Skip to main content

Integration Test Analysis

· 7 min read
Kunshuai Zhu
Apache ShenYu Committer

This article will provide an in-depth analysis of Apache ShenYu's integration tests.

What are integration tests?#

Integration testing is also called E2E (End To End) testing in some projects. It is mainly used to test whether each module can meet expectations after being assembled into a system.

Apache ShenYu puts integration tests in continuous integration, using GitHub Actions to trigger each time a Pull Request or Merge is submitted to the main branch. This can greatly reduce the maintenance cost of the project and improve the stability of Apache ShenYu.

How to automate integration testing?#

In Apache ShenYu, the main steps of integration testing are embodied in the script of the GitHub Action workflow, as shown below, which is located at [~/.github/workflows]( /master/.github/workflows) directory.

name: iton:  pull_request:  push:    branches:      - masterjobs:  build:    strategy:      matrix:        case:          - shenyu-integrated-test-alibaba-dubbo          ...    runs-on: ubuntu-latest    steps:      - uses: actions/checkout@v2        with:          submodules: true      ...

Next, I will start from this yaml file and take you to analyze the entire process of automated integration testing.

Triggering the workflow#

Since we specified pull_request and push.branch: master in on, this workflow will be triggered when we submit pull_request or merge branch to master (push).

For more usage of GitHub Action, you can refer to the documentation of GitHub Action, which will not be introduced in detail here.

Initialize the environment#

  • pull code
- uses: actions/checkout@v2  with:        submodules: true
  • set skip flag
- name: Set Skip Env Var      uses: ./.github/actions/skip-ci

When something unrelated to functionality occurs (such as changing documentation), integration tests are skipped to save resources.

  • Cache maven repos, install Java
- name: Cache Maven Repos...- uses: actions/setup-java@v1

Build the entire project while building the docker image#

./mvnw -B clean install -Prelease,docker -Dmaven.javadoc.skip=true -Dmaven.test.skip=true

In the above command, -P is followed by release,docker, which means that the relevant profile configuration in the pom file will be activated.

The two profiles, release and docker, currently only exist in several submodules under shenyu-dist. The following will take the shenyu-dist-admin module as an example to introduce profiles as release and docker The specific content of the configuration. Also, integration tests only use the shenyu-admin image built in this step.

  • First is release

    <profile>    <id>release</id>    <activation>        <activeByDefault>false</activeByDefault>    </activation>    <build>        <finalName>apache-shenyu-incubating-${project.version}</finalName>        <plugins>            <plugin>                <groupId>org.apache.maven.plugins</groupId>                <artifactId>maven-assembly-plugin</artifactId>                <executions>                    <execution>                        <id>admin-bin</id>                        <phase>package</phase>                        <goals>                            <goal>single</goal>                        </goals>                    </execution>                </executions>                <configuration>                    <descriptors>                        <descriptor>${project.basedir}/src/main/assembly/binary.xml</descriptor>                    </descriptors>                    <tarLongFileMode>posix</tarLongFileMode>                </configuration>            </plugin>        </plugins>    </build></profile>

    When -P is followed by release, the above maven-assembly-plugin plugin is activated. In executions, the execution timing of the plugin is bound to the maven life cycle package, which means that it will be triggered when we execute mvn install.

    The binary.xml we wrote is specified in the configuration, and the maven-assembly-plugin plugin will copy the required files and package them according to this file. You can click the link to view the file: [shenyu-dist/shenyu-admin-dist/src/main/assembly/binary.xml]( dist/shenyu-admin-dist/src/main/assembly/binary.xml)

    According to this file, the plugin will "copy" the packaged jar packages, configuration files, startup scripts, etc. under other modules, and finally make them into a compressed package in tar.gz format.

  • then docker

    <profile>    <id>docker</id>    <activation>        <activeByDefault>false</activeByDefault>    </activation>    <build>        <plugins>            <plugin>                <groupId>com.spotify</groupId>                <artifactId>dockerfile-maven-plugin</artifactId>                <version>${dockerfile-maven-plugin.version}</version>                <executions>                    <execution>                        <id>tag-latest</id>                        <goals>                            <goal>build</goal>                        </goals>                        <configuration>                            <tag>latest</tag>                        </configuration>                    </execution>                    <execution>                        <id>tag-version</id>                        <goals>                            <goal>build</goal>                        </goals>                        <configuration>                            <tag>${project.version}</tag>                        </configuration>                    </execution>                </executions>                <configuration>                    <repository>apache/shenyu-admin</repository>                    <buildArgs>                        <APP_NAME>apache-shenyu-incubating-${project.version}-admin-bin</APP_NAME>                    </buildArgs>                </configuration>            </plugin>        </plugins>    </build></profile>

    Similar to the release above, here is the activation of the dockerfile-maven-plugin plugin. When mvn install -Pdocker, the plugin will use the dockerfile we wrote to build the docker image.

It should be noted that the dockerfile-maven-plugin currently has limited support for aarch64 architecture devices, and the following error will occur when running the plugin on aarch64 architecture machines. And when I wrote this article, it has not been maintained for a long time, which means that the problem of aarch64 architecture devices using this plugin will not be solved in the short term.

[ERROR] Failed to execute goal com.spotify:dockerfile-maven-plugin:1.4.6:build (tag-latest) on project shenyu-admin-dist: Could not build image: java.util.concurrent.ExecutionException: java.lang.UnsatisfiedLinkError: could not load FFI provider jnr.ffi.provider.jffi.Provider: ExceptionInInitializerError: Can't overwrite cause with java.lang.UnsatisfiedLinkError: java.lang.UnsatisfiedLinkError: /private/var/folders/w2/j27f16yj7cvf_1cxbgqb89vh0000gn/T/jffi4972193792308935312.dylib: dlopen(/private/var/folders/w2/j27f16yj7cvf_1cxbgqb89vh0000gn/T/jffi4972193792308935312.dylib, 1): no suitable image found.  Did find:[ERROR]         /private/var/folders/w2/j27f16yj7cvf_1cxbgqb89vh0000gn/T/jffi4972193792308935312.dylib: no matching architecture in universal wrapper[ERROR]         /private/var/folders/w2/j27f16yj7cvf_1cxbgqb89vh0000gn/T/jffi4972193792308935312.dylib: no matching architecture in universal wrapper...

Here is a temporary solution:

  1. Open a new shell, enter the following command, and use socat to route the unix socket to the tcp port

    socat TCP-LISTEN:2375,range=,reuseaddr,fork UNIX-CLIENT:/var/run/docker.sock
  2. Set environment variables

    export DOCKER_HOST=tcp://

Build the examples module#

- name: Build examples  if: env.SKIP_CI != 'true'  run: ./mvnw -B clean install -Pexample -Dmaven.javadoc.skip=true -Dmaven.test.skip=true -f ./shenyu-examples/pom.xml

Considering the need for release, the current pom file in the project root directory does not contain the example submodule, so the examples module is additionally built in the above step.

Similar to the above, this line of command will also use the maven plugin to build an image for our subsequent docker orchestration.

Build custom gateways#

- name: Build integrated tests  if: env.SKIP_CI != 'true'  run: ./mvnw -B clean install -Pit -DskipTests -f ./shenyu-integrated-test/pom.xml

In order to subdivide the integration tests of different functions of Apache ShenYu, we will build a gateway customized by the integration test module in this step. The so-called "customization" is to introduce the minimum required dependencies in the pom file, and then replace the default shenyu-bootstrap. Similar to the above two steps, this step will also build the docker image.

It is worth noting that the way of packaging and building here is slightly different from that of the shenyu-dist module, which you can find by comparing the pom file.

Run docker compose#

- name: Start docker compose  if: env.SKIP_CI != 'true'  run: docker-compose -f ./shenyu-integrated-test/${{ }}/docker-compose.yml up -d

In this step, docker will be arranged according to the different docker-compose.yml files written under the integration test module.

version: "3.9"services:  shenyu-zk:    container_name: shenyu-zk    image: zookeeper:3.5    ...  shenyu-redis:    image: redis:6.0-alpine    container_name: shenyu-redis    ...
  shenyu-examples-http:    deploy:      resources:        limits:          memory: 2048M    container_name: shenyu-examples-http    image: shenyu-examples-http:latest    ...
  shenyu-admin:    image: apache/shenyu-admin:latest    container_name: shenyu-admin    ...
  shenyu-integrated-test-http:    container_name: shenyu-integrated-test-http    image: apache/shenyu-integrated-test-http:latest    ...    depends_on:      shenyu-admin:        condition: service_healthy    healthcheck:      test: [ "CMD", "wget", "http://shenyu-integrated-test-http:9195/actuator/health" ]      timeout: 2s      retries: 30
networks:  shenyu:    name: shenyu

For example, the docker-compose.yml under the shenyu-integrated-test-http module starts zookeeper, redis, example, admin, gateway and other services in sequence. Among them, the mirrors of example, admin, and gateway are built by us before.

Among them, docker-compose uses depends_on to determine the topological relationship between services, and most services have corresponding health checks, and the next service will not be started until the health check passes.

Run the health check and wait for docker-compose to start#

- name: Wait for docker compose start up completely  if: env.SKIP_CI != 'true'  run: bash ./shenyu-integrated-test/${{ }}/script/

In this step, the host will run the script, and then use the curl command to access the health status interface /actuator/health of each service list (in the services.list file), until the service status is normal. will continue.

run tests#

- name: Run test  id: test  if: env.SKIP_CI != 'true'  run: ./mvnw test -Pit -f ./shenyu-integrated-test/${{ }}/pom.xml  continue-on-error: true

This step is to use the maven test command to execute the test classes in the /src/test/ directory one by one.

View Docker Compose logs#

- name: Check test result  if: env.SKIP_CI != 'true'  run: |    docker-compose -f ./shenyu-integrated-test/${{ }}/docker-compose.yml logs --tail="all"    if [[ ${{steps.test.outcome}} == "failure" ]]; then      echo "Test Failed"      exit 1    else      echo "Test Successful"      exit 0    fi

When there is an error in the workflow, the log of docker compose can help us to better troubleshoot the problem, so in this step, we will print the log of docker compose.

Code Analysis For Context-Path Plugin

· 3 min read
Kunshuai Zhu
Apache ShenYu Contributor

Before starting, you can refer to this article to start the gateway


First, look at the ContextPathPlugin#doExecute method, which is the core of this plugin.

protected Mono<Void> doExecute(final ServerWebExchange exchange, final ShenyuPluginChain chain, final SelectorData selector, final RuleData rule) {    ...    // 1. get the contextMappingHandle from the JVM cache    ContextMappingHandle contextMappingHandle = ContextPathPluginDataHandler.CACHED_HANDLE.get().obtainHandle(CacheKeyUtils.INST.getKey(rule));    ...    // 2. set shenyu context according to contextMappingHandle    buildContextPath(shenyuContext, contextMappingHandle);    return chain.execute(exchange);}
  1. Get the contextMappingHandle from the JVM cache

    The contextMappingHandle here is an instance of the ContextMappingHandle class, which has two member variables: contextPath and addPrefix

    These two variables have appeared in the Rules form in the Admin before, and they are updated when the data is synchronized.

  2. Set shenyu context according to contextMappingHandle

    Below is the source code of the ContextPathPlugin#buildContextPath method

    private void buildContextPath(final ShenyuContext context, final ContextMappingHandle handle) {    String realURI = "";    // 1. set the context path of shenyu, remove the prefix of the real URI according to the length of the contextPath    if (StringUtils.isNoneBlank(handle.getContextPath())) {        context.setContextPath(handle.getContextPath());        context.setModule(handle.getContextPath());        realURI = context.getPath().substring(handle.getContextPath().length());    }    // add prefix    if (StringUtils.isNoneBlank(handle.getAddPrefix())) {        if (StringUtils.isNotBlank(realURI)) {            realURI = handle.getAddPrefix() + realURI;        } else {            realURI = handle.getAddPrefix() + context.getPath();        }    }    context.setRealUrl(realURI);}
    • Set the context path of shenyu, remove the prefix of the real URI according to the length of the contextPath

      You may be wondering whether there is a problem with the so-called "according to the length of the contextPath" here?

      In fact, such a judgment is not a problem, because the request will be processed by the plugin only after it is matched by the Selector and Rules. Therefore, under the premise of setting up Selector and Rules, it is completely possible to meet the needs of converting a specific contextPath.

Then, the ContextPathPlugin class has a more important method skip, part of the code is shown below. We can find: If it is a call to the RPC service, the context_path plugin will be skipped directly.

public Boolean skip(final ServerWebExchange exchange) {    ...    return Objects.equals(rpcType, RpcTypeEnum.DUBBO.getName())            || Objects.equals(rpcType, RpcTypeEnum.GRPC.getName())            || Objects.equals(rpcType, RpcTypeEnum.TARS.getName())            || Objects.equals(rpcType, RpcTypeEnum.MOTAN.getName())            || Objects.equals(rpcType, RpcTypeEnum.SOFA.getName());}

Finally, the context-path plugin has another class ContextPathPluginDataHandler. The function of this class is to subscribe to the data of the plug-in. When the plugin configuration is modified, deleted, or added, the data is modified, deleted, or added to the JVM cache.

Code Analysis For Param-Mapping Plugin

· 5 min read
Kunshuai Zhu
Apache ShenYu Contributor

Before starting, you can refer to this article to start the gateway


Let's take a look at the structure of this plugin first, as shown in the figure below.


Guess: handler is used for data synchronization; strategy may be adapted to various request bodies, which should be the focus of this plugin; ParamMappingPlugin should be the implementation of ShenyuPlugin.

First, take a look at the ParamMappingPlugin, the focus is on the override of the doExecute method.

public Mono<Void> doExecute(final ServerWebExchange exchange, final ShenyuPluginChain chain, final SelectorData selector, final RuleData rule) {    ... // judge whether paramMappingHandle is null    // Determine the request body type according to the contentType in the header line    HttpHeaders headers = exchange.getRequest().getHeaders();    MediaType contentType = headers.getContentType();    // *    return match(contentType).apply(exchange, chain, paramMappingHandle);}
  • The match method returns the corresponding Operator according to contentType

    private Operator match(final MediaType mediaType) {    if (MediaType.APPLICATION_JSON.isCompatibleWith(mediaType)) {        return operatorMap.get(MediaType.APPLICATION_JSON.toString());    } else if (MediaType.APPLICATION_FORM_URLENCODED.isCompatibleWith(mediaType)) {        return operatorMap.get(MediaType.APPLICATION_FORM_URLENCODED.toString());    } else {        return operatorMap.get(Constants.DEFAULT);    }}

    As can be seen from the code of the match method, there are currently three types of DefaultOperator, FormDataOperator, and JsonOperator, which support the request body in two formats: x-www-form-urlencoded and json.

