// startFakeServerGossips creates local gossip instances and remote
// faked gossip instance. The remote gossip instance launches its
// faked gossip service just for check the client message.
func startFakeServerGossips(t *testing.T) (local *Gossip, remote *fakeGossipServer, stopper *stop.Stopper) {
stopper = stop.NewStopper()
lRPCContext := rpc.NewContext(&base.Context{Insecure: true}, nil, stopper)
lserver := rpc.NewServer(lRPCContext)
lln, err := netutil.ListenAndServeGRPC(stopper, lserver, util.TestAddr)
if err != nil {
t.Fatal(err)
}
local = New(lRPCContext, nil, stopper)
local.start(lserver, lln.Addr())
rRPCContext := rpc.NewContext(&base.Context{Insecure: true}, nil, stopper)
rserver := rpc.NewServer(rRPCContext)
rln, err := netutil.ListenAndServeGRPC(stopper, rserver, util.TestAddr)
if err != nil {
t.Fatal(err)
}
remote = newFakeGossipServer(rserver, stopper)
addr := rln.Addr()
remote.nodeAddr = util.MakeUnresolvedAddr(addr.Network(), addr.String())
return
}
// CreateNode creates a simulation node and starts an RPC server for it.
func (n *Network) CreateNode() (*Node, error) {
server := rpc.NewServer(n.rpcContext)
ln, err := netutil.ListenAndServeGRPC(n.Stopper, server, util.TestAddr)
if err != nil {
return nil, err
}
node := &Node{Server: server, Addr: ln.Addr()}
node.Gossip = gossip.New(n.rpcContext, nil, n.Stopper)
n.Nodes = append(n.Nodes, node)
return node, nil
}
// createTestNode creates an rpc server using the specified address,
// gossip instance, KV database and a node using the specified slice
// of engines. The server, clock and node are returned. If gossipBS is
// not nil, the gossip bootstrap address is set to gossipBS.
func createTestNode(addr net.Addr, engines []engine.Engine, gossipBS net.Addr, t *testing.T) (
*grpc.Server, net.Addr, *hlc.Clock, *Node, *stop.Stopper) {
ctx := storage.StoreContext{}
stopper := stop.NewStopper()
ctx.Clock = hlc.NewClock(hlc.UnixNano)
nodeRPCContext := rpc.NewContext(nodeTestBaseContext, ctx.Clock, stopper)
ctx.ScanInterval = 10 * time.Hour
ctx.ConsistencyCheckInterval = 10 * time.Hour
grpcServer := rpc.NewServer(nodeRPCContext)
serverCtx := makeTestContext()
g := gossip.New(
context.Background(),
nodeRPCContext,
grpcServer,
serverCtx.GossipBootstrapResolvers,
stopper,
metric.NewRegistry())
ln, err := netutil.ListenAndServeGRPC(stopper, grpcServer, addr)
if err != nil {
t.Fatal(err)
}
if gossipBS != nil {
// Handle possibility of a :0 port specification.
if gossipBS.Network() == addr.Network() && gossipBS.String() == addr.String() {
gossipBS = ln.Addr()
}
r, err := resolver.NewResolverFromAddress(gossipBS)
if err != nil {
t.Fatalf("bad gossip address %s: %s", gossipBS, err)
}
g.SetResolvers([]resolver.Resolver{r})
g.Start(ln.Addr())
}
ctx.Gossip = g
retryOpts := base.DefaultRetryOptions()
retryOpts.Closer = stopper.ShouldQuiesce()
distSender := kv.NewDistSender(&kv.DistSenderConfig{
Clock: ctx.Clock,
RPCContext: nodeRPCContext,
RPCRetryOptions: &retryOpts,
}, g)
ctx.Ctx = tracing.WithTracer(context.Background(), tracing.NewTracer())
sender := kv.NewTxnCoordSender(ctx.Ctx, distSender, ctx.Clock, false, stopper,
kv.MakeTxnMetrics())
ctx.DB = client.NewDB(sender)
ctx.Transport = storage.NewDummyRaftTransport()
node := NewNode(ctx, status.NewMetricsRecorder(ctx.Clock), metric.NewRegistry(), stopper,
kv.MakeTxnMetrics(), sql.MakeEventLogger(nil))
roachpb.RegisterInternalServer(grpcServer, node)
return grpcServer, ln.Addr(), ctx.Clock, node, stopper
}
// AddNodeWithoutGossip registers a node with the cluster. Nodes must
// be added before they can be used in other methods of
// raftTransportTestContext. Unless you are testing the effects of
// delaying gossip, use AddNode instead.
