// New creates an instance of a gossip node.
func New(rpcContext *rpc.Context, resolvers []resolver.Resolver) *Gossip {
g := &Gossip{
Connected: make(chan struct{}),
server: newServer(),
outgoing: makeNodeSet(1),
bootstrapping: map[string]struct{}{},
clients: []*client{},
disconnected: make(chan *client, 10),
stalled: make(chan struct{}, 1),
resolverIdx: len(resolvers) - 1,
resolvers: resolvers,
}
// The gossip RPC context doesn't measure clock offsets, isn't
// shared with the other RPC clients which the node may be using,
// and disables reconnects to make it possible to know for certain
// which other nodes in the cluster are incoming via gossip.
if rpcContext != nil {
g.rpcContext = rpcContext.Copy()
g.rpcContext.DisableCache = true
g.rpcContext.DisableReconnects = true
g.rpcContext.RemoteClocks = nil
}
// Add ourselves as a SystemConfig watcher.
g.is.registerCallback(KeySystemConfig, g.updateSystemConfig)
return g
}
// start dials the remote addr and commences gossip once connected. Upon exit,
// the client is sent on the disconnected channel. This method starts client
// processing in a goroutine and returns immediately.
func (c *client) start(g *Gossip, disconnected chan *client, ctx *rpc.Context, stopper *stop.Stopper) {
stopper.RunWorker(func() {
defer func() {
disconnected <- c
}()
// Note: avoid using `grpc.WithBlock` here. This code is already
// asynchronous from the caller's perspective, so the only effect of
// `WithBlock` here is blocking shutdown - at the time of this writing,
// that ends ups up making `kv` tests take twice as long.
conn, err := ctx.GRPCDial(c.addr.String())
if err != nil {
log.Errorf("failed to dial: %v", err)
return
}
// Start gossiping.
if err := c.gossip(g, NewGossipClient(conn), stopper); err != nil {
if !grpcutil.IsClosedConnection(err) {
g.mu.Lock()
peerID := c.peerID
g.mu.Unlock()
if peerID != 0 {
log.Infof("closing client to node %d (%s): %s", peerID, c.addr, err)
} else {
log.Infof("closing client to %s: %s", c.addr, err)
}
}
}
})
}
// New creates an instance of a gossip node.
func New(rpcContext *rpc.Context, gossipInterval time.Duration, resolvers []resolver.Resolver) *Gossip {
g := &Gossip{
Connected: make(chan struct{}),
RPCContext: rpcContext,
server: newServer(gossipInterval),
outgoing: makeNodeSet(MaxPeers),
bootstrapping: map[string]struct{}{},
clients: []*client{},
disconnected: make(chan *client, MaxPeers),
stalled: make(chan struct{}, 1),
resolvers: resolvers,
}
// Create the bootstrapping RPC context. This context doesn't
// measure clock offsets and doesn't cache clients because bootstrap
// connections may go through a load balancer.
if rpcContext != nil {
g.bsRPCContext = rpcContext.Copy()
g.bsRPCContext.DisableCache = true
g.bsRPCContext.RemoteClocks = nil
}
// Add ourselves as a SystemConfig watcher.
g.is.registerCallback(KeySystemConfig, g.updateSystemConfig)
return g
}
// NewSender returns an implementation of Sender which exposes the Key-Value
// database provided by a Cockroach cluster by connecting via RPC to a
// Cockroach node.
func NewSender(ctx *rpc.Context, target string) (Sender, error) {
conn, err := ctx.GRPCDial(target)
if err != nil {
return nil, err
}
return sender{roachpb.NewExternalClient(conn)}, nil
}
// start dials the remote addr and commences gossip once connected. Upon exit,
// the client is sent on the disconnected channel. This method starts client
// processing in a goroutine and returns immediately.
func (c *client) start(g *Gossip, disconnected chan *client, ctx *rpc.Context, stopper *stop.Stopper) {
stopper.RunWorker(func() {
defer func() {
disconnected <- c
}()
conn, err := ctx.GRPCDial(c.addr.String(), grpc.WithBlock())
if err != nil {
log.Errorf("failed to dial: %v", err)
return
}
// Start gossiping.
