// TestRejectFutureCommand verifies that lease holders reject commands that
// would cause a large time jump.
func TestRejectFutureCommand(t *testing.T) {
	defer leaktest.AfterTest(t)()

	const maxOffset = 100 * time.Millisecond
	manual := hlc.NewManualClock(0)
	clock := hlc.NewClock(manual.UnixNano)
	clock.SetMaxOffset(maxOffset)
	mtc := multiTestContext{
		clock: clock,
	}
	mtc.Start(t, 1)
	defer mtc.Stop()

	// First do a write. The first write will advance the clock by MaxOffset
	// because of the read cache's low water mark.
	getArgs := putArgs([]byte("b"), []byte("b"))
	if _, err := client.SendWrapped(rg1(mtc.stores[0]), nil, &getArgs); err != nil {
		t.Fatal(err)
	}
	if now := clock.Now(); now.WallTime != int64(maxOffset) {
		t.Fatalf("expected clock to advance to 100ms; got %s", now)
	}
	// The logical clock has advanced past the physical clock; increment
	// the "physical" clock to catch up.
	manual.Increment(int64(maxOffset))

	startTime := manual.UnixNano()

	// Commands with a future timestamp that is within the MaxOffset
	// bound will be accepted and will cause the clock to advance.
	for i := int64(0); i < 3; i++ {
		incArgs := incrementArgs([]byte("a"), 5)
		ts := hlc.ZeroTimestamp.Add(startTime+((i+1)*30)*int64(time.Millisecond), 0)
		if _, err := client.SendWrappedWith(rg1(mtc.stores[0]), nil, roachpb.Header{Timestamp: ts}, &incArgs); err != nil {
			t.Fatal(err)
		}
	}
	if now := clock.Now(); now.WallTime != int64(190*time.Millisecond) {
		t.Fatalf("expected clock to advance to 190ms; got %s", now)
	}

	// Once the accumulated offset reaches MaxOffset, commands will be rejected.
	incArgs := incrementArgs([]byte("a"), 11)
	ts := hlc.ZeroTimestamp.Add(int64((time.Duration(startTime)+maxOffset+1)*time.Millisecond), 0)
	if _, err := client.SendWrappedWith(rg1(mtc.stores[0]), nil, roachpb.Header{Timestamp: ts}, &incArgs); err == nil {
		t.Fatalf("expected clock offset error but got nil")
	}

	// The clock remained at 190ms and the final command was not executed.
	if now := clock.Now(); now.WallTime != int64(190*time.Millisecond) {
		t.Errorf("expected clock to advance to 190ms; got %s", now)
	}
	val, _, err := engine.MVCCGet(context.Background(), mtc.engines[0], roachpb.Key("a"), clock.Now(), true, nil)
	if err != nil {
		t.Fatal(err)
	}
	if v := mustGetInt(val); v != 15 {
		t.Errorf("expected 15, got %v", v)
	}
}

// TestTxnMultipleCoord checks that a coordinator uses the Writing flag to
// enforce that only one coordinator can be used for transactional writes.
func TestTxnMultipleCoord(t *testing.T) {
	defer leaktest.AfterTest(t)()
	s, sender := createTestDB(t)
	defer s.Stop()

	testCases := []struct {
		args    roachpb.Request
		writing bool
		ok      bool
	}{
		{roachpb.NewGet(roachpb.Key("a")), true, false},
		{roachpb.NewGet(roachpb.Key("a")), false, true},
		{roachpb.NewPut(roachpb.Key("a"), roachpb.Value{}), false, false}, // transactional write before begin
		{roachpb.NewPut(roachpb.Key("a"), roachpb.Value{}), true, false},  // must have switched coordinators
	}

	for i, tc := range testCases {
		txn := roachpb.NewTransaction("test", roachpb.Key("a"), 1, enginepb.SERIALIZABLE,
			s.Clock.Now(), s.Clock.MaxOffset().Nanoseconds())
		txn.Writing = tc.writing
		reply, pErr := client.SendWrappedWith(sender, nil, roachpb.Header{
			Txn: txn,
		}, tc.args)
		if pErr == nil != tc.ok {
			t.Errorf("%d: %T (writing=%t): success_expected=%t, but got: %v",
				i, tc.args, tc.writing, tc.ok, pErr)
		}
		if pErr != nil {
			continue
		}

		txn = reply.Header().Txn
		// The transaction should come back rw if it started rw or if we just
		// wrote.
		isWrite := roachpb.IsTransactionWrite(tc.args)
		if (tc.writing || isWrite) != txn.Writing {
			t.Errorf("%d: unexpected writing state: %s", i, txn)
		}
		if !isWrite {
			continue
		}
		// Abort for clean shutdown.
		if _, pErr := client.SendWrappedWith(sender, nil, roachpb.Header{
			Txn: txn,
		}, &roachpb.EndTransactionRequest{
			Commit: false,
		}); pErr != nil {
			t.Fatal(pErr)
		}
	}
}

