Merge pull request #8102 from heyitsanthony/txn-nested

api: nested txns

Documentation/dev-guide/api_reference_v3.md

@@ -669,6 +669,7 @@ Empty field.
 | request_range |  | RangeRequest |
 | request_put |  | PutRequest |
 | request_delete_range |  | DeleteRangeRequest |
+| request_txn |  | TxnRequest |
 
 
 
@@ -691,6 +692,7 @@ Empty field.
 | response_range |  | RangeResponse |
 | response_put |  | PutResponse |
 | response_delete_range |  | DeleteRangeResponse |
+| response_txn |  | TxnResponse |
 
 
 

Documentation/dev-guide/apispec/swagger/rpc.swagger.json

@@ -1954,6 +1954,9 @@
         },
         "request_range": {
           "$ref": "#/definitions/etcdserverpbRangeRequest"
+        },
+        "request_txn": {
+          "$ref": "#/definitions/etcdserverpbTxnRequest"
         }
       }
     },
@@ -1993,6 +1996,9 @@
         },
         "response_range": {
           "$ref": "#/definitions/etcdserverpbRangeResponse"
+        },
+        "response_txn": {
+          "$ref": "#/definitions/etcdserverpbTxnResponse"
         }
       }
     },

clientv3/integration/txn_test.go

@@ -106,27 +106,35 @@ func TestTxnReadRetry(t *testing.T) {
 	defer clus.Terminate(t)
 
 	kv := clus.Client(0)
-	clus.Members[0].Stop(t)
-	<-clus.Members[0].StopNotify()
 
-	donec := make(chan struct{})
+	thenOps := [][]clientv3.Op{
-	go func() {
+		{clientv3.OpGet("foo")},
-		ctx := context.TODO()
+		{clientv3.OpTxn(nil, []clientv3.Op{clientv3.OpGet("foo")}, nil)},
-		_, err := kv.Txn(ctx).Then(clientv3.OpGet("foo")).Commit()
+		{clientv3.OpTxn(nil, nil, nil)},
-		if err != nil {
+		{},
-			t.Fatalf("expected response, got error %v", err)
+	}
+	for i := range thenOps {
+		clus.Members[0].Stop(t)
+		<-clus.Members[0].StopNotify()
+
+		donec := make(chan struct{})
+		go func() {
+			_, err := kv.Txn(context.TODO()).Then(thenOps[i]...).Commit()
+			if err != nil {
+				t.Fatalf("expected response, got error %v", err)
+			}
+			donec <- struct{}{}
+		}()
+		// wait for txn to fail on disconnect
+		time.Sleep(100 * time.Millisecond)
+
+		// restart node; client should resume
+		clus.Members[0].Restart(t)
+		select {
+		case <-donec:
+		case <-time.After(2 * clus.Members[1].ServerConfig.ReqTimeout()):
+			t.Fatalf("waited too long")
 		}
-		donec <- struct{}{}
-	}()
-	// wait for txn to fail on disconnect
-	time.Sleep(100 * time.Millisecond)
-
-	// restart node; client should resume
-	clus.Members[0].Restart(t)
-	select {
-	case <-donec:
-	case <-time.After(2 * clus.Members[1].ServerConfig.ReqTimeout()):
-		t.Fatalf("waited too long")
 	}
 }
 
@@ -179,3 +187,41 @@ func TestTxnCompareRange(t *testing.T) {
 		t.Fatal("expected prefix compare to false, got compares as true")
 	}
 }
+
+func TestTxnNested(t *testing.T) {
+	defer testutil.AfterTest(t)
+
+	clus := integration.NewClusterV3(t, &integration.ClusterConfig{Size: 3})
+	defer clus.Terminate(t)
+
+	kv := clus.Client(0)
+
+	tresp, err := kv.Txn(context.TODO()).
+		If(clientv3.Compare(clientv3.Version("foo"), "=", 0)).
+		Then(
+			clientv3.OpPut("foo", "bar"),
+			clientv3.OpTxn(nil, []clientv3.Op{clientv3.OpPut("abc", "123")}, nil)).
+		Else(clientv3.OpPut("foo", "baz")).Commit()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(tresp.Responses) != 2 {
+		t.Errorf("expected 2 top-level txn responses, got %+v", tresp.Responses)
+	}
+
+	// check txn writes were applied
+	resp, err := kv.Get(context.TODO(), "foo")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(resp.Kvs) != 1 || string(resp.Kvs[0].Value) != "bar" {
+		t.Errorf("unexpected Get response %+v", resp)
+	}
+	resp, err = kv.Get(context.TODO(), "abc")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(resp.Kvs) != 1 || string(resp.Kvs[0].Value) != "123" {
+		t.Errorf("unexpected Get response %+v", resp)
+	}
+}

clientv3/kv.go

@@ -66,11 +66,13 @@ type OpResponse struct {
 	put *PutResponse
 	get *GetResponse
 	del *DeleteResponse
+	txn *TxnResponse
 }
 
 func (op OpResponse) Put() *PutResponse    { return op.put }
 func (op OpResponse) Get() *GetResponse    { return op.get }
 func (op OpResponse) Del() *DeleteResponse { return op.del }
+func (op OpResponse) Txn() *TxnResponse    { return op.txn }
 
 type kv struct {
 	remote pb.KVClient
@@ -134,7 +136,6 @@ func (kv *kv) Do(ctx context.Context, op Op) (OpResponse, error) {
 func (kv *kv) do(ctx context.Context, op Op) (OpResponse, error) {
 	var err error
 	switch op.t {
-	// TODO: handle other ops
 	case tRange:
 		var resp *pb.RangeResponse
 		resp, err = kv.remote.Range(ctx, op.toRangeRequest(), grpc.FailFast(false))
@@ -155,6 +156,12 @@ func (kv *kv) do(ctx context.Context, op Op) (OpResponse, error) {
 		if err == nil {
 			return OpResponse{del: (*DeleteResponse)(resp)}, nil
 		}
+	case tTxn:
+		var resp *pb.TxnResponse
+		resp, err = kv.remote.Txn(ctx, op.toTxnRequest())
+		if err == nil {
+			return OpResponse{txn: (*TxnResponse)(resp)}, nil
+		}
 	default:
 		panic("Unknown op")
 	}