if err := c.gossip(g, NewGossipClient(conn), stopper); err != nil {
if !grpcutil.IsClosedConnection(err) {
g.mu.Lock()
peerID := c.peerID
g.mu.Unlock()
if peerID != 0 {
log.Infof("closing client to node %d (%s): %s", peerID, c.addr, err)
} else {
log.Infof("closing client to %s: %s", c.addr, err)
}
}
}
})
}
// New creates an instance of a gossip node.
func New(rpcContext *rpc.Context, resolvers []resolver.Resolver, stopper *stop.Stopper) *Gossip {
g := &Gossip{
Connected: make(chan struct{}),
server: newServer(stopper),
outgoing: makeNodeSet(minPeers),
bootstrapping: map[string]struct{}{},
clients: []*client{},
disconnected: make(chan *client, 10),
stalled: make(chan struct{}, 1),
stallInterval: defaultStallInterval,
bootstrapInterval: defaultBootstrapInterval,
cullInterval: defaultCullInterval,
nodeDescs: map[roachpb.NodeID]*roachpb.NodeDescriptor{},
}
g.SetResolvers(resolvers)
// The gossip RPC context doesn't measure clock offsets, isn't
// shared with the other RPC clients which the node may be using,
// and disables reconnects to make it possible to know for certain
// which other nodes in the cluster are incoming via gossip.
if rpcContext != nil {
g.rpcContext = rpcContext.Copy()
g.rpcContext.DisableCache = true
g.rpcContext.DisableReconnects = true
g.rpcContext.RemoteClocks = nil
}
// Add ourselves as a SystemConfig watcher.
g.is.registerCallback(KeySystemConfig, g.updateSystemConfig)
// Add ourselves as a node descriptor watcher.
g.is.registerCallback(MakePrefixPattern(KeyNodeIDPrefix), g.updateNodeAddress)
return g
}
// grpcTransportFactory is the default TransportFactory, using GRPC.
func grpcTransportFactory(
opts SendOptions,
rpcContext *rpc.Context,
replicas ReplicaSlice,
args roachpb.BatchRequest,
) (Transport, error) {
clients := make([]batchClient, 0, len(replicas))
for _, replica := range replicas {
conn, err := rpcContext.GRPCDial(replica.NodeDesc.Address.String())
if err != nil {
return nil, err
}
argsCopy := args
argsCopy.Replica = replica.ReplicaDescriptor
remoteAddr := replica.NodeDesc.Address.String()
clients = append(clients, batchClient{
remoteAddr: remoteAddr,
conn: conn,
client: roachpb.NewInternalClient(conn),
args: argsCopy,
healthy: rpcContext.IsConnHealthy(remoteAddr),
})
}
// Put known-unhealthy clients last.
splitHealthy(clients)
return &grpcTransport{
opts: opts,
rpcContext: rpcContext,
orderedClients: clients,
}, nil
}
func newTestServer(t *testing.T, ctx *rpc.Context) (*rpc.Server, net.Listener) {
s := rpc.NewServer(ctx)
tlsConfig, err := ctx.GetServerTLSConfig()
if err != nil {
t.Fatal(err)
}
addr := util.CreateTestAddr("tcp")
ln, err := util.ListenAndServe(ctx.Stopper, s, addr, tlsConfig)
if err != nil {
t.Fatal(err)
}
return s, ln
}
func (lt *localRPCTransport) Listen(id roachpb.StoreID, handler RaftMessageHandler) error {
ctx := crpc.Context{
Context: base.Context{
Insecure: true,
},
Stopper: lt.stopper,
DisableCache: true,
}
rpcServer := crpc.NewServer(&ctx)
err := rpcServer.RegisterAsync(raftMessageName, false, /*not public*/
func(argsI proto.Message, callback func(proto.Message, error)) {
defer func() {
// TODO(bdarnell): the http/rpc code is swallowing panics somewhere.
if p := recover(); p != nil {
log.Fatalf("caught panic: %s", p)
}
}()
args := argsI.(*RaftMessageRequest)
err := handler(args)
callback(&RaftMessageResponse{}, err)
}, &RaftMessageRequest{})
if err != nil {
return err
}
tlsConfig, err := ctx.GetServerTLSConfig()
if err != nil {
return err
}
addr := util.CreateTestAddr("tcp")
ln, err := util.ListenAndServe(ctx.Stopper, rpcServer, addr, tlsConfig)
if err != nil {
return err
}
lt.mu.Lock()
if _, ok := lt.servers[id]; ok {
log.Fatalf("node %d already listening", id)
}
lt.servers[id] = serverWithAddr{rpcServer, ln.Addr()}
lt.mu.Unlock()
return nil
}
// start dials the remote addr and commences gossip once connected.
