(item.cachedRegion.StartKey(), r.EndKey()) >= 0 {
 			return false
@@ -106,10 +112,13 @@ func (s *SortedRegions) removeIntersecting(r *Region) []*btreeItem {
 		deleted = append(deleted, item)
 		return true
 	})
+	if stale {
+		return nil, true
+	}
 	for _, item := range deleted {
 		s.b.Delete(item)
 	}
-	return deleted
+	return deleted, false
 }
 
 // Clear removes all items from the btree.
diff --git a/internal/mockstore/mocktikv/cluster.go b/internal/mockstore/mocktikv/cluster.go
index e46b02d3d..7b5f1cc3f 100644
--- a/internal/mockstore/mocktikv/cluster.go
+++ b/internal/mockstore/mocktikv/cluster.go
@@ -413,6 +413,14 @@ func (c *Cluster) Bootstrap(regionID uint64, storeIDs, peerIDs []uint64, leaderP
 	c.regions[regionID] = newRegion(regionID, storeIDs, peerIDs, leaderPeerID)
 }
 
+// PutRegion adds or replaces a region.
+func (c *Cluster) PutRegion(regionID, confVer, ver uint64, storeIDs, peerIDs []uint64, leaderPeerID uint64) {
+	c.Lock()
+	defer c.Unlock()
+
+	c.regions[regionID] = newRegion(regionID, storeIDs, peerIDs, leaderPeerID, confVer, ver)
+}
+
 // AddPeer adds a new Peer for the Region on the Store.
 func (c *Cluster) AddPeer(regionID, storeID, peerID uint64) {
 	c.Lock()
@@ -634,7 +642,7 @@ func newPeerMeta(peerID, storeID uint64) *metapb.Peer {
 	}
 }
 
-func newRegion(regionID uint64, storeIDs, peerIDs []uint64, leaderPeerID uint64) *Region {
+func newRegion(regionID uint64, storeIDs, peerIDs []uint64, leaderPeerID uint64, epoch ...uint64) *Region {
 	if len(storeIDs) != len(peerIDs) {
 		panic("len(storeIDs) != len(peerIds)")
 	}
@@ -647,6 +655,10 @@ func newRegion(regionID uint64, storeIDs, peerIDs []uint64, leaderPeerID uint64)
 		Peers:       peers,
 		RegionEpoch: &metapb.RegionEpoch{},
 	}
+	if len(epoch) == 2 {
+		meta.RegionEpoch.ConfVer = epoch[0]
+		meta.RegionEpoch.Version = epoch[1]
+	}
 	return &Region{
 		Meta:   meta,
 		leader: leaderPeerID,
diff --git a/internal/mockstore/mocktikv/pd.go b/internal/mockstore/mocktikv/pd.go
index 833b914c0..b2b01aa41 100644
--- a/internal/mockstore/mocktikv/pd.go
+++ b/internal/mockstore/mocktikv/pd.go
@@ -61,6 +61,16 @@ var tsMu = struct {
 
 const defaultResourceGroupName = "default"
 
+var _ pd.Client = (*pdClient)(nil)
+
+type MockPDOption func(*pdClient)
+
+func WithDelay(delay *atomic.Bool) MockPDOption {
+	return func(pc *pdClient) {
+		pc.delay = delay
+	}
+}
+
 type pdClient struct {
 	cluster *Cluster
 	// SafePoint set by `UpdateGCSafePoint`. Not to be confused with SafePointKV.
@@ -73,11 +83,13 @@ type pdClient struct {
 	externalTimestamp atomic.Uint64
 
 	groups map[string]*rmpb.ResourceGroup
+
+	delay *atomic.Bool
 }
 
 // NewPDClient creates a mock pd.Client that uses local timestamp and meta data
 // from a Cluster.
-func NewPDClient(cluster *Cluster) pd.Client {
+func NewPDClient(cluster *Cluster, ops ...MockPDOption) *pdClient {
 	mockCli := &pdClient{
 		cluster:           cluster,
 		serviceSafePoints: make(map[string]uint64),
@@ -97,6 +109,9 @@ func NewPDClient(cluster *Cluster) pd.Client {
 		},
 		Priority: 8,
 	}
+	for _, op := range ops {
+		op(mockCli)
+	}
 	return mockCli
 }
 
@@ -206,6 +221,12 @@ func (c *pdClient) GetRegion(ctx context.Context, key []byte, opts ...pd.GetRegi
 	if len(opts) == 0 {
 		buckets = nil
 	}
+	if c.delay != nil && c.delay.Load() {
+		select {
+		case <-ctx.Done():
+		case <-time.After(200 * time.Millisecond):
+		}
+	}
 	return &pd.Region{Meta: region, Leader: peer, Buckets: buckets, DownPeers: downPeers}, nil
 }