// fillRange writes keys with the given prefix and associated values
// until bytes bytes have been written or the given range has split.
func fillRange(store *storage.Store, rangeID roachpb.RangeID, prefix roachpb.Key, bytes int64, t *testing.T) {
	src := rand.New(rand.NewSource(0))
	for {
		var ms enginepb.MVCCStats
		if err := engine.MVCCGetRangeStats(context.Background(), store.Engine(), rangeID, &ms); err != nil {
			t.Fatal(err)
		}
		keyBytes, valBytes := ms.KeyBytes, ms.ValBytes
		if keyBytes+valBytes >= bytes {
			return
		}
		key := append(append([]byte(nil), prefix...), randutil.RandBytes(src, 100)...)
		key = keys.MakeRowSentinelKey(key)
		val := randutil.RandBytes(src, int(src.Int31n(1<<8)))
		pArgs := putArgs(key, val)
		_, pErr := client.SendWrappedWith(store, nil, roachpb.Header{
			RangeID: rangeID,
		}, &pArgs)
		// When the split occurs in the background, our writes may start failing.
		// We know we can stop writing when this happens.
		if _, ok := pErr.GetDetail().(*roachpb.RangeKeyMismatchError); ok {
			return
		} else if pErr != nil {
			t.Fatal(pErr)
		}
	}
}

func TestComputeStatsForKeySpan(t *testing.T) {
	defer leaktest.AfterTest(t)()
	mtc := startMultiTestContext(t, 3)
	defer mtc.Stop()

	// Create a number of ranges using splits.
	splitKeys := []string{"a", "c", "e", "g", "i"}
	for _, k := range splitKeys {
		key := []byte(k)
		repl := mtc.stores[0].LookupReplica(key, roachpb.RKeyMin)
		args := adminSplitArgs(key, key)
		header := roachpb.Header{
			RangeID: repl.RangeID,
		}
		if _, err := client.SendWrappedWith(mtc.stores[0], nil, header, &args); err != nil {
			t.Fatal(err)
		}
	}

	// Wait for splits to finish.
	util.SucceedsSoon(t, func() error {
		repl := mtc.stores[0].LookupReplica(roachpb.RKey("z"), nil)
		if actualRSpan := repl.Desc().RSpan(); !actualRSpan.Key.Equal(roachpb.RKey("i")) {
			return errors.Errorf("expected range %s to begin at key 'i'", repl)
		}
		return nil
	})

	// Create some keys across the ranges.
	incKeys := []string{"b", "bb", "bbb", "d", "dd", "h"}
	for _, k := range incKeys {
		if _, err := mtc.dbs[0].Inc([]byte(k), 5); err != nil {
			t.Fatal(err)
		}
	}

	// Verify stats across different spans.
	for _, tcase := range []struct {
		startKey       string
		endKey         string
		expectedRanges int
		expectedKeys   int64
	}{
		{"a", "i", 4, 6},
		{"a", "c", 1, 3},
		{"b", "e", 2, 5},
		{"e", "i", 2, 1},
	} {
		start, end := tcase.startKey, tcase.endKey
		stats, count := mtc.stores[0].ComputeStatsForKeySpan(
			roachpb.RKey(start), roachpb.RKey(end))
		if a, e := count, tcase.expectedRanges; a != e {
			t.Errorf("Expected %d ranges in span [%s - %s], found %d", e, start, end, a)
		}
		if a, e := stats.LiveCount, tcase.expectedKeys; a != e {
			t.Errorf("Expected %d keys in span [%s - %s], found %d", e, start, end, a)
		}
	}
}

// TestRangeCommandClockUpdate verifies that followers update their
// clocks when executing a command, even if the lease holder's clock is far
// in the future.
func TestRangeCommandClockUpdate(t *testing.T) {
	defer leaktest.AfterTest(t)()

	const numNodes = 3
	var manuals []*hlc.ManualClock
	var clocks []*hlc.Clock
	for i := 0; i < numNodes; i++ {
		manuals = append(manuals, hlc.NewManualClock(1))
		clocks = append(clocks, hlc.NewClock(manuals[i].UnixNano))
		clocks[i].SetMaxOffset(100 * time.Millisecond)
	}
	mtc := multiTestContext{
		clocks: clocks,
	}
	mtc.Start(t, numNodes)
	defer mtc.Stop()
	mtc.replicateRange(1, 1, 2)

	// Advance the lease holder's clock ahead of the followers (by more than
	// MaxOffset but less than the range lease) and execute a command.
	manuals[0].Increment(int64(500 * time.Millisecond))
	incArgs := incrementArgs([]byte("a"), 5)
	ts := clocks[0].Now()
	if _, err := client.SendWrappedWith(rg1(mtc.stores[0]), nil, roachpb.Header{Timestamp: ts}, &incArgs); err != nil {
		t.Fatal(err)
	}

	// Wait for that command to execute on all the followers.
	util.SucceedsSoon(t, func() error {
		values := []int64{}
		for _, eng := range mtc.engines {
			val, _, err := engine.MVCCGet(context.Background(), eng, roachpb.Key("a"), clocks[0].Now(), true, nil)
			if err != nil {
				return err
			}
			values = append(values, mustGetInt(val))
		}
		if !reflect.DeepEqual(values, []int64{5, 5, 5}) {
			return errors.Errorf("expected (5, 5, 5), got %v", values)
		}
		return nil
	})