// Upon exit, the client is sent on the disconnected channel.
// If the client experienced an error, its err field will
// be set. This method starts client processing in a goroutine and
// returns immediately.
func (c *client) start(g *Gossip, disconnected chan *client, ctx *rpc.Context, stopper *stop.Stopper) {
stopper.RunWorker(func() {
defer func() {
disconnected <- c
}()
var dialOpt grpc.DialOption
if ctx.Insecure {
dialOpt = grpc.WithInsecure()
} else {
tlsConfig, err := ctx.GetClientTLSConfig()
if err != nil {
log.Error(err)
return
}
dialOpt = grpc.WithTransportCredentials(credentials.NewTLS(tlsConfig))
}
conn, err := grpc.Dial(c.addr.String(), dialOpt)
if err != nil {
log.Errorf("failed to dial: %v", err)
return
}
defer func() {
if err := conn.Close(); err != nil {
log.Error(err)
}
}()
c.rpcClient = NewGossipClient(conn)
// Start gossiping.
if err := c.gossip(g, stopper); err != nil {
if !grpcutil.IsClosedConnection(err) {
if c.peerID != 0 {
log.Infof("closing client to node %d (%s): %s", c.peerID, c.addr, err)
} else {
log.Infof("closing client to %s: %s", c.addr, err)
}
}
}
})
}
func newTestServer(t *testing.T, ctx *rpc.Context) (*rpc.Server, net.Listener) {
s := rpc.NewServer(ctx)
tlsConfig, err := ctx.GetServerTLSConfig()
if err != nil {
t.Fatal(err)
}
// We may be called in a loop, meaning tlsConfig is may be in used by a
// running server during a call to `util.ListenAndServe`, which may
// mutate it (due to http2.ConfigureServer). Make a copy to avoid trouble.
tlsConfigCopy := *tlsConfig
addr := util.CreateTestAddr("tcp")
ln, err := util.ListenAndServe(ctx.Stopper, s, addr, &tlsConfigCopy)
if err != nil {
t.Fatal(err)
}
return s, ln
}
func newTestServer(t *testing.T, ctx *rpc.Context, manual bool) (*rpc.Server, net.Listener) {
var s *rpc.Server
if manual {
s = rpc.NewManualServer(ctx)
} else {
s = rpc.NewServer(ctx)
}
tlsConfig, err := ctx.GetServerTLSConfig()
if err != nil {
t.Fatal(err)
}
addr := util.CreateTestAddr("tcp")
ln, err := util.ListenAndServe(ctx.Stopper, s, addr, tlsConfig)
if err != nil {
t.Fatal(err)
}
return s, ln
}
func (lt *localRPCTransport) Listen(id roachpb.StoreID, server ServerInterface) error {
ctx := crpc.Context{
Context: base.Context{
Insecure: true,
},
Stopper: lt.stopper,
DisableCache: true,
}
rpcServer := crpc.NewServer(&ctx)
err := rpcServer.RegisterAsync(raftMessageName, false, /*not public*/
func(argsI proto.Message, callback func(proto.Message, error)) {
args := argsI.(*RaftMessageRequest)
resp, err := server.RaftMessage(args)
callback(resp, err)
}, &RaftMessageRequest{})
if err != nil {
return err
}
tlsConfig, err := ctx.GetServerTLSConfig()
if err != nil {
return err
}
addr := util.CreateTestAddr("tcp")
ln, err := util.ListenAndServe(ctx.Stopper, rpcServer, addr, tlsConfig)
if err != nil {
return err
}
lt.mu.Lock()
if _, ok := lt.servers[id]; ok {
log.Fatalf("node %d already listening", id)
}
lt.servers[id] = serverWithAddr{rpcServer, ln.Addr()}
lt.mu.Unlock()
return nil
}
// sendOne invokes the specified RPC on the supplied client when the
// client is ready. On success, the reply is sent on the channel;
// otherwise an error is sent.
