灰度发布（又名金丝雀发布）是指在黑与白之间，能够平滑过渡的一种发布方式。在其上可以进行A/B testing，即让一部分用户继续用产品特性A，一部分用户开始用产品特性B，如果用户对B没有什么反对意见，那么逐步扩大范围，把所有用户都迁移到B上面来。灰度发布可以保证整体系统的稳定，在初始灰度的时候就可以发现、调整问题，以保证其影响度。
本文以springcloud gateway + nacos来演示如何实现灰度发布，如果对springcloud gateway和nacos还不熟悉的朋友，可以先阅读如下文章，然后再阅读本文。
springcloud gateway官方介绍
nacos官方介绍
实现的整体思路：
- 编写带权重的灰度路由
- 编写自定义filter
- nacos服务配置需要灰度发布的服务的元数据信息以及权重
- 灰度路由从nacos服务拉取元数据信息以及权重，然后根据权重算法，返回符合要求的服务实例给自定义的filter
- 网关配置文件配置需要灰度路由的服务（因为本文代码没有网关实现动态路由，不然灰度路由可以配置在配置中心，从配置中心拉取）
- filter通过责任链模式，把服务实例透传给其他filter比如NettyRoutingFilter
下边进入实战
正文1、所使用的开发版本
1.8
Hoxton.SR3
2.2.5.RELEASE
2.2.1.RELEASE
2、pom.xml引入
org.springframework.cloud
spring-cloud-starter-gateway
org.springframework.boot
spring-boot-starter-webflux
com.alibaba.cloud
spring-cloud-starter-alibaba-nacos-discovery
org.springframework.cloud
spring-cloud-starter-netflix-ribbon
org.springframework.cloud
spring-cloud-loadbalancer
org.apache.commons
commons-lang3
ps：nacos的jar注意排除ribbon依赖，不然loadbalancer无法生效
3、编写权重路由
public class GrayLoadBalancer implements ReactorServiceInstanceLoadBalancer {
private static final Log log = LogFactory.getLog(GrayLoadBalancer.class);
private ObjectProvider serviceInstanceListSupplierProvider;
private String serviceId;
public GrayLoadBalancer(ObjectProvider serviceInstanceListSupplierProvider, String serviceId) {
this.serviceId = serviceId;
this.serviceInstanceListSupplierProvider = serviceInstanceListSupplierProvider;
}
@Override
public Mono> choose(Request request) {
HttpHeaders headers = (HttpHeaders) request.getContext();
if (this.serviceInstanceListSupplierProvider != null) {
ServiceInstanceListSupplier supplier = (ServiceInstanceListSupplier)this.serviceInstanceListSupplierProvider.getIfAvailable(NoopServiceInstanceListSupplier::new);
return ((Flux)supplier.get()).next().map(list->getInstanceResponse((List)list,headers));
}
return null;
}
private Response getInstanceResponse(List instances,HttpHeaders headers) {
if (instances.isEmpty()) {
return getServiceInstanceEmptyResponse();
} else {
return getServiceInstanceResponseWithWeight(instances);
}
}
private Response getServiceInstanceResponseByVersion(List instances, HttpHeaders headers) {
String versionNo = headers.getFirst("version");
System.out.println(versionNo);
Map versionMap = new HashMap<>();
versionMap.put("version",versionNo);
final Set> attributes =
Collections.unmodifiableSet(versionMap.entrySet());
ServiceInstance serviceInstance = null;
for (ServiceInstance instance : instances) {
Map metadata = instance.getmetadata();
if(metadata.entrySet().containsAll(attributes)){
serviceInstance = instance;
break;
}
}
if(ObjectUtils.isEmpty(serviceInstance)){
return getServiceInstanceEmptyResponse();
}
return new DefaultResponse(serviceInstance);
}
private Response getServiceInstanceResponseWithWeight(List instances) {
Map weightMap = new HashMap<>();
for (ServiceInstance instance : instances) {
Map metadata = instance.getmetadata();
System.out.println(metadata.get("version")+"-->weight:"+metadata.get("weight"));
if(metadata.containsKey("weight")){
weightMap.put(instance,Integer.valueOf(metadata.get("weight")));
}
}
Weightmeta weightmeta = WeightRandomUtils.buildWeightmeta(weightMap);
if(ObjectUtils.isEmpty(weightmeta)){
return getServiceInstanceEmptyResponse();
}
ServiceInstance serviceInstance = weightmeta.random();
if(ObjectUtils.isEmpty(serviceInstance)){
return getServiceInstanceEmptyResponse();
}
System.out.println(serviceInstance.getmetadata().get("version"));
return new DefaultResponse(serviceInstance);
}
private Response getServiceInstanceEmptyResponse() {
log.warn("No servers available for service: " + this.serviceId);
return new EmptyResponse();
}
4、自定义filter
public class GrayReactiveLoadBalancerClientFilter implements GlobalFilter, Ordered {
private static final Log log = LogFactory.getLog(ReactiveLoadBalancerClientFilter.class);
private static final int LOAD_BALANCER_CLIENT_FILTER_ORDER = 10150;
private final LoadBalancerClientFactory clientFactory;
private LoadBalancerProperties properties;