	// Verify that all the followers have accepted the clock update from
	// node 0 even though it comes from outside the usual max offset.
	now := clocks[0].Now()
	for i, clock := range clocks {
		// Only compare the WallTimes: it's normal for clock 0 to be a few logical ticks ahead.
		if clock.Now().WallTime < now.WallTime {
			t.Errorf("clock %d is behind clock 0: %s vs %s", i, clock.Now(), now)
		}
	}
}

func writeRandomTimeSeriesDataToRange(
	t testing.TB,
	store *storage.Store,
	rangeID roachpb.RangeID,
	keyPrefix []byte,
) (midpoint []byte) {
	src := rand.New(rand.NewSource(0))
	r := ts.Resolution10s
	for i := 0; i < 20; i++ {
		var data []tspb.TimeSeriesData
		for j := int64(0); j <= src.Int63n(5); j++ {
			d := tspb.TimeSeriesData{
				Name:   "test.random.metric",
				Source: "cpu01",
			}
			for k := int64(0); k <= src.Int63n(10); k++ {
				d.Datapoints = append(d.Datapoints, tspb.TimeSeriesDatapoint{
					TimestampNanos: src.Int63n(200) * r.KeyDuration(),
					Value:          src.Float64(),
				})
			}
			data = append(data, d)
		}
		for _, d := range data {
			idatas, err := d.ToInternal(r.KeyDuration(), r.SampleDuration())
			if err != nil {
				t.Fatal(err)
			}
			for _, idata := range idatas {
				var value roachpb.Value
				if err := value.SetProto(&idata); err != nil {
					t.Fatal(err)
				}
				mArgs := roachpb.MergeRequest{
					Span: roachpb.Span{
						Key: encoding.EncodeVarintAscending(keyPrefix, idata.StartTimestampNanos),
					},
					Value: value,
				}
				if _, pErr := client.SendWrappedWith(rg1(store), nil, roachpb.Header{
					RangeID: rangeID,
				}, &mArgs); pErr != nil {
					t.Fatal(pErr)
				}
			}
		}
	}
	// Return approximate midway point (100 is midway between random timestamps in range [0,200)).
	midKey := append([]byte(nil), keyPrefix...)
	midKey = encoding.EncodeVarintAscending(midKey, 100*r.KeyDuration())
	return keys.MakeRowSentinelKey(midKey)
}

func fillTestRange(t testing.TB, rep *Replica, size int) {
	src := rand.New(rand.NewSource(0))
	for i := 0; i < snapSize/(keySize+valSize); i++ {
		key := keys.MakeRowSentinelKey(randutil.RandBytes(src, keySize))
		val := randutil.RandBytes(src, valSize)
		pArgs := putArgs(key, val)
		if _, pErr := client.SendWrappedWith(rep, nil, roachpb.Header{
			RangeID: rangeID,
		}, &pArgs); pErr != nil {
			t.Fatal(pErr)
		}
	}
}

// getTxn fetches the requested key and returns the transaction info.
func getTxn(coord *TxnCoordSender, txn *roachpb.Transaction) (*roachpb.Transaction, *roachpb.Error) {
	hb := &roachpb.HeartbeatTxnRequest{
		Span: roachpb.Span{
			Key: txn.Key,
		},
	}
	reply, pErr := client.SendWrappedWith(coord, nil, roachpb.Header{
		Txn: txn,
	}, hb)
	if pErr != nil {
		return nil, pErr
	}
	return reply.(*roachpb.HeartbeatTxnResponse).Txn, nil
}

func writeRandomDataToRange(t testing.TB, store *storage.Store, rangeID roachpb.RangeID, keyPrefix []byte) {
	src := rand.New(rand.NewSource(0))
	for i := 0; i < 100; i++ {
		key := append([]byte(nil), keyPrefix...)
		key = append(key, randutil.RandBytes(src, int(src.Int31n(1<<7)))...)
		key = keys.MakeRowSentinelKey(key)
		val := randutil.RandBytes(src, int(src.Int31n(1<<8)))
		pArgs := putArgs(key, val)
		if _, pErr := client.SendWrappedWith(rg1(store), nil, roachpb.Header{
			RangeID: rangeID,
		}, &pArgs); pErr != nil {
			t.Fatal(pErr)
		}
	}
}

// FindRangeLeaseHolder returns the current lease holder for the given range. If
// there is no lease at the time of the call, a replica is gets one as a
// side-effect of calling this; if hint is not nil, that replica will be the
// one.
//
// One of the Stores in the cluster is used as a Sender to send a dummy read
// command, which will either result in success (if a replica on that Node has
// the lease), in a NotLeaseHolderError pointing to the current lease holder (if
// there is an active lease), or in the replica on that store acquiring the
// lease (if there isn't an active lease).
// If an active lease existed for the range, it's extended as a side-effect.
func (tc *TestCluster) FindRangeLeaseHolder(
	rangeDesc *roachpb.RangeDescriptor,
	hint *ReplicationTarget,
) (ReplicationTarget, error) {
	var hintReplicaDesc roachpb.ReplicaDescriptor
	if hint != nil {
		var ok bool
		if hintReplicaDesc, ok = rangeDesc.GetReplicaDescriptor(hint.StoreID); !ok {
			return ReplicationTarget{}, errors.Errorf(
				"bad hint; store doesn't have a replica of the range")
		}
	} else {
		hint = &ReplicationTarget{
			NodeID:  rangeDesc.Replicas[0].NodeID,
			StoreID: rangeDesc.Replicas[0].StoreID}
		hintReplicaDesc = rangeDesc.Replicas[0]
	}
	// TODO(andrei): Using a dummy GetRequest for the purpose of figuring out the
	// lease holder is a hack. Instead, we should have a dedicate admin command.
	getReq := roachpb.GetRequest{
		Span: roachpb.Span{
			Key: rangeDesc.StartKey.AsRawKey(),
		},
	}