clientv3/namespace/kv.go

@@ -74,7 +74,7 @@ func (kv *kvPrefix) Delete(ctx context.Context, key string, opts ...clientv3.OpO
 }
 
 func (kv *kvPrefix) Do(ctx context.Context, op clientv3.Op) (clientv3.OpResponse, error) {
-	if len(op.KeyBytes()) == 0 {
+	if len(op.KeyBytes()) == 0 && !op.IsTxn() {
 		return clientv3.OpResponse{}, rpctypes.ErrEmptyKey
 	}
 	r, err := kv.KV.Do(ctx, kv.prefixOp(op))
@@ -88,6 +88,8 @@ func (kv *kvPrefix) Do(ctx context.Context, op clientv3.Op) (clientv3.OpResponse
 		kv.unprefixPutResponse(r.Put())
 	case r.Del() != nil:
 		kv.unprefixDeleteResponse(r.Del())
+	case r.Txn() != nil:
+		kv.unprefixTxnResponse(r.Txn())
 	}
 	return r, nil
 }
@@ -102,34 +104,17 @@ func (kv *kvPrefix) Txn(ctx context.Context) clientv3.Txn {
 }
 
 func (txn *txnPrefix) If(cs ...clientv3.Cmp) clientv3.Txn {
-	newCmps := make([]clientv3.Cmp, len(cs))
+	txn.Txn = txn.Txn.If(txn.kv.prefixCmps(cs)...)
-	for i := range cs {
-		newCmps[i] = cs[i]
-		pfxKey, endKey := txn.kv.prefixInterval(cs[i].KeyBytes(), cs[i].RangeEnd)
-		newCmps[i].WithKeyBytes(pfxKey)
-		if len(cs[i].RangeEnd) != 0 {
-			newCmps[i].RangeEnd = endKey
-		}
-	}
-	txn.Txn = txn.Txn.If(newCmps...)
 	return txn
 }
 
 func (txn *txnPrefix) Then(ops ...clientv3.Op) clientv3.Txn {
-	newOps := make([]clientv3.Op, len(ops))
+	txn.Txn = txn.Txn.Then(txn.kv.prefixOps(ops)...)
-	for i := range ops {
-		newOps[i] = txn.kv.prefixOp(ops[i])
-	}
-	txn.Txn = txn.Txn.Then(newOps...)
 	return txn
 }
 
 func (txn *txnPrefix) Else(ops ...clientv3.Op) clientv3.Txn {
-	newOps := make([]clientv3.Op, len(ops))
+	txn.Txn = txn.Txn.Else(txn.kv.prefixOps(ops)...)
-	for i := range ops {
-		newOps[i] = txn.kv.prefixOp(ops[i])
-	}
-	txn.Txn = txn.Txn.Else(newOps...)
 	return txn
 }
 
@@ -143,10 +128,14 @@ func (txn *txnPrefix) Commit() (*clientv3.TxnResponse, error) {
 }
 
 func (kv *kvPrefix) prefixOp(op clientv3.Op) clientv3.Op {
-	begin, end := kv.prefixInterval(op.KeyBytes(), op.RangeBytes())
+	if !op.IsTxn() {
-	op.WithKeyBytes(begin)
+		begin, end := kv.prefixInterval(op.KeyBytes(), op.RangeBytes())
-	op.WithRangeBytes(end)
+		op.WithKeyBytes(begin)
-	return op
+		op.WithRangeBytes(end)
+		return op
+	}
+	cmps, thenOps, elseOps := op.Txn()
+	return clientv3.OpTxn(kv.prefixCmps(cmps), kv.prefixOps(thenOps), kv.prefixOps(elseOps))
 }
 
 func (kv *kvPrefix) unprefixGetResponse(resp *clientv3.GetResponse) {
@@ -182,6 +171,10 @@ func (kv *kvPrefix) unprefixTxnResponse(resp *clientv3.TxnResponse) {
 			if tv.ResponseDeleteRange != nil {
 				kv.unprefixDeleteResponse((*clientv3.DeleteResponse)(tv.ResponseDeleteRange))
 			}
+		case *pb.ResponseOp_ResponseTxn:
+			if tv.ResponseTxn != nil {
+				kv.unprefixTxnResponse((*clientv3.TxnResponse)(tv.ResponseTxn))
+			}
 		default:
 		}
 	}
@@ -190,3 +183,24 @@ func (kv *kvPrefix) unprefixTxnResponse(resp *clientv3.TxnResponse) {
 func (p *kvPrefix) prefixInterval(key, end []byte) (pfxKey []byte, pfxEnd []byte) {
 	return prefixInterval(p.pfx, key, end)
 }
+
+func (kv *kvPrefix) prefixCmps(cs []clientv3.Cmp) []clientv3.Cmp {
+	newCmps := make([]clientv3.Cmp, len(cs))
+	for i := range cs {
+		newCmps[i] = cs[i]
+		pfxKey, endKey := kv.prefixInterval(cs[i].KeyBytes(), cs[i].RangeEnd)
+		newCmps[i].WithKeyBytes(pfxKey)
+		if len(cs[i].RangeEnd) != 0 {
+			newCmps[i].RangeEnd = endKey
+		}
+	}
+	return newCmps
+}
+
+func (kv *kvPrefix) prefixOps(ops []clientv3.Op) []clientv3.Op {
+	newOps := make([]clientv3.Op, len(ops))
+	for i := range ops {
+		newOps[i] = kv.prefixOp(ops[i])
+	}
+	return newOps
+}

clientv3/op.go

@@ -23,6 +23,7 @@ const (
 	tRange opType = iota + 1
 	tPut
 	tDeleteRange
+	tTxn
 )
 
 var (
@@ -67,10 +68,18 @@ type Op struct {
 	// for put
 	val     []byte
 	leaseID LeaseID
+
+	// txn
+	cmps    []Cmp
+	thenOps []Op
+	elseOps []Op
 }
 
