# grpc外部负载均衡器测试
(金庆的专栏 2020.4)
grpc 对每个请求进行负载均衡。负载均衡的方式有：
* 代理模式
* 客户端实现
* 外部负载均衡
参考：gRPC LB https://blog.csdn.net/xiaojia1100/article/details/78842295
gRPC 中负载均衡的主要机制是外部负载均衡。
gRPC 定义了外部负载均衡服务的接口：https://github.com/grpc/grpc/tree/master/src/proto/grpc/lb/v1
* load_balancer.proto 客户端向 lb 服查询后端列表
* load_reporter.proto lb 服向后端服查询负载
https://github.com/bsm/grpclb 实现了一个 grpc 的外部负载均衡服。
因为其实现早于负载均衡服的接口规范，所以接口定义与 grpc 规范不同。
见 issue#26: https://github.com/bsm/grpclb/issues/26#issuecomment-613873655
grpclb 目前仅支持 consul 服务发现。
标准的 grpclb 实现目前好像只有 https://github.com/joa/jawlb。
jawlb 通过 Kubernetes API 来发现服务。
以下测试 grpc 客户端从 jawlb 服查询服务器列表，然后请求服务。
首先在本机开了多个 greeter 服实例，端口不同。
然后更改 greeter 客户端，不要直接连 greeter 服地址，而是配一个 jawlb 服地址。
同时更改 jawlb, 删除服务发现，改为固定输出本机服务列表，定时切换。
greeter 是指 grpc-go 中的例子：grpc-go\examples\helloworld\greeter
## greeter 服更改
添加参数指定服务端口。
```
package main
import (
"fmt"
"log"
"net"
"github.com/spf13/pflag"
"github.com/spf13/viper"
"golang.org/x/net/context"
"google.golang.org/grpc"
pb "google.golang.org/grpc/examples/helloworld/helloworld"
)
// GreeterServer is used to implement helloworld.GreeterServer.
type GreeterServer struct {
}
// SayHello implements helloworld.GreeterServer
func (s *GreeterServer) SayHello(ctx context.Context, in *pb.HelloRequest) (*pb.HelloReply, error) {
msg := fmt.Sprintf("Hello %s from server-%d", in.Name, viper.GetInt("port"))
return &pb.HelloReply{Message: msg}, nil
}
func main() {
pflag.Int("port", 8000, "server bind port")
pflag.Parse()
viper.BindPFlags(pflag.CommandLine)
port := viper.GetInt("port")
addr := fmt.Sprintf(":%d", port)
lis, err := net.Listen("tcp", addr)
if err != nil {
log.Fatalf("failed to listen: %v", err)
}
s := grpc.NewServer()
pb.RegisterGreeterServer(s, &GreeterServer{})
s.Serve(lis)
}
```
## greeter 客户端更改
```
package main
import (
"context"
"log"
"os"
"time"
"github.com/sirupsen/logrus"
"google.golang.org/grpc"
_ "google.golang.org/grpc/balancer/grpclb"
pb "google.golang.org/grpc/examples/helloworld/helloworld"
"google.golang.org/grpc/grpclog"
"google.golang.org/grpc/resolver"
"google.golang.org/grpc/resolver/manual"
)
const (
defaultName = "world"
)
func init() {
grpclog.SetLogger(logrus.New())
}
func main() {
rb := manual.NewBuilderWithScheme("whatever")
rb.InitialState(resolver.State{Addresses: []resolver.Address{
{Addr: "127.0.0.1:8888", Type: resolver.GRPCLB},
}})
conn, err := grpc.Dial("whatever:///this-gets-overwritten", grpc.WithInsecure(), grpc.WithBlock(),
grpc.WithResolvers(rb))
if err != nil {
log.Fatalf("did not connect: %v", err)
}
defer conn.Close()
c := pb.NewGreeterClient(conn)
name := defaultName
if len(os.Args) > 1 {
name = os.Args[1]
}
for {
ctx, cancel := context.WithTimeout(context.Background(), 3*time.Second)
r, err := c.SayHello(ctx, &pb.HelloRequest{Name: name})
cancel()
if err != nil {
log.Fatalf("could not greet: %v", err)
time.Sleep(time.Second)
continue
}
log.Printf("Greeting: %s", r.GetMessage())
time.Sleep(time.Second)
}
}
```
有以下更改：
* import _ "google.golang.org/grpc/balancer/grpclb"
* grpc.Dial("whatever:///this-gets-overwritten", grpc.WithResolvers(rb))
+ 采用一个自定义解析器，用来获取 jawlb 地址
+ Scheme("whatever") 可以任意，用作解析器名字
+ 目标 this-gets-overwritten 可以任意，因为 jawlb 忽略了该名字
+ 127.0.0.1:8888 是 jawlb 地址
* 改为每秒请求一次
正常的 grpclb 是在 DNS 中设置 SRV 记录，
此处测试避免设置 DNS, 采用了一个自定义解析器，
代码上多了几行。
用 DNS 设置的好处是, 可以直接解析为后端 IP, 也可以添加 grpclb, 代码上如同直接连接后端：
```
conn, err := grpc.Dial("dns:///myservice.domain.com", grpc.WithInsecure())