	store, err := tc.findMemberStore(hint.StoreID)
	if err != nil {
		return ReplicationTarget{}, err
	}
	_, pErr := client.SendWrappedWith(
		store, nil,
		roachpb.Header{RangeID: rangeDesc.RangeID, Replica: hintReplicaDesc},
		&getReq)
	if pErr != nil {
		if nle, ok := pErr.GetDetail().(*roachpb.NotLeaseHolderError); ok {
			if nle.LeaseHolder == nil {
				return ReplicationTarget{}, errors.Errorf(
					"unexpected NotLeaseHolderError with lease holder unknown")
			}
			return ReplicationTarget{
				NodeID: nle.LeaseHolder.NodeID, StoreID: nle.LeaseHolder.StoreID}, nil
		}
		return ReplicationTarget{}, pErr.GoError()
	}
	// The replica we sent the request to either was already or just became
	// the lease holder.
	return *hint, nil
}

func fillTestRange(t testing.TB, rep *Replica, size int64) {
	src := rand.New(rand.NewSource(0))
	for i := int64(0); i < size/int64(keySize+valSize); i++ {
		key := keys.MakeRowSentinelKey(randutil.RandBytes(src, keySize))
		val := randutil.RandBytes(src, valSize)
		pArgs := putArgs(key, val)
		if _, pErr := client.SendWrappedWith(rep, nil, roachpb.Header{
			RangeID: rangeID,
		}, &pArgs); pErr != nil {
			t.Fatal(pErr)
		}
	}
	rep.mu.Lock()
	after := rep.mu.state.Stats.Total()
	rep.mu.Unlock()
	if after < size {
		t.Fatalf("range not full after filling: wrote %d, but range at %d", size, after)
	}
}

// process synchronously invokes admin split for each proposed split key.
func (sq *splitQueue) process(
	ctx context.Context,
	now hlc.Timestamp,
	rng *Replica,
	sysCfg config.SystemConfig,
) error {
	// First handle case of splitting due to zone config maps.
	desc := rng.Desc()
	splitKeys := sysCfg.ComputeSplitKeys(desc.StartKey, desc.EndKey)
	if len(splitKeys) > 0 {
		log.Infof("splitting %s at keys %v", rng, splitKeys)
		log.Trace(ctx, fmt.Sprintf("splitting at keys %v", splitKeys))
		for _, splitKey := range splitKeys {
			if err := sq.db.AdminSplit(splitKey.AsRawKey()); err != nil {
				return errors.Errorf("unable to split %s at key %q: %s", rng, splitKey, err)
			}
		}
		return nil
	}

	// Next handle case of splitting due to size.
	zone, err := sysCfg.GetZoneConfigForKey(desc.StartKey)
	if err != nil {
		return err
	}
	size := rng.GetMVCCStats().Total()
	// FIXME: why is this implementation not the same as the one above?
	if float64(size)/float64(zone.RangeMaxBytes) > 1 {
		log.Infof("splitting %s size=%d max=%d", rng, size, zone.RangeMaxBytes)
		log.Trace(ctx, fmt.Sprintf("splitting size=%d max=%d", size, zone.RangeMaxBytes))
		if _, pErr := client.SendWrappedWith(rng, ctx, roachpb.Header{
			Timestamp: now,
		}, &roachpb.AdminSplitRequest{
			Span: roachpb.Span{Key: desc.StartKey.AsRawKey()},
		}); pErr != nil {
			return pErr.GoError()
		}
	}
	return nil
}

func BenchmarkReplicaSnapshot(b *testing.B) {
	defer tracing.Disable()()
	defer config.TestingDisableTableSplits()()
	store, stopper, _ := createTestStore(b)
	// We want to manually control the size of the raft log.
	store.DisableRaftLogQueue(true)
	defer stopper.Stop()

	const rangeID = 1
	const keySize = 1 << 7   // 128 B
	const valSize = 1 << 10  // 1 KiB
	const snapSize = 1 << 25 // 32 MiB

	rep, err := store.GetReplica(rangeID)
	if err != nil {
		b.Fatal(err)
	}

	src := rand.New(rand.NewSource(0))
	for i := 0; i < snapSize/(keySize+valSize); i++ {
		key := keys.MakeRowSentinelKey(randutil.RandBytes(src, keySize))
		val := randutil.RandBytes(src, valSize)
		pArgs := putArgs(key, val)
		if _, pErr := client.SendWrappedWith(rep, nil, roachpb.Header{
			RangeID: rangeID,
		}, &pArgs); pErr != nil {
			b.Fatal(pErr)
		}
	}

	b.ResetTimer()
	for i := 0; i < b.N; i++ {
		if _, err := rep.GetSnapshot(); err != nil {
			b.Fatal(err)
		}
	}
}

func writeRandomDataToRange(
	t testing.TB,
	store *storage.Store,
	rangeID roachpb.RangeID,
	keyPrefix []byte,
) (midpoint []byte) {
	src := rand.New(rand.NewSource(0))
	for i := 0; i < 100; i++ {
		key := append([]byte(nil), keyPrefix...)
		key = append(key, randutil.RandBytes(src, int(src.Int31n(1<<7)))...)
		key = keys.MakeRowSentinelKey(key)
		val := randutil.RandBytes(src, int(src.Int31n(1<<8)))
		pArgs := putArgs(key, val)
		if _, pErr := client.SendWrappedWith(rg1(store), nil, roachpb.Header{
			RangeID: rangeID,
		}, &pArgs); pErr != nil {
			t.Fatal(pErr)
		}
	}
	// Return approximate midway point ("Z" in string "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz").
	midKey := append([]byte(nil), keyPrefix...)
	midKey = append(midKey, []byte("Z")...)
	return keys.MakeRowSentinelKey(midKey)
}