So let's take a look at what the above three operators are like.

1. DefaultOperator#

Nothing happens, its apply method just continues to execute the plug-in chain, and has no real function. When the request body does not match the Operator, it will be skipped by DefaultOperator.

2. FormDataOperator#

This class is used to process the request body in the format of x-www-form-urlencoded.

Mainly depends on the apply method, but it looks a bit strange.

public Mono<Void> apply(final ServerWebExchange exchange, final ShenyuPluginChain shenyuPluginChain, final ParamMappingHandle paramMappingHandle) {    return exchange.getFormData()            .switchIfEmpty(Mono.defer(() -> Mono.just(new LinkedMultiValueMap<>())))            .flatMap(multiValueMap -> {                ...            });}

The code in the ellipsis is the processing of the request body, as follows.

// judge whether it is emptyif (Objects.isNull(multiValueMap) || multiValueMap.isEmpty()) {    return shenyuPluginChain.execute(exchange);}// convert form-data to jsonString original = GsonUtils.getInstance().toJson(multiValueMap);"get from data success data:{}", original);// *modify request body*String modify = operation(original, paramMappingHandle);if (StringUtils.isEmpty(modify)) {    return shenyuPluginChain.execute(exchange);}...// Convert the modified json into LinkedMultiValueMap. Pay attention to this line, it will be mentioned later!LinkedMultiValueMap<String, String> modifyMap = GsonUtils.getInstance().toLinkedMultiValueMap(modify); BodyInserter bodyInserter = BodyInserters.fromValue(modifyMap);...// modify the request body in the exchange, and then continue to execute the plugin chainreturn bodyInserter.insert(cachedBodyOutputMessage, new BodyInserterContext())        .then(Mono.defer(() -> shenyuPluginChain.execute(exchange.mutate()                .request(new ModifyServerHttpRequestDecorator(httpHeaders, exchange.getRequest(), cachedBodyOutputMessage))                .build())        )).onErrorResume((Function<Throwable, Mono<Void>>) throwable -> release(cachedBodyOutputMessage, throwable));

PS: The omitted part is to set the request first and other operations.

The more important thing above should be the modification request body of the star, that is, the call of the operation method. Here, because of the parameter type, the default method of the Operator interface will be called first (instead of being overridden by the FormDataOperator).

default String operation(final String jsonValue, final ParamMappingHandle paramMappingHandle) {    DocumentContext context = JsonPath.parse(jsonValue);    // call the override operation method and add addParameterKey    operation(context, paramMappingHandle);    // replace the related replacedParameterKey    if (!CollectionUtils.isEmpty(paramMappingHandle.getReplaceParameterKeys())) {        paramMappingHandle.getReplaceParameterKeys().forEach(info -> {            context.renameKey(info.getPath(), info.getKey(), info.getValue());        });    }    // Delete the related removeParameterKey    if (!CollectionUtils.isEmpty(paramMappingHandle.getRemoveParameterKeys())) {        paramMappingHandle.getRemoveParameterKeys().forEach(info -> {            context.delete(info);        });    }    return context.jsonString();}

After sorting it out, we can find that the json tool JsonPath imported here makes the processing of the request body much simpler and clearer.

In addition, we can notice that the FormDataOperator overrides the operation(DocumentContext, ParamMappingHandle) method.

Why override it? There is a default method for handling addParameterKey in the interface.

// Default method in Operator interfacedefault void operation(final DocumentContext context, final ParamMappingHandle paramMappingHandle) {    if (!CollectionUtils.isEmpty(paramMappingHandle.getAddParameterKeys())) {        paramMappingHandle.getAddParameterKeys().forEach(info -> {            context.put(info.getPath(), info.getKey(), info.getValue()); //不同之处        });    }}
// method overridden by FormDataOperator@Overridepublic void operation(final DocumentContext context, final ParamMappingHandle paramMappingHandle) {    if (!CollectionUtils.isEmpty(paramMappingHandle.getAddParameterKeys())) {        paramMappingHandle.getAddParameterKeys().forEach(info -> {            context.put(info.getPath(), info.getKey(), Arrays.asList(info.getValue()));        });    }}

In fact, there is such a line in FormDataOperator#apply (mentioned earlier): LinkedMultiValueMap<String, String> modifyMap = GsonUtils.getInstance().toLinkedMultiValueMap(modify);

This line converts the modified json into LinkedMultiValueMap, GsonUtils#toLinkedMultiValueMap is as follows.

public LinkedMultiValueMap<String, String> toLinkedMultiValueMap(final String json) {    return GSON.fromJson(json, new TypeToken<LinkedMultiValueMap<String, String>>() {    }.getType());}

The attribute targetMap in the LinkedMultiValueMap class is defined as: private final Map<K, List<V>> targetMap

Therefore, the value in the json string must be in the form of a list, otherwise Gson will throw a conversion error exception, which is why the FormDataOperator must override the operator method.

But why use LinkedMultiValueMap?

Go back to the first line exchange.getFormData of the FormDataOperator#apply method. In SpringMVC, the return value type of DefaultServerWebExchange#getFormData is Mono<MultiValueMap<String, String>>, and LinkedMultiValueMap is a subclass of MultiValueMap. And, the getFormData method is for the request body in the format of x-www-form-urlencoded.



Obviously, this class is used to process the request body in Json format.

public Mono<Void> apply(final ServerWebExchange exchange, final ShenyuPluginChain shenyuPluginChain, final ParamMappingHandle paramMappingHandle) {    ServerRequest serverRequest = ServerRequest.create(exchange, MESSAGE_READERS);    Mono<String> mono = serverRequest.bodyToMono(String.class).switchIfEmpty(Mono.defer(() -> Mono.just(""))).flatMap(originalBody -> {"get body data success data:{}", originalBody);        // call the default operation method to modify the request body        String modify = operation(originalBody, paramMappingHandle);        return Mono.just(modify);    });    BodyInserter bodyInserter = BodyInserters.fromPublisher(mono, String.class);    ... //process the header line    CachedBodyOutputMessage outputMessage = new CachedBodyOutputMessage(exchange, headers);    // modify the request body in the exchange, and then continue to execute the plugin chain    return bodyInserter.insert(outputMessage, new BodyInserterContext())            .then(Mono.defer(() -> {                ServerHttpRequestDecorator decorator = new ModifyServerHttpRequestDecorator(headers, exchange.getRequest(), outputMessage);                return shenyuPluginChain.execute(exchange.mutate().request(decorator).build());            })).onErrorResume((Function<Throwable, Mono<Void>>) throwable -> release(outputMessage, throwable));}

The processing flow of JsonOperator is roughly similar to that of FormDataOperator.


Finally, use a picture to briefly summarize.


Register Center Source Code Analysis of Http Register

· 29 min read
Apache ShenYu Committer

Apache ShenYu is an asynchronous, high-performance, cross-language, responsive API gateway.

In ShenYu gateway, the registration center is used to register the client information to shenyu-admin, admin then synchronizes this information to the gateway through data synchronization, and the gateway completes traffic filtering through these data. The client information mainly includes interface information and URI information.

This article is based on shenyu-2.4.1 version for source code analysis, please refer to Client Access Principles for the introduction of the official website.

1. Registration Center Principle#

When the client starts, it reads the interface information and uri information, and sends the data to shenyu-admin by the specified registration type.

The registration center in the figure requires the user to specify which registration type to use. ShenYu currently supports Http, Zookeeper, Etcd, Consul and Nacos for registration. Please refer to Client Access Configuration for details on how to configure them.

ShenYu introduces Disruptor in the principle design of the registration center, in which the Disruptor queue plays a role in decoupling data and operations, which is conducive to expansion. If too many registration requests lead to registration exceptions, it also has a data buffering role.

As shown in the figure, the registration center is divided into two parts, one is the registration center client register-client, the load processing client data reading. The other is the registration center server register-server, which is loaded to handle the server side (that is shenyu-admin) data writing. Data is sent and received by specifying the registration type.

  • Client: Usually it is a microservice, which can be springmvc, spring-cloud, dubbo, grpc, etc.
  • register-client: register the central client, read the client interface and uri information.
  • Disruptor: decoupling data from operations, data buffering role.
  • register-server: registry server, here is shenyu-admin, receive data, write to database, send data synchronization events.
  • registration-type: specify the registration type, complete data registration, currently supports Http, Zookeeper, Etcd, Consul and Nacos.

This article analyzes the use of Http for registration, so the specific processing flow is as follows.

On the client side, after the data is out of the queue, the data is transferred via http and on the server side, the corresponding interface is provided to receive the data and then write it to the queue.

2. Client Registration Process#

When the client starts, it reads the attribute information according to the relevant configuration, and then writes it to the queue. Let's take the official shenyu-examples-http as an example and start the source code analysis . The official example is a microservice built by springboot. For the configuration of the registration center, please refer to the official website client access configuration .

2.1 Load configuration, read properties#

Let's start with a diagram that ties together the initialization process of the registry client.

We are analyzing registration by means of http, so the following configuration is required.

shenyu:  register:    registerType: http    serverLists: http://localhost:9095  client:    http:        props:          contextPath: /http          appName: http          port: 8189            isFull: false

Each attribute indicates the following meaning.

  • registerType: the service registration type, fill in http.
  • serverList: The address of the Shenyu-Admin project to fill in for the http registration type, note the addition of http:// and separate multiple addresses with English commas.
  • port: the start port of your project, currently springmvc/tars/grpc needs to be filled in.
  • contextPath: the routing prefix for your mvc project in shenyu gateway, such as /order, /product, etc. The gateway will route according to your prefix.
  • appName: the name of your application, if not configured, it will take the value of by default.
  • isFull: set true to proxy your entire service, false to proxy one of your controllers; currently applies to springmvc/springcloud.

After the project starts, it will first load the configuration file, read the property information and generate the corresponding Bean.

The first configuration file read is ShenyuSpringMvcClientConfiguration, which is the http registration configuration class for the shenyu client, indicated by @Configuration which is a configuration class, and by @ImportAutoConfiguration which is a configuration class. to introduce other configuration classes. Create SpringMvcClientBeanPostProcessor, which mainly handles metadata. Create ContextRegisterListener, which mainly handles URI information.

/** * Shenyu SpringMvc Client Configuration */@Configuration@ImportAutoConfiguration(ShenyuClientCommonBeanConfiguration.class)public class ShenyuSpringMvcClientConfiguration {
    //create SpringMvcClientBeanPostProcessor to handle metadata    @Bean    public SpringMvcClientBeanPostProcessor springHttpClientBeanPostProcessor(final ShenyuClientConfig clientConfig,final ShenyuClientRegisterRepository shenyuClientRegisterRepository) {        return new SpringMvcClientBeanPostProcessor(clientConfig.getClient().get(RpcTypeEnum.HTTP.getName()), shenyuClientRegisterRepository);    }       //create ContextRegisterListener to handle URI    @Bean    public ContextRegisterListener contextRegisterListener(final ShenyuClientConfig clientConfig) {        return new ContextRegisterListener(clientConfig.getClient().get(RpcTypeEnum.HTTP.getName()));    }}

ShenyuClientCommonBeanConfiguration is a shenyu client common configuration class that will create the bean common to the registry client.

  • Create ShenyuClientRegisterRepository, which is created by factory class.
  • Create ShenyuRegisterCenterConfig, which reads the shenyu.register property configuration.
  • Create ShenyuClientConfig, read the shenyu.client property configuration.

/** * Shenyu Client Common Bean Configuration */@Configurationpublic class ShenyuClientCommonBeanConfiguration {       // create ShenyuClientRegisterRepository by factory     @Bean    public ShenyuClientRegisterRepository shenyuClientRegisterRepository(final ShenyuRegisterCenterConfig config) {        return ShenyuClientRegisterRepositoryFactory.newInstance(config);    }        // create ShenyuRegisterCenterConfig to read shenyu.register properties    @Bean    @ConfigurationProperties(prefix = "shenyu.register")    public ShenyuRegisterCenterConfig shenyuRegisterCenterConfig() {        return new ShenyuRegisterCenterConfig();    }      // create ShenyuClientConfig to read shenyu.client properties    @Bean    @ConfigurationProperties(prefix = "shenyu")    public ShenyuClientConfig shenyuClientConfig() {        return new ShenyuClientConfig();    }}

2.2 HttpClientRegisterRepository#

The ShenyuClientRegisterRepository generated in the configuration file above is a concrete implementation of the client registration, which is an interface with the following implementation class.

  • HttpClientRegisterRepository: registration via http.
  • ConsulClientRegisterRepository: registration via Consul.
  • EtcdClientRegisterRepository: registration via Etcd; EtcdClientRegisterRepository: registration via Etcd.
  • NacosClientRegisterRepository: registration via nacos; NacosClientRegisterRepository: registration via nacos.
  • ZookeeperClientRegisterRepository: registration through Zookeeper.

The specific way which is achieved by loading through SPI, the implementation logic is as follows.

/** * load ShenyuClientRegisterRepository */public final class ShenyuClientRegisterRepositoryFactory {        private static final Map<String, ShenyuClientRegisterRepository> REPOSITORY_MAP = new ConcurrentHashMap<>();        /**     * create ShenyuClientRegisterRepository     */    public static ShenyuClientRegisterRepository newInstance(final ShenyuRegisterCenterConfig shenyuRegisterCenterConfig) {        if (!REPOSITORY_MAP.containsKey(shenyuRegisterCenterConfig.getRegisterType())) {            // Loading by means of SPI, type determined by registerType            ShenyuClientRegisterRepository result = ExtensionLoader.getExtensionLoader(ShenyuClientRegisterRepository.class).getJoin(shenyuRegisterCenterConfig.getRegisterType());            //init ShenyuClientRegisterRepository            result.init(shenyuRegisterCenterConfig);            ShenyuClientShutdownHook.set(result, shenyuRegisterCenterConfig.getProps());            REPOSITORY_MAP.put(shenyuRegisterCenterConfig.getRegisterType(), result);            return result;        }        return REPOSITORY_MAP.get(shenyuRegisterCenterConfig.getRegisterType());    }}

The load type is specified by registerType, which is the type we specify in the configuration file at

shenyu:  register:    registerType: http    serverLists: http://localhost:9095

We specified http, so it will go to load HttpClientRegisterRepository. After the object is successfully created, the initialization method init() is executed as follows.