 // accesors / mutators
 
+func (op Op) IsTxn() bool              { return op.t == tTxn }
+func (op Op) Txn() ([]Cmp, []Op, []Op) { return op.cmps, op.thenOps, op.elseOps }
+
 // KeyBytes returns the byte slice holding the Op's key.
 func (op Op) KeyBytes() []byte { return op.key }
 
@@ -113,6 +122,22 @@ func (op Op) toRangeRequest() *pb.RangeRequest {
 	return r
 }
 
+func (op Op) toTxnRequest() *pb.TxnRequest {
+	thenOps := make([]*pb.RequestOp, len(op.thenOps))
+	for i, tOp := range op.thenOps {
+		thenOps[i] = tOp.toRequestOp()
+	}
+	elseOps := make([]*pb.RequestOp, len(op.elseOps))
+	for i, eOp := range op.elseOps {
+		elseOps[i] = eOp.toRequestOp()
+	}
+	cmps := make([]*pb.Compare, len(op.cmps))
+	for i := range op.cmps {
+		cmps[i] = (*pb.Compare)(&op.cmps[i])
+	}
+	return &pb.TxnRequest{Compare: cmps, Success: thenOps, Failure: elseOps}
+}
+
 func (op Op) toRequestOp() *pb.RequestOp {
 	switch op.t {
 	case tRange:
@@ -123,12 +148,27 @@ func (op Op) toRequestOp() *pb.RequestOp {
 	case tDeleteRange:
 		r := &pb.DeleteRangeRequest{Key: op.key, RangeEnd: op.end, PrevKv: op.prevKV}
 		return &pb.RequestOp{Request: &pb.RequestOp_RequestDeleteRange{RequestDeleteRange: r}}
+	case tTxn:
+		return &pb.RequestOp{Request: &pb.RequestOp_RequestTxn{RequestTxn: op.toTxnRequest()}}
 	default:
 		panic("Unknown Op")
 	}
 }
 
 func (op Op) isWrite() bool {
+	if op.t == tTxn {
+		for _, tOp := range op.thenOps {
+			if tOp.isWrite() {
+				return true
+			}
+		}
+		for _, tOp := range op.elseOps {
+			if tOp.isWrite() {
+				return true
+			}
+		}
+		return false
+	}
 	return op.t != tRange
 }
 
@@ -194,6 +234,10 @@ func OpPut(key, val string, opts ...OpOption) Op {
 	return ret
 }
 
+func OpTxn(cmps []Cmp, thenOps []Op, elseOps []Op) Op {
+	return Op{t: tTxn, cmps: cmps, thenOps: thenOps, elseOps: elseOps}
+}
+
 func opWatch(key string, opts ...OpOption) Op {
 	ret := Op{t: tRange, key: []byte(key)}
 	ret.applyOpts(opts)

etcdserver/api/v3rpc/key.go

@@ -16,11 +16,10 @@
 package v3rpc
 
 import (
-	"sort"
-
 	"github.com/coreos/etcd/etcdserver"
 	"github.com/coreos/etcd/etcdserver/api/v3rpc/rpctypes"
 	pb "github.com/coreos/etcd/etcdserver/etcdserverpb"
+	"github.com/coreos/etcd/pkg/adt"
 	"github.com/coreos/pkg/capnslog"
 	"golang.org/x/net/context"
 )
@@ -89,6 +88,13 @@ func (s *kvServer) Txn(ctx context.Context, r *pb.TxnRequest) (*pb.TxnResponse,
 	if err := checkTxnRequest(r, int(s.maxTxnOps)); err != nil {
 		return nil, err
 	}
+	// check for forbidden put/del overlaps after checking request to avoid quadratic blowup
+	if _, _, err := checkIntervals(r.Success); err != nil {
+		return nil, err
+	}
+	if _, _, err := checkIntervals(r.Failure); err != nil {
+		return nil, err
+	}
 
 	resp, err := s.kv.Txn(ctx, r)
 	if err != nil {
@@ -137,7 +143,14 @@ func checkDeleteRequest(r *pb.DeleteRangeRequest) error {
 }
 
 func checkTxnRequest(r *pb.TxnRequest, maxTxnOps int) error {
-	if len(r.Compare) > maxTxnOps || len(r.Success) > maxTxnOps || len(r.Failure) > maxTxnOps {
+	opc := len(r.Compare)
+	if opc < len(r.Success) {
+		opc = len(r.Success)
+	}
+	if opc < len(r.Failure) {
+		opc = len(r.Failure)
+	}
+	if opc > maxTxnOps {
 		return rpctypes.ErrGRPCTooManyOps
 	}
 
@@ -146,58 +159,29 @@ func checkTxnRequest(r *pb.TxnRequest, maxTxnOps int) error {
 			return rpctypes.ErrGRPCEmptyKey
 		}
 	}
-
 	for _, u := range r.Success {
-		if err := checkRequestOp(u); err != nil {
+		if err := checkRequestOp(u, maxTxnOps-opc); err != nil {
 			return err
 		}
 	}
-	if err := checkRequestDupKeys(r.Success); err != nil {
+	for _, u := range r.Failure {
-		return err
+		if err := checkRequestOp(u, maxTxnOps-opc); err != nil {
+			return err
+		}
 	}
 
-	for _, u := range r.Failure {
+	return nil
-		if err := checkRequestOp(u); err != nil {
-			return err
-		}
-	}
-	return checkRequestDupKeys(r.Failure)
 }
 