//.........这里部分代码省略.........
	err := store.DB().Put(key, initVal)
	if err != nil {
		t.Fatalf("failed to put: %s", err)
	}

	waitPut := make(chan struct{})
	waitFirstGet := make(chan struct{})
	waitTxnRestart := make(chan struct{})
	waitSecondGet := make(chan struct{})
	waitTxnComplete := make(chan struct{})

	// Start the Writer.
	go func() {
		epoch := -1
		// Start a txn that does read-after-write.
		// The txn will be restarted twice, and the out-of-order put
		// will happen in the second epoch.
		if err := store.DB().Txn(func(txn *client.Txn) error {
			epoch++

			if epoch == 1 {
				// Wait until the second get operation is issued.
				close(waitTxnRestart)
				<-waitSecondGet
			}

			updatedVal := []byte("updatedVal")
			if err := txn.Put(key, updatedVal); err != nil {
				return err
			}

			// Make sure a get will return the value that was just written.
			actual, err := txn.Get(key)
			if err != nil {
				return err
			}
			if !bytes.Equal(actual.ValueBytes(), updatedVal) {
				t.Fatalf("unexpected get result: %s", actual)
			}

			if epoch == 0 {
				// Wait until the first get operation will push the txn timestamp.
				close(waitPut)
				<-waitFirstGet
			}

			b := txn.NewBatch()
			return txn.CommitInBatch(b)
		}); err != nil {
			t.Fatal(err)
		}

		if epoch != 2 {
			t.Fatalf("unexpected number of txn retries: %d", epoch)
		}

		close(waitTxnComplete)
	}()

	<-waitPut

	// Start the Reader.

	// Advance the clock and send a get operation with higher
	// priority to trigger the txn restart.
	manualClock.Increment(100)

	priority := roachpb.UserPriority(-math.MaxInt32)
	requestHeader := roachpb.Span{
		Key: roachpb.Key(key),
	}
	ts := clock.Now()
	if _, err := client.SendWrappedWith(rg1(store), nil, roachpb.Header{
		Timestamp:    ts,
		UserPriority: priority,
	}, &roachpb.GetRequest{Span: requestHeader}); err != nil {
		t.Fatalf("failed to get: %s", err)
	}

	// Wait until the writer restarts the txn.
	close(waitFirstGet)
	<-waitTxnRestart

	// Advance the clock and send a get operation again. This time
	// we use TestingCommandFilter so that a get operation is not
	// processed after the write intent is resolved (to prevent the
	// timestamp cache from being updated).
	manualClock.Increment(100)

	ts = clock.Now()
	if _, err := client.SendWrappedWith(rg1(store), nil, roachpb.Header{
		Timestamp:    ts,
		UserPriority: priority,
	}, &roachpb.GetRequest{Span: requestHeader}); err == nil {
		t.Fatal("unexpected success of get")
	}

	close(waitSecondGet)
	<-waitTxnComplete
}

// TestStoreRangeMergeMetadataCleanup tests that all metadata of a
// subsumed range is cleaned up on merge.
func TestStoreRangeMergeMetadataCleanup(t *testing.T) {
	defer leaktest.AfterTest(t)()
	sCtx := storage.TestStoreContext()
	sCtx.TestingKnobs.DisableSplitQueue = true
	store, stopper, _ := createTestStoreWithContext(t, sCtx)
	defer stopper.Stop()

	scan := func(f func(roachpb.KeyValue) (bool, error)) {
		if _, err := engine.MVCCIterate(context.Background(), store.Engine(), roachpb.KeyMin, roachpb.KeyMax, hlc.ZeroTimestamp, true, nil, false, f); err != nil {
			t.Fatal(err)
		}
	}
	content := roachpb.Key("testing!")

	// Write some values left of the proposed split key.
	pArgs := putArgs([]byte("aaa"), content)
	if _, err := client.SendWrapped(rg1(store), nil, &pArgs); err != nil {
		t.Fatal(err)
	}

	// Collect all the keys.
	preKeys := make(map[string]struct{})
	scan(func(kv roachpb.KeyValue) (bool, error) {
		preKeys[string(kv.Key)] = struct{}{}
		return false, nil
	})

	// Split the range.
	_, bDesc, err := createSplitRanges(store)
	if err != nil {
		t.Fatal(err)
	}

	// Write some values right of the split key.
	pArgs = putArgs([]byte("ccc"), content)
	if _, err := client.SendWrappedWith(rg1(store), nil, roachpb.Header{
		RangeID: bDesc.RangeID,
	}, &pArgs); err != nil {
		t.Fatal(err)
	}

	// Merge the b range back into the a range.
	args := adminMergeArgs(roachpb.KeyMin)
	if _, err := client.SendWrapped(rg1(store), nil, &args); err != nil {
		t.Fatal(err)
	}