@Joinpublic class HttpClientRegisterRepository implements ShenyuClientRegisterRepository {        @Override    public void init(final ShenyuRegisterCenterConfig config) {        this.serverList = Lists.newArrayList(Splitter.on(",").split(config.getServerLists()));    }    // ......}

Read serverLists from the configuration file, the address of sheenyu-admin, in preparation for subsequent data sending. The class annotation @Join is used for SPI loading.

SPI, known as Service Provider Interface, is a service provider discovery feature built into the JDK, a mechanism for dynamic replacement discovery.

shenyu-spi is a custom SPI extension implementation for the Apache ShenYu gateway, designed and implemented with reference to Dubbo SPI extension implementation.

2.3 SpringMvcClientBeanPostProcessor#

Create SpringMvcClientBeanPostProcessor, which is responsible for metadata construction and registration, with the following constructor logic.

/** *  spring mvc client BeanPostProcessor */public class SpringMvcClientBeanPostProcessor implements BeanPostProcessor {
    /**     * Instantiation by constructor     */    public SpringMvcClientBeanPostProcessor(final PropertiesConfig clientConfig,                                            final ShenyuClientRegisterRepository shenyuClientRegisterRepository) {        // read Properties        Properties props = clientConfig.getProps();        // get port information and verify        int port = Integer.parseInt(props.getProperty(ShenyuClientConstants.PORT));        if (port <= 0) {            String errorMsg = "http register param must config the port must > 0";            LOG.error(errorMsg);            throw new ShenyuClientIllegalArgumentException(errorMsg);        }        // get appName        this.appName = props.getProperty(ShenyuClientConstants.APP_NAME);        // get contextPath        this.contextPath = props.getProperty(ShenyuClientConstants.CONTEXT_PATH);        // check appName and contextPath        if (StringUtils.isBlank(appName) && StringUtils.isBlank(contextPath)) {            String errorMsg = "http register param must config the appName or contextPath";            LOG.error(errorMsg);            throw new ShenyuClientIllegalArgumentException(errorMsg);        }        // get isFull        this.isFull = Boolean.parseBoolean(props.getProperty(ShenyuClientConstants.IS_FULL, Boolean.FALSE.toString()));        // start publisher        publisher.start(shenyuClientRegisterRepository);    }
    // ......    @Override    public Object postProcessAfterInitialization(@NonNull final Object bean, @NonNull final String beanName) throws BeansException {      // ......    }

In the constructor, the main purpose is to read the property information and then perform the checksum.

shenyu:  client:    http:        props:          contextPath: /http          appName: http          port: 8189            isFull: false

Finally, publisher.start() is executed to start event publishing and prepare for registration.

  • ShenyuClientRegisterEventPublisher

ShenyuClientRegisterEventPublisher is implemented via singleton pattern, mainly generating metadata and URI subscribers (subsequently used for data publishing), and then starting the Disruptor queue. A common method publishEvent() is provided to publish events and send data to the Disruptor queue.

public class ShenyuClientRegisterEventPublisher {    private static final ShenyuClientRegisterEventPublisher INSTANCE = new ShenyuClientRegisterEventPublisher();        private DisruptorProviderManage providerManage;        private RegisterClientExecutorFactory factory;        public static ShenyuClientRegisterEventPublisher getInstance() {        return INSTANCE;    }        public void start(final ShenyuClientRegisterRepository shenyuClientRegisterRepository) {        factory = new RegisterClientExecutorFactory();        factory.addSubscribers(new ShenyuClientMetadataExecutorSubscriber(shenyuClientRegisterRepository));        factory.addSubscribers(new ShenyuClientURIExecutorSubscriber(shenyuClientRegisterRepository));        providerManage = new DisruptorProviderManage(factory);        providerManage.startup();    }        public <T> void publishEvent(final T data) {        DisruptorProvider<Object> provider = providerManage.getProvider();        provider.onData(f -> f.setData(data));    }}

The logic of the constructor of SpringMvcClientBeanPostProcessor is analyzed, it mainly reads the property configuration, creates metadata and URI subscribers, and starts the Disruptor queue. It is important to note that it implements BeanPostProcessor, an interface provided by Spring that executes the postProcessAfterInitialization() method of the post-processor before it actually starts to be used in the Bean lifecycle.

  • postProcessAfterInitialization()

SpringMvcClientBeanPostProcessor acts as a post-processor which does the following: reads the metadata in the annotation and registers it with admin.

public class SpringMvcClientBeanPostProcessor implements BeanPostProcessor {   // ......        // reads the metadata in the annotation and registers it with admin    @Override    public Object postProcessAfterInitialization(@NonNull final Object bean, @NonNull final String beanName) throws BeansException {        // Configuration attribute, if isFull=true, means register the whole microservice        if (isFull) {            return bean;        }        // get Controller annotation        Controller controller = AnnotationUtils.findAnnotation(bean.getClass(), Controller.class);         // get RequestMapping annotation        RequestMapping requestMapping = AnnotationUtils.findAnnotation(bean.getClass(), RequestMapping.class);        if (controller != null || requestMapping != null) {             // get the ShenyuSpringMvcClient annotation for the current bean            ShenyuSpringMvcClient clazzAnnotation = AnnotationUtils.findAnnotation(bean.getClass(), ShenyuSpringMvcClient.class);            String prePath = "";            // If there is no ShenyuSpringMvcClient annotation, it is returned, indicating that the interface does not need to be registered            if (Objects.isNull(clazzAnnotation)) {                return bean;            }             // If the path attribute of the ShenyuSpringMvcClient annotation includes *, it means that the entire interface is registered            if (clazzAnnotation.path().indexOf("*") > 1) {                // build metadata, publish registration event                publisher.publishEvent(buildMetaDataDTO(clazzAnnotation, prePath));                return bean;            }                        prePath = clazzAnnotation.path();            // get all methods of the current bean            final Method[] methods = ReflectionUtils.getUniqueDeclaredMethods(bean.getClass());            for (Method method : methods) {                // get ShenyuSpringMvcClient annotation                ShenyuSpringMvcClient shenyuSpringMvcClient = AnnotationUtils.findAnnotation(method, ShenyuSpringMvcClient.class);                // If the method has the annotation ShenyuSpringMvcClient, it means that the method needs to be registered                if (Objects.nonNull(shenyuSpringMvcClient)) {                    // build metadata, publish registration event                    publisher.publishEvent(buildMetaDataDTO(shenyuSpringMvcClient, prePath));                }            }        }        return bean;    }
    // build metadata    private MetaDataRegisterDTO buildMetaDataDTO(final ShenyuSpringMvcClient shenyuSpringMvcClient, final String prePath) {        // contextPath        String contextPath = this.contextPath;        // appName        String appName = this.appName;        // path        String path;        if (StringUtils.isEmpty(contextPath)) {            path = prePath + shenyuSpringMvcClient.path();        } else {            path = contextPath + prePath + shenyuSpringMvcClient.path();        }        // desc info        String desc = shenyuSpringMvcClient.desc();        // ruleName        String configRuleName = shenyuSpringMvcClient.ruleName();        String ruleName = StringUtils.isBlank(configRuleName) ? path : configRuleName;        return MetaDataRegisterDTO.builder()                .contextPath(contextPath)                .appName(appName)                .path(path)                .pathDesc(desc)                .rpcType(RpcTypeEnum.HTTP.getName())                .enabled(shenyuSpringMvcClient.enabled())                .ruleName(ruleName)                .registerMetaData(shenyuSpringMvcClient.registerMetaData())                .build();    }}

In the post-processor, you need to read the configuration property, if isFull=true, it means register the whole microservice. Get the Controller annotation, RequestMapping annotation, ShenyuSpringMvcClient annotation of the current bean and determine if the current bean is an interface by reading these annotations? Does the interface need to be registered? Does the method need to be registered? Then build metadata based on the properties in the ShenyuSpringMvcClient annotation, and finally publish the event for registration via publisher.publishEvent().

The Controller annotation and the RequestMapping annotation are provided by Spring, which you should be familiar with, so I won't go into details. The ShenyuSpringMvcClient annotation is provided by Apache ShenYu to register the SpringMvc client, which is defined as follows.

/** * ShenyuSpringMvcClient */@Retention(RetentionPolicy.RUNTIME)@Target({ElementType.TYPE, ElementType.METHOD})public @interface ShenyuSpringMvcClient {    // path     String path();        // ruleName     String ruleName() default "";        // desc info    String desc() default "";
    // enabled    boolean enabled() default true;        // register MetaData     boolean  registerMetaData() default false;}

It is used as follows.

  • register the entire interface
@RestController@RequestMapping("/test")@ShenyuSpringMvcClient(path = "/test/**")  // register the entire interfacepublic class HttpTestController { //......}
  • register current method
@RestController@RequestMapping("/order")@ShenyuSpringMvcClient(path = "/order")public class OrderController {
    /**     * Save order dto.     *     * @param orderDTO the order dto     * @return the order dto     */    @PostMapping("/save")    @ShenyuSpringMvcClient(path = "/save", desc = "Save order") // register current method    public OrderDTO save(@RequestBody final OrderDTO orderDTO) {        orderDTO.setName("hello world save order");        return orderDTO;    }
  • publisher.publishEvent()

This method sends the data to the Disruptor queue. More details about the Disruptor queue are not described here, which does not affect the flow of analyzing the registration.

When the data is sent, the consumers of the Disruptor queue will process the data for consumption.

This method sends the data to the Disruptor queue. More details about the Disruptor queue are not described here, which does not affect the flow of analyzing the registration.

  • QueueConsumer

QueueConsumer is a consumer that implements the WorkHandler interface, which is created in the providerManage.startup() logic. The WorkHandler interface is the data consumption interface for disruptor, and the only method is onEvent().

package com.lmax.disruptor;
public interface WorkHandler<T> {    void onEvent(T var1) throws Exception;}

The QueueConsumer overrides the onEvent() method, and the main logic is to generate the consumption task and then go to the thread pool to execute it.

/** *  * QueueConsumer */public class QueueConsumer<T> implements WorkHandler<DataEvent<T>> {        // ......
    @Override    public void onEvent(final DataEvent<T> t) {        if (t != null) {            // create queue consumption tasks via factory            QueueConsumerExecutor<T> queueConsumerExecutor = factory.create();            // set data            queueConsumerExecutor.setData(t.getData());            // help gc            t.setData(null);            // put in the thread pool to execute the consumption task            executor.execute(queueConsumerExecutor);        }    }}

QueueConsumerExecutor is the task that is executed in the thread pool, it implements the Runnable interface, and there are two specific implementation classes.

  • RegisterClientConsumerExecutor: the client-side consumer executor.
  • RegisterServerConsumerExecutor: server-side consumer executor.

As the name implies, one is responsible for handling client-side tasks, and one is responsible for handling server-side tasks (the server side is admin, which is analyzed below).

  • RegisterClientConsumerExecutor

The logic of the rewritten run() is as follows.

public final class RegisterClientConsumerExecutor extends QueueConsumerExecutor<DataTypeParent> {        //...... 
    @Override    public void run() {        // get data        DataTypeParent dataTypeParent = getData();        // call the appropriate processor for processing according to the data type        subscribers.get(dataTypeParent.getType()).executor(Lists.newArrayList(dataTypeParent));    }    }

Different processors are called to perform the corresponding tasks based on different data types. There are two types of data, one is metadata, which records the client registration information. One is the URI data, which records the client service information.

public enum DataType {     META_DATA,      URI,}
  • ExecutorSubscriber#executor()

The actuator subscribers are also divided into two categories, one that handles metadata and one that handles URIs. There are two on the client side and two on the server side, so there are four in total.

  • ShenyuClientMetadataExecutorSubscriber#executor()

The metadata processing logic on the client side is: iterate through the metadata information and call the interface method persistInterface() to finish publishing the data.

public class ShenyuClientMetadataExecutorSubscriber implements ExecutorTypeSubscriber<MetaDataRegisterDTO> {       //......        @Override    public DataType getType() {        return DataType.META_DATA;     }        @Override    public void executor(final Collection<MetaDataRegisterDTO> metaDataRegisterDTOList) {        for (MetaDataRegisterDTO metaDataRegisterDTO : metaDataRegisterDTOList) {            // call the interface method persistInterface() to finish publishing the data            shenyuClientRegisterRepository.persistInterface(metaDataRegisterDTO);        }    }}
  • ShenyuClientRegisterRepository#persistInterface()

ShenyuClientRegisterRepository is an interface to represent client-side data registration, and it has five implementation classes at present, each of which represents a registration method.

  • ConsulClientRegisterRepository: client registration is achieved through Consul.
  • EtcdClientRegisterRepository: client registration through Etcd.
  • HttpClientRegisterRepository: client registration via Http; NacosClientRegisterRepository: client registration via Http.
  • NacosClientRegisterRepository: client registration via Nacos; ZookeeperClientRegisterRepository: client registration via Nacos.
  • ZookeeperClientRegisterRepository: client registration via Zookeeper.

As you can see from the diagram, the loading of the registry is done by means of SPI. This was mentioned earlier, and the specific class loading is done in the client-side generic configuration file by specifying the properties in the configuration file.

/** * load ShenyuClientRegisterRepository */public final class ShenyuClientRegisterRepositoryFactory {        private static final Map<String, ShenyuClientRegisterRepository> REPOSITORY_MAP = new ConcurrentHashMap<>();        /**     * create ShenyuClientRegisterRepository     */    public static ShenyuClientRegisterRepository newInstance(final ShenyuRegisterCenterConfig shenyuRegisterCenterConfig) {        if (!REPOSITORY_MAP.containsKey(shenyuRegisterCenterConfig.getRegisterType())) {            // loading by means of SPI, type determined by registerType            ShenyuClientRegisterRepository result = ExtensionLoader.getExtensionLoader(ShenyuClientRegisterRepository.class).getJoin(shenyuRegisterCenterConfig.getRegisterType());            // perform initialization operations            result.init(shenyuRegisterCenterConfig);            ShenyuClientShutdownHook.set(result, shenyuRegisterCenterConfig.getProps());            REPOSITORY_MAP.put(shenyuRegisterCenterConfig.getRegisterType(), result);            return result;        }        return REPOSITORY_MAP.get(shenyuRegisterCenterConfig.getRegisterType());    }}

The source code analysis in this article is based on the Http way of registration, so we first analyze the HttpClientRegisterRepository, and the other registration methods will be analyzed afterwards.

