在云原生与微服务主导架构模式的时代,内部服务间交互所采用的通信协议选型无非就是两类:HTTP API(RESTful API)和RPC。在如今的硬件配置与网络条件下,现代RPC实现的性能一般都是好于HTTP API的。我们以json over http与gRPC[1](insecure)作比较,分别使用ghz[2]和hey[3]压测gRPC和json over http的实现,gRPC的性能(Requests/sec: 59924.34)要比http api性能(Requests/sec: 49969.9234)高出20%。实测gPRC使用的protobuf的编解码性能更是最快的json编解码的2-3倍,是Go标准库json包编解码性能的10倍以上(具体数据见本文附录)。

对于性能敏感并且内部通信协议较少变动的系统来说,内部服务使用RPC可能是多数人的选择。而gRPC虽然不是性能最好的RPC实现,但作为有谷歌大厂背书且是CNCF唯一的RPC项目[4],gRPC自然得到了开发人员最广泛的关注与使用。

本文也来说说gRPC,不过我们更多关注一下gRPC的客户端,我们来看看使用gRPC客户端时都会考虑的那些事情(本文所有代码基于gRPC v1.40.0版本,Go 1.17版本)。

1. 默认的gRPC的客户端

gRPC支持四种通信模式,它们是(以下四张图截自《gRPC: Up and Running》[5]一书):

  • 简单RPC(Simple RPC):最简单的,也是最常用的gRPC通信模式,简单来说就是一请求一应答

grpc连接的数目 grpc客户端连接数_办公软件

  • 服务端流RPC(Server-streaming RPC):一请求,多应答

grpc连接的数目 grpc客户端连接数_proxy_02

  • 客户端流RPC(Client-streaming RPC):多请求,一应答

grpc连接的数目 grpc客户端连接数_办公软件_03

  • 双向流RPC(Bidirectional-Streaming RPC):多请求,多应答

grpc连接的数目 grpc客户端连接数_go_04

我们以最常用的Simple RPC(也称Unary RPC)为例来看一下如何实现一个gRPC版的helloworld。

我们无需自己从头来编写helloworld.proto并生成相应的gRPC代码,gRPC官方提供了一个helloworld的例子[6],我们仅需对其略微改造一下即可。

helloworld例子的IDL文件helloworld.proto如下:

// https://github.com/grpc/grpc-go/tree/master/examples/helloworld/helloworld/helloworld.proto

syntax = "proto3";

option go_package = "google.golang.org/grpc/examples/helloworld/helloworld";
option java_multiple_files = true;
option java_package = "io.grpc.examples.helloworld";
option java_outer_classname = "HelloWorldProto";

package helloworld;

// The greeting service definition.
service Greeter {
  // Sends a greeting
  rpc SayHello (HelloRequest) returns (HelloReply) {}
}

// The request message containing the user's name.
message HelloRequest {
  string name = 1;
}

// The response message containing the greetings
message HelloReply {
  string message = 1;
}

对.proto文件的规范讲解大家可以参考grpc官方文档[7],这里不赘述。显然上面这个IDL是极致简单的。这里定义了一个service:Greeter,它仅包含一个方法SayHello,并且这个方法的参数与返回值都是一个仅包含一个string字段的结构体。

我们无需手工执行protoc命令来基于该.proto文件生成对应的Greeter service的实现以及HelloRequest、HelloReply的protobuf编解码实现,因为gRPC在example下已经放置了生成后的Go源文件,我们直接引用即可。这里要注意,最新的grpc-go项目仓库[8]采用了多module的管理模式,examples作为一个独立的go module而存在,因此我们需要将其单独作为一个module导入到其使用者的项目中。以gRPC客户端greeter_client为例,它的go.mod要这样来写:

// https://github.com/bigwhite/experiments/tree/master/grpc-client/demo1/greeter_client/go.mod
module github.com/bigwhite/grpc-client/demo1

go 1.17

require (
 google.golang.org/grpc v1.40.0
 google.golang.org/grpc/examples v1.40.0
)

require (
 github.com/golang/protobuf v1.4.3 // indirect
 golang.org/x/net v0.0.0-20201021035429-f5854403a974 // indirect
 golang.org/x/sys v0.0.0-20200930185726-fdedc70b468f // indirect
 golang.org/x/text v0.3.3 // indirect
 google.golang.org/genproto v0.0.0-20200806141610-86f49bd18e98 // indirect
 google.golang.org/protobuf v1.25.0 // indirect
)

replace google.golang.org/grpc v1.40.0 => /Users/tonybai/Go/src/github.com/grpc/grpc-go

replace google.golang.org/grpc/examples v1.40.0 => /Users/tonybai/Go/src/github.com/grpc/grpc-go/examples

注:grpc-go项目的标签(tag)似乎打的有问题,由于没有打grpc/examples/v1.40.0标签,go命令在grpc-go的v1.40.0标签中找不到examples,因此上面的go.mod中使用了一个replace trick(example module的v1.40.0版本是假的哦),将examples module指向本地的代码。

gRPC通信的两端我们也稍作改造。原greeter_client仅发送一个请求便退出,这里我们将其改为每隔2s发送请求(便于后续观察),如下面代码所示:

// https://github.com/bigwhite/experiments/tree/master/grpc-client/demo1/greeter_client/main.go
... ...
func main() {
    // Set up a connection to the server.
    ctx, cf1 := context.WithTimeout(context.Background(), time.Second*3)
    defer cf1()
    conn, err := grpc.DialContext(ctx, address, grpc.WithInsecure(), grpc.WithBlock())
    if err != nil {
        log.Fatalf("did not connect: %v", err)
    }
    defer conn.Close()
    c := pb.NewGreeterClient(conn)

