前言
grpc默认支持两种负载均衡算法pick_first 和 round_robin
轮询法round_robin不能满足因服务器配置不同而承担不同负载量,这篇文章将介绍如何实现自定义负载均衡策略--加权随机。
加权随机法可以根据服务器的处理能力而分配不同的权重,从而实现处理能力高的服务器可承担更多的请求,处理能力低的服务器少承担请求。
上一篇我们实现resolverBuilder接口,用来解析服务器地址,而负载均衡算法就是在这个接口基础上筛选出一个地址,这个过程是在客户端发送请求的时候进行。
以下示例基于grpc v1.26.0版本,更高的版本不兼容etcd,不方便测试,而接口名或者参数也会不同,不过原理是相似的
自定义负载均衡策略
使用自定义的负载均衡策略主要实现V2PickerBuilder
和V2Picker
这两个接口
type V2PickerBuilder interface {
Build(info PickerBuildInfo) balancer.V2Picker
}
Build
方法:返回一个V2选择器,将用于gRPC选择子连接。
type V2Picker interface {
Pick(info PickInfo) (PickResult, error)
}
Pick
方法:子连接选择,具体的算法在这个方法内实现
完整代码:
package rpc
import (
"google.golang.org/grpc/balancer"
"google.golang.org/grpc/balancer/base"
"google.golang.org/grpc/grpclog"
"math/rand"
"sync"
)
const WEIGHT_LB_NAME = "weight"
func init() {
balancer.Register(
base.NewBalancerBuilderV2(WEIGHT_LB_NAME,&LocalBuilder{},base.Config{HealthCheck: false}))
}
type LocalBuilder struct {
}
func (*LocalBuilder) Build(info base.PickerBuildInfo) balancer.V2Picker {
if len(info.ReadySCs) == 0 {
return base.NewErrPickerV2(balancer.ErrNoSubConnAvailable)
}
var scs []balancer.SubConn
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 &localPick{
subConn: scs,
}
}
type localPick struct {
subConn []balancer.SubConn
mu sync.Mutex
}
func (l *localPick)Pick(info balancer.PickInfo) (r balancer.PickResult, err error){
l.mu.Lock()
index := rand.Intn(len(l.subConn))
r.SubConn = l.subConn[index]
l.mu.Lock()
return r,nil
}
结合示例再看一下Build(info PickerBuildInfo) balancer.V2Picker
这个方法的作用:
这个函数的返回值就是我们还要实现的第二个接口,所以这里主要看一下参数
type PickerBuildInfo struct {
// ReadySCs is a map from all ready SubConns to the Addresses used to
// create them.
ReadySCs map[balancer.SubConn]
}
type SubConnInfo struct {
Address resolver.Address // the address used to create this SubConn
}
ReadySCs
:是所有可用的子连接
balancer.SubConn
:是一个子连接的结构,这里我们不用关心这个值,在pick里面直接返回就可以了。
SubConnInfo
:里面是一个Address的结构
type Address struct {
Addr string
ServerName string
// Attributes contains arbitrary data about this address intended for
// consumption by the load balancing policy.
Attributes *attributes.Attributes
Type AddressType
Metadata interface{}
}
...
...
type Attributes struct {
m map[interface{}]interface{}
}
这个Address就是上一篇中resolverBuilder中设置的值,所以这两个功能是有联系的,Attributes的m是一个map,刚刚好保存我们需要的权重weight
结下来就是Pick方法
Pick(info balancer.PickInfo) (r balancer.PickResult, err error)
type PickInfo struct {
// FullMethodName is the method name that NewClientStream() is called
// with. The canonical format is /service/Method.
FullMethodName string //eg : /User/GetInfo
// Ctx is the RPC's context, and may contain relevant RPC-level information
// like the outgoing header metadata.
Ctx context.Context
}
type PickResult struct {
SubConn SubConn
Done func(DoneInfo)
}
FullMethodName
:对应XXXpb.go中生成的FullMethod: "/Memo/GetMemo",
Ctx
: 如注释所说,包含metadata
PickResult
: 返回的子连接
Done
:rpc完成之后调用Done
使用新建的策略
func main() {
grpc.UseCompressor(gzip.Name)
conn, err := grpc.Dial(
fmt.Sprintf("%s:///%s", "game", baseService),
//grpc.WithDefaultServiceConfig(fmt.Sprintf(`{"LoadBalancingPolicy": "%s"}`, roundrobin.Name)),
grpc.WithDefaultServiceConfig(fmt.Sprintf(`{"LoadBalancingPolicy": "%s"}`, rpc.WEIGHT_LB_NAME)),
grpc.WithInsecure(),
)
...
}