	// Collect all the keys again.
	postKeys := make(map[string]struct{})
	scan(func(kv roachpb.KeyValue) (bool, error) {
		postKeys[string(kv.Key)] = struct{}{}
		return false, nil
	})

	// Compute the new keys.
	for k := range preKeys {
		delete(postKeys, k)
	}

	// Keep only the subsumed range's local keys.
	localRangeKeyPrefix := string(keys.MakeRangeIDPrefix(bDesc.RangeID))
	for k := range postKeys {
		if !strings.HasPrefix(k, localRangeKeyPrefix) {
			delete(postKeys, k)
		}
	}

	if numKeys := len(postKeys); numKeys > 0 {
		var buf bytes.Buffer
		fmt.Fprintf(&buf, "%d keys were not cleaned up:\n", numKeys)
		for k := range postKeys {
			fmt.Fprintf(&buf, "%q\n", k)
		}
		t.Fatal(buf.String())
	}
}

// TestStoreRangeMergeWithData attempts to merge two collocate ranges
// each containing data.
func TestStoreRangeMergeWithData(t *testing.T) {
	defer leaktest.AfterTest(t)()
	sCtx := storage.TestStoreContext()
	sCtx.TestingKnobs.DisableSplitQueue = true
	store, stopper, _ := createTestStoreWithContext(t, sCtx)
	defer stopper.Stop()

	content := roachpb.Key("testing!")

	aDesc, bDesc, err := createSplitRanges(store)
	if err != nil {
		t.Fatal(err)
	}

	// Write some values left and right of the proposed split key.
	pArgs := putArgs([]byte("aaa"), content)
	if _, err := client.SendWrapped(rg1(store), nil, &pArgs); err != nil {
		t.Fatal(err)
	}
	pArgs = putArgs([]byte("ccc"), content)
	if _, err := client.SendWrappedWith(rg1(store), nil, roachpb.Header{
		RangeID: bDesc.RangeID,
	}, &pArgs); err != nil {
		t.Fatal(err)
	}

	// Confirm the values are there.
	gArgs := getArgs([]byte("aaa"))
	if reply, err := client.SendWrapped(rg1(store), nil, &gArgs); err != nil {
		t.Fatal(err)
	} else if replyBytes, err := reply.(*roachpb.GetResponse).Value.GetBytes(); err != nil {
		t.Fatal(err)
	} else if !bytes.Equal(replyBytes, content) {
		t.Fatalf("actual value %q did not match expected value %q", replyBytes, content)
	}
	gArgs = getArgs([]byte("ccc"))
	if reply, err := client.SendWrappedWith(rg1(store), nil, roachpb.Header{
		RangeID: bDesc.RangeID,
	}, &gArgs); err != nil {
		t.Fatal(err)
	} else if replyBytes, err := reply.(*roachpb.GetResponse).Value.GetBytes(); err != nil {
		t.Fatal(err)
	} else if !bytes.Equal(replyBytes, content) {
		t.Fatalf("actual value %q did not match expected value %q", replyBytes, content)
	}

	// Merge the b range back into the a range.
	args := adminMergeArgs(roachpb.KeyMin)
	if _, err := client.SendWrapped(rg1(store), nil, &args); err != nil {
		t.Fatal(err)
	}

	// Verify no intents remains on range descriptor keys.
	for _, key := range []roachpb.Key{keys.RangeDescriptorKey(aDesc.StartKey), keys.RangeDescriptorKey(bDesc.StartKey)} {
		if _, _, err := engine.MVCCGet(context.Background(), store.Engine(), key, store.Clock().Now(), true, nil); err != nil {
			t.Fatal(err)
		}
	}

	// Verify the merge by looking up keys from both ranges.
	rangeA := store.LookupReplica([]byte("a"), nil)
	rangeB := store.LookupReplica([]byte("c"), nil)
	rangeADesc := rangeA.Desc()
	rangeBDesc := rangeB.Desc()

	if !reflect.DeepEqual(rangeA, rangeB) {
		t.Fatalf("ranges were not merged %+v=%+v", rangeADesc, rangeBDesc)
	}
	if !bytes.Equal(rangeADesc.StartKey, roachpb.RKeyMin) {
		t.Fatalf("The start key is not equal to KeyMin %q=%q", rangeADesc.StartKey, roachpb.RKeyMin)
	}
	if !bytes.Equal(rangeADesc.EndKey, roachpb.RKeyMax) {
		t.Fatalf("The end key is not equal to KeyMax %q=%q", rangeADesc.EndKey, roachpb.RKeyMax)
	}

	// Try to get values from after the merge.
	gArgs = getArgs([]byte("aaa"))
	if reply, err := client.SendWrapped(rg1(store), nil, &gArgs); err != nil {
		t.Fatal(err)
	} else if replyBytes, err := reply.(*roachpb.GetResponse).Value.GetBytes(); err != nil {
		t.Fatal(err)
	} else if !bytes.Equal(replyBytes, content) {
		t.Fatalf("actual value %q did not match expected value %q", replyBytes, content)
	}
	gArgs = getArgs([]byte("ccc"))
	if reply, err := client.SendWrappedWith(rg1(store), nil, roachpb.Header{
		RangeID: rangeB.RangeID,
	}, &gArgs); err != nil {
		t.Fatal(err)
	} else if replyBytes, err := reply.(*roachpb.GetResponse).Value.GetBytes(); err != nil {
		t.Fatal(err)
	} else if !bytes.Equal(replyBytes, content) {
		t.Fatalf("actual value %q did not match expected value %q", replyBytes, content)
	}

	// Put new values after the merge on both sides.
	pArgs = putArgs([]byte("aaaa"), content)
	if _, err := client.SendWrapped(rg1(store), nil, &pArgs); err != nil {
//.........这里部分代码省略.........