Registration by way of http is very simple, it is to call the tool class to send http requests. The registration metadata and URI are both called by the same method doRegister(), specifying the interface and type.

  • /shenyu-client/register-metadata: the interface provided by the server for registering metadata.
  • /shenyu-client/register-uri: Server-side interface for registering URIs.
@Joinpublic class HttpClientRegisterRepository implements ShenyuClientRegisterRepository {    // server-side provided interface for registering metadata        private static final String META_PATH = "/shenyu-client/register-metadata";
    // the interface provided by the server for registering URIs    private static final String URI_PATH = "/shenyu-client/register-uri";
    @Override    public void persistURI(final URIRegisterDTO registerDTO) {        doRegister(registerDTO, URI_PATH, Constants.URI);    }        @Override    public void persistInterface(final MetaDataRegisterDTO metadata) {        doRegister(metadata, META_PATH, META_TYPE);    }        // do register    private <T> void doRegister(final T t, final String path, final String type) {        // iterate through the list of admin services (admin may be clustered)        for (String server : serverList) {            try {                // calling the tool class to send http requests                RegisterUtils.doRegister(GsonUtils.getInstance().toJson(t), server + path, type);                return;            } catch (Exception e) {                LOGGER.error("register admin url :{} is fail, will retry", server);            }        }    }}

Serialize the data and send it via OkHttp.

public final class RegisterUtils {      //...... 
    // Sending data via OkHttp    public static void doRegister(final String json, final String url, final String type) throws IOException {        String result = OkHttpTools.getInstance().post(url, json);        if (Objects.equals(SUCCESS, result)) {  "{} client register success: {} ", type, json);        } else {            LOGGER.error("{} client register error: {} ", type, json);        }    }}

At this point, the logic of the client registering metadata by means of http is finished. To summarize: construct metadata by reading custom annotation information, send the data to the Disruptor queue, then consume the data from the queue, put the consumer into the thread pool to execute, and finally send an http request to the admin.

The source code analysis process of the client-side metadata registration process is completed and described in a flowchart as follows.

2.4 ContextRegisterListener#

Create ContextRegisterListener, which is responsible for the construction and registration of client-side URI data, and its creation is done in the configuration file.

@Configuration@ImportAutoConfiguration(ShenyuClientCommonBeanConfiguration.class)public class ShenyuSpringMvcClientConfiguration {     // ......        // create ContextRegisterListener    @Bean    public ContextRegisterListener contextRegisterListener(final ShenyuClientConfig clientConfig) {        return new ContextRegisterListener(clientConfig.getClient().get(RpcTypeEnum.HTTP.getName()));    }}

ContextRegisterListener implements the ApplicationListener interface and overrides the onApplicationEvent() method, which is executed when a Spring event occurs.

public class ContextRegisterListener implements ApplicationListener<ContextRefreshedEvent> {
    //Instantiation is done through the constructor    public ContextRegisterListener(final PropertiesConfig clientConfig) {        // read shenyu.client.http properties        Properties props = clientConfig.getProps();        // isFull        this.isFull = Boolean.parseBoolean(props.getProperty(ShenyuClientConstants.IS_FULL, Boolean.FALSE.toString()));        // contextPath        String contextPath = props.getProperty(ShenyuClientConstants.CONTEXT_PATH);        this.contextPath = contextPath;        if (isFull) {            if (StringUtils.isBlank(contextPath)) {                String errorMsg = "http register param must config the contextPath";                LOG.error(errorMsg);                throw new ShenyuClientIllegalArgumentException(errorMsg);            }            this.contextPath = contextPath + "/**";        }        // port         int port = Integer.parseInt(props.getProperty(ShenyuClientConstants.PORT));        // appName         this.appName = props.getProperty(ShenyuClientConstants.APP_NAME);        // host = props.getProperty(ShenyuClientConstants.HOST);        this.port = port;    }
    // This method is executed when a context refresh event(ContextRefreshedEvent), occurs    @Override    public void onApplicationEvent(@NonNull final ContextRefreshedEvent contextRefreshedEvent) {        // The contents of the method are guaranteed to be executed only once        if (!registered.compareAndSet(false, true)) {            return;        }        // If isFull=true means register the entire service, build the metadata and register        if (isFull) {            publisher.publishEvent(buildMetaDataDTO());        }                // build URI data and register it        publisher.publishEvent(buildURIRegisterDTO());    }
    // build URI data     private URIRegisterDTO buildURIRegisterDTO() {        String host = IpUtils.isCompleteHost( ? : IpUtils.getHost(;        return URIRegisterDTO.builder()                .contextPath(this.contextPath)                .appName(appName)                .host(host)                .port(port)                .rpcType(RpcTypeEnum.HTTP.getName())                .build();    }
    // build metadata    private MetaDataRegisterDTO buildMetaDataDTO() {        String contextPath = this.contextPath;        String appName = this.appName;        return MetaDataRegisterDTO.builder()                .contextPath(contextPath)                .appName(appName)                .path(contextPath)                .rpcType(RpcTypeEnum.HTTP.getName())                .enabled(true)                .ruleName(contextPath)                .build();    }}

The main thing in the constructor is to read the property configuration.

The onApplicationEvent() method is executed when a Spring event occurs, the parameter here is ContextRefreshedEvent, which means the context refresh event. The logic here is executed when the Spring container is ready: if isFull=true means register the whole service, build the metadata and register it, the case isFull=true is not handled in the post-processor SpringMvcClientBeanPostProcessor analyzed earlier, so here it is processing. Then the URI data is constructed and registered.

ContextRefreshedEvent is a Spring built-in event. It is fired when the ApplicationContext is initialized or refreshed. This can also happen in the ConfigurableApplicationContext interface using the refresh() method. Initialization here means that all Beans have been successfully loaded, post-processing Beans have been detected and activated, all Singleton Beans have been pre-instantiated, and the ApplicationContext container is ready to be used.

The registration logic is done through publisher.publishEvent(). It has been analyzed earlier: data is written to the Disruptor queue, consumed from it, and finally processed by the ExecutorSubscriber.

  • ExecutorSubscriber#executor()

The actuator subscribers are divided into two categories, one that handles metadata and one that handles URIs. There are two on the client side and two on the server side, so there are four in total.

Here is the registration URI information, so the execution class is ShenyuClientURIExecutorSubscriber.

  • ShenyuClientURIExecutorSubscriber#executor()

The main logic is to iterate through the URI data collection and implement data registration through the persistURI() method.

public class ShenyuClientURIExecutorSubscriber implements ExecutorTypeSubscriber<URIRegisterDTO> {        //......        @Override    public DataType getType() {        return DataType.URI;     }        // register URI    @Override    public void executor(final Collection<URIRegisterDTO> dataList) {        for (URIRegisterDTO uriRegisterDTO : dataList) {            Stopwatch stopwatch = Stopwatch.createStarted();            while (true) {                try (Socket ignored = new Socket(uriRegisterDTO.getHost(), uriRegisterDTO.getPort())) {                    break;                } catch (IOException e) {                    long sleepTime = 1000;                    // maybe the port is delay exposed                    if (stopwatch.elapsed(TimeUnit.SECONDS) > 5) {                        LOG.error("host:{}, port:{} connection failed, will retry",                                uriRegisterDTO.getHost(), uriRegisterDTO.getPort());                        // If the connection fails for a long time, Increase sleep time                        if (stopwatch.elapsed(TimeUnit.SECONDS) > 180) {                            sleepTime = 10000;                        }                    }                    try {                        TimeUnit.MILLISECONDS.sleep(sleepTime);                    } catch (InterruptedException ex) {                        ex.printStackTrace();                    }                }            }            ShenyuClientShutdownHook.delayOtherHooks();                        shenyuClientRegisterRepository.persistURI(uriRegisterDTO);        }    }}

The while(true) loop in the code is to ensure that the client has been successfully started and can connect via host and port.

The logic behind it is: add the hook function for gracefully stopping the client .

Data registration is achieved through the persistURI() method. The whole logic is also analyzed in the previous section, and ultimately it is the OkHttp client that initiates http to shenyu-admin and registers the URI by way of http.

The analysis of the registration logic of the client is finished here, and the metadata and URI data constructed are sent to the Disruptor queue, from which they are then consumed, read, and sent to admin via http.

The source code analysis of the client-side URI registration process is complete, with the following flow chart.

3. Server-side registration process#

3.1 ShenyuHttpRegistryController#

From the previous analysis, we know that the server side provides two interfaces for registration.

  • /shenyu-client/register-metadata: The interface provided by the server side is used to register metadata.
  • /shenyu-client/register-uri: The server-side interface is provided for registering URIs.

These two interfaces are located in ShenyuHttpRegistryController, which implements the ShenyuServerRegisterRepository interface and is the implementation class for server-side registration. It is marked with @Join to indicate loading via SPI.

@RequestMapping("/shenyu-client")@Joinpublic class ShenyuHttpRegistryController implements ShenyuServerRegisterRepository {
    private ShenyuServerRegisterPublisher publisher;
    @Override    public void init(final ShenyuServerRegisterPublisher publisher, final ShenyuRegisterCenterConfig config) {        this.publisher = publisher;    }        // register Metadata    @PostMapping("/register-metadata")    @ResponseBody    public String registerMetadata(@RequestBody final MetaDataRegisterDTO metaDataRegisterDTO) {        publish(metaDataRegisterDTO);        return ShenyuResultMessage.SUCCESS;    }           // register URI    @PostMapping("/register-uri")    @ResponseBody    public String registerURI(@RequestBody final URIRegisterDTO uriRegisterDTO) {        publish(uriRegisterDTO);        return ShenyuResultMessage.SUCCESS;    }
    // publish event    private <T> void publish(final T t) {        publisher.publish(Collections.singletonList(t));    }}

The two registration interfaces get the data well and call the publish() method to publish the data to the Disruptor queue.

  • ShenyuServerRegisterRepository

The ShenyuServerRegisterRepository interface is a service registration interface, which has five implementation classes, indicating five types of registration.

  • ConsulServerRegisterRepository: registration is achieved through Consul;
  • EtcdServerRegisterRepository: registration through Etcd.
  • NacosServerRegisterRepository: registration through Nacos.
  • ShenyuHttpRegistryController: registration via Http; ShenyuHttpRegistryController: registration via Http.
  • ZookeeperServerRegisterRepository: registration through Zookeeper.

The exact method used is specified by the configuration file and then loaded via SPI.

In the application.yml file in shenyu-admin configure the registration method, registerType specify the registration type, when registering with http, serverLists do not need to be filled in, for more configuration instructions you can refer to the official website Client Access Configuration.

shenyu:  register:    registerType: http     serverLists: 
  • RegisterCenterConfiguration

After introducing the relevant dependencies and properties configuration, when starting shenyu-admin, the configuration file will be loaded first, and the configuration file class related to the registration center is RegisterCenterConfiguration.

@Configurationpublic class RegisterCenterConfiguration {    @Bean    @ConfigurationProperties(prefix = "shenyu.register")    public ShenyuRegisterCenterConfig shenyuRegisterCenterConfig() {        return new ShenyuRegisterCenterConfig();    }        //create ShenyuServerRegisterRepository to register in admin    @Bean    public ShenyuServerRegisterRepository shenyuServerRegisterRepository(final ShenyuRegisterCenterConfig shenyuRegisterCenterConfig, final List<ShenyuClientRegisterService> shenyuClientRegisterService) {        // 1. get the registration type from the configuration property        String registerType = shenyuRegisterCenterConfig.getRegisterType();        // 2. load the implementation class by registering the type with the SPI method        ShenyuServerRegisterRepository registerRepository = ExtensionLoader.getExtensionLoader(ShenyuServerRegisterRepository.class).getJoin(registerType);        // 3. get the publisher and write data to the Disruptor queue        RegisterServerDisruptorPublisher publisher = RegisterServerDisruptorPublisher.getInstance();        // 4. ShenyuClientRegisterService, rpcType -> registerService        Map<String, ShenyuClientRegisterService> registerServiceMap =, e -> e));        // 5. start publisher        publisher.start(registerServiceMap);        // 6. init registerRepository        registerRepository.init(publisher, shenyuRegisterCenterConfig);        return registerRepository;    }}

Two beans are generated in the configuration class.

  • shenyuRegisterCenterConfig: to read the attribute configuration.

  • shenyuServerRegisterRepository: for server-side registration.

In the process of creating shenyuServerRegisterRepository, a series of preparations are also performed.

    1. get the registration type from the configuration property.
    1. Load the implementation class by the registration type with the SPI method: for example, if the specified type is http, ShenyuHttpRegistryController will be loaded.
    1. Get publisher and write data to the Disruptor queue.
    1. Register Service, rpcType -> registerService: get the registered Service, each rpc has a corresponding Service. The client for this article is built through springboot, which belongs to the http type, and other client types: dubbo, Spring Cloud, gRPC, etc.
    1. Preparation for event publishing: add server-side metadata and URI subscribers, process the data. And start the Disruptor queue.
    1. Initialization operation for registration: http type registration initialization operation is to save publisher.
  • RegisterServerDisruptorPublisher#publish()

The server-side publisher that writes data to the Disruptor queue , built via the singleton pattern.

public class RegisterServerDisruptorPublisher implements ShenyuServerRegisterPublisher {    private static final RegisterServerDisruptorPublisher INSTANCE = new RegisterServerDisruptorPublisher();
    public static RegisterServerDisruptorPublisher getInstance() {        return INSTANCE;    }       //prepare for event publishing, add server-side metadata and URI subscribers, process data. And start the Disruptor queue.    public void start(final Map<String, ShenyuClientRegisterService> shenyuClientRegisterService) {        factory = new RegisterServerExecutorFactory();        // add URI data subscriber        factory.addSubscribers(new URIRegisterExecutorSubscriber(shenyuClientRegisterService));        // add Metadata subscriber        factory.addSubscribers(new MetadataExecutorSubscriber(shenyuClientRegisterService));        //start Disruptor        providerManage = new DisruptorProviderManage(factory);        providerManage.startup();    }        // write data to queue    @Override    public <T> void publish(final T data) {        DisruptorProvider<Object> provider = providerManage.getProvider();        provider.onData(f -> f.setData(data));    }        @Override    public void close() {        providerManage.getProvider().shutdown();    }}

The loading of the configuration file, which can be seen as the initialization process of the registry server, is described in the following diagram.