//
// Do not call directly, but instead use sendOneFn. Tests mock out this method
// via sendOneFn in order to test various error cases.
func sendOne(client batchClient, timeout time.Duration,
rpcContext *rpc.Context, done chan batchCall) {
addr := client.remoteAddr
if log.V(2) {
log.Infof("sending request to %s: %+v", addr, client.args)
}
// TODO(tamird/tschottdorf): pass this in from DistSender.
ctx := context.TODO()
if timeout != 0 {
ctx, _ = context.WithTimeout(ctx, timeout)
}
if localServer := rpcContext.GetLocalInternalServerForAddr(addr); enableLocalCalls && localServer != nil {
reply, err := localServer.Batch(ctx, &client.args)
done <- batchCall{reply: reply, err: err}
return
}
go func() {
c := client.conn
for state, err := c.State(); state != grpc.Ready; state, err = c.WaitForStateChange(ctx, state) {
if err != nil {
done <- batchCall{err: newRPCError(
util.Errorf("rpc to %s failed: %s", addr, err))}
return
}
if state == grpc.Shutdown {
done <- batchCall{err: newRPCError(
util.Errorf("rpc to %s failed as client connection was closed", addr))}
return
}
}
reply, err := client.client.Batch(ctx, &client.args)
done <- batchCall{reply: reply, err: err}
}()
}
// sendOne invokes the specified RPC on the supplied client when the
// client is ready. On success, the reply is sent on the channel;
// otherwise an error is sent.
//
// Do not call directly, but instead use sendOneFn. Tests mock out this method
// via sendOneFn in order to test various error cases.
func sendOne(opts SendOptions, rpcContext *rpc.Context, client batchClient, done chan batchCall) {
addr := client.remoteAddr
if log.V(2) {
log.Infof("sending request to %s: %+v", addr, client.args)
}
if localServer := rpcContext.GetLocalInternalServerForAddr(addr); enableLocalCalls && localServer != nil {
ctx, cancel := opts.contextWithTimeout()
defer cancel()
reply, err := localServer.Batch(ctx, &client.args)
done <- batchCall{reply: reply, err: err}
return
}
go func() {
ctx, cancel := opts.contextWithTimeout()
defer cancel()
c := client.conn
for state, err := c.State(); state != grpc.Ready; state, err = c.WaitForStateChange(ctx, state) {
if err != nil {
done <- batchCall{err: newRPCError(
util.Errorf("rpc to %s failed: %s", addr, err))}
return
}
if state == grpc.Shutdown {
done <- batchCall{err: newRPCError(
util.Errorf("rpc to %s failed as client connection was closed", addr))}
return
}
}
reply, err := client.client.Batch(ctx, &client.args)
done <- batchCall{reply: reply, err: err}
}()
}
// Send sends one or more RPCs to clients specified by the slice of
// replicas. On success, Send returns the first successful reply. Otherwise,
// Send returns an error if and as soon as the number of failed RPCs exceeds
// the available endpoints less the number of required replies.
func send(opts SendOptions, replicas ReplicaSlice,
args roachpb.BatchRequest, rpcContext *rpc.Context) (*roachpb.BatchResponse, error) {
if len(replicas) < 1 {
return nil, roachpb.NewSendError(
fmt.Sprintf("insufficient replicas (%d) to satisfy send request of %d",
len(replicas), 1), false)
}
done := make(chan batchCall, len(replicas))
clients := make([]batchClient, 0, len(replicas))
for _, replica := range replicas {
conn, err := rpcContext.GRPCDial(replica.NodeDesc.Address.String())
if err != nil {
return nil, err
}
argsCopy := args
argsCopy.Replica = replica.ReplicaDescriptor
clients = append(clients, batchClient{
remoteAddr: replica.NodeDesc.Address.String(),
conn: conn,
client: roachpb.NewInternalClient(conn),
args: argsCopy,
})
}
// Put known-unhealthy clients last.
nHealthy, err := splitHealthy(clients)
if err != nil {
return nil, err
}
var orderedClients []batchClient
switch opts.Ordering {
case orderStable:
orderedClients = clients
case orderRandom:
// Randomly permute order, but keep known-unhealthy clients last.
shuffleClients(clients[:nHealthy])
shuffleClients(clients[nHealthy:])
orderedClients = clients
}
// TODO(spencer): going to need to also sort by affinity; closest
// ping time should win. Makes sense to have the rpc client/server
// heartbeat measure ping times. With a bit of seasoning, each
// node will be able to order the healthy replicas based on latency.