// TestMultiRangeScanWithMaxResults tests that commands which access multiple
// ranges with MaxResults parameter are carried out properly.
func TestMultiRangeScanWithMaxResults(t *testing.T) {
	defer leaktest.AfterTest(t)()
	testCases := []struct {
		splitKeys []roachpb.Key
		keys      []roachpb.Key
	}{
		{[]roachpb.Key{roachpb.Key("m")},
			[]roachpb.Key{roachpb.Key("a"), roachpb.Key("z")}},
		{[]roachpb.Key{roachpb.Key("h"), roachpb.Key("q")},
			[]roachpb.Key{roachpb.Key("b"), roachpb.Key("f"), roachpb.Key("k"),
				roachpb.Key("r"), roachpb.Key("w"), roachpb.Key("y")}},
	}

	for i, tc := range testCases {
		s, _, _ := serverutils.StartServer(t, base.TestServerArgs{})
		defer s.Stopper().Stop()
		ts := s.(*TestServer)
		retryOpts := base.DefaultRetryOptions()
		retryOpts.Closer = ts.stopper.ShouldQuiesce()
		ds := kv.NewDistSender(&kv.DistSenderConfig{
			Clock:           s.Clock(),
			RPCContext:      s.RPCContext(),
			RPCRetryOptions: &retryOpts,
		}, ts.Gossip())
		ctx := tracing.WithTracer(context.Background(), tracing.NewTracer())
		tds := kv.NewTxnCoordSender(ctx, ds, ts.Clock(), ts.Ctx.Linearizable,
			ts.stopper, kv.MakeTxnMetrics())

		for _, sk := range tc.splitKeys {
			if err := ts.node.ctx.DB.AdminSplit(sk); err != nil {
				t.Fatal(err)
			}
		}

		for _, k := range tc.keys {
			put := roachpb.NewPut(k, roachpb.MakeValueFromBytes(k))
			if _, err := client.SendWrapped(tds, nil, put); err != nil {
				t.Fatal(err)
			}
		}

		// Try every possible ScanRequest startKey.
		for start := 0; start < len(tc.keys); start++ {
			// Try every possible maxResults, from 1 to beyond the size of key array.
			for maxResults := 1; maxResults <= len(tc.keys)-start+1; maxResults++ {
				scan := roachpb.NewScan(tc.keys[start], tc.keys[len(tc.keys)-1].Next())
				reply, err := client.SendWrappedWith(
					tds, nil, roachpb.Header{MaxSpanRequestKeys: int64(maxResults)}, scan,
				)
				if err != nil {
					t.Fatal(err)
				}
				rows := reply.(*roachpb.ScanResponse).Rows
				if start+maxResults <= len(tc.keys) && len(rows) != maxResults {
					t.Errorf("%d: start=%s: expected %d rows, but got %d", i, tc.keys[start], maxResults, len(rows))
				} else if start+maxResults == len(tc.keys)+1 && len(rows) != maxResults-1 {
					t.Errorf("%d: expected %d rows, but got %d", i, maxResults-1, len(rows))
				}
			}
		}
	}
}

// TestMultiRangeScanDeleteRange tests that commands which access multiple
// ranges are carried out properly.
func TestMultiRangeScanDeleteRange(t *testing.T) {
	defer leaktest.AfterTest(t)()
	s, _, _ := serverutils.StartServer(t, base.TestServerArgs{})
	defer s.Stopper().Stop()
	ts := s.(*TestServer)
	retryOpts := base.DefaultRetryOptions()
	retryOpts.Closer = ts.stopper.ShouldQuiesce()
	ds := kv.NewDistSender(&kv.DistSenderConfig{
		Clock:           s.Clock(),
		RPCContext:      s.RPCContext(),
		RPCRetryOptions: &retryOpts,
	}, ts.Gossip())
	ctx := tracing.WithTracer(context.Background(), tracing.NewTracer())
	tds := kv.NewTxnCoordSender(ctx, ds, s.Clock(), ts.Ctx.Linearizable,
		ts.stopper, kv.MakeTxnMetrics())

	if err := ts.node.ctx.DB.AdminSplit("m"); err != nil {
		t.Fatal(err)
	}
	writes := []roachpb.Key{roachpb.Key("a"), roachpb.Key("z")}
	get := &roachpb.GetRequest{
		Span: roachpb.Span{Key: writes[0]},
	}
	get.EndKey = writes[len(writes)-1]
	if _, err := client.SendWrapped(tds, nil, get); err == nil {
		t.Errorf("able to call Get with a key range: %v", get)
	}
	var delTS hlc.Timestamp
	for i, k := range writes {
		put := roachpb.NewPut(k, roachpb.MakeValueFromBytes(k))
		reply, err := client.SendWrapped(tds, nil, put)
		if err != nil {
			t.Fatal(err)
		}
		scan := roachpb.NewScan(writes[0], writes[len(writes)-1].Next())
		reply, err = client.SendWrapped(tds, nil, scan)
		if err != nil {
			t.Fatal(err)
		}
		sr := reply.(*roachpb.ScanResponse)
		if sr.Txn != nil {
			// This was the other way around at some point in the past.
			// Same below for Delete, etc.
			t.Errorf("expected no transaction in response header")
		}
		if rows := sr.Rows; len(rows) != i+1 {
			t.Fatalf("expected %d rows, but got %d", i+1, len(rows))
		}
	}