func (rttc *raftTransportTestContext) AddNodeWithoutGossip(
nodeID roachpb.NodeID,
) (*storage.RaftTransport, net.Addr) {
grpcServer := rpc.NewServer(rttc.nodeRPCContext)
ln, err := netutil.ListenAndServeGRPC(rttc.stopper, grpcServer, util.TestAddr)
if err != nil {
rttc.t.Fatal(err)
}
transport := storage.NewRaftTransport(storage.GossipAddressResolver(rttc.gossip),
grpcServer, rttc.nodeRPCContext)
rttc.transports[nodeID] = transport
return transport, ln.Addr()
}
func newTestServer(t *testing.T, ctx *Context, manual bool) (*grpc.Server, net.Listener) {
tlsConfig, err := ctx.GetServerTLSConfig()
if err != nil {
t.Fatal(err)
}
s := grpc.NewServer(grpc.Creds(credentials.NewTLS(tlsConfig)))
ln, err := netutil.ListenAndServeGRPC(ctx.Stopper, s, util.TestAddr)
if err != nil {
t.Fatal(err)
}
return s, ln
}
// TestClientRegisterInitNodeID verifies two client's gossip request with NodeID 0.
func TestClientRegisterWithInitNodeID(t *testing.T) {
defer leaktest.AfterTest(t)()
stopper := stop.NewStopper()
defer stopper.Stop()
// Create three gossip nodes, and connect to the first with NodeID 0.
var g []*Gossip
var gossipAddr string
for i := 0; i < 3; i++ {
RPCContext := rpc.NewContext(&base.Context{Insecure: true}, nil, stopper)
server := rpc.NewServer(RPCContext)
ln, err := netutil.ListenAndServeGRPC(stopper, server, util.TestAddr)
if err != nil {
t.Fatal(err)
}
// Connect to the first gossip node.
if gossipAddr == "" {
gossipAddr = ln.Addr().String()
}
var resolvers []resolver.Resolver
resolver, err := resolver.NewResolver(RPCContext.Context, gossipAddr)
if err != nil {
t.Fatal(err)
}
resolvers = append(resolvers, resolver)
gnode := New(RPCContext, server, resolvers, stopper, metric.NewRegistry())
// node ID must be non-zero
gnode.SetNodeID(roachpb.NodeID(i + 1))
g = append(g, gnode)
gnode.Start(ln.Addr())
}
util.SucceedsSoon(t, func() error {
// The first gossip node should have two gossip client address
// in nodeMap if these three gossip nodes registered success.
g[0].mu.Lock()
defer g[0].mu.Unlock()
if a, e := len(g[0].nodeMap), 2; a != e {
return errors.Errorf("expected %s to contain %d nodes, got %d", g[0].nodeMap, e, a)
}
return nil
})
}
// startGossip creates and starts a gossip instance.
func startGossip(nodeID roachpb.NodeID, stopper *stop.Stopper, t *testing.T) *Gossip {
rpcContext := rpc.NewContext(&base.Context{Insecure: true}, nil, stopper)
server := rpc.NewServer(rpcContext)
ln, err := netutil.ListenAndServeGRPC(stopper, server, util.TestAddr)
if err != nil {
t.Fatal(err)
}
addr := ln.Addr()
g := New(rpcContext, nil, stopper)
g.SetNodeID(nodeID)
if err := g.SetNodeDescriptor(&roachpb.NodeDescriptor{
NodeID: nodeID,
Address: util.MakeUnresolvedAddr(addr.Network(), addr.String()),
}); err != nil {
t.Fatal(err)
}
g.start(server, addr)
time.Sleep(time.Millisecond)
return g
}
func TestSendAndReceive(t *testing.T) {
defer leaktest.AfterTest(t)()
stopper := stop.NewStopper()
defer stopper.Stop()
nodeRPCContext := rpc.NewContext(testutils.NewNodeTestBaseContext(), nil, stopper)
g := gossip.New(nodeRPCContext, nil, stopper)
g.SetNodeID(roachpb.NodeID(1))
// Create several servers, each of which has two stores (A raft
// node ID addresses a store). Node 1 has stores 1 and 2, node 2 has
// stores 3 and 4, etc.