```
## jawlb 更改
### main.go
删除所有配置，改为固定本机 8888 端口监听。
* 删除 `envconfig.MustProcess("JAWLB", &cfg)`
* listen() 改为
```
func listen() (conn net.Listener, err error) {
conn, err = net.Listen("tcp", ":8888")
return
}
```
### watch.go
```
package main
import (
"context"
"fmt"
"net"
"time"
)
func watchService(ctx context.Context) (_ <-chan ServerList, err error) {
ch := make(chan ServerList)
go func() {
ticker := time.NewTicker(10 * time.Second)
i := 0
for {
select {
case <-ctx.Done():
ticker.Stop()
close(ch)
return
case <-ticker.C:
i += 1
fmt.Printf("i = %d\n", i)
ports := []int32{8010, 8020}
var servers []Server
for _, port := range ports {
servers = append(servers, Server{IP: net.ParseIP("127.0.0.1"), Port: port + int32(i%2)})
}
ch <- servers
} // select
} // for
}()
return ch, nil
}
```
删除所有服务发现代码，改为每10秒切换端口：8010,8020 <-> 8011,8021
## 运行
### jawlb
```
λ jawlb.exe
2020/04/16 15:35:17 waiting for TERM
i = 1
2020/04/16 15:35:27 endpoints:
2020/04/16 15:35:27 127.0.0.1:8011
2020/04/16 15:35:27 127.0.0.1:8021
i = 2
2020/04/16 15:35:37 endpoints:
2020/04/16 15:35:37 127.0.0.1:8010
2020/04/16 15:35:37 127.0.0.1:8020
```
### server
运行 4 个实例：
```
server --port 8010
server --port 8020
server --port 8011
server --port 8021
```
### client
```
λ client
INFO[0002] lbBalancer: handle SubConn state change: 0xc00008a590, CONNECTING
INFO[0002] Channel Connectivity change to CONNECTING
INFO[0002] lbBalancer: handle SubConn state change: 0xc00008a5f0, CONNECTING
INFO[0002] Subchannel picks a new address "127.0.0.1:8021" to connect
INFO[0002] Subchannel Connectivity change to READY
INFO[0002] lbBalancer: handle SubConn state change: 0xc00008a590, READY
INFO[0002] Channel Connectivity change to READY
INFO[0002] Subchannel Connectivity change to READY
INFO[0002] lbBalancer: handle SubConn state change: 0xc00008a5f0, READY
2020/04/16 15:37:47 Greeting: Hello world from server-8021
2020/04/16 15:37:48 Greeting: Hello world from server-8011
2020/04/16 15:37:49 Greeting: Hello world from server-8021
2020/04/16 15:37:50 Greeting: Hello world from server-8011
2020/04/16 15:37:51 Greeting: Hello world from server-8021
2020/04/16 15:37:52 Greeting: Hello world from server-8011
2020/04/16 15:37:53 Greeting: Hello world from server-8021
2020/04/16 15:37:54 Greeting: Hello world from server-8011
2020/04/16 15:37:55 Greeting: Hello world from server-8021
2020/04/16 15:37:56 Greeting: Hello world from server-8011
INFO[0012] lbBalancer: processing server list: servers:<ip_address:"\000\000\000\000\000\000\000\000\000\000\377\377\177\000\000\001" port:8020 > servers:<ip_address:"\000\000\000\000\000\000\000\000\000\000\377\377\177\000\000\001" port:8010 >
INFO[0012] lbBalancer: server list entry[0]: ipStr:|127.0.0.1|, port:|8020|, load balancer token:||
INFO[0012] lbBalancer: server list entry[1]: ipStr:|127.0.0.1|, port:|8010|, load balancer token:||
2020/04/16 15:37:57 Greeting: Hello world from server-8020
2020/04/16 15:37:58 Greeting: Hello world from server-8010
2020/04/16 15:37:59 Greeting: Hello world from server-8020
2020/04/16 15:38:00 Greeting: Hello world from server-8010
2020/04/16 15:38:01 Greeting: Hello world from server-8020
2020/04/16 15:38:02 Greeting: Hello world from server-8010
2020/04/16 15:38:03 Greeting: Hello world from server-8020
2020/04/16 15:38:04 Greeting: Hello world from server-8010
2020/04/16 15:38:05 Greeting: Hello world from server-8020
2020/04/16 15:38:06 Greeting: Hello world from server-8010
INFO[0022] lbBalancer: processing server list: servers:<ip_address:"\000\000\000\000\000\000\000\000\000\000\377\377\177\000\000\001" port:8021 > servers:<ip_address:"\000\000\000\000\000\000\000\000\000\000\377\377\177\000\000\001" port:8011 >
INFO[0022] lbBalancer: server list entry[0]: ipStr:|127.0.0.1|, port:|8021|, load balancer token:||
INFO[0022] lbBalancer: server list entry[1]: ipStr:|127.0.0.1|, port:|8011|, load balancer token:||
2020/04/16 15:38:07 Greeting: Hello world from server-8011
2020/04/16 15:38:08 Greeting: Hello world from server-8021
2020/04/16 15:38:09 Greeting: Hello world from server-8011
```
## 结论
客户端应用一个自定义 resolver 解析 "whatever:///this-gets-overwritten"，
获取到 `{Addr: "127.0.0.1:8888", Type: resolver.GRPCLB}`,
知道这是一个 grpclb，于是按 load_balancer.proto 的定义查询 jawlb 来获取后端地址列表。
jawlb 每 10s 更新一次服务器列表，每次输出多个地址。客户端在多个地址间轮换请求。
## 其他测试
* 不开 jawlb，客户端将无法成功请求，直到 jawlb 开启才成功
* 中途关闭 jawlb, 请求仍会成功，但是保持为最后的服务器列表
+ 同时会不断尝试重连 jawlb, 但是重连成功后没有切换服务，应该是个错误
* Dial() 不加 grpc.WithBlock() 参数, 报错：all SubConns are in TransientFailure
```
λ client
INFO[0000] parsed scheme: "whatever"
INFO[0000] ccResolverWrapper: sending update to cc: {[{127.0.0.1:8888 <nil> 1 <nil>}] <nil> <nil>}
INFO[0000] ClientConn switching balancer to "grpclb"
INFO[0000] Channel switches to new LB policy "grpclb"
INFO[0000] lbBalancer: UpdateClientConnState: {ResolverState:{Addresses:[{Addr:127.0.0.1:8888 ServerName: Attributes:<nil> Type:1 Metadata:<nil>}] ServiceConfig:<nil> Attributes:<nil>} BalancerConfig:<nil>}
INFO[0000] parsed scheme: "grpclb-internal"
INFO[0000] ccResolverWrapper: sending update to cc: {[{127.0.0.1:8888 <nil> 0 <nil>}] <nil> <nil>}
INFO[0000] ClientConn switching balancer to "pick_first"
INFO[0000] Channel switches to new LB policy "pick_first"
INFO[0000] Subchannel Connectivity change to CONNECTING
INFO[0000] blockingPicker: the picked transport is not ready, loop back to repick
INFO[0000] pickfirstBalancer: HandleSubConnStateChange: 0xc00003fb10, {CONNECTING <nil>}
INFO[0000] Channel Connectivity change to CONNECTING
INFO[0000] Subchannel picks a new address "127.0.0.1:8888" to connect
INFO[0000] CPU time info is unavailable on non-linux or appengine environment.
INFO[0000] Subchannel Connectivity change to READY
INFO[0000] pickfirstBalancer: HandleSubConnStateChange: 0xc00003fb10, {READY <nil>}
INFO[0000] Channel Connectivity change to READY
INFO[0000] lbBalancer: processing server list: servers:<ip_address:"\000\000\000\000\000\000\000\000\000\000\377\377\177\000\000\001" port:8010 > servers:<ip_address:"\000\000\000\000\000\000\000\000\000\000\377\377\177\000\000\001" port:8020 >
INFO[0000] lbBalancer: server list entry[0]: ipStr:|127.0.0.1|, port:|8010|, load balancer token:||
INFO[0000] lbBalancer: server list entry[1]: ipStr:|127.0.0.1|, port:|8020|, load balancer token:||
INFO[0000] Subchannel Connectivity change to CONNECTING
INFO[0000] Subchannel Connectivity change to CONNECTING
INFO[0000] Channel Connectivity change to TRANSIENT_FAILURE
INFO[0000] lbBalancer: handle SubConn state change: 0xc00008a220, CONNECTING
INFO[0000] Channel Connectivity change to CONNECTING
INFO[0000] lbBalancer: handle SubConn state change: 0xc00008a280, CONNECTING
2020/04/16 16:40:06 could not greet: rpc error: code = Unavailable desc = all SubConns are in TransientFailure
```