3.2 QueueConsumer#

In the previous analysis of the client-side disruptor queue consumption of data over. The server side has the same logic, except that the executor performing the task changes.

The QueueConsumer is a consumer that implements the WorkHandler interface, which is created in the providerManage.startup() logic. The WorkHandler interface is the data consumption interface for disruptor, and the only method is onEvent().

package com.lmax.disruptor;
public interface WorkHandler<T> {    void onEvent(T var1) throws Exception;}

The QueueConsumer overrides the onEvent() method, and the main logic is to generate the consumption task and then go to the thread pool to execute it.

/** *  * QueueConsumer */public class QueueConsumer<T> implements WorkHandler<DataEvent<T>> {        // ......
    @Override    public void onEvent(final DataEvent<T> t) {        if (t != null) {            // create queue consumption tasks via factory            QueueConsumerExecutor<T> queueConsumerExecutor = factory.create();            // set data            queueConsumerExecutor.setData(t.getData());            // help gc            t.setData(null);            // put in the thread pool to execute the consumption task            executor.execute(queueConsumerExecutor);        }    }}

QueueConsumerExecutor is the task that is executed in the thread pool, it implements the Runnable interface, and there are two specific implementation classes.

  • RegisterClientConsumerExecutor: the client-side consumer executor.
  • RegisterServerConsumerExecutor: server-side consumer executor.

As the name implies, one is responsible for handling client-side tasks and one is responsible for handling server-side tasks.

  • RegisterServerConsumerExecutor#run()

RegisterServerConsumerExecutor is a server-side consumer executor that indirectly implements the Runnable interface via QueueConsumerExecutor and overrides the run() method.

public final class RegisterServerConsumerExecutor extends QueueConsumerExecutor<List<DataTypeParent>> {   // ...
    @Override    public void run() {        //get the data from the disruptor queue        List<DataTypeParent> results = getData();        // check data        results = -> isValidData(data)).collect(Collectors.toList());        if (CollectionUtils.isEmpty(results)) {            return;        }        //execute operations according to type        getType(results).executor(results);    }        // get subscribers by type    private ExecutorSubscriber getType(final List<DataTypeParent> list) {        DataTypeParent result = list.get(0);        return subscribers.get(result.getType());    }}
  • ExecutorSubscriber#executor()

The actuator subscribers are divided into two categories, one that handles metadata and one that handles URIs. There are two on the client side and two on the server side, so there are four in total.

  • MetadataExecutorSubscriber#executor()

In case of registering metadata, this is achieved by MetadataExecutorSubscriber#executor(): get the registered Service according to the type and call register().

public class MetadataExecutorSubscriber implements ExecutorTypeSubscriber<MetaDataRegisterDTO> {     //......
    @Override    public DataType getType() {        return DataType.META_DATA;     }
    @Override    public void executor(final Collection<MetaDataRegisterDTO> metaDataRegisterDTOList) {        // Traversing the metadata list        for (MetaDataRegisterDTO metaDataRegisterDTO : metaDataRegisterDTOList) {            // Get registered Service by type            ShenyuClientRegisterService shenyuClientRegisterService = this.shenyuClientRegisterService.get(metaDataRegisterDTO.getRpcType());            Objects.requireNonNull(shenyuClientRegisterService);            // Registration of metadata, locking to ensure sequential execution and prevent concurrent errors            synchronized (ShenyuClientRegisterService.class) {                shenyuClientRegisterService.register(metaDataRegisterDTO);            }        }    }}
  • URIRegisterExecutorSubscriber#executor()

In case of registration metadata, this is achieved by URIRegisterExecutorSubscriber#executor(): construct URI data, find Service according to the registration type, and achieve registration by the registerURI method.

public class URIRegisterExecutorSubscriber implements ExecutorTypeSubscriber<URIRegisterDTO> {    //......        @Override    public DataType getType() {        return DataType.URI;     }        @Override    public void executor(final Collection<URIRegisterDTO> dataList) {        if (CollectionUtils.isEmpty(dataList)) {            return;        }        // Build URI data types and register them with the registerURI method        findService(dataList).ifPresent(service -> {            Map<String, List<URIRegisterDTO>> listMap = buildData(dataList);            listMap.forEach(service::registerURI);        });    }        // Find Service by type    private Optional<ShenyuClientRegisterService> findService(final Collection<URIRegisterDTO> dataList) {        return -> shenyuClientRegisterService.get(dto.getRpcType())).findFirst();    }}
  • ShenyuClientRegisterService#register()

ShenyuClientRegisterService is the registration method interface, which has several implementation classes.

  • AbstractContextPathRegisterService: abstract class, handling part of the public logic.
  • AbstractShenyuClientRegisterServiceImpl: : abstract class, handles part of the public logic.
  • ShenyuClientRegisterDivideServiceImpl: divide class, handles http registration types.
  • ShenyuClientRegisterDubboServiceImpl: dubbo class, handles dubbo registration types.
  • ShenyuClientRegisterGrpcServiceImpl: gRPC class, handles gRPC registration types.
  • ShenyuClientRegisterMotanServiceImpl: Motan class, handles Motan registration types.
  • ShenyuClientRegisterSofaServiceImpl: Sofa class, handles Sofa registration types.
  • ShenyuClientRegisterSpringCloudServiceImpl: SpringCloud class, handles SpringCloud registration types.
  • ShenyuClientRegisterTarsServiceImpl: Tars class, handles Tars registration types.

From the above, we can see that each microservice has a corresponding registration implementation class. The source code analysis in this article is based on the official shenyu-examples-http as an example, it is of http registration type, so the registration implementation class for metadata and URI data is ShenyuClientRegisterDivideServiceImpl: ShenyuClientRegisterDivideServiceImpl.

  • register():
public String register(final MetaDataRegisterDTO dto) {        // 1.register selector information        String selectorHandler = selectorHandler(dto);        String selectorId = selectorService.registerDefault(dto, PluginNameAdapter.rpcTypeAdapter(rpcType()), selectorHandler);        // 2.register rule information        String ruleHandler = ruleHandler();        RuleDTO ruleDTO = buildRpcDefaultRuleDTO(selectorId, dto, ruleHandler);        ruleService.registerDefault(ruleDTO);        // 3.register metadata information        registerMetadata(dto);        // 4.register contextPath        String contextPath = dto.getContextPath();        if (StringUtils.isNotEmpty(contextPath)) {            registerContextPath(dto);        }        return ShenyuResultMessage.SUCCESS;    }

The whole registration logic can be divided into 4 steps.

    1. Register selector information
    1. Register rule information
    1. Register metadata information
    1. Register `contextPath

This side of admin requires the construction of selectors, rules, metadata and ContextPath through the metadata information of the client. The specific registration process and details of processing are related to the rpc type. We will not continue to track down the logical analysis of the registration center, tracking to this point is enough.

The source code of the server-side metadata registration process is analyzed and the flow chart is described as follows.

  • registerURI()
public String registerURI(final String selectorName, final List<URIRegisterDTO> uriList) {        if (CollectionUtils.isEmpty(uriList)) {            return "";        }        // Does the corresponding selector exist        SelectorDO selectorDO = selectorService.findByNameAndPluginName(selectorName, PluginNameAdapter.rpcTypeAdapter(rpcType()));        if (Objects.isNull(selectorDO)) {            return "";        }        // Handle handler information in the selector        String handler = buildHandle(uriList, selectorDO);        selectorDO.setHandle(handler);        SelectorData selectorData = selectorService.buildByName(selectorName, PluginNameAdapter.rpcTypeAdapter(rpcType()));        selectorData.setHandle(handler);               // Update records in the database        selectorService.updateSelective(selectorDO);        // publish Event to gateway        eventPublisher.publishEvent(new DataChangedEvent(ConfigGroupEnum.SELECTOR, DataEventTypeEnum.UPDATE, Collections.singletonList(selectorData)));        return ShenyuResultMessage.SUCCESS;    }

After admin gets the URI data, it mainly updates the handler information in the selector, then writes it to the database, and finally publishes the event notification gateway. The logic of notifying the gateway is done by the data synchronization operation, which has been analyzed in the previous article, so we will not repeat it.

The source code analysis of the server-side URI registration process is complete and is described in the following diagram.

At this point, the server-side registration process is also analyzed, mainly through the interface provided externally, accept the registration information from the client, and then write to the Disruptor queue, and then consume data from it, and update the admin selector, rules, metadata and selector handler according to the received metadata and URI data.

4. Summary#

This article focuses on the http registration module of the Apache ShenYu gateway for source code analysis. The main knowledge points involved are summarized as follows.

  • The register center is for registering client information to admin to facilitate traffic filtering.
  • http registration is to register client metadata information and URI information to admin.
  • http service access is identified by the annotation @ShenyuSpringMvcClient.
  • construction of the registration information mainly through the Spring post-processor BeanPostProcessor and the application listener ApplicationListener.
  • loading of the registration type is done through SPI.
  • The Disruptor queue was introduced to decouple data from operations, and data buffering.
  • The implementation of the registry uses interface-oriented programming, using design patterns such as template methods, singleton, and observer.

LoadBalance SPI Source Code Analysis

· 13 min read
Apache ShenYu Contributor

Gateway applications need to support a variety of load balancing strategies, including random,Hashing, RoundRobin and so on. In Apache Shenyu gateway, it not only realizes such traditional algorithms, but also makes smoother traffic processing for the entry of server nodes through detailed processing such as traffic warm-up, so as to obtain better overall stability. In this article, let's walk through how Apache Shenyu is designed and implemented this part of the function.

This article based on shenyu-2.4.0 version of the source code analysis.


LoadBalance SPI#

The implementation of LoadBalance is in shenyu-plugin-divide module. It has based on its SPI creation mechanism. The core interface code is shown as follows. This interface well explains the concept: load balancing is to select the most appropriate node from a series of server nodes. Routing, traffic processing and load balancing is the basic function of LoadBalance SPI.

@SPIpublic interface LoadBalance {
    /**     * @param upstreamList upstream list     * @param ip ip     * @return divide upstream     */    DivideUpstream select(List<DivideUpstream> upstreamList, String ip);}

Where upstreamList represents the server nodes list available for routing. DivideUpstream is the data structure of server node, the important elements including protocol, upstreamUrl , weight, timestamp, warmup.

public class DivideUpstream implements Serializable {    private String upstreamHost;    /**     * this is http protocol.     */    private String protocol;    private String upstreamUrl;    private int weight;    /**     * false close/ true open.     */    @Builder.Default    private boolean status = true;    private long timestamp;    private int warmup;}

Design of LoadBalance module#

The class diagram of LoadBalance moduleisshown as follows.


We can draw the outline of LoadBalance module from the class diagram:

  1. The abstract class AbstractLoadBalance implements the SPI LoadBalance interface,and supplies the template methods for selection related, such as select(), selector(),and gives the calculation of weight.

  2. Three implementation classes which inherit AbstractLoadBalance to realize their own logic:

    • RandomLoadBalance - Weight Random
    • HashLoadBalance - Consistent Hashing
    • RoundRobinLoadBalance -Weight Round Robin per-packet
  3. The utility class LoadBalanceUtil provides public static method to be called.

    The implementation classes and algorithms are configurable. According to its specification, by adding profile in SHENYU_DIERECTORY directory, the data in profile should be key=value-class format, where the value-class will be load by the Apache Shenyu SPI class loader, and key value should be an name defined in LoadBalanceEnum.


The code of LoadBalanceEnum is as follows:

public enum LoadBalanceEnum {    /**     * Hash load balance enum.     */    HASH(1, "hash", true),
    /**     * Random load balance enum.     */    RANDOM(2, "random", true),
    /**     * Round robin load balance enum.     */    ROUND_ROBIN(3, "roundRobin", true);
    private final int code;    private final String name;    private final boolean support;}


This abstract class implements the LoadBalance interface and define the abstract method doSelect() to be processed by the implementation classes. In the template method select(), It will do validation first then call the doSelect() method.

   /**      * Do select divide upstream.     *     * @param upstreamList the upstream list     * @param ip           the ip     * @return the divide upstream     */    protected abstract DivideUpstream doSelect(List<DivideUpstream> upstreamList, String ip);
    @Override    public DivideUpstream select(final List<DivideUpstream> upstreamList, final String ip) {        if (CollectionUtils.isEmpty(upstreamList)) {            return null;        }        if (upstreamList.size() == 1) {            return upstreamList.get(0);        }        return doSelect(upstreamList, ip);    }

When the timestamp of server node is not null, and the interval between current time and timestamp is within the traffic warm-up time, the formula for weight calculation is. $$ {1-1} ww = min(1,uptime/(warmup/weight)) $$ It can be seen from the formula that the final weight(ww) is proportional to the original-weight value. The closer the time interval is to the warmup time, the greater the final ww. That is, the longer the waiting time of the request, the higher the final weight. When there is no timestamp or other conditions, the ww is equal to the weight value of DivideUpstream object.

The central of thinking about warm-upis to avoid bad performance when adding new server and the new JVMs starting up.

Let's see how the load balancing with Random, Hashing and RoundRobin strategy is implemented.


The RandomLoadBalance can handle two situations:

  1. Each node without weight, or every node has the same weight, randomly choose one.
  2. Server Nodes with different weight, choose one randomly by weight.

Following is the random() method of RandomLoadBalance. When traversing server node list, if the randomly generated value is less than the weight of node, then the current node will be chosen. If after one round traversing, there's is no server node match, then it will return the first item of the list. The getWeight(DivideUpstream upstream) is defined in AbstractLoadBalance class.

    private DivideUpstream random(final int totalWeight, final List<DivideUpstream> upstreamList) {        // If the weights are not the same and the weights are greater than 0, then random by the total number of weights        int offset = RANDOM.nextInt(totalWeight);        // Determine which segment the random value falls on        for (DivideUpstream divideUpstream : upstreamList) {            offset -= getWeight(divideUpstream);            if (offset < 0) {                return divideUpstream;            }        }        return upstreamList.get(0);    }


In HashLoadBalance, it takes the advantages of consistent hashing , that maps both the input traffic and the servers to a unit circle, or name as hash ring. For the requestedip address, with its hash value to find the node closest in clockwise order as the node to be routed. Let's see how consistent hashing is implemented in HashLoadBalance.