    // Contact the server and print out its response.
    name := defaultName
    if len(os.Args) > 1 {
        name = os.Args[1]
    }

    for i := 0; ; i++ {
        ctx, _ := context.WithTimeout(context.Background(), time.Second)
        r, err := c.SayHello(ctx, &pb.HelloRequest{Name: fmt.Sprintf("%s-%d", name, i+1)})
        if err != nil {
            log.Fatalf("could not greet: %v", err)
        }
        log.Printf("Greeting: %s", r.GetMessage())
        time.Sleep(2 * time.Second)
    }
}

greeter_server加了一个命令行选项-port并支持gRPC server的优雅退出[9]

// https://github.com/bigwhite/experiments/tree/master/grpc-client/demo1/greeter_server/main.go
... ...

var port int

func init() {
    flag.IntVar(&port, "port", 50051, "listen port")
}

func main() {
    flag.Parse()
    lis, err := net.Listen("tcp", fmt.Sprintf("localhost:%d", port))
    if err != nil {
        log.Fatalf("failed to listen: %v", err)
    }
    s := grpc.NewServer()
    pb.RegisterGreeterServer(s, &server{})

    go func() {
        if err := s.Serve(lis); err != nil {
            log.Fatalf("failed to serve: %v", err)
        }
    }()

    var c = make(chan os.Signal)
    signal.Notify(c, os.Interrupt, os.Kill)
    <-c
    s.Stop()
    fmt.Println("exit")
}

搞定go.mod以及对client和server进行改造ok后,我们就可以来构建和运行greeter_client和greeter_server了:

编译和启动server:

$cd grpc-client/demo1/greeter_server
$make
$./demo1-server -port 50051
2021/09/11 12:10:33 Received: world-1
2021/09/11 12:10:35 Received: world-2
2021/09/11 12:10:37 Received: world-3
... ...

编译和启动client:
$cd grpc-client/demo1/greeter_client
$make
$./demo1-client
2021/09/11 12:10:33 Greeting: Hello world-1
2021/09/11 12:10:35 Greeting: Hello world-2
2021/09/11 12:10:37 Greeting: Hello world-3
... ...

我们看到:greeter_client和greeter_server启动后可以正常的通信!我们重点看一下greeter_client。

greeter_client在Dial服务端时传给DialContext的target参数是一个静态的服务地址:

const (
      address     = "localhost:50051"
)

这个形式的target经过google.golang.org/grpc/internal/grpcutil.ParseTarget的解析后返回一个值为nil的resolver.Target。于是gRPC采用默认的scheme:"passthrough"(github.com/grpc/grpc-go/resolver/resolver.go),默认的"passthrough" scheme下,gRPC将使用内置的passthrough resolver(google.golang.org/grpc/internal/resolver/passthrough)。默认的这个passthrough resolver是如何设置要连接的service地址的呢?下面是passthrough resolver的代码摘录:

// github.com/grpc/grpc-go/internal/resolver/passthrough/passthrough.go

func (r *passthroughResolver) start() {
    r.cc.UpdateState(resolver.State{Addresses: []resolver.Address{{Addr: r.target.Endpoint}}})
}

我们看到它将target.Endpoint,即localhost:50051直接传给了ClientConnection(上面代码的r.cc),后者将向这个地址建立tcp连接。这正应了该resolver的名字:passthrough

上面greeter_client连接的仅仅是service的一个实例(instance),如果我们同时启动了该service的三个实例,比如使用goreman[10]通过加载脚本文件来启动多个service实例:

// https://github.com/bigwhite/experiments/tree/master/grpc-client/demo1/greeter_server/Procfile

# Use goreman to run `go get github.com/mattn/goreman`
demo1-server1: ./demo1-server -port 50051
demo1-server2: ./demo1-server -port 50052
demo1-server3: ./demo1-server -port 50053

同时启动多实例:

$goreman start
15:22:12 demo1-server3 | Starting demo1-server3 on port 5200
15:22:12 demo1-server2 | Starting demo1-server2 on port 5100
15:22:12 demo1-server1 | Starting demo1-server1 on port 5000

那么我们应该如何告诉greeter_client去连接这三个实例呢?是否可以将address改为下面这样就可以了呢:

const (
 address     = "localhost:50051,localhost:50052,localhost:50053"
 defaultName = "world"
)

我们来改改试试,修改后重新编译greeter_client,启动greeter_client,我们看到下面结果:

$./demo1-client
2021/09/11 15:26:32 did not connect: context deadline exceeded

greeter_client连接server超时!也就是说像上面这样简单的传入多个实例的地址是不行的!那问题来了!我们该怎么让greeter_client去连接一个service的多个实例呢?我们继续向下看。

2. 连接一个Service的多个实例(instance)

grpc.Dial/grpc.DialContext的参数target可不仅仅是service实例的服务地址这么简单,它的实参(argument)形式决定了gRPC client将采用哪一个resolver来确定service实例的地址集合

下面我们以一个返回service实例地址静态集合(即service的实例数量固定且服务地址固定)的StaticResolver为例,来看如何让gRPC client连接一个Service的多个实例。

1) StaticResolver

我们首先来设计一下传给grpc.DialContext的target形式。关于gRPC naming resolution,gRPC有专门文档说明[11]。在这里,我们也创建一个新的scheme:static,多个service instance的服务地址通过逗号分隔的字符串传入,如下面代码:

// https://github.com/bigwhite/experiments/tree/master/grpc-client/demo2/greeter_client/main.go

const (
      address = "static:///localhost:50051,localhost:50052,localhost:50053"
)

当address被作为target的实参传入grpc.DialContext后,它会被grpcutil.ParseTarget解析为一个resolver.Target结构体,该结构体包含三个字段:

// github.com/grpc/grpc-go/resolver/resolver.go 
type Target struct {
    Scheme    string
    Authority string
    Endpoint  string
}

其中Scheme为"static",Authority为空,Endpoint为"localhost:50051,localhost:50052,localhost:50053"。

接下来,gRPC会根据Target.Scheme的值到resolver包中的builder map中查找是否有对应的Resolver Builder实例。到目前为止gRPC内置的的resolver Builder都无法匹配该Scheme值。是时候自定义一个StaticResolver的Builder了!

grpc的resolve包定义了一个Builder实例需要实现的接口:

// github.com/grpc/grpc-go/resolver/resolver.go 

// Builder creates a resolver that will be used to watch name resolution updates.
type Builder interface {
    // Build creates a new resolver for the given target.
    //
    // gRPC dial calls Build synchronously, and fails if the returned error is
    // not nil.
    Build(target Target, cc ClientConn, opts BuildOptions) (Resolver, error)
    // Scheme returns the scheme supported by this resolver.
    // Scheme is defined at https://github.com/grpc/grpc/blob/master/doc/naming.md.
    Scheme() string
}

Scheme方法返回这个Builder对应的scheme,而Build方法则是真正用于构建Resolver实例的方法,我们来看一下StaticBuilder的实现:

// https://github.com/bigwhite/experiments/tree/master/grpc-client/demo2/greeter_client/builder.go

func init() {
    resolver.Register(&StaticBuilder{}) //在init函数中将StaticBuilder实例注册到resolver包的Resolver map中
}

type StaticBuilder struct{}

func (sb *StaticBuilder) Build(target resolver.Target, cc resolver.ClientConn,
    opts resolver.BuildOptions) (resolver.Resolver, error) {

    // 解析target.Endpoint (例如:localhost:50051,localhost:50052,localhost:50053)
    endpoints := strings.Split(target.Endpoint, ",")

    r := &StaticResolver{
        endpoints: endpoints,
        cc:        cc,
    }
    r.ResolveNow(resolver.ResolveNowOptions{})
    return r, nil
}

func (sb *StaticBuilder) Scheme() string {
    return "static" // 返回StaticBuilder对应的scheme字符串
}

在这个StaticBuilder实现中,init函数在包初始化是就将一个StaticBuilder实例注册到resolver包的Resolver map中。这样gRPC在Dial时就能通过target中的scheme找到该builder。Build方法是StaticBuilder的关键,在这个方法中,它首先解析传入的target.Endpoint,得到三个service instance的服务地址并存到新创建的StaticResolver实例中,并调用StaticResolver实例的ResolveNow方法确定即将连接的service instance集合。

和Builder一样,grpc的resolver包也定义了每个resolver需要实现的Resolver接口:

// github.com/grpc/grpc-go/resolver/resolver.go 

// Resolver watches for the updates on the specified target.
// Updates include address updates and service config updates.
type Resolver interface {
    // ResolveNow will be called by gRPC to try to resolve the target name
    // again. It's just a hint, resolver can ignore this if it's not necessary.
    //
    // It could be called multiple times concurrently.
    ResolveNow(ResolveNowOptions)
    // Close closes the resolver.
    Close()
}

从这个接口注释我们也能看出,Resolver的实现负责监视(watch)服务测的地址与配置变化,并将变化更新给grpc的ClientConn。我们来看看我们的StaticResolver的实现:

// https://github.com/bigwhite/experiments/tree/master/grpc-client/demo2/greeter_client/resolver.go

type StaticResolver struct {
    endpoints []string
    cc        resolver.ClientConn
    sync.Mutex
}

func (r *StaticResolver) ResolveNow(opts resolver.ResolveNowOptions) {
    r.Lock()
    r.doResolve()
    r.Unlock()
}

func (r *StaticResolver) Close() {
}

func (r *StaticResolver) doResolve() {
    var addrs []resolver.Address
    for i, addr := range r.endpoints {
        addrs = append(addrs, resolver.Address{
            Addr:       addr,
            ServerName: fmt.Sprintf("instance-%d", i+1),
        })
    }

    newState := resolver.State{
        Addresses: addrs,
    }

    r.cc.UpdateState(newState)
}

注:resolver.Resolver接口的注释要求ResolveNow方法是要支持并发安全的,所以这里我们通过sync.Mutex来实现同步。

由于服务侧的服务地址数量与信息都是不变的,因此这里并没有watch和update的过程,而只是在实现了ResolveNow(并在Builder中的Build方法中调用),在ResolveNow中将service instance的地址集合更新给ClientConnection(r.cc)。

接下来我们来编译与运行一下demo2的client与server:

$cd grpc-client/demo2/greeter_server
$make
$goreman start
22:58:21 demo2-server1 | Starting demo2-server1 on port 5000
22:58:21 demo2-server2 | Starting demo2-server2 on port 5100
22:58:21 demo2-server3 | Starting demo2-server3 on port 5200


$cd grpc-client/demo2/greeter_client
$make
$./demo2-client

执行一段时间后,你会在server端的日志中发现一个问题,如下日志所示:

22:57:16 demo2-server1 | 2021/09/11 22:57:16 Received: world-1
22:57:18 demo2-server1 | 2021/09/11 22:57:18 Received: world-2
22:57:20 demo2-server1 | 2021/09/11 22:57:20 Received: world-3
22:57:22 demo2-server1 | 2021/09/11 22:57:22 Received: world-4
22:57:24 demo2-server1 | 2021/09/11 22:57:24 Received: world-5
22:57:26 demo2-server1 | 2021/09/11 22:57:26 Received: world-6
22:57:28 demo2-server1 | 2021/09/11 22:57:28 Received: world-7
22:57:30 demo2-server1 | 2021/09/11 22:57:30 Received: world-8
22:57:32 demo2-server1 | 2021/09/11 22:57:32 Received: world-9

我们的Service instance集合中明明有三个地址,为何只有server1收到了rpc请求,其他两个server都处于空闲状态呢?这是客户端的负载均衡策略在作祟!默认情况下,grpc会为客户端选择内置的“pick_first”负载均衡策略,即在service instance集合中选择第一个intance进行请求。在这个例子中,在pick_first策略的作用下,grpc总是会选择demo2-server1发起rpc请求。

如果要将请求发到各个server上,我们可以将负载均衡策略改为另外一个内置的策略:round_robin,就像下面代码这样:

// https://github.com/bigwhite/experiments/tree/master/grpc-client/demo2/greeter_client/main.go

conn, err := grpc.DialContext(ctx, address, grpc.WithInsecure(), grpc.WithBlock(), grpc.WithBalancerName("round_robin"))

重新编译运行greeter_client后,在server测我们就可以看到rpc请求被轮询地发到了每个server instance上了。

2) Resolver原理

我们再来用一幅图来梳理一下Builder以及Resolver的工作原理:

grpc连接的数目 grpc客户端连接数_proxy_05

图中的SchemeResolver泛指实现了某一特定scheme的resolver。如图所示,service instance集合resolve过程的步骤大致如下:

  1. SchemeBuilder将自身实例注册到resolver包的map中;
  2. grpc.Dial/DialContext时使用特定形式的target参数
  3. 对target解析后,根据target.Scheme到resolver包的map中查找Scheme对应的Buider;
  4. 调用Buider的Build方法
  5. Build方法构建出SchemeResolver实例;
  6. 后续由SchemeResolver实例监视service instance变更状态并在有变更的时候更新ClientConnection。
3) NacosResolver

在生产环境中,考虑到服务的高可用、可伸缩等,我们很少使用固定地址、固定数量的服务实例集合,更多是通过服务注册和发现机制自动实现服务实例集合的更新。这里我们再来实现一个基于nacos[12]的NacosResolver,实现服务实例变更时grpc Client的自动调整(注:nacos的本地单节点安装方案见文本附录),让示例具实战意义^_^。

由于有了上面关于Resolver原理的描述,这里简化了一些描述。

首先和StaticResolver一样,我们也来设计一下target的形式。nacos有namespace, group的概念,因此我们将target设计为如下形式:

nacos://[authority]/host:port/namespace/group/serviceName

具体到我们的greeter_client中,其address为:

// https://github.com/bigwhite/experiments/tree/master/grpc-client/demo3/greeter_client/main.go

const (
      address = "nacos:///localhost:8848/public/group-a/demo3-service" //no authority
)

接下来我们来看NacosBuilder:

// https://github.com/bigwhite/experiments/tree/master/grpc-client/demo3/greeter_client/builder.go

func (nb *NacosBuilder) Build(target resolver.Target,
    cc resolver.ClientConn,
    opts resolver.BuildOptions) (resolver.Resolver, error) {

    // use info in target to access naming service
    // parse the target.endpoint
    // target.Endpoint - localhost:8848/public/DEFAULT_GROUP/serviceName, the addr of naming service :nacos endpoint
    sl := strings.Split(target.Endpoint, "/")
    nacosAddr := sl[0]
    namespace := sl[1]
    group := sl[2]
    serviceName := sl[3]
    sl1 := strings.Split(nacosAddr, ":")
    host := sl1[0]
    port := sl1[1]
    namingClient, err := initNamingClient(host, port, namespace, group)
    if err != nil {
        return nil, err
    }

    r := &NacosResolver{
        namingClient: namingClient,
        cc:           cc,
        namespace:    namespace,
        group:        group,
        serviceName:  serviceName,
    }

    // initialize the cc's states
    r.ResolveNow(resolver.ResolveNowOptions{})

    // subscribe and watch
    r.watch()
    return r, nil
}

func (nb *NacosBuilder) Scheme() string {
    return "nacos"
}

NacosBuilder的Build方法流程也StaticBuilder并无二致,首先我们也是解析传入的target的Endpoint,即"localhost:8848/public/group-a/demo3-service",并将解析后的各段信息存入新创建的NacosResolver实例中备用。NacosResolver还需要一个信息,那就是与nacos的连接,这里用initNamingClient创建一个nacos client端实例(调用nacos提供的go sdk[13])。

接下来我们调用NacosResolver的ResolveNow获取一次nacos上demo3-service的服务实例列表并初始化ClientConn,最后我们调用NacosResolver的watch方法来订阅并监视demo3-service的实例变化。下面是NacosResolver的部分实现:

// https://github.com/bigwhite/experiments/tree/master/grpc-client/demo3/greeter_client/resolver.go

func (r *NacosResolver) doResolve(opts resolver.ResolveNowOptions) {
 instances, err := r.namingClient.SelectAllInstances(vo.SelectAllInstancesParam{
  ServiceName: r.serviceName,
  GroupName:   r.group,
 })
 if err != nil {
  fmt.Println(err)
  return
 }

 if len(instances) == 0 {
  fmt.Printf("service %s has zero instance\n", r.serviceName)
  return
 }

 // update cc.States
 var addrs []resolver.Address
 for i, inst := range instances {
  if (!inst.Enable) || (inst.Weight == 0) {
   continue
  }

  addrs = append(addrs, resolver.Address{
   Addr:       fmt.Sprintf("%s:%d", inst.Ip, inst.Port),
   ServerName: fmt.Sprintf("instance-%d", i+1),
  })
 }

 if len(addrs) == 0 {
  fmt.Printf("service %s has zero valid instance\n", r.serviceName)
 }

 newState := resolver.State{
  Addresses: addrs,
 }

 r.Lock()
 r.cc.UpdateState(newState)
 r.Unlock()
}

func (r *NacosResolver) ResolveNow(opts resolver.ResolveNowOptions) {
 r.doResolve(opts)
}

func (r *NacosResolver) Close() {
 r.namingClient.Unsubscribe(&vo.SubscribeParam{
  ServiceName: r.serviceName,
  GroupName:   r.group,
 })
}

func (r *NacosResolver) watch() {
 r.namingClient.Subscribe(&vo.SubscribeParam{
  ServiceName: r.serviceName,
  GroupName:   r.group,
  SubscribeCallback: func(services []model.SubscribeService, err error) {
   fmt.Printf("subcallback: %#v\n", services)
   r.doResolve(resolver.ResolveNowOptions{})
  },
 })
}

这里的一个重要实现是ResolveNow和watch都调用的doResolve方法,该方法通过nacos-go sdk中的SelectAllInstances获取demo-service3的所有实例,并将得到的enabled(=true)和权重(weight)不为0的合法实例集合更新给ClientConn(r.cc.UpdateState)。

在NacosResolver的watch方法中,我们通过nacos-go sdk中的Subscribe方法订阅demo3-service并提供了一个回调函数。这样每当demo3-service的实例发生变化时,该回调会被调用。在该回调中我们可以基于传回的最新的service实例集合(services []model.SubscribeService)来更新ClientConn,但在这里我们复用了doResolve方法,即又去nacos获取一次demo-service3的实例。

编译运行demo3下greeter_server:

$cd grpc-client/demo3/greeter_server
$make
$goreman start
06:06:02 demo3-server3 | Starting demo3-server3 on port 5200
06:06:02 demo3-server1 | Starting demo3-server1 on port 5000
06:06:02 demo3-server2 | Starting demo3-server2 on port 5100
06:06:02 demo3-server3 | 2021-09-12T06:06:02.913+0800 INFO nacos_client/nacos_client.go:87 logDir:</tmp/nacos/log/50053>   cacheDir:</tmp/nacos/cache/50053>
06:06:02 demo3-server2 | 2021-09-12T06:06:02.913+0800 INFO nacos_client/nacos_client.go:87 logDir:</tmp/nacos/log/50052>   cacheDir:</tmp/nacos/cache/50052>
06:06:02 demo3-server1 | 2021-09-12T06:06:02.913+0800 INFO nacos_client/nacos_client.go:87 logDir:</tmp/nacos/log/50051>   cacheDir:</tmp/nacos/cache/50051>

运行greeter_server后,我们在nacos dashboard上会看到demo-service3的所有实例信息:

grpc连接的数目 grpc客户端连接数_proxy_06

grpc连接的数目 grpc客户端连接数_go_07

编译运行demo3下greeter_client:

$cd grpc-client/demo3/greeter_client
$make
$./demo3-client
2021-09-12T06:08:25.551+0800 INFO nacos_client/nacos_client.go:87 logDir:</Users/tonybai/go/src/github.com/bigwhite/experiments/grpc-client/demo3/greeter_client/log>   cacheDir:</Users/tonybai/go/src/github.com/bigwhite/experiments/grpc-client/demo3/greeter_client/cache>
2021/09/12 06:08:25 Greeting: Hello world-1
2021/09/12 06:08:27 Greeting: Hello world-2
2021/09/12 06:08:29 Greeting: Hello world-3
2021/09/12 06:08:31 Greeting: Hello world-4
2021/09/12 06:08:33 Greeting: Hello world-5
2021/09/12 06:08:35 Greeting: Hello world-6
... ...

由于采用了round robin负载策略,greeter_server侧每个server(权重都为1)都会平等的收到rpc请求:

06:06:36 demo3-server1 | 2021/09/12 06:06:36 Received: world-1
06:06:38 demo3-server3 | 2021/09/12 06:06:38 Received: world-2
06:06:40 demo3-server2 | 2021/09/12 06:06:40 Received: world-3
06:06:42 demo3-server1 | 2021/09/12 06:06:42 Received: world-4
06:06:44 demo3-server3 | 2021/09/12 06:06:44 Received: world-5
06:06:46 demo3-server2 | 2021/09/12 06:06:46 Received: world-6
... ...

这时我们可以通过nacos dashboard调整demo3-service的实例权重或下线某个实例,比如下线service instance-2(端口50052),之后我们会看到greeter_client回调函数执行,之后greeter_server侧将只有实例1和实例3收到rpc请求。重新上线service instance-2后,一切会恢复正常。

3. 自定义客户端balancer

现实中服务端的实例所部署的主机(虚拟机/容器)算力可能不同,如果所有实例都使用相同权重1,那么肯定是不科学且存在算力浪费。但grpc-go内置的balancer实现有限,不能满足我们需求,我们就需要自定义一个可以满足我们需求的balancer了。

这里我们以自定义一个Weighted Round Robin(wrr) Balancer为例,看看自定义balancer的步骤(我们参考grpc-go中内置round_robin的实现[14])。

和resolver包相似,balancer也是通过一个Builder(创建模式)来实例化的,并且balancer的Balancer接口与resolver.Balancer差不多:

// github.com/grpc/grpc-go/balancer/balancer.go 

// Builder creates a balancer.
type Builder interface {
    // Build creates a new balancer with the ClientConn.
    Build(cc ClientConn, opts BuildOptions) Balancer
    // Name returns the name of balancers built by this builder.
    // It will be used to pick balancers (for example in service config).
    Name() string
}

通过Builder.Build方法我们构建一个Balancer接口的实现,Balancer接口定义如下:

// github.com/grpc/grpc-go/balancer/balancer.go 

type Balancer interface {
    // UpdateClientConnState is called by gRPC when the state of the ClientConn
    // changes.  If the error returned is ErrBadResolverState, the ClientConn
    // will begin calling ResolveNow on the active name resolver with
    // exponential backoff until a subsequent call to UpdateClientConnState
    // returns a nil error.  Any other errors are currently ignored.
    UpdateClientConnState(ClientConnState) error
    // ResolverError is called by gRPC when the name resolver reports an error.
    ResolverError(error)
    // UpdateSubConnState is called by gRPC when the state of a SubConn
    // changes.
    UpdateSubConnState(SubConn, SubConnState)
    // Close closes the balancer. The balancer is not required to call
    // ClientConn.RemoveSubConn for its existing SubConns.
    Close()
}

可以看到,Balancer要比Resolver要复杂很多。gRPC的核心开发者们也看到了这一点,于是他们提供了一个可简化自定义Balancer创建的包:google.golang.org/grpc/balancer/base。gRPC内置的round_robin Balancer也是基于base包实现的。

base包提供了NewBalancerBuilder可以快速返回一个balancer.Builder的实现:

// github.com/grpc/grpc-go/balancer/base/base.go 

// NewBalancerBuilder returns a base balancer builder configured by the provided config.
func NewBalancerBuilder(name string, pb PickerBuilder, config Config) balancer.Builder {
    return &baseBuilder{
        name:          name,
        pickerBuilder: pb,
        config:        config,
    }
}

我们看到,这个函数接收一个参数:pb,它的类型是PikcerBuilder,这个接口类型则比较简单:

// github.com/grpc/grpc-go/balancer/base/base.go 

// PickerBuilder creates balancer.Picker.
type PickerBuilder interface {
    // Build returns a picker that will be used by gRPC to pick a SubConn.
    Build(info PickerBuildInfo) balancer.Picker
}

我们仅需要提供一个PickerBuilder的实现以及一个balancer.Picker的实现即可,而Picker则是仅有一个方法的接口类型:

// github.com/grpc/grpc-go/balancer/balancer.go 

type Picker interface {
    Pick(info PickInfo) (PickResult, error)
}

嵌套的有些多,我们用下面这幅图来直观看一下balancer的创建和使用流程:

grpc连接的数目 grpc客户端连接数_grpc连接的数目_08

再简述一下大致流程:

  • 首先要注册一个名为"my_weighted_round_robin"的balancer Builder:wrrBuilder,该Builder由base包的NewBalancerBuilder构建;
  • base包的NewBalancerBuilder函数需要传入一个PickerBuilder实现,于是我们需要自定义一个返回Picker接口实现的PickerBuilder。
  • grpc.Dial调用时传入一个WithBalancerName("my_weighted_round_robin"),grpc通过balancer Name从已注册的balancer builder中选出我们实现的wrrBuilder,并调用wrrBuilder创建Picker:wrrPicker。
  • 在grpc实施rpc调用SayHello时,wrrPicker的Pick方法会被调用,选出一个Connection,并在该connection上发送rpc请求。

由于用到的权重值,我们的resolver实现需要做一些变动,主要是在doResolve方法时将service instance的权重(weight)通过Attribute设置到ClientConnection中:

// https://github.com/bigwhite/experiments/tree/master/grpc-client/demo4/greeter_client/resolver.go

func (r *NacosResolver) doResolve(opts resolver.ResolveNowOptions) {
 instances, err := r.namingClient.SelectAllInstances(vo.SelectAllInstancesParam{
  ServiceName: r.serviceName,
  GroupName:   r.group,
 })
 if err != nil {
  fmt.Println(err)
  return
 }

 if len(instances) == 0 {
  fmt.Printf("service %s has zero instance\n", r.serviceName)
  return
 }

 // update cc.States
 var addrs []resolver.Address
 for i, inst := range instances {
  if (!inst.Enable) || (inst.Weight == 0) {
   continue
  }

  addr := resolver.Address{
   Addr:       fmt.Sprintf("%s:%d", inst.Ip, inst.Port),
   ServerName: fmt.Sprintf("instance-%d", i+1),
  }
  addr.Attributes = addr.Attributes.WithValues("weight", int(inst.Weight)) //考虑权重并纳入cc的状态中
  addrs = append(addrs, addr)
 }

 if len(addrs) == 0 {
  fmt.Printf("service %s has zero valid instance\n", r.serviceName)
 }

 newState := resolver.State{
  Addresses: addrs,
 }

 r.Lock()
 r.cc.UpdateState(newState)
 r.Unlock()
}

接下来我们重点看看greeter_client中wrrPickerBuilder与wrrPicker的实现:

// https://github.com/bigwhite/experiments/tree/master/grpc-client/demo4/greeter_client/balancer.go

type wrrPickerBuilder struct{}

func (*wrrPickerBuilder) Build(info base.PickerBuildInfo) balancer.Picker {
 if len(info.ReadySCs) == 0 {
  return base.NewErrPicker(balancer.ErrNoSubConnAvailable)
 }

 var scs []balancer.SubConn
 // 提取已经就绪的connection的权重信息,作为Picker实例的输入
 for subConn, addr := range info.ReadySCs {
  weight := addr.Address.Attributes.Value("weight").(int)
  if weight <= 0 {
   weight = 1
  }
  for i := 0; i < weight; i++ {
   scs = append(scs, subConn)
  }
 }

 return &wrrPicker{
  subConns: scs,
  // Start at a random index, as the same RR balancer rebuilds a new
  // picker when SubConn states change, and we don't want to apply excess
  // load to the first server in the list.
  next: rand.Intn(len(scs)),
 }
}

type wrrPicker struct {
 // subConns is the snapshot of the roundrobin balancer when this picker was
 // created. The slice is immutable. Each Get() will do a round robin
 // selection from it and return the selected SubConn.
 subConns []balancer.SubConn

 mu   sync.Mutex
 next int
}

// 选出一个Connection
func (p *wrrPicker) Pick(info balancer.PickInfo) (balancer.PickResult, error) {
 p.mu.Lock()
 sc := p.subConns[p.next]
 p.next = (p.next + 1) % len(p.subConns)
 p.mu.Unlock()
 return balancer.PickResult{SubConn: sc}, nil
}

这是一个简单的Weighted Round Robin实现,加权算法十分简单,如果一个conn的权重为n,那么就在加权结果集中加入n个conn,这样在后续Pick时不需要考虑加权的问题,只需向普通Round Robin那样逐个Pick出来即可。

运行demo4 greeter_server后,我们在nacos将instance-1的权重改为5,我们后续就会看到如下输出:

$goreman start
09:20:18 demo4-server3 | Starting demo4-server3 on port 5200
09:20:18 demo4-server2 | Starting demo4-server2 on port 5100
09:20:18 demo4-server1 | Starting demo4-server1 on port 5000
09:20:18 demo4-server2 | 2021-09-12T09:20:18.633+0800 INFO nacos_client/nacos_client.go:87 logDir:</tmp/nacos/log/50052>   cacheDir:</tmp/nacos/cache/50052>
09:20:18 demo4-server1 | 2021-09-12T09:20:18.633+0800 INFO nacos_client/nacos_client.go:87 logDir:</tmp/nacos/log/50051>   cacheDir:</tmp/nacos/cache/50051>
09:20:18 demo4-server3 | 2021-09-12T09:20:18.633+0800 INFO nacos_client/nacos_client.go:87 logDir:</tmp/nacos/log/50053>   cacheDir:</tmp/nacos/cache/50053>
09:20:23 demo4-server2 | 2021/09/12 09:20:23 Received: world-1
09:20:25 demo4-server3 | 2021/09/12 09:20:25 Received: world-2
09:20:27 demo4-server1 | 2021/09/12 09:20:27 Received: world-3
09:20:29 demo4-server2 | 2021/09/12 09:20:29 Received: world-4
09:20:31 demo4-server3 | 2021/09/12 09:20:31 Received: world-5
09:20:33 demo4-server1 | 2021/09/12 09:20:33 Received: world-6
09:20:35 demo4-server2 | 2021/09/12 09:20:35 Received: world-7
09:20:37 demo4-server3 | 2021/09/12 09:20:37 Received: world-8
09:20:39 demo4-server1 | 2021/09/12 09:20:39 Received: world-9
09:20:41 demo4-server2 | 2021/09/12 09:20:41 Received: world-10
09:20:43 demo4-server1 | 2021/09/12 09:20:43 Received: world-11
09:20:45 demo4-server2 | 2021/09/12 09:20:45 Received: world-12
09:20:47 demo4-server3 | 2021/09/12 09:20:47 Received: world-13
//这里将权重改为5后
09:20:49 demo4-server1 | 2021/09/12 09:20:49 Received: world-14
09:20:51 demo4-server1 | 2021/09/12 09:20:51 Received: world-15
09:20:53 demo4-server1 | 2021/09/12 09:20:53 Received: world-16
09:20:55 demo4-server1 | 2021/09/12 09:20:55 Received: world-17
09:20:57 demo4-server1 | 2021/09/12 09:20:57 Received: world-18
09:20:59 demo4-server2 | 2021/09/12 09:20:59 Received: world-19
09:21:01 demo4-server3 | 2021/09/12 09:21:01 Received: world-20
09:21:03 demo4-server1 | 2021/09/12 09:21:03 Received: world-21

注意:每次nacos的service instance发生变化后,balancer都会重新build一个新Picker实例,后续会使用新Picker实例在其Connection集合中Pick出一个conn。

4. 小结

在本文中我们了解了gRPC的四种通信模式。我们重点关注了在最常用的simple RPC(unary RPC)模式下gRPC Client侧需要考虑的事情,包括:

  • 如何实现一个helloworld的一对一的通信
  • 如何实现一个自定义的Resolver以实现一个client到一个静态服务实例集合的通信
  • 如何实现一个自定义的Resolver以实现一个client到一个动态服务实例集合的通信
  • 如何自定义客户端Balancer

本文代码仅做示例使用,并未考虑太多异常处理。

本文涉及的所有代码可以从这里下载[15]:https://github.com/bigwhite/experiments/tree/master/grpc-client

5. 参考资料

  • gRPC Name Resolution - https://github.com/grpc/grpc/blob/master/doc/naming.md
  • Load Balancing in gRPC - https://github.com/grpc/grpc/blob/master/doc/load-balancing.md
  • 基于 gRPC的服务发现与负载均衡(基础篇)- https://pandaychen.github.io/2019/07/11/GRPC-SERVICE-DISCOVERY/
  • 比较 gRPC服务和HTTP API - https://docs.microsoft.com/zh-cn/aspnet/core/grpc/comparison

6. 附录

1) json vs. protobuf编解码性能基准测试结果

测试源码位于这里[16]:https://github.com/bigwhite/experiments/tree/master/grpc-client/grpc-vs-httpjson/codec

我们使用了Go标准库json编解码、字节开源的sonic json编解码包[17]以及minio[18]开源的simdjson-go[19]高性能json解析库与protobuf作对比的结果如下:

$go test -bench .
goos: darwin
goarch: amd64
pkg: github.com/bigwhite/codec
cpu: Intel(R) Core(TM) i5-8257U CPU @ 1.40GHz
BenchmarkSimdJsonUnmarshal-8        43304      28177 ns/op   113209 B/op       19 allocs/op
BenchmarkJsonUnmarshal-8           153214       7187 ns/op     1024 B/op        6 allocs/op
BenchmarkJsonMarshal-8             601590       2057 ns/op     2688 B/op        2 allocs/op
BenchmarkSonicJsonUnmarshal-8     1394211        861.1 ns/op     2342 B/op        2 allocs/op
BenchmarkSonicJsonMarshal-8       1592898        765.2 ns/op     2239 B/op        4 allocs/op
BenchmarkProtobufUnmarshal-8      3823441        317.0 ns/op     1208 B/op        3 allocs/op
BenchmarkProtobufMarshal-8        4461583        274.8 ns/op     1152 B/op        1 allocs/op
PASS
ok   github.com/bigwhite/codec 10.901s

benchmark测试结果印证了protobuf的编解码性能要远高于json编解码。但是在benchmark结果中,一个结果让我很意外,那就是号称高性能的simdjson-go的数据难看到离谱。谁知道为什么吗?simd指令没生效?字节开源的sonic的确性能很好,与pb也就2-3倍的差距,没有数量级的差距。

2) gRPC(insecure) vs. json over http

测试源码位于这里[20]:https://github.com/bigwhite/experiments/tree/master/grpc-client/grpc-vs-httpjson/protocol

使用ghz对gRPC实现的server进行压测结果如下:

$ghz --insecure -n 100000 -c 500 --proto publish.proto --call proto.PublishService.Publish -D data.json localhost:10000

Summary:
  Count:    100000
  Total:    1.67 s
  Slowest:    48.49 ms
  Fastest:    0.13 ms
  Average:    6.34 ms
  Requests/sec:    59924.34

Response time histogram:
  0.133  [1]     |
  4.968  [40143] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
  9.803  [47335] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
  14.639 [11306] |∎∎∎∎∎∎∎∎∎∎
  19.474 [510]   |
  24.309 [84]    |
  29.144 [89]    |
  33.980 [29]    |
  38.815 [3]     |
  43.650 [8]     |
  48.485 [492]   |

Latency distribution:
  10 % in 3.07 ms
  25 % in 4.12 ms
  50 % in 5.49 ms
  75 % in 7.94 ms
  90 % in 10.24 ms
  95 % in 11.28 ms
  99 % in 15.52 ms

Status code distribution:
  [OK]   100000 responses

使用hey对使用fasthttp[21]与sonic实现的http server进行压测结果如下:

$hey -n 100000 -c 500  -m POST -D ./data.json http://127.0.0.1:10001/

Summary:
  Total:    2.0012 secs
  Slowest:    0.1028 secs
  Fastest:    0.0001 secs
  Average:    0.0038 secs
  Requests/sec:    49969.9234
 

Response time histogram:
  0.000 [1]     |
  0.010 [96287] |■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■
  0.021 [2639]  |■
  0.031 [261]   |
  0.041 [136]   |
  0.051 [146]   |
  0.062 [128]   |
  0.072 [43]    |
  0.082 [24]    |
  0.093 [10]    |
  0.103 [4]     |


Latency distribution:
  10% in 0.0013 secs
  25% in 0.0020 secs
  50% in 0.0031 secs
  75% in 0.0040 secs
  90% in 0.0062 secs
  95% in 0.0089 secs
  99% in 0.0179 secs

Details (average, fastest, slowest):
  DNS+dialup:    0.0000 secs, 0.0001 secs, 0.1028 secs
  DNS-lookup:    0.0000 secs, 0.0000 secs, 0.0000 secs
  req write:    0.0000 secs, 0.0000 secs, 0.0202 secs
  resp wait:    0.0031 secs, 0.0000 secs, 0.0972 secs
  resp read:    0.0005 secs, 0.0000 secs, 0.0575 secs

Status code distribution:
  [200]    99679 responses

我们看到:gRPC的性能(Requests/sec: 59924.34)要比http api性能(Requests/sec: 49969.9234)高出20%。

3) nacos docker安装

单机容器版nacos安装步骤如下:

$git clone https://github.com/nacos-group/nacos-docker.git
$cd nacos-docker
$docker-compose -f example/standalone-derby.yaml up

nacos相关容器启动成功后,可以打开浏览器访问http://localhost:8848/nacos,打开nacos仪表盘登录页面,输入nacos/nacos即可进入nacos web操作界面。