embed: wait up to election-timeout for pending RPCs when closing

Both grpc.Server.Stop and grpc.Server.GracefulStop close the listeners
first, to stop accepting the new connections. GracefulStop blocks until
all clients close their open transports(connections). Unary RPCs
only take a few seconds to finish. Stream RPCs, like watch, might never
close the connections from client side, thus making gRPC server wait
forever.

This patch still calls GracefulStop, but waits up to 10s before manually
closing the open transports.

Address etcd-io#8224.

Signed-off-by: Gyu-Ho Lee <gyuhox@gmail.com>

embed/etcd.go

@@ -189,11 +189,30 @@ func (e *Etcd) Config() Config {
 func (e *Etcd) Close() {
 	e.closeOnce.Do(func() { close(e.stopc) })
 
-	// (gRPC server) stops accepting new connections,
-	// RPCs, and blocks until all pending RPCs are finished
+	// 2s for RPC finish time
+	timeout := 2 * time.Second
+	if e.Server != nil {
+		timeout += e.Server.Cfg.ElectionTimeout()
+	}
 	for _, sctx := range e.sctxs {
 		for gs := range sctx.grpcServerC {
-			gs.GracefulStop()
+			ch := make(chan struct{})
+			go func() {
+				defer close(ch)
+				// close listeners to stop accepting new connections,
+				// will block on any existing transports
+				gs.GracefulStop()
+			}()
+			// wait until all pending RPCs are finished
+			select {
+			case <-ch:
+			case <-time.After(timeout):
+				// took too long, manually close open transports
+				// e.g. watch streams
+				gs.Stop()
+				// concurrent GracefulStop should be interrupted
+				<-ch
+			}
 		}
 	}
 

etcdserver/config.go

@@ -175,7 +175,7 @@ func (c *ServerConfig) ReqTimeout() time.Duration {
 	return 5*time.Second + 2*time.Duration(c.ElectionTicks)*time.Duration(c.TickMs)*time.Millisecond
 }
 
-func (c *ServerConfig) electionTimeout() time.Duration {
+func (c *ServerConfig) ElectionTimeout() time.Duration {
 	return time.Duration(c.ElectionTicks) * time.Duration(c.TickMs) * time.Millisecond
 }
 

etcdserver/server.go

@@ -1313,7 +1313,7 @@ func (s *EtcdServer) applyEntryNormal(e *raftpb.Entry) {
 		// promote lessor when the local member is leader and finished
 		// applying all entries from the last term.
 		if s.isLeader() {
-			s.lessor.Promote(s.Cfg.electionTimeout())
+			s.lessor.Promote(s.Cfg.ElectionTimeout())
 		}
 		return
 	}