As to the hash algorithms, HashLoadBalance uses MD5 hash, which has the advantage of mixing the input in an unpredictable but deterministic way. The output is a 32-bit integer. the code is shown as follows:

private static long hash(final String key) {    // md5 byte    MessageDigest md5;    try {        md5 = MessageDigest.getInstance("MD5");    } catch (NoSuchAlgorithmException e) {        throw new ShenyuException("MD5 not supported", e);    }    md5.reset();    byte[] keyBytes;    keyBytes = key.getBytes(StandardCharsets.UTF_8);    md5.update(keyBytes);    byte[] digest = md5.digest();    // hash code, Truncate to 32-bits    long hashCode = (long) (digest[3] & 0xFF) << 24            | ((long) (digest[2] & 0xFF) << 16)            | ((long) (digest[1] & 0xFF) << 8)            | (digest[0] & 0xFF);    return hashCode & 0xffffffffL;}

Importantly, how to generate the hash ring and avoid skewness? Let's thedoSelect() method inHashLoadBalance as follows:

    private static final int VIRTUAL_NODE_NUM = 5;
    @Override    public DivideUpstream doSelect(final List<DivideUpstream> upstreamList, final String ip) {        final ConcurrentSkipListMap<Long, DivideUpstream> treeMap = new ConcurrentSkipListMap<>();        for (DivideUpstream address : upstreamList) {            for (int i = 0; i < VIRTUAL_NODE_NUM; i++) {                long addressHash = hash("SOUL-" + address.getUpstreamUrl() + "-HASH-" + i);                treeMap.put(addressHash, address);            }        }        long hash = hash(String.valueOf(ip));        SortedMap<Long, DivideUpstream> lastRing = treeMap.tailMap(hash);        if (!lastRing.isEmpty()) {            return lastRing.get(lastRing.firstKey());        }        return treeMap.firstEntry().getValue();    }

In this method, duplicated labels are used which are called "virtual nodes" (i.e. 5 virtual nodes point to a single "real" server). It will make the distribution in hash ring more evenly, and reduce the occurrence of data skewness.

In order to rescue the data sorted in the hash ring, and can be accessed quickly, we use ConcurrentSkipListMap of Java to store the server node lists ( with virtual nodes) and its hash value as key. This class a member of Java Collections Framework, providing expected average log(n) time cost for retrieve and access operations safely execute concurrent by multiple threads.

Furthermore, the method tailMap(K fromKey) of ConcurrentSkipListMap can return a view of portion of the map whose keys are greater or equal to the fromKey, and not need to navigate the whole map.

In the above code section, after the hash ring is generated, it uses tailMap(K fromKey) of ConcurrentSkipListMap to find the subset that the elements greater, or equal to the hash value of the requested ip, its first element is just the node to be routed. With the suitable data structure, the code looks particularly clear and concise.

Consistent hashing resolved the poor scalability of the traditional hashing by modular operation.


The original Round-robin selection is to select server nodes one by one from the candidate list. Whenever some nodes has crash ( ex, cannot be connected after 1 minute), it will be removed from the candidate list, and do not attend the next round, until the server node is recovered and it will be add to the candidate list again. In RoundRobinLoadBalance,the weight Round Robin per-packet schema is implemented.

In order to work in concurrent system, it provides an inner static class WeigthRoundRobin to store and calculate the rolling selections of each server node. Following is the main section of this class( removed remark )

protected static class WeightedRoundRobin {
    private int weight;    private final AtomicLong current = new AtomicLong(0);
    private long lastUpdate;
    void setWeight(final int weight) {        this.weight = weight;        current.set(0);    }    long increaseCurrent() {        return current.addAndGet(weight);    }
    void sel(final int total) {        current.addAndGet(-1 * total);    }    void setLastUpdate(final long lastUpdate) {        this.lastUpdate = lastUpdate;    }}

Please focus on the these method:

  • setWeight(final int weight) : set the current value by weight

  • increaseCurrent(): Increment the current value by weight, and current set to 0.

  • sel(final int total): decrement the current value by total

    Let's see how the weight factor being used in this round-robin selection?

    First it defines a two-level ConcurrentMap type variable named as methodWeightMap , to cache the server node lists and the rolling selection data about each server node.

private final ConcurrentMap<String, ConcurrentMap<String, WeightedRoundRobin>> methodWeightMap = new ConcurrentHashMap<>(16);

In this map, the key of first level is set to upstreamUrl of first element in server node list. The type of second object is ConcurrentMap<String, WeightedRoundRobin>, the key of this inner Map is the value upstreamUrlvariable of each server node in this server list, the value object is WeightedRoundRobin, used to trace the rolling selection data about each server node. As to the implementation class for the Map object, we use ConcurrentHashMap of JUC, a hash table supporting full concurrency of retrievals and high expected concurrency for updates.

In the second level of the map, the embedded static class - WeighedRoundRobin of each node is thread-safe, implementing the weighted RoundRobin per bucket. The following is the code of the doselect() method of this class.

@Overridepublic DivideUpstream doSelect(final List<DivideUpstream> upstreamList, final String ip) {    String key = upstreamList.get(0).getUpstreamUrl();    ConcurrentMap<String, WeightedRoundRobin> map = methodWeightMap.get(key);    if (map == null) {        methodWeightMap.putIfAbsent(key, new ConcurrentHashMap<>(16));        map = methodWeightMap.get(key);    }    int totalWeight = 0;    long maxCurrent = Long.MIN_VALUE;    long now = System.currentTimeMillis();    DivideUpstream selectedInvoker = null;    WeightedRoundRobin selectedWRR = null;    for (DivideUpstream upstream : upstreamList) {        String rKey = upstream.getUpstreamUrl();        WeightedRoundRobin weightedRoundRobin = map.get(rKey);        int weight = getWeight(upstream);        if (weightedRoundRobin == null) {            weightedRoundRobin = new WeightedRoundRobin();            weightedRoundRobin.setWeight(weight);            map.putIfAbsent(rKey, weightedRoundRobin);        }        if (weight != weightedRoundRobin.getWeight()) {            //weight changed            weightedRoundRobin.setWeight(weight);        }        long cur = weightedRoundRobin.increaseCurrent();        weightedRoundRobin.setLastUpdate(now);        if (cur > maxCurrent) {            maxCurrent = cur;            selectedInvoker = upstream;            selectedWRR = weightedRoundRobin;        }        totalWeight += weight;    }    ......  //erase the section which handles the time-out upstreams.     if (selectedInvoker != null) {        selectedWRR.sel(totalWeight);        return selectedInvoker;    }    // should not happen here    return upstreamList.get(0);}

For example we assume upstreamUrl values of three server nodes is: LIST = [upstream-20, upstream-50, upstream-30]. After a round of execution, the data in newly created methodWeightMap is as follows:


For the above example LIST, assumes the weight array is [20,50,30]. the following figure shows the value change and polling selection process of the current array in WeighedRoundRobin object.


In each round, it will choose the server node with max current value.

  • Round1:
    • Traverse the server node list, initialize the weightedRoundRobin instance of each server node or update the weight value of server nodes object DivideUpstream
    • say, in this case, after traverse, the current array of the node list changes to [20, 50,30],so according to rule, the node Stream-50 would be chosen, and then the static object WeightedRoundRobin of Stream-50 executes sel(-total) , the current array is now [20,-50, 30].
  • Round 2: after traverse, the current array should be [40,0,60], so the Stream-30 node would be chosen, current array is now [40,0,-40].
  • Round 3: after traverse, current array changes to [60,50,-10], Stream-20 would be chosen,and current array is now [-40,50,-10].

When there is any inconsistence or some server crashed, for example, the lists size does not match with the elements in map, it would copy and modify the element with lock mechanism, and remove the timeout server node, the data in Map updated. Following is the fault tolerance code segment.

    if (!updateLock.get() && upstreamList.size() != map.size() && updateLock.compareAndSet(false, true)) {        try {            // copy -> modify -> update reference            ConcurrentMap<String, WeightedRoundRobin> newMap = new ConcurrentHashMap<>(map);            newMap.entrySet().removeIf(item -> now - item.getValue().getLastUpdate() > recyclePeriod);            methodWeightMap.put(key, newMap);        } finally {            updateLock.set(false);        }    }
    if (selectedInvoker != null) {        selectedWRR.sel(totalWeight);        return selectedInvoker;    }    // should not happen here    return upstreamList.get(0);


In this class, a static method calling LoadBalance is provided, whereExtensionLoader is the entry point of Apache Shenyu SPI. That is to say, LoadBalance module is configurable and extensible. The algorithm variable in this static method is the name enumeration type defined in LoadBalanceEnum.

    /**     * Selector divide upstream.     *     * @param upstreamList the upstream list     * @param algorithm    the loadBalance algorithm     * @param ip           the ip     * @return the divide upstream     */    public static DivideUpstream selector(final List<DivideUpstream> upstreamList, final String algorithm, final String ip) {        LoadBalance loadBalance = ExtensionLoader.getExtensionLoader(LoadBalance.class).getJoin(algorithm);        return, ip);    }

Using of LoadBalance module#

In the above section, we describe the LoadBalance SPI and three implementation classes. Let's take a look at how the LoadBalance to be used in Apache Shenyu. DividePlugin is a plugin in Apache Shenyu responsible for routing http request. when enable to use this plugin, it will transfer traffic according to selection data and rule data, and deliver to next plugin downstream.

@SneakyThrows@Overrideprotected Mono<Void> doExecute(final ServerWebExchange exchange, final ShenyuPluginChain chain, final SelectorData selector, final RuleData rule) {   ......}

The type of second parameter of doExecute() is ShenyuPluginChain, which represents the execution chain of plugins. For details, see the mechanism of Apache Shenyu Plugins. The third one is SelectorData type, and the fourth is RuleData type working as the rule data.

In doExecute() of DividePlugin, first verify the size of header, content length, etc, then preparing for load balancing.

Following is a code fragment usingLoadBalance in the doExecute() method:

   // find the routing server node list   List<DivideUpstream> upstreamList = UpstreamCacheManager.getInstance().findUpstreamListBySelectorId(selector.getId());    ...     // the requested ip    String ip =   Objects.requireNonNull(exchange.getRequest().getRemoteAddress()).getAddress().getHostAddress();
    //calling the Utility class and invoke the LoadBalance processing.    DivideUpstream divideUpstream = LoadBalanceUtils.selector(upstreamList, ruleHandle.getLoadBalance(), ip);

In the above code, the output ofruleHandle.getLoadBalance() is the name variable defined in LoadBalanceEnum, that is random, hash, roundRobin, etc. It is very convenient to use LoadBalance by LoadBalanceUtils. When adding more LoadBalance implementing classes, the interface in plugin module will not be affect at all.


After reading through the code of LoadBalance module, from the design perspective, it is concluded that this module has the following characteristics:

  1. Extensibility: Interface oriented design and implemented on Apache Shenyu SPI mechanism, it can be easily extended to other dynamic load balancing algorithms (for example, least connection, fastest mode, etc), and supports cluster processing.
  2. Scalability: Every load balancing implementation, weighted Random, consistency Hashing and weighted RoundRobin can well support increase or decrease cluster overall capacity.
  3. More detailed design such as warm-up can bring better performance and obtain better overall stability.

MatchStrategy -- analyze the design based on SPI

· 5 min read
Apache ShenYu Contributor

In most of the plugins ( such as Dubbo, gRPC,Spring-cloud, etc) of Apache Shenyu, the routingparameters are designed to support the combination of multiple conditions. In order to realize such requirements, the parameters and behaviors are abstracted to three parts according to its SPI mechanism, and implemented in shenyu-plugin-base module.

  • ParameterData-parameters
  • PredictJudge-predicate
  • MatchStrategy-matching strategy

Relatively speaking, the MatchStrategy is the part that needs the least extension points. For the combination judgement of multiple conditions, the common select rules are: All conditions are matched, at least one is match, at least the first is met, or most of conditions satisfied. As well need to handle various type of parameters, for example: IP, header, uri, etc. How to make the MatchStrategy to be simple to use and extensible?


The implementation of MatchStrategy is in shenyu-plugin-base module. It has based on the SPI creation mechanism, and used factory pattern and strategy design pattern. The class diagram of MatchStrategy is showed as follows.


Base on the interface MatchStrategy to design the implementation classes, and the abstract class AbstractMatchStrategy supplies common method, while the factory class MatchStrategyFactory provides creation functions.

MatchStrategy Interface#

First, let's look at the MatchStrategy SPI interface

@SPIpublic interface MatchStrategy {
    Boolean match(List<ConditionData> conditionDataList, ServerWebExchange exchange);}

The annotation @SPI means that this is an SPI interface. Where ServerWebExchange is org.springframework.web.server.ServerWebExchange, represents the request-response interactive content of of HTTP. Following is the code of ConditionData, the more detail about this class can refer to code analysis of PredicteJudge

public class ConditionData {
    private String paramType;    private String operator;
    private String paramName;    private String paramValue;}


Second, let's look at the abstract class AbstractMatchStrategy,it has defined a buildRealData method,in this method wrapped various parameters to a unified interface through the functionality of ParameterDataFactory, which is the factory class of ParameterData. There supports a variety types of parameters , such as Ip, Cookie, Header,uri, etc. Modifications of such parameters will not impact the calling of matching rules of MatchStrategy.

public abstract class AbstractMatchStrategy {
    public String buildRealData(final ConditionData condition, final ServerWebExchange exchange) {        return ParameterDataFactory.builderData(condition.getParamType(), condition.getParamName(), exchange);    }}

Implementation class and profile#

Then, let's look at the two implementation class based on the above interface in shenyu-plugin-base module , that is:

  • AndMatchStrategy- AND -All conditions are matched

  • OrMatchStrategy- OR -at least one is match

    The profile is shown as follows, which locates at the SHENYU_DIRECTORYdirectory. When starting up, the top-level SPI classes will read the key-value and load the classes and cache them.


These two implementation classes inherit AbstractMatchStrategy and implement MatchStrategy.

AndMatchStrategy- “AND” relation#

Because PredicateJudge encapsulates the diversity of Predicate , and ConditionData and ParameData can present variety of parameters, for treating of multiple conditions, usingstream and lambda expression, It can be very simple and efficient to process "AND" logic- all conditions must be matched.