//
// We suppose that range 1 is replicated across the odd-numbered
// stores in reverse order to ensure that the various IDs are not
// equal: replica 1 is store 5, replica 2 is store 3, and replica 3
// is store 1.
const numNodes = 3
const storesPerNode = 2
nextNodeID := roachpb.NodeID(2)
nextStoreID := roachpb.StoreID(2)
// Per-node state.
transports := map[roachpb.NodeID]*storage.RaftTransport{}
// Per-store state.
storeNodes := map[roachpb.StoreID]roachpb.NodeID{}
channels := map[roachpb.StoreID]channelServer{}
replicaIDs := map[roachpb.StoreID]roachpb.ReplicaID{
1: 3,
3: 2,
5: 1,
}
messageTypes := []raftpb.MessageType{
raftpb.MsgSnap,
raftpb.MsgHeartbeat,
}
for nodeIndex := 0; nodeIndex < numNodes; nodeIndex++ {
nodeID := nextNodeID
nextNodeID++
grpcServer := rpc.NewServer(nodeRPCContext)
ln, err := netutil.ListenAndServeGRPC(stopper, grpcServer, util.TestAddr)
if err != nil {
t.Fatal(err)
}
addr := ln.Addr()
// Have to call g.SetNodeID before call g.AddInfo.
g.ResetNodeID(roachpb.NodeID(nodeID))
if err := g.AddInfoProto(gossip.MakeNodeIDKey(nodeID),
&roachpb.NodeDescriptor{
Address: util.MakeUnresolvedAddr(addr.Network(), addr.String()),
},
time.Hour); err != nil {
t.Fatal(err)
}
transports[nodeID] = storage.NewRaftTransport(storage.GossipAddressResolver(g), grpcServer, nodeRPCContext)
// This channel is normally unbuffered, but it is also normally serviced by
// the raft goroutine. Since we don't have that goroutine in this test, we
// must buffer the channel to prevent snapshots from blocking while we
// iterate through the recipients in an order that may differ from the
// sending order.
sendersPerNode := storesPerNode
recipientsPerSender := numNodes * storesPerNode
outboundSnapshotsPerNode := sendersPerNode * recipientsPerSender
transports[nodeID].SnapshotStatusChan = make(chan storage.RaftSnapshotStatus, outboundSnapshotsPerNode)
for storeIndex := 0; storeIndex < storesPerNode; storeIndex++ {
storeID := nextStoreID
nextStoreID++
storeNodes[storeID] = nodeID
sendersPerRecipient := numNodes * storesPerNode
inboundMessagesPerStore := sendersPerRecipient * len(messageTypes)
channels[storeID] = newChannelServer(inboundMessagesPerStore, 0)
transports[nodeID].Listen(storeID, channels[storeID].RaftMessage)
}
}
messageTypeCounts := make(map[roachpb.StoreID]map[raftpb.MessageType]int)
// Each store sends one snapshot and one heartbeat to each store, including
// itself.
for toStoreID, toNodeID := range storeNodes {
if _, ok := messageTypeCounts[toStoreID]; !ok {
messageTypeCounts[toStoreID] = make(map[raftpb.MessageType]int)
}
for fromStoreID, fromNodeID := range storeNodes {
baseReq := storage.RaftMessageRequest{
RangeID: 1,
Message: raftpb.Message{
From: uint64(fromStoreID),
To: uint64(toStoreID),
},
FromReplica: roachpb.ReplicaDescriptor{
NodeID: fromNodeID,
StoreID: fromStoreID,
},
ToReplica: roachpb.ReplicaDescriptor{
NodeID: toNodeID,
StoreID: toStoreID,
},
}
for _, messageType := range messageTypes {
req := baseReq
req.Message.Type = messageType
if !transports[fromNodeID].MakeSender(func(err error, _ roachpb.ReplicaDescriptor) {
if err != nil && !grpcutil.IsClosedConnection(err) {
panic(err)
}
}).SendAsync(&req) {
t.Errorf("unable to send %s from %d to %d", req.Message.Type, fromNodeID, toNodeID)
}
messageTypeCounts[toStoreID][req.Message.Type]++
}
}
}
// Read all the messages from the channels. Note that the transport
// does not guarantee in-order delivery between independent
// transports, so we just verify that the right number of messages
// end up in each channel.
for toStoreID := range storeNodes {
func() {
for len(messageTypeCounts[toStoreID]) > 0 {
req := <-channels[toStoreID].ch
if req.Message.To != uint64(toStoreID) {
t.Errorf("got unexpected message %v on channel %d", req, toStoreID)
}
// Each MsgSnap should have a corresponding entry on the
// sender's SnapshotStatusChan.
if req.Message.Type == raftpb.MsgSnap {
st := <-transports[req.FromReplica.NodeID].SnapshotStatusChan
if st.Err != nil {
t.Errorf("unexpected error sending snapshot: %s", st.Err)
}
}
if typeCounts, ok := messageTypeCounts[toStoreID]; ok {
if _, ok := typeCounts[req.Message.Type]; ok {
typeCounts[req.Message.Type]--
if typeCounts[req.Message.Type] == 0 {
delete(typeCounts, req.Message.Type)
}
} else {
t.Errorf("expected %v to have key %v, but it did not", typeCounts, req.Message.Type)
}
} else {
t.Errorf("expected %v to have key %v, but it did not", messageTypeCounts, toStoreID)
}
}
delete(messageTypeCounts, toStoreID)
}()
select {
case req := <-channels[toStoreID].ch:
t.Errorf("got unexpected message %v on channel %d", req, toStoreID)
case <-time.After(100 * time.Millisecond):
}
}
if len(messageTypeCounts) > 0 {
t.Errorf("remaining messages expected: %v", messageTypeCounts)
}
// Real raft messages have different node/store/replica IDs.