// Send the first request.
sendOneFn(opts, rpcContext, orderedClients[0], done)
orderedClients = orderedClients[1:]
var errors, retryableErrors int
// Wait for completions.
var sendNextTimer util.Timer
defer sendNextTimer.Stop()
for {
sendNextTimer.Reset(opts.SendNextTimeout)
select {
case <-sendNextTimer.C:
sendNextTimer.Read = true
// On successive RPC timeouts, send to additional replicas if available.
if len(orderedClients) > 0 {
log.Trace(opts.Context, "timeout, trying next peer")
sendOneFn(opts, rpcContext, orderedClients[0], done)
orderedClients = orderedClients[1:]
}
case call := <-done:
err := call.err
if err == nil {
if log.V(2) {
log.Infof("successful reply: %+v", call.reply)
}
return call.reply, nil
}
// Error handling.
if log.V(1) {
log.Warningf("error reply: %s", err)
}
errors++
// Since we have a reconnecting client here, disconnect errors are retryable.
disconnected := err == io.ErrUnexpectedEOF
if retryErr, ok := err.(retry.Retryable); disconnected || (ok && retryErr.CanRetry()) {
retryableErrors++
}
if remainingNonErrorRPCs := len(replicas) - errors; remainingNonErrorRPCs < 1 {
return nil, roachpb.NewSendError(
fmt.Sprintf("too many errors encountered (%d of %d total): %v",
errors, len(clients), err), remainingNonErrorRPCs+retryableErrors >= 1)
}
// Send to additional replicas if available.
if len(orderedClients) > 0 {
log.Trace(opts.Context, "error, trying next peer")
sendOneFn(opts, rpcContext, orderedClients[0], done)
orderedClients = orderedClients[1:]
}
}
}
}
// start dials the remote addr and commences gossip once connected. Upon exit,
// the client is sent on the disconnected channel. This method starts client
// processing in a goroutine and returns immediately.
func (c *client) start(
g *Gossip,
disconnected chan *client,
rpcCtx *rpc.Context,
stopper *stop.Stopper,
nodeID roachpb.NodeID,
breaker *circuit.Breaker,
) {
stopper.RunWorker(func() {
ctx, cancel := context.WithCancel(c.ctx)
var wg sync.WaitGroup
defer func() {
// This closes the outgoing stream, causing any attempt to send or
// receive to return an error.
//
// Note: it is still possible for incoming gossip to be processed after
// this point.
cancel()
// The stream is closed, but there may still be some incoming gossip
// being processed. Wait until that is complete to avoid racing the
// client's removal against the discovery of its remote's node ID.
wg.Wait()
disconnected <- c
}()
consecFailures := breaker.ConsecFailures()
var stream Gossip_GossipClient
if err := breaker.Call(func() error {
// Note: avoid using `grpc.WithBlock` here. This code is already
// asynchronous from the caller's perspective, so the only effect of
// `WithBlock` here is blocking shutdown - at the time of this writing,
// that ends ups up making `kv` tests take twice as long.
conn, err := rpcCtx.GRPCDial(c.addr.String())
if err != nil {
return err
}
if stream, err = NewGossipClient(conn).Gossip(ctx); err != nil {
return err
}
return c.requestGossip(g, stream)
}, 0); err != nil {
if consecFailures == 0 {
log.Warningf(ctx, "node %d: failed to start gossip client: %s", nodeID, err)
}
return
}
// Start gossiping.
log.Infof(ctx, "node %d: started gossip client to %s", nodeID, c.addr)
if err := c.gossip(ctx, g, stream, stopper, &wg); err != nil {
if !grpcutil.IsClosedConnection(err) {
g.mu.Lock()
peerID := c.peerID
g.mu.Unlock()
if peerID != 0 {
log.Infof(ctx, "node %d: closing client to node %d (%s): %s", nodeID, peerID, c.addr, err)
} else {
log.Infof(ctx, "node %d: closing client to %s: %s", nodeID, c.addr, err)
}
}
}
})
}