	del := &roachpb.DeleteRangeRequest{
		Span: roachpb.Span{
			Key:    writes[0],
			EndKey: roachpb.Key(writes[len(writes)-1]).Next(),
		},
		ReturnKeys: true,
	}
	reply, err := client.SendWrappedWith(tds, nil, roachpb.Header{Timestamp: delTS}, del)
	if err != nil {
		t.Fatal(err)
	}
	dr := reply.(*roachpb.DeleteRangeResponse)
	if dr.Txn != nil {
		t.Errorf("expected no transaction in response header")
	}
	if !reflect.DeepEqual(dr.Keys, writes) {
		t.Errorf("expected %d keys to be deleted, but got %d instead", writes, dr.Keys)
	}

	scan := roachpb.NewScan(writes[0], writes[len(writes)-1].Next())
	txn := &roachpb.Transaction{Name: "MyTxn"}
	reply, err = client.SendWrappedWith(tds, nil, roachpb.Header{Txn: txn}, scan)
	if err != nil {
		t.Fatal(err)
	}
	sr := reply.(*roachpb.ScanResponse)
	if txn := sr.Txn; txn == nil || txn.Name != "MyTxn" {
		t.Errorf("wanted Txn to persist, but it changed to %v", txn)
	}
	if rows := sr.Rows; len(rows) > 0 {
		t.Fatalf("scan after delete returned rows: %v", rows)
	}
}

// TestStoreRangeSplitStatsWithMerges starts by splitting the system keys from
// user-space keys and verifying that the user space side of the split (which is empty),
// has all zeros for stats. It then issues a number of Merge requests to the user
// space side, simulating TimeSeries data. Finally, the test splits the user space
// side halfway and verifies the stats on either side of the split are equal to a
// recomputation.
//
// Note that unlike TestStoreRangeSplitStats, we do not check if the two halves of the
// split's stats are equal to the pre-split stats when added, because this will not be
// true of ranges populated with Merge requests. The reason for this is that Merge
// requests' impact on MVCCStats are only estimated. See updateStatsOnMerge.
func TestStoreRangeSplitStatsWithMerges(t *testing.T) {
	defer leaktest.AfterTest(t)()
	sCtx := storage.TestStoreContext()
	sCtx.TestingKnobs.DisableSplitQueue = true
	store, stopper, manual := createTestStoreWithContext(t, sCtx)
	defer stopper.Stop()

	// Split the range after the last table data key.
	keyPrefix := keys.MakeTablePrefix(keys.MaxReservedDescID + 1)
	keyPrefix = keys.MakeRowSentinelKey(keyPrefix)
	args := adminSplitArgs(roachpb.KeyMin, keyPrefix)
	if _, pErr := client.SendWrapped(rg1(store), nil, &args); pErr != nil {
		t.Fatal(pErr)
	}
	// Verify empty range has empty stats.
	rng := store.LookupReplica(keyPrefix, nil)
	// NOTE that this value is expected to change over time, depending on what
	// we store in the sys-local keyspace. Update it accordingly for this test.
	empty := enginepb.MVCCStats{LastUpdateNanos: manual.UnixNano()}
	if err := verifyRangeStats(store.Engine(), rng.RangeID, empty); err != nil {
		t.Fatal(err)
	}

	// Write random TimeSeries data.
	midKey := writeRandomTimeSeriesDataToRange(t, store, rng.RangeID, keyPrefix)
	manual.Increment(100)

	// Split the range at approximate halfway point.
	args = adminSplitArgs(keyPrefix, midKey)
	if _, pErr := client.SendWrappedWith(rg1(store), nil, roachpb.Header{
		RangeID: rng.RangeID,
	}, &args); pErr != nil {
		t.Fatal(pErr)
	}

	snap := store.Engine().NewSnapshot()
	defer snap.Close()
	var msLeft, msRight enginepb.MVCCStats
	if err := engine.MVCCGetRangeStats(context.Background(), snap, rng.RangeID, &msLeft); err != nil {
		t.Fatal(err)
	}
	rngRight := store.LookupReplica(midKey, nil)
	if err := engine.MVCCGetRangeStats(context.Background(), snap, rngRight.RangeID, &msRight); err != nil {
		t.Fatal(err)
	}

	// Stats should both have the new timestamp.
	now := manual.UnixNano()
	if lTs := msLeft.LastUpdateNanos; lTs != now {
		t.Errorf("expected left range stats to have new timestamp, want %d, got %d", now, lTs)
	}
	if rTs := msRight.LastUpdateNanos; rTs != now {
		t.Errorf("expected right range stats to have new timestamp, want %d, got %d", now, rTs)
	}

	// Stats should agree with recomputation.
	if err := verifyRecomputedStats(snap, rng.Desc(), msLeft, now); err != nil {
		t.Fatalf("failed to verify left range's stats after split: %v", err)
	}
	if err := verifyRecomputedStats(snap, rngRight.Desc(), msRight, now); err != nil {
		t.Fatalf("failed to verify right range's stats after split: %v", err)
	}
}