public GrayReactiveLoadBalancerClientFilter(LoadBalancerClientFactory clientFactory, LoadBalancerProperties properties) {
this.clientFactory = clientFactory;
this.properties = properties;
}
@Override
public int getOrder() {
return LOAD_BALANCER_CLIENT_FILTER_ORDER;
}
@Override
public Mono filter(ServerWebExchange exchange, GatewayFilterChain chain) {
URI url = (URI)exchange.getAttribute(ServerWebExchangeUtils.GATEWAY_REQUEST_URL_ATTR);
String schemePrefix = (String)exchange.getAttribute(ServerWebExchangeUtils.GATEWAY_SCHEME_PREFIX_ATTR);
if (url != null && ("grayLb".equals(url.getScheme()) || "grayLb".equals(schemePrefix))) {
ServerWebExchangeUtils.addOriginalRequestUrl(exchange, url);
if (log.isTraceEnabled()) {
log.trace(ReactiveLoadBalancerClientFilter.class.getSimpleName() + " url before: " + url);
}
return this.choose(exchange).doOnNext((response) -> {
if (!response.hasServer()) {
throw NotFoundException.create(this.properties.isUse404(), "Unable to find instance for " + url.getHost());
} else {
URI uri = exchange.getRequest().getURI();
String overrideScheme = null;
if (schemePrefix != null) {
overrideScheme = url.getScheme();
}
DelegatingServiceInstance serviceInstance = new DelegatingServiceInstance((ServiceInstance)response.getServer(), overrideScheme);
URI requestUrl = this.reconstructURI(serviceInstance, uri);
if (log.isTraceEnabled()) {
log.trace("LoadBalancerClientFilter url chosen: " + requestUrl);
}
exchange.getAttributes().put(ServerWebExchangeUtils.GATEWAY_REQUEST_URL_ATTR, requestUrl);
}
}).then(chain.filter(exchange));
} else {
return chain.filter(exchange);
}
}
protected URI reconstructURI(ServiceInstance serviceInstance, URI original) {
return LoadBalancerUriTools.reconstructURI(serviceInstance, original);
}
private Mono> choose(ServerWebExchange exchange) {
URI uri = (URI)exchange.getAttribute(ServerWebExchangeUtils.GATEWAY_REQUEST_URL_ATTR);
GrayLoadBalancer loadBalancer = new GrayLoadBalancer(clientFactory.getLazyProvider(uri.getHost(), ServiceInstanceListSupplier.class), uri.getHost());
if (loadBalancer == null) {
throw new NotFoundException("No loadbalancer available for " + uri.getHost());
} else {
return loadBalancer.choose(this.createRequest(exchange));
}
}
private Request createRequest(ServerWebExchange exchange) {
HttpHeaders headers = exchange.getRequest().getHeaders();
Request request = new DefaultRequest<>(headers);
return request;
}
}
5、配置自定义filter给spring管理
@Configuration
public class GrayGatewayReactiveLoadBalancerClientAutoConfiguration {
public GrayGatewayReactiveLoadBalancerClientAutoConfiguration() {
}
@Bean
@ConditionalOnMissingBean({GrayReactiveLoadBalancerClientFilter.class})
public GrayReactiveLoadBalancerClientFilter grayReactiveLoadBalancerClientFilter(LoadBalancerClientFactory clientFactory, LoadBalancerProperties properties) {
return new GrayReactiveLoadBalancerClientFilter(clientFactory, properties);
}
}
6、编写网关application.yml配置
server:
port: 9082
# 配置输出日志
logging:
level:
org.springframework.cloud.gateway: TRACE
org.springframework.http.server.reactive: DEBUG
org.springframework.web.reactive: DEBUG
reactor.ipc.netty: DEBUG
#开启端点
management:
endpoints:
web:
exposure:
include: '*'
spring:
application:
name: gateway-reactor-gray
cloud:
nacos:
discovery:
server-addr: localhost:8848
gateway:
discovery:
locator:
enabled: true
lower-case-service-id: true
routes:
- id: hello-consumer
uri: grayLb://hello-consumer
predicates:
- Path=/hello/**
uri中的grayLb配置，代表该服务需要进行灰度发布
7、在注册中心nacos配置灰度发布的服务版本以及权重值
weight代表权重，version代表版本
总结上述就是实现灰度发布的过程，实现灰度发布的方法有很多种，文章中只是提供一种思路。虽然springcloud官方推荐使用loadbalancer来代替ribbon。因为ribbon是阻塞的，但从官方的loadbalancer的负载均衡算法来看，目前loadbalancer默认只支持轮询算法，要其他算法得自己扩展实现，而ribbon默认支持7种算法，用默认的算法基本上就可以满足我们的需求了。其次ribbon支持懒加载处理，超时以及重试与断路器hystrix集成等配置，loadbalancer目前就支持重试。所以如果正式环境要自己实现灰度发布，可以考虑对ribbon进行扩展。本文的实现只是作为一种扩展补充，毕竟springcloud推荐loadbalancer，索性就写个demo实现下。
最后灰度发布的实现，业内也有开源的实现–Discovery，感兴趣的朋友可以通过如下链接进行查看
demo链接https://github.com/Nepxion/Discovery
https://github.com/lyb-geek/gateway