@Joinpublic class AndMatchStrategy extends AbstractMatchStrategy implements MatchStrategy {
    @Override    public Boolean match(final List<ConditionData> conditionDataList, final ServerWebExchange exchange) {        return conditionDataList                .stream()                .allMatch(condition -> PredicateJudgeFactory.judge(condition, buildRealData(condition, exchange)));    }}

The OrMatchStrategy similarly implements the "OR" logic- at least one is match.


This is the factory class of MatchStrategy,there are two methods, one is newInstance(), which will return the MatchStrategy implementation class instance cached by the SPI ExtensionLoader indexed by the key-value.

    public static MatchStrategy newInstance(final Integer strategy) {        String matchMode = MatchModeEnum.getMatchModeByCode(strategy);        return ExtensionLoader.getExtensionLoader(MatchStrategy.class).getJoin(matchMode);    }

the matchMode will be the name of strategy, the value will be "and" or "or". The MatchModeEnum defines the code and name of match strategy as follows.

AND(0, "and"), OR(1, "or");

Another method is match() method, which will invoke the match() method of implementation class.

    public static boolean match(final Integer strategy, final List<ConditionData> conditionDataList, final ServerWebExchange exchange) {        return newInstance(strategy).match(conditionDataList, exchange);    }

How it works#

AbstractShenyuPlugin is the base class of plugins in shenyu-plugin module. In this class two selection method are defined: filterSelector() and filterRule() , Both of them call the match() method of MatchStrategyFactory. The code of filterSelector() is shown as follows.

    private Boolean filterSelector(final SelectorData selector, final ServerWebExchange exchange) {        if (selector.getType() == SelectorTypeEnum.CUSTOM_FLOW.getCode()) {            if (CollectionUtils.isEmpty(selector.getConditionList())) {                return false;            }            return MatchStrategyFactory.match(selector.getMatchMode(), selector.getConditionList(), exchange);        }        return true;    }

In filterSelector() method, after validation of the SelectorData, calls the match method of MatchStrategyFactory, and then this factory class will invokes the match method of corresponding implementation class.

    private Boolean filterRule(final RuleData ruleData, final ServerWebExchange exchange) {        return ruleData.getEnabled() && MatchStrategyFactory.match(ruleData.getMatchMode(), ruleData.getConditionDataList(), exchange);    }

In filterRule() it is also calls the match() method of MatchStrategyFactory. Does it look particularly concise or even simple? In the code analysis of PredicteJudge , you can see more detail about parameter processing in shenyu-plugin.


Due to the use of SPI mechanism of Apache Shen, the parameter selection module has the characteristic of loose coupling and extensibility. In terms of the combination of multiple conditions, MatchStrategy provides a good design. Although currently only two implementation classes are present, it can be easily develop more complex MatchStrategy rules in the future, such as "firstOf"-first condition must matched, or "mostOf"- most of the conditions must be matched, etc.

Interested readers can read the source code of 'shenyu-plugin' to learn more.

PredicateJudge -- analyze the design based on SPI

· 6 min read
Apache ShenYu Contributor

Apache Shenyu has been identified as a gateway application which supports a variety of protocols and microservice frameworks such as Dubbo, gRPC, Spring-Cloud, etc. To do this, the product has accomplished an elegant SPI (Service Provider Interface) as its foundation, and make the Rule data parsing and predicting program very simple , resiliency and security. As to rule data parsing processing, the SPI design increases the product's scalability. When appending new plugin, in most cases, the existing module is enough for rule data parsing , otherwise it can be rapidly carry out with tiny effort.

Top level design of SPI#

In Apache Shenyu, the SPI archtecure is defined in shenyu-spi module and composed of three parts: SPI interface, factory design pattern, and configuration file. There is two interface defined as annotation: @SPI and @Join. When class file with @Join annotation, it means that it will join as an SPI extension class, in other words, it is an application or registration. The @SPI denotes that the class is an SPI extension class.

Fig 1 classes in the shenyu-spi


The SPI configuration directory is META-INF/shenyu/. that is specified:


When starting the gateway system , the ExtensionLoader will scan the profiles under SHENYU_DIRECTORY, in turn, load and validate and then initialize each configed class. The configuration file uses "Key = class-file" format. During operation of the system, the corresponding SPI implementation class will be invoked through the factory mechanism.

Implementation of shenyu-plugin SPI#

In shenyu-plugin module, various plugins for HTTP routing are implemented according to the plugin mechanism, including request, redirect, response and rewrite, etc. Plugins for microservice frameworks such as Dubbo, gRPC , Spring-Cloud and Tars have been developed in the gateway product. And plugins are still increasing. If no such dependent module fo parsing and judge routing parameters and data, each plugin is necessary to implement the parsing functions, and has to frequently modify to support their matching rules, such as wildcard, regular expression, SpEL expression, etc. Therefore, they made a high level abstraction for routing parameter data following the SPI framework in shenyu-plugin module. The rule analysis consists of three parts:

  • ParameterData- parameter data

  • PredicatJudge- predicate whether the actural data match the rule

  • MatchStrategy- combine multiple conditions, the final used strategy

These implementation classes are defined in shenyu-plugin-base module. In each plugin, resolution and predication of the routing parameter can be realized through AbstractShenyuPlugin using the above SPIs. That is dedicated and easy to extend, in line with SOLID principle.

​ This section analyzes the PredictJudge in detail. You can find the dependency to shenyu-spi in the pom.xml of this module.

<dependency>    <groupId>org.apache.shenyu</groupId>    <artifactId>shenyu-spi</artifactId>    <version>${project.version}</version></dependency>

Design of PredicateJudge SPI#

PredicateJudge SPI is used to analyze and judge various routing rules configed in Apache Shenyu gateway. The name and functions of this SPI are similar to Predicate in Java, but the acceptance behavior is further abstracted applying for routing aspect. This SPI is implemented through the Factory pattern. Let's look at the Predictejudge SPI interface:

@SPI@FunctionalInterfacepublic interface PredicateJudge {
    /**     * judge conditionData and realData is match.     *     * @param conditionData {@linkplain ConditionData}     * @param realData       realData     * @return true is pass  false is not pass.     */    Boolean judge(ConditionData conditionData, String realData);}

The class diagram is as follows:

Fig 2-Predicate class diagram


The important methods of PredicateJudgeFactory are shown as follows:

Whenever need to parsing and matching routing data, you can use

    public static PredicateJudge newInstance(final String operator) {        return ExtensionLoader.getExtensionLoader(PredicateJudge.class).getJoin(processSpecialOperator(operator));    }
    public static Boolean judge(final ConditionData conditionData, final String realData) {        if (Objects.isNull(conditionData) || StringUtils.isBlank(realData)) {            return false;        }        return newInstance(conditionData.getOperator()).judge(conditionData, realData);    }

ConditionData contains of four attributes of String type: paramType, operator,paramName,paramValue


Where paramType must be the enumeration type ParamTypeEnum. The default supported paramType are:

post, uri,query, host, ip,header, cookie,req_method


operator must be the enumeration type OperatorEnum, currently supported operators are:

   match, =,regex, >,<, contains, SpEL,  Groovy, TimeBefore,TimeAfter

Base on the above defination , the plugin module provides the following eight PredicateJudge implemetation classes to realize the logic of these operators respectively.

Implementation classLogic descriptioncorespondece operator
ContainsPredicateJudge"contain" relation, the actual data needs contain the specified stringcontains
EqualsPredicateJudgeequals "="=
MatchPredicateJudgeused for URI context path matchingmatch
TimerAfterPredicateJudgeWhether the local time is after the specified timeTimeBefore
TimerBeforePredicateJudgeWhether the local time is before the specified timeTimeAfter
GroovyPredicateJudgeused Groovy syntax toe set ParamName and value dataGroovy
RegexPredicateJudgeused Regex to matchregex

How to use PredicateJudge#

When you want to parse parameters, you only need to call PredicateJudgeFactory as follows.

PredicateJudgeFactory.judge(final ConditionData conditionData, final String realData);

SPI profile#

The implementation class is configed in the file under directory SHENYU_DIRECTORY . It will be loaded and cached at startup.


The usage of PredicateJudge SPI in Shenyu gateway#

Most plugins in Apache Shenyu are inherited from AbstractShenyuPlugin. In this abstract class, the filter functions (selection and matching) are achieved through MatchStrategy SPI, and PredicateJudge will be invoked from MatchStrategy to predicate each condition data.

Fig 3- class diagram of plugins with PredicateJudge and MatchStrategy SPI


The process from client request calling the routing parsing moodule is showed as following chart.

Fig 4- flow chart for Shenyu gateway filter with parameter processing


  • When startup, the system will load SPI classes from profile and cache them.
  • When the client sends a new request to the Shenyu gateway, will call the corresponding plugin within the gateway.
  • When analyzing real data with routing rules, the PredicateJudge implementation class will be invoked according to the contained operator.


Examples of PredicateJudge judgement#

ContainsPredicateJudge- " contains“ rule#

For example, giving a ConditionData with: paramType="uri", paramValue 是 "/http/**", when using the "contains" relation: ContainsPredicateJudge, the matching result is as follows.

ConditionData (operator="contains")real datajudge result
paramType="uri", "/http/**""/http/**/test"true

About other PredicateJudge implemetantion classes, you can refer to the code and test classes.

RateLimiter SPI code analysis

· 16 min read
Apache ShenYu Contributor

Rate limiter is a very important integral of gateway application, to deal with high traffic. When the system is attacked abnormally by a large number of traffic gathered in a short time; When there are a large number of lower priority request need to be slow down or else it will effect your high priority transactions; Or sometimes your system can not afford the regular traffic; in these scenarios, we need to start rate limiter component to protect our system, through rejection, wait, load shedding,etc, limit the requests to an acceptable quantities, or only certain domains (or services) requests can get through.

Facing above scenarios, following need to be considered when designing the rate limiter component of an gateway.

  1. Supports a variety of rate limiter algorithms and easy to extends.
  2. Resilient resolvers which can distinguish traffic by different way, such as ip, url, even user group etc.
  3. High availability, can quickly get allow or reject result from rate limiter
  4. With fault tolerance against when rate limiter is down, the gateway can continue work.

This article will first introduce the overall architecture of the rate limiter module in Apache Shenyu, and then focus on the code analysis of rate limiter SPI.

This article based on shenyu-2.4.0 version of the source code analysis.

Overall design of RateLimiter#

Spring WebFlux is reactive and non-blocking web framework, which can benefit throughput and make applications more resilient. The plugin of Apache Shenyu is based on WebFlux,its rate limiter component is implemented in ratelimiter-plugin. In rate limiter process, the commonly used algorithms are token bucket, leaky bucket, etc. To speed up concurrency performance, the counting and calculation logic is treated in Redis, and Java code is responsible for the transmission of parameters. When applying Redis, the Lua script can be resident memory, and be executed as a whole, so it is atomic. Let alone the reducing of network overhead. Redis commands abstraction and automatic serialization/deserialization with Redis store is provided in Spring Data Redis. Because of based on reactive framework, the Spring Redis Reactive is used in ratelimiter-plugin.

The class diagram of this plugin is as follows, highlighting two packages related to RateLimiter SPI: resolver 和algorithm.


Design of RateLimiter SPI#

High performance issue is achieved through the architecture of Spring data+ Redis+Lua , two SPI are supplied in ratelimiter-plugin for the extension of algorithm and key resolver。

  • RateLimiterAlgorithm:used for algorithms expansion.
  • RateLimiterKeyResolver: used for resolver expansion, to distinguish requests by various information, including ip, url, ect.

The profile of SPI is located at directory of SHENYU_DIRECTORY (default/META-INF/shenyu).


Obtain the critical info of the request used for packet rate limiter,the interface of RateLimiterKeyResolver is follows:

@SPIpublic interface RateLimiterKeyResolver {
    /**     * get Key resolver's name.     *     * @return Key resolver's name     */    String getKeyResolverName();
    /**     * resolve.     *     * @param exchange exchange the current server exchange {@linkplain ServerWebExchange}     * @return rate limiter key     */    String resolve(ServerWebExchange exchange);}

@SPI registers the current interface as Apache Shenyu SPI. Method resolve(ServerWebExchange exchange) is used to provide the resolution way. Currently there are two key resolvers in RateLimiterKeyResolver SPI:WholeKeyResolve and RemoteAddrKeyResolver. The resolve method of RemoteAddrKeyResolveris as follows:

    @Override    public String resolve(final ServerWebExchange exchange) {        return Objects.requireNonNull(exchange.getRequest().getRemoteAddress()).getAddress().getHostAddress();    }

Where the resolved key is ip of request. Based on SPI mechanism and its factory pattern, new resolver can be easily developed.

RateLimiterAlgorithm SPI#

RateLimiterAlgorithm SPI is used to identify and define different rate limiter algorithms, following is the class diagram of this module.


In this module, factory pattern is used , providing interface, abstract class and factory class, and four implementation classes. The lua script corresponding to the implementation class is enumerated in RateLimitEnum and located in /META-INF/scripts.

@SPIpublic interface RateLimiterAlgorithm<T> {        RedisScript<T> getScript();    List<String> getKeys(String id);        /**     * Callback string.     *     * @param script the script     * @param keys the keys     * @param scriptArgs the script args     */    default void callback(final RedisScript<?> script, final List<String> keys, final List<String> scriptArgs) {    }}

@SPI registers the current interface as Apache Shenyu SPI. There are three methods:

  • getScript() returns a RedisScript object, which will be passed to Redis.
  • getKeys(String id) returns a List contains with keys.
  • callback() the callback function which will be executed asynchronously later on, and default is an empty method.


The template method is implemented in this abstract class, and the reified generics used is List<Long>. Two abstract methods getScriptName() and getKeyName() are left for the implementation class. Following is the code to load lua script.

    public RedisScript<List<Long>> getScript() {        if (!this.initialized.get()) {            DefaultRedisScript redisScript = new DefaultRedisScript<>();            String scriptPath = "/META-INF/scripts/" + getScriptName();            redisScript.setScriptSource(new ResourceScriptSource(new ClassPathResource(scriptPath)));            redisScript.setResultType(List.class);            this.script = redisScript;            initialized.compareAndSet(false, true);            return redisScript;        }        return script;    }

initialized is an AtomicBoolean type variable used to indicate whether the lua script is loaded. If has not been loaded, the system will read specified scripts form META-INF/scripts; After reading, specify the result with List type, and set initialized=true, then returning RedisScriptobject.