// Send a message from replica 2 (on store 3, node 2) to replica 1 (on store 5, node 3)
fromStoreID := roachpb.StoreID(3)
toStoreID := roachpb.StoreID(5)
expReq := &storage.RaftMessageRequest{
RangeID: 1,
Message: raftpb.Message{
Type: raftpb.MsgApp,
From: uint64(replicaIDs[fromStoreID]),
To: uint64(replicaIDs[toStoreID]),
},
FromReplica: roachpb.ReplicaDescriptor{
NodeID: storeNodes[fromStoreID],
StoreID: fromStoreID,
ReplicaID: replicaIDs[fromStoreID],
},
ToReplica: roachpb.ReplicaDescriptor{
NodeID: storeNodes[toStoreID],
StoreID: toStoreID,
ReplicaID: replicaIDs[toStoreID],
},
}
if !transports[storeNodes[fromStoreID]].MakeSender(func(err error, _ roachpb.ReplicaDescriptor) {
if err != nil && !grpcutil.IsClosedConnection(err) {
panic(err)
}
}).SendAsync(expReq) {
t.Errorf("unable to send message from %d to %d", fromStoreID, toStoreID)
}
if req := <-channels[toStoreID].ch; !proto.Equal(req, expReq) {
t.Errorf("got unexpected message %v on channel %d", req, toStoreID)
}
select {
case req := <-channels[toStoreID].ch:
t.Errorf("got unexpected message %v on channel %d", req, toStoreID)
default:
}
}
// TestInOrderDelivery verifies that for a given pair of nodes, raft
// messages are delivered in order.
func TestInOrderDelivery(t *testing.T) {
defer leaktest.AfterTest(t)()
stopper := stop.NewStopper()
defer stopper.Stop()
nodeRPCContext := rpc.NewContext(testutils.NewNodeTestBaseContext(), nil, stopper)
g := gossip.New(nodeRPCContext, nil, stopper)
grpcServer := rpc.NewServer(nodeRPCContext)
ln, err := netutil.ListenAndServeGRPC(stopper, grpcServer, util.TestAddr)
if err != nil {
t.Fatal(err)
}
const numMessages = 100
nodeID := roachpb.NodeID(roachpb.NodeID(2))
serverTransport := storage.NewRaftTransport(storage.GossipAddressResolver(g), grpcServer, nodeRPCContext)
serverChannel := newChannelServer(numMessages, 10*time.Millisecond)
serverTransport.Listen(roachpb.StoreID(nodeID), serverChannel.RaftMessage)
addr := ln.Addr()
// Have to set gossip.NodeID before calling gossip.AddInfoXXX.
g.SetNodeID(nodeID)
if err := g.AddInfoProto(gossip.MakeNodeIDKey(nodeID),
&roachpb.NodeDescriptor{
Address: util.MakeUnresolvedAddr(addr.Network(), addr.String()),
},
time.Hour); err != nil {
t.Fatal(err)
}
clientNodeID := roachpb.NodeID(2)
clientTransport := storage.NewRaftTransport(storage.GossipAddressResolver(g), nil, nodeRPCContext)
for i := 0; i < numMessages; i++ {
req := &storage.RaftMessageRequest{
RangeID: 1,
Message: raftpb.Message{
To: uint64(nodeID),
From: uint64(clientNodeID),
Commit: uint64(i),
},
ToReplica: roachpb.ReplicaDescriptor{
NodeID: nodeID,
StoreID: roachpb.StoreID(nodeID),
ReplicaID: roachpb.ReplicaID(nodeID),
},
FromReplica: roachpb.ReplicaDescriptor{
NodeID: clientNodeID,
StoreID: roachpb.StoreID(clientNodeID),
ReplicaID: roachpb.ReplicaID(clientNodeID),
},
}
if !clientTransport.MakeSender(func(err error, _ roachpb.ReplicaDescriptor) {
if err != nil && !grpcutil.IsClosedConnection(err) {
panic(err)
}
}).SendAsync(req) {
t.Errorf("failed to send message %d", i)
}
}
for i := 0; i < numMessages; i++ {
req := <-serverChannel.ch
if req.Message.Commit != uint64(i) {
t.Errorf("messages out of order: got %d while expecting %d", req.Message.Commit, i)
}
}
}