-// checkRequestDupKeys gives rpctypes.ErrGRPCDuplicateKey if the same key is modified twice
+// checkIntervals tests whether puts and deletes overlap for a list of ops. If
-func checkRequestDupKeys(reqs []*pb.RequestOp) error {
+// there is an overlap, returns an error. If no overlap, return put and delete
-	// check put overlap
+// sets for recursive evaluation.
-	keys := make(map[string]struct{})
+func checkIntervals(reqs []*pb.RequestOp) (map[string]struct{}, adt.IntervalTree, error) {
-	for _, requ := range reqs {
+	var dels adt.IntervalTree
-		tv, ok := requ.Request.(*pb.RequestOp_RequestPut)
-		if !ok {
-			continue
-		}
-		preq := tv.RequestPut
-		if preq == nil {
-			continue
-		}
-		if _, ok := keys[string(preq.Key)]; ok {
-			return rpctypes.ErrGRPCDuplicateKey
-		}
-		keys[string(preq.Key)] = struct{}{}
-	}
 
-	// no need to check deletes if no puts; delete overlaps are permitted
+	// collect deletes from this level; build first to check lower level overlapped puts
-	if len(keys) == 0 {
+	for _, req := range reqs {
-		return nil
+		tv, ok := req.Request.(*pb.RequestOp_RequestDeleteRange)
-	}
-
-	// sort keys for range checking
-	sortedKeys := []string{}
-	for k := range keys {
-		sortedKeys = append(sortedKeys, k)
-	}
-	sort.Strings(sortedKeys)
-
-	// check put overlap with deletes
-	for _, requ := range reqs {
-		tv, ok := requ.Request.(*pb.RequestOp_RequestDeleteRange)
 		if !ok {
 			continue
 		}
@@ -205,41 +189,87 @@ func checkRequestDupKeys(reqs []*pb.RequestOp) error {
 		if dreq == nil {
 			continue
 		}
-		if dreq.RangeEnd == nil {
+		var iv adt.Interval
-			if _, found := keys[string(dreq.Key)]; found {
+		if len(dreq.RangeEnd) != 0 {
-				return rpctypes.ErrGRPCDuplicateKey
+			iv = adt.NewStringAffineInterval(string(dreq.Key), string(dreq.RangeEnd))
-			}
 		} else {
-			lo := sort.SearchStrings(sortedKeys, string(dreq.Key))
+			iv = adt.NewStringAffinePoint(string(dreq.Key))
-			hi := sort.SearchStrings(sortedKeys, string(dreq.RangeEnd))
-			if lo != hi {
-				// element between lo and hi => overlap
-				return rpctypes.ErrGRPCDuplicateKey
-			}
 		}
+		dels.Insert(iv, struct{}{})
 	}
 
-	return nil
+	// collect children puts/deletes
+	puts := make(map[string]struct{})
+	for _, req := range reqs {
+		tv, ok := req.Request.(*pb.RequestOp_RequestTxn)
+		if !ok {
+			continue
+		}
+		putsThen, delsThen, err := checkIntervals(tv.RequestTxn.Success)
+		if err != nil {
+			return nil, dels, err
+		}
+		putsElse, delsElse, err := checkIntervals(tv.RequestTxn.Failure)
+		if err != nil {
+			return nil, dels, err
+		}
+		for k := range putsThen {
+			if _, ok := puts[k]; ok {
+				return nil, dels, rpctypes.ErrGRPCDuplicateKey
+			}
+			if dels.Intersects(adt.NewStringAffinePoint(k)) {
+				return nil, dels, rpctypes.ErrGRPCDuplicateKey
+			}
+			puts[k] = struct{}{}
+		}
+		for k := range putsElse {
+			if _, ok := puts[k]; ok {
+				// if key is from putsThen, overlap is OK since
+				// either then/else are mutually exclusive
+				if _, isSafe := putsThen[k]; !isSafe {
+					return nil, dels, rpctypes.ErrGRPCDuplicateKey
+				}
+			}
+			if dels.Intersects(adt.NewStringAffinePoint(k)) {
+				return nil, dels, rpctypes.ErrGRPCDuplicateKey
+			}
+			puts[k] = struct{}{}
+		}
+		dels.Union(delsThen, adt.NewStringAffineInterval("\x00", ""))
+		dels.Union(delsElse, adt.NewStringAffineInterval("\x00", ""))
+	}
+
+	// collect and check this level's puts
+	for _, req := range reqs {
+		tv, ok := req.Request.(*pb.RequestOp_RequestPut)
+		if !ok || tv.RequestPut == nil {
+			continue
+		}
+		k := string(tv.RequestPut.Key)
+		if _, ok := puts[k]; ok {
+			return nil, dels, rpctypes.ErrGRPCDuplicateKey
+		}
+		if dels.Intersects(adt.NewStringAffinePoint(k)) {
+			return nil, dels, rpctypes.ErrGRPCDuplicateKey
+		}
+		puts[k] = struct{}{}
+	}
+	return puts, dels, nil
 }
 
-func checkRequestOp(u *pb.RequestOp) error {
+func checkRequestOp(u *pb.RequestOp, maxTxnOps int) error {
 	// TODO: ensure only one of the field is set.
 	switch uv := u.Request.(type) {
 	case *pb.RequestOp_RequestRange:
-		if uv.RequestRange != nil {
+		return checkRangeRequest(uv.RequestRange)
-			return checkRangeRequest(uv.RequestRange)
-		}
 	case *pb.RequestOp_RequestPut:
-		if uv.RequestPut != nil {
+		return checkPutRequest(uv.RequestPut)
-			return checkPutRequest(uv.RequestPut)
-		}
 	case *pb.RequestOp_RequestDeleteRange:
-		if uv.RequestDeleteRange != nil {
+		return checkDeleteRequest(uv.RequestDeleteRange)
-			return checkDeleteRequest(uv.RequestDeleteRange)
+	case *pb.RequestOp_RequestTxn:
-		}
+		return checkTxnRequest(uv.RequestTxn, maxTxnOps)
 	default:
 		// empty op / nil entry
 		return rpctypes.ErrGRPCKeyNotFound
 	}
-	return nil
 }

etcdserver/apply.go

@@ -76,14 +76,30 @@ type applierV3 interface {
 	RoleList(ua *pb.AuthRoleListRequest) (*pb.AuthRoleListResponse, error)
 }
 
+type checkReqFunc func(mvcc.ReadView, *pb.RequestOp) error
+
 type applierV3backend struct {
 	s *EtcdServer
+
+	checkPut   checkReqFunc
+	checkRange checkReqFunc
+}
+
+func (s *EtcdServer) newApplierV3Backend() applierV3 {
+	base := &applierV3backend{s: s}
+	base.checkPut = func(rv mvcc.ReadView, req *pb.RequestOp) error {
+		return base.checkRequestPut(rv, req)
+	}
+	base.checkRange = func(rv mvcc.ReadView, req *pb.RequestOp) error {
+		return base.checkRequestRange(rv, req)
+	}
+	return base
 }
 
 func (s *EtcdServer) newApplierV3() applierV3 {
 	return newAuthApplierV3(
 		s.AuthStore(),
-		newQuotaApplierV3(s, &applierV3backend{s}),
+		newQuotaApplierV3(s, s.newApplierV3Backend()),
 		s.lessor,
 	)
 }
@@ -315,24 +331,19 @@ func (a *applierV3backend) Txn(rt *pb.TxnRequest) (*pb.TxnResponse, error) {
 	isWrite := !isTxnReadonly(rt)
 	txn := mvcc.NewReadOnlyTxnWrite(a.s.KV().Read())
 
-	reqs, ok := a.compareToOps(txn, rt)
+	txnPath := compareToPath(txn, rt)
 	if isWrite {
-		if err := a.checkRequestPut(txn, reqs); err != nil {
+		if _, err := checkRequests(txn, rt, txnPath, a.checkPut); err != nil {
 			txn.End()
 			return nil, err
 		}
 	}
-	if err := checkRequestRange(txn, reqs); err != nil {
+	if _, err := checkRequests(txn, rt, txnPath, a.checkRange); err != nil {
 		txn.End()
 		return nil, err
 	}
 
-	resps := make([]*pb.ResponseOp, len(reqs))
+	txnResp, _ := newTxnResp(rt, txnPath)
-	txnResp := &pb.TxnResponse{
-		Responses: resps,
-		Succeeded: ok,
-		Header:    &pb.ResponseHeader{},
-	}
 
 	// When executing mutable txn ops, etcd must hold the txn lock so
 	// readers do not see any intermediate results. Since writes are
@@ -342,9 +353,7 @@ func (a *applierV3backend) Txn(rt *pb.TxnRequest) (*pb.TxnResponse, error) {
 		txn.End()
 		txn = a.s.KV().Write()
 	}
-	for i := range reqs {
+	a.applyTxn(txn, rt, txnPath, txnResp)
-		resps[i] = a.applyUnion(txn, reqs[i])
-	}
 	rev := txn.Rev()
 	if len(txn.Changes()) != 0 {
 		rev++
@@ -355,13 +364,60 @@ func (a *applierV3backend) Txn(rt *pb.TxnRequest) (*pb.TxnResponse, error) {
 	return txnResp, nil
 }
 
-func (a *applierV3backend) compareToOps(rv mvcc.ReadView, rt *pb.TxnRequest) ([]*pb.RequestOp, bool) {
+// newTxnResp allocates a txn response for a txn request given a path.
-	for _, c := range rt.Compare {
+func newTxnResp(rt *pb.TxnRequest, txnPath []bool) (txnResp *pb.TxnResponse, txnCount int) {
-		if !applyCompare(rv, c) {
+	reqs := rt.Success
-			return rt.Failure, false
+	if !txnPath[0] {
+		reqs = rt.Failure
+	}
+	resps := make([]*pb.ResponseOp, len(reqs))
+	txnResp = &pb.TxnResponse{
+		Responses: resps,
+		Succeeded: txnPath[0],
+		Header:    &pb.ResponseHeader{},
+	}
+	for i, req := range reqs {
+		switch tv := req.Request.(type) {
+		case *pb.RequestOp_RequestRange:
+			resps[i] = &pb.ResponseOp{Response: &pb.ResponseOp_ResponseRange{}}
+		case *pb.RequestOp_RequestPut:
+			resps[i] = &pb.ResponseOp{Response: &pb.ResponseOp_ResponsePut{}}
+		case *pb.RequestOp_RequestDeleteRange:
+			resps[i] = &pb.ResponseOp{Response: &pb.ResponseOp_ResponseDeleteRange{}}
+		case *pb.RequestOp_RequestTxn:
+			resp, txns := newTxnResp(tv.RequestTxn, txnPath[1:])
+			resps[i] = &pb.ResponseOp{Response: &pb.ResponseOp_ResponseTxn{ResponseTxn: resp}}
+			txnPath = txnPath[1+txns:]
+			txnCount += txns + 1
+		default:
 		}
 	}
-	return rt.Success, true
+	return txnResp, txnCount
+}
+
+func compareToPath(rv mvcc.ReadView, rt *pb.TxnRequest) []bool {
+	txnPath := make([]bool, 1)
+	ops := rt.Success
+	if txnPath[0] = applyCompares(rv, rt.Compare); !txnPath[0] {
+		ops = rt.Failure
+	}
+	for _, op := range ops {
+		tv, ok := op.Request.(*pb.RequestOp_RequestTxn)
+		if !ok || tv.RequestTxn == nil {
+			continue
+		}
+		txnPath = append(txnPath, compareToPath(rv, tv.RequestTxn)...)
+	}
+	return txnPath
+}
+
+func applyCompares(rv mvcc.ReadView, cmps []*pb.Compare) bool {
+	for _, c := range cmps {
+		if !applyCompare(rv, c) {
+			return false
+		}
+	}
+	return true
 }
 
 // applyCompare applies the compare request.
@@ -431,38 +487,42 @@ func compareKV(c *pb.Compare, ckv mvccpb.KeyValue) bool {
 	return true
 }
 
-func (a *applierV3backend) applyUnion(txn mvcc.TxnWrite, union *pb.RequestOp) *pb.ResponseOp {
+func (a *applierV3backend) applyTxn(txn mvcc.TxnWrite, rt *pb.TxnRequest, txnPath []bool, tresp *pb.TxnResponse) (txns int) {
-	switch tv := union.Request.(type) {
+	reqs := rt.Success
-	case *pb.RequestOp_RequestRange:
+	if !txnPath[0] {
-		if tv.RequestRange != nil {
+		reqs = rt.Failure
+	}
+	for i, req := range reqs {
+		respi := tresp.Responses[i].Response
+		switch tv := req.Request.(type) {
+		case *pb.RequestOp_RequestRange:
 			resp, err := a.Range(txn, tv.RequestRange)
 			if err != nil {
 				plog.Panicf("unexpected error during txn: %v", err)
 			}
-			return &pb.ResponseOp{Response: &pb.ResponseOp_ResponseRange{ResponseRange: resp}}
+			respi.(*pb.ResponseOp_ResponseRange).ResponseRange = resp
-		}
+		case *pb.RequestOp_RequestPut:
-	case *pb.RequestOp_RequestPut:
-		if tv.RequestPut != nil {
 			resp, err := a.Put(txn, tv.RequestPut)
 			if err != nil {
 				plog.Panicf("unexpected error during txn: %v", err)
 			}
-			return &pb.ResponseOp{Response: &pb.ResponseOp_ResponsePut{ResponsePut: resp}}
+			respi.(*pb.ResponseOp_ResponsePut).ResponsePut = resp
-		}
+		case *pb.RequestOp_RequestDeleteRange:
-	case *pb.RequestOp_RequestDeleteRange:
-		if tv.RequestDeleteRange != nil {
 			resp, err := a.DeleteRange(txn, tv.RequestDeleteRange)
 			if err != nil {
 				plog.Panicf("unexpected error during txn: %v", err)
 			}
-			return &pb.ResponseOp{Response: &pb.ResponseOp_ResponseDeleteRange{ResponseDeleteRange: resp}}
+			respi.(*pb.ResponseOp_ResponseDeleteRange).ResponseDeleteRange = resp
+		case *pb.RequestOp_RequestTxn:
+			resp := respi.(*pb.ResponseOp_ResponseTxn).ResponseTxn
+			applyTxns := a.applyTxn(txn, tv.RequestTxn, txnPath[1:], resp)
+			txns += applyTxns + 1
+			txnPath = txnPath[applyTxns+1:]
+		default:
+			// empty union
 		}
-	default:
-		// empty union
-		return nil
 	}
-	return nil
+	return txns
-
 }
 
 func (a *applierV3backend) Compaction(compaction *pb.CompactionRequest) (*pb.CompactionResponse, <-chan struct{}, error) {
@@ -768,53 +828,66 @@ func (s *kvSortByValue) Less(i, j int) bool {
 	return bytes.Compare(s.kvs[i].Value, s.kvs[j].Value) < 0
 }
 
-func (a *applierV3backend) checkRequestPut(rv mvcc.ReadView, reqs []*pb.RequestOp) error {
+func checkRequests(rv mvcc.ReadView, rt *pb.TxnRequest, txnPath []bool, f checkReqFunc) (int, error) {
-	for _, requ := range reqs {
+	txnCount := 0
-		tv, ok := requ.Request.(*pb.RequestOp_RequestPut)
+	reqs := rt.Success
-		if !ok {
+	if !txnPath[0] {
-			continue
+		reqs = rt.Failure
-		}
+	}
-		preq := tv.RequestPut
+	for _, req := range reqs {
-		if preq == nil {
+		if tv, ok := req.Request.(*pb.RequestOp_RequestTxn); ok && tv.RequestTxn != nil {
-			continue
+			txns, err := checkRequests(rv, tv.RequestTxn, txnPath[1:], f)
-		}
-		if preq.IgnoreValue || preq.IgnoreLease {
-			// expects previous key-value, error if not exist
-			rr, err := rv.Range(preq.Key, nil, mvcc.RangeOptions{})
 			if err != nil {
-				return err
+				return 0, err
 			}
-			if rr == nil || len(rr.KVs) == 0 {
+			txnCount += txns + 1
-				return ErrKeyNotFound
+			txnPath = txnPath[txns+1:]
-			}
-		}
-		if lease.LeaseID(preq.Lease) == lease.NoLease {
 			continue
 		}
-		if l := a.s.lessor.Lookup(lease.LeaseID(preq.Lease)); l == nil {
+		if err := f(rv, req); err != nil {
+			return 0, err
+		}
+	}
+	return txnCount, nil
+}
+
+func (a *applierV3backend) checkRequestPut(rv mvcc.ReadView, reqOp *pb.RequestOp) error {
+	tv, ok := reqOp.Request.(*pb.RequestOp_RequestPut)
+	if !ok || tv.RequestPut == nil {
+		return nil
+	}
+	req := tv.RequestPut
+	if req.IgnoreValue || req.IgnoreLease {
+		// expects previous key-value, error if not exist
+		rr, err := rv.Range(req.Key, nil, mvcc.RangeOptions{})
+		if err != nil {
+			return err
+		}
+		if rr == nil || len(rr.KVs) == 0 {
+			return ErrKeyNotFound
+		}
+	}
+	if lease.LeaseID(req.Lease) != lease.NoLease {
+		if l := a.s.lessor.Lookup(lease.LeaseID(req.Lease)); l == nil {
 			return lease.ErrLeaseNotFound
 		}
 	}
 	return nil
 }
 
-func checkRequestRange(rv mvcc.ReadView, reqs []*pb.RequestOp) error {
+func (a *applierV3backend) checkRequestRange(rv mvcc.ReadView, reqOp *pb.RequestOp) error {
-	for _, requ := range reqs {
+	tv, ok := reqOp.Request.(*pb.RequestOp_RequestRange)
-		tv, ok := requ.Request.(*pb.RequestOp_RequestRange)
+	if !ok || tv.RequestRange == nil {
-		if !ok {
+		return nil
-			continue
+	}
-		}
+	req := tv.RequestRange
-		greq := tv.RequestRange
+	switch {
-		if greq == nil || greq.Revision == 0 {
+	case req.Revision == 0:
-			continue
+		return nil
-		}
+	case req.Revision > rv.Rev():
-
+		return mvcc.ErrFutureRev
-		if greq.Revision > rv.Rev() {
+	case req.Revision < rv.FirstRev():
-			return mvcc.ErrFutureRev
+		return mvcc.ErrCompacted
-		}
-		if greq.Revision < rv.FirstRev() {
-			return mvcc.ErrCompacted
-		}
 	}
 	return nil
 }

etcdserver/etcdserverpb/rpc.proto

@@ -469,6 +469,7 @@ message RequestOp {
     RangeRequest request_range = 1;
     PutRequest request_put = 2;
     DeleteRangeRequest request_delete_range = 3;
+    TxnRequest request_txn = 4;
   }
 }
 
@@ -478,6 +479,7 @@ message ResponseOp {
     RangeResponse response_range = 1;
     PutResponse response_put = 2;
     DeleteRangeResponse response_delete_range = 3;
+    TxnResponse response_txn = 4;
   }
 }
 

etcdserver/server.go

@@ -474,7 +474,7 @@ func NewServer(cfg ServerConfig) (srv *EtcdServer, err error) {
 		srv.compactor.Run()
 	}
 
-	srv.applyV3Base = &applierV3backend{srv}
+	srv.applyV3Base = srv.newApplierV3Backend()
 	if err = srv.restoreAlarms(); err != nil {
 		return nil, err
 	}

integration/v3_grpc_test.go

@@ -192,11 +192,22 @@ func TestV3TxnTooManyOps(t *testing.T) {
 				},
 			})
 	}
+	addTxnOps := func(txn *pb.TxnRequest) {
+		newTxn := &pb.TxnRequest{}
+		addSuccessOps(newTxn)
+		txn.Success = append(txn.Success,
+			&pb.RequestOp{Request: &pb.RequestOp_RequestTxn{
+				RequestTxn: newTxn,
+			},
+			},
+		)
+	}
 
 	tests := []func(txn *pb.TxnRequest){
 		addCompareOps,
 		addSuccessOps,
 		addFailureOps,
+		addTxnOps,
 	}
 
 	for i, tt := range tests {
@@ -236,6 +247,27 @@ func TestV3TxnDuplicateKeys(t *testing.T) {
 		},
 	},
 	}
+	txnDelReq := &pb.RequestOp{Request: &pb.RequestOp_RequestTxn{
+		RequestTxn: &pb.TxnRequest{Success: []*pb.RequestOp{delInRangeReq}},
+	},
+	}
+	txnDelReqTwoSide := &pb.RequestOp{Request: &pb.RequestOp_RequestTxn{
+		RequestTxn: &pb.TxnRequest{
+			Success: []*pb.RequestOp{delInRangeReq},
+			Failure: []*pb.RequestOp{delInRangeReq}},
+	},
+	}
+
+	txnPutReq := &pb.RequestOp{Request: &pb.RequestOp_RequestTxn{
+		RequestTxn: &pb.TxnRequest{Success: []*pb.RequestOp{putreq}},
+	},
+	}
+	txnPutReqTwoSide := &pb.RequestOp{Request: &pb.RequestOp_RequestTxn{
+		RequestTxn: &pb.TxnRequest{
+			Success: []*pb.RequestOp{putreq},
+			Failure: []*pb.RequestOp{putreq}},
+	},
+	}
 
 	kvc := toGRPC(clus.RandClient()).KV
 	tests := []struct {
@@ -258,6 +290,36 @@ func TestV3TxnDuplicateKeys(t *testing.T) {
 
 			werr: rpctypes.ErrGRPCDuplicateKey,
 		},
+		// Then(Put(a), Then(Del(a)))
+		{
+			txnSuccess: []*pb.RequestOp{putreq, txnDelReq},
+
+			werr: rpctypes.ErrGRPCDuplicateKey,
+		},
+		// Then(Del(a), Then(Put(a)))
+		{
+			txnSuccess: []*pb.RequestOp{delInRangeReq, txnPutReq},
+
+			werr: rpctypes.ErrGRPCDuplicateKey,
+		},
+		// Then((Then(Put(a)), Else(Put(a))), (Then(Put(a)), Else(Put(a)))
+		{
+			txnSuccess: []*pb.RequestOp{txnPutReqTwoSide, txnPutReqTwoSide},
+
+			werr: rpctypes.ErrGRPCDuplicateKey,
+		},
+		// Then(Del(x), (Then(Put(a)), Else(Put(a))))
+		{
+			txnSuccess: []*pb.RequestOp{delOutOfRangeReq, txnPutReqTwoSide},
+
+			werr: nil,
+		},
+		// Then(Then(Del(a)), (Then(Del(a)), Else(Del(a))))
+		{
+			txnSuccess: []*pb.RequestOp{txnDelReq, txnDelReqTwoSide},
+
+			werr: nil,
+		},
 		{
 			txnSuccess: []*pb.RequestOp{delKeyReq, delInRangeReq, delKeyReq, delInRangeReq},
 
@@ -469,6 +531,59 @@ func TestV3TxnRangeCompare(t *testing.T) {
 	}
 }
 
+// TestV3TxnNested tests nested txns follow paths as expected.
+func TestV3TxnNestedPath(t *testing.T) {
+	defer testutil.AfterTest(t)
+	clus := NewClusterV3(t, &ClusterConfig{Size: 1})
+	defer clus.Terminate(t)
+
+	kvc := toGRPC(clus.RandClient()).KV
+
+	cmpTrue := &pb.Compare{
+		Result:      pb.Compare_EQUAL,
+		Target:      pb.Compare_VERSION,
+		Key:         []byte("k"),
+		TargetUnion: &pb.Compare_Version{Version: int64(0)},
+	}
+	cmpFalse := &pb.Compare{
+		Result:      pb.Compare_EQUAL,
+		Target:      pb.Compare_VERSION,
+		Key:         []byte("k"),
+		TargetUnion: &pb.Compare_Version{Version: int64(1)},
+	}
+
+	// generate random path to eval txns
+	topTxn := &pb.TxnRequest{}
+	txn := topTxn
+	txnPath := make([]bool, 10)
+	for i := range txnPath {
+		nextTxn := &pb.TxnRequest{}
+		op := &pb.RequestOp{Request: &pb.RequestOp_RequestTxn{RequestTxn: nextTxn}}
+		txnPath[i] = rand.Intn(2) == 0
+		if txnPath[i] {
+			txn.Compare = append(txn.Compare, cmpTrue)
+			txn.Success = append(txn.Success, op)
+		} else {
+			txn.Compare = append(txn.Compare, cmpFalse)
+			txn.Failure = append(txn.Failure, op)
+		}
+		txn = nextTxn
+	}
+
+	tresp, err := kvc.Txn(context.TODO(), topTxn)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	curTxnResp := tresp
+	for i := range txnPath {
+		if curTxnResp.Succeeded != txnPath[i] {
+			t.Fatalf("expected path %+v, got response %+v", txnPath, *tresp)
+		}
+		curTxnResp = curTxnResp.Responses[0].Response.(*pb.ResponseOp_ResponseTxn).ResponseTxn
+	}
+}
+
 // TestV3PutIgnoreValue ensures that writes with ignore_value overwrites with previous key-value pair.
 func TestV3PutIgnoreValue(t *testing.T) {
 	defer testutil.AfterTest(t)

pkg/adt/interval_tree.go

@@ -485,6 +485,15 @@ func (ivt *IntervalTree) Stab(iv Interval) (ivs []*IntervalValue) {
 	return ivs
 }
 
+// Union merges a given interval tree into the receiver.
+func (ivt *IntervalTree) Union(inIvt IntervalTree, ivl Interval) {
+	f := func(n *IntervalValue) bool {
+		ivt.Insert(n.Ivl, n.Val)
+		return true
+	}
+	inIvt.Visit(ivl, f)
+}
+
 type StringComparable string
 
 func (s StringComparable) Compare(c Comparable) int {

proxy/grpcproxy/kv.go

@@ -99,28 +99,13 @@ func (p *kvProxy) txnToCache(reqs []*pb.RequestOp, resps []*pb.ResponseOp) {
 }
 
 func (p *kvProxy) Txn(ctx context.Context, r *pb.TxnRequest) (*pb.TxnResponse, error) {
-	txn := p.kv.Txn(ctx)
+	op := TxnRequestToOp(r)
-	cmps := make([]clientv3.Cmp, len(r.Compare))
+	opResp, err := p.kv.Do(ctx, op)
-	thenops := make([]clientv3.Op, len(r.Success))
-	elseops := make([]clientv3.Op, len(r.Failure))
-
-	for i := range r.Compare {
-		cmps[i] = (clientv3.Cmp)(*r.Compare[i])
-	}
-
-	for i := range r.Success {
-		thenops[i] = requestOpToOp(r.Success[i])
-	}
-
-	for i := range r.Failure {
-		elseops[i] = requestOpToOp(r.Failure[i])
-	}
-
-	resp, err := txn.If(cmps...).Then(thenops...).Else(elseops...).Commit()
-
 	if err != nil {
 		return nil, err
 	}
+	resp := opResp.Txn()
+
 	// txn may claim an outdated key is updated; be safe and invalidate
 	for _, cmp := range r.Compare {
 		p.cache.Invalidate(cmp.Key, cmp.RangeEnd)
@@ -167,6 +152,10 @@ func requestOpToOp(union *pb.RequestOp) clientv3.Op {
 		if tv.RequestDeleteRange != nil {
 			return DelRequestToOp(tv.RequestDeleteRange)
 		}
+	case *pb.RequestOp_RequestTxn:
+		if tv.RequestTxn != nil {
+			return TxnRequestToOp(tv.RequestTxn)
+		}
 	}
 	panic("unknown request")
 }
@@ -219,3 +208,19 @@ func DelRequestToOp(r *pb.DeleteRangeRequest) clientv3.Op {
 	}
 	return clientv3.OpDelete(string(r.Key), opts...)
 }
+
+func TxnRequestToOp(r *pb.TxnRequest) clientv3.Op {
+	cmps := make([]clientv3.Cmp, len(r.Compare))
+	thenops := make([]clientv3.Op, len(r.Success))
+	elseops := make([]clientv3.Op, len(r.Failure))
+	for i := range r.Compare {
+		cmps[i] = (clientv3.Cmp)(*r.Compare[i])
+	}
+	for i := range r.Success {
+		thenops[i] = requestOpToOp(r.Success[i])
+	}
+	for i := range r.Failure {
+		elseops[i] = requestOpToOp(r.Failure[i])
+	}
+	return clientv3.OpTxn(cmps, thenops, elseops)
+}