The code of getKeys() in AbstractRateLimiterAlgorithm is as follows:

    @Override    public List<String> getKeys(final String id) {        String prefix = getKeyName() + ".{" + id;        String tokenKey = prefix + "}.tokens";        String timestampKey = prefix + "}.timestamp";        return Arrays.asList(tokenKey, timestampKey);    }

Two strings are generated in this template method, where the tokenKey will work as Key to a Sorted map in Redis.

We can observe from above class diagram that ConcurrentRateLimiterAlgorithm and SlidingWindowRateLimiterAlgorithm override getKeys(String id) method but another two implementation classes not, and use template method in AbstractRateLimiterAlgorithm. Only in ConcurrentRateLimiterAlgorithm has override callback() method, the others not. We will do further analysis in the following.


The method getsRateLimiterAlgorithm instance by name in RateLimiterAlgorithmFactory is as follows:

public static RateLimiterAlgorithm<?> newInstance(final String name) {    return Optional.ofNullable(ExtensionLoader.getExtensionLoader(RateLimiterAlgorithm.class).getJoin(name)).orElse(new TokenBucketRateLimiterAlgorithm());}

ExtensionLoader of SPI is responsible for loading SPI classes by "name", if cannot find the specified algorithm class, it will return TokenBucketRateLimiterAlgorithm by default.

Data access with Redis#

Above detailed the extension interface in RateLimiter SPI. In Apache Shenyu, we use ReactiveRedisTemplate to perform Redis processing reactively, which is implemented inisAllowed() method of RedisRateLimiter class.

    public Mono<RateLimiterResponse> isAllowed(final String id, final RateLimiterHandle limiterHandle) {        // get parameters that will pass to redis from RateLimiterHandle Object        double replenishRate = limiterHandle.getReplenishRate();        double burstCapacity = limiterHandle.getBurstCapacity();        double requestCount = limiterHandle.getRequestCount();        // get the current used RateLimiterAlgorithm        RateLimiterAlgorithm<?> rateLimiterAlgorithm = RateLimiterAlgorithmFactory.newInstance(limiterHandle.getAlgorithmName());                ........        Flux<List<Long>> resultFlux = Singleton.INST.get(ReactiveRedisTemplate.class).execute(script, keys, scriptArgs);        return resultFlux.onErrorResume(throwable -> Flux.just(Arrays.asList(1L, -1L)))                .reduce(new ArrayList<Long>(), (longs, l) -> {                    longs.addAll(l);                    return longs;                }).map(results -> {                    boolean allowed = results.get(0) == 1L;                    Long tokensLeft = results.get(1);                    return new RateLimiterResponse(allowed, tokensLeft);                })                .doOnError(throwable -> log.error("Error occurred while judging if user is allowed by RedisRateLimiter:{}", throwable.getMessage()))                .doFinally(signalType -> rateLimiterAlgorithm.callback(script, keys, scriptArgs));    }

The POJO class RateLimiterHandle wraps the parameters needed in rate limiter, they are algorithName, replenishRate, burstCapacity, requestCount, etc. First, gets the parameters that need to be passed into Redis from RateLimiterHandle class. Then obtain the current implementation class from RateLimiterAlgorithmFactory.

For convenience, we give an flow image to show the parameters I/O and execution procedure in Java and Redis respectively. On the left is the second half of isAllowed() , and on the right is the processing of Lua script.

Following is the execution process of the JAVA code.

  1. Get two keys value in List<String> type from the getKeys() method, the first element will map to a sorted set in Redis.

  2. Set four parameters, replenishRate , burstCapacity, timestamp (EpochSecond) and requestcount.

  3. Calling ReactiveRedisTemplate with the scripts, keys and parameters, the return a Flux<List<Long>>

  4. The return value is converted from Flux<ArrayList<Long>> to Mono<ArrayList<Long>> the through reduce() of Flux ,and then transform it to Mono<RateLimiterResponse> via map() function. Returned two data, one is allowed (1-allow, 0- not allowed), the other is tokensLeft, the number of available remaining request.

  5. As for the fault tolerance, due to using of reactor non-blocking model, when an error occurs, the fallback function onErrorResume() will be executed and a new stream (1L, -1L) will generated by Flux.just, which means allow the request getting through, and log the error on the side.

  6. After that, performs the doFinally() method, that is to execute the callback() method of the implementation class.


Four rate limiter algorithms#

From above we know that how the java code works with Redis in the gateway. In this chapter we briefly analysis some code of the four rate limiter algorithms, to understand how to develop the interface of RateLimiter SPI and work efficiently with Redis.

Four rate limiter algorithms are supplied in Apache Shenyu Ratelimit SPI:

Algorithm nameJava classLua script file
Request rate limiterTokenBucketRateLimiterAlgorithmrequest_rate_limiter.lua
Slide window rate limiterSlidingWindowRateLimiterAlgorithmliding_window_request_rate_limiter.lua
Concurrent rate limiterConcurrentRateLimiterAlgorithmconcurrent_request_rate_limiter.lua
Leaky bucket algorithmLeakyBucketRateLimiterAlgorithmrequest_leaky_rate_limiter.lua
  1. Token bucket rate limiter: Limiting the traffic according to the number of requests. Assuming that N requests can be passed per second, when requests exceeding N will be rejected. In implementing of the algorithm, the requests will be grouped by bucket, the tokens will be generated at an evenly rate. If the number of requests is less than the tokens in the bucket, then it is allowed to pass. The time window is 2* capacity/rate.
  2. Slide window rate limiter: Different from token bucket algorithm, its window size is smaller than that of token bucket rate limiter, which is a capacity/rate. And move backward one time window at a time. Other rate limiter principles are similar to token bucket.
  3. Concurrent rate limiter: Strictly limit the concurrent requests to N. Each time when there is a new request, it will check whether the number of concurrent requests is greater than N. If it is less than N, it is allowed to pass through, and the count is increased by 1. When the requests call ends, the signal is released (count minus 1).
  4. Leaky bucket rate limiter: In contrast with token bucket algorithm, the leaky bucket algorithm can help to smooths the burst of requests and only allows a pre-defined N number of requests. This limiter can force the output flow at a constant rate of N. It is based on a leaky bucket model, the leaky water quantity is time interval*rate. if the leaky water quantity is greater than the number of has used (represented by key_bucket_count), then clear the bucket, that is, set the key_bucket_count to 0. Otherwise, set key_bucket_count minus the leaky water quantity. If the number (requests + key_bucket_count ) is less than the capacity, then allow the requests passing through.

Let's understand the functionality of callback() by reading concurrent rate limiter code, and understand the usage of getKeys() through reading the Lua script of token rate limiter and slide window rate limiter.

callback() used in Concurrent requests limiter#

The getKeys() method of ConcurrentRateLimiterAlgorithm overrides the template method in AbstractRateLimiterAlgorithm

    @Override    public List<String> getKeys(final String id) {        String tokenKey = getKeyName() + ".{" + id + "}.tokens";        String requestKey = UUIDUtils.getInstance().generateShortUuid();        return Arrays.asList(tokenKey, requestKey);    }

The second element, requestKey is a long type and non-duplicate value (generated by a distributed ID generator,it is incremented and smaller than the current time Epochsecond value). The corresponding Lua script in concurrent_request_rate_limiter.lua:

local key = KEYS[1]
local capacity = tonumber(ARGV[2])local timestamp = tonumber(ARGV[3])local id = KEYS[2]

Here id is requestKey generated by getKeys() method, it is an uuid(unique value). Subsequent process is as follows:

local count ="zcard", key)local allowed = 0
if count < capacity then"zadd", key, timestamp, id)  allowed = 1  count = count + 1endreturn { allowed, count }

First, using zcard command to obtain the cardinality of the sorted set, and set count equals the cardinality , if the cardinality is less than the capacity, we will add a new member id (it is an uuid) to the sorted set, with the score of current time(in seconds) . then count =count+1, the cardinality is also incremented by 1 in reality.

All of the code above is executed in Redis as an atomic transaction. If there are a large number of concurrent requests from the same key( such as ip) , the cardinality of the sorted set of this key will increasing sharply, when then capacity limit is exceeded, the service will be denied, that is allowed =0

In concurrent requests limiter, It is required to release the semaphore when the request is completed. However, it is not included in Lua script.

Let's see the callback function of ConcurrentRateLimiterAlgorithm

    @Override    @SuppressWarnings("unchecked")    public void callback(final RedisScript<?> script, final List<String> keys, final List<String> scriptArgs) {        Singleton.INST.get(ReactiveRedisTemplate.class).opsForZSet().remove(keys.get(0), keys.get(1)).subscribe();    }

Here gives asynchronous subscription, using ReactiveRedisTemplate to delete the elements (key,id) in Redis store. That is once the request operation ends, the semaphore will be released. This remove operation cannot be executed in Lua script. This is just what design intention of callback in RateLimiterAlgorithm SPI .

getKeys() used in token bucket rate limiter#

Following is the corresponding Lua code:

local tokens_key = KEYS[1]local timestamp_key = KEYS[2]

Here we omit the code that get the parameters of rate ,capacity, etc.

local fill_time = capacity/ratelocal ttl = math.floor(fill_time*2)

The window size variable(ttl) is approximately two times of capacity/rate.

local last_tokens = tonumber("get", tokens_key))if last_tokens == nil then  last_tokens = capacityend

Get last_tokens from the sorted set, if it not exist, then last_tokens equals capacity.

local last_refreshed = tonumber("get", timestamp_key))if last_refreshed == nil then  last_refreshed = 0end

Get the last refreshed time by the key =timestamp_key from the sorted set, and default 0.

local delta = math.max(0, now-last_refreshed)local filled_tokens = math.min(capacity, last_tokens+(delta*rate))local allowed = filled_tokens >= requestedlocal allowed_num = 0if allowed then  new_tokens = filled_tokens - requested  allowed_num = 1end

The filled_tokens is produced evenly by time interval * rate,if the number of tokens greater than requests, then allowed=1, and update new_tokens."setex", tokens_key, ttl, new_tokens)"setex", timestamp_key, ttl, now)
return { allowed_num, new_tokens }

Here now is current time parameters passed in, set tokens_key to hold the string new_tokens and settokens_key to timeout after ttl of seconds. Set timestamp_key to hold the string value now, and expires after ttl seconds.

getKeys() used in sliding window rate limiter#

The getKeys() in SlidingWindowRateLimiterAlgorithm also overrides the parent class, and the code is consistent with the method in ConcurrentRateLimiterAlgorithm

Following is the Lua code of slide window rate limiter, the receiving of other parameters is omitted.

local timestamp_key = KEYS[2]...... local window_size = tonumber(capacity / rate)local window_time = 1

Here set the window_size to capacity/rate.

local last_requested = 0local exists_key ='exists', tokens_key)if (exists_key == 1) then    last_requested ='zcard', tokens_key)end

Obtain the cardinality(last_requested) of the tokens_key in the sorted set.

local remain_request = capacity - last_requestedlocal allowed_num = 0if (last_requested < capacity) then    allowed_num = 1'zadd', tokens_key, now, timestamp_key)end

Calculate remaining available remain_request equals capacity minus last_requested . If last_requested less than capacity ,then allow current requests passing through,add element in the sorted set with (key=timestamp_key, value=now) .'zremrangebyscore', tokens_key, 0, now - window_size / window_time)'expire', tokens_key, window_size)
return { allowed_num, remain_request }

Previously has set window_time=1, using zremrangebyscore command of Redis to remove all the elements in the sorted set stored at tokens_key with a score in [0,now - window_size / window_time] , that is, move the window a window size. Set the expire time of tokens_key to window_size.

In the template method getKeys(final String id) of AbstractRateLimiterAlgorithm,the second key ( represented y secondKey) is a fixed string which concat the input parameter{id}. As we can see from the above three algorithm codes, in the token bucket algorithm, secondKey will be updated to the latest time in the Lua code, so it doesn't matter what value is passed in. In the concurrent rate limiter, secondKey will be used as the key to remove Redis data in the java callback method. In the sliding window algorithm, the secondKey will be added to the sorted set as the key of a new element, and will be removed during window sliding.

That's all, when in a new rate limiter algorithm, the getKeys(final String id)method should be carefully designed according to the logic of the algorithm.

How to use RateLimiter SPI#

The three parameters in doExecute() method of RateLimiter plugin, exchange is an web request, chain is the execution chain of the plugins,selector is the selection parameters,rule is the policies or rules of rate limiter setting in the system.

protected Mono<Void> doExecute(final ServerWebExchange exchange, final ShenyuPluginChain chain, final SelectorData selector, final RuleData rule) {    //get  the `RateLimiterHandle` parameters from cache     RateLimiterHandle limiterHandle = RatelimiterRuleHandleCache.getInstance()        .obtainHandle(CacheKeyUtils.INST.getKey(rule));    //find the resolver name     String resolverKey = Optional.ofNullable(limiterHandle.getKeyResolverName())        .flatMap(name -> Optional.of("-" + RateLimiterKeyResolverFactory.newInstance(name).resolve(exchange)))        .orElse("");    return redisRateLimiter.isAllowed(rule.getId() + resolverKey, limiterHandle)        .flatMap(response -> {            if (!response.isAllowed()) {                exchange.getResponse().setStatusCode(HttpStatus.TOO_MANY_REQUESTS);                Object error = ShenyuResultWrap.error(ShenyuResultEnum.TOO_MANY_REQUESTS.getCode(), ShenyuResultEnum.TOO_MANY_REQUESTS.getMsg(), null);                return WebFluxResultUtils.result(exchange, error);            }            return chain.execute(exchange);        });}
  1. Firstly get the RateLimiterHandle parameters from cache.

  2. Obtains the corresponding Key resolver by RateLimiterHandle instance.

  3. Reactively executes isAllowed() method of RedisRateLimiter.

  4. If not allowed, error handling is performed.

  5. If the request is allowed, dispatch it to the next process of execution chain.


RateLimiter plugin is based on Spring WebFlux,and with Apache Shen SPI, with Redis and Lua script to responsible for the critical algorithm and logic process, make it with characteristic of high concurrency and elastic. As for the RateLimiter SPI.

  1. RateLimiter SPI provides two SPI interface, with interface oriented design and various design patterns, it's easy to develop new rate limiter algorithm and key resolver rule.
  2. RateLimiterAlgorithm SPI supplies four rate limiter algorithms, token bucket,concurrency rate limiter, leaky bucket and sliding window rate limiter. When designing rate limiter algorithm, the KEY generation need to be carefully designed according to the algorithm characteristic. Using Lua script to realize the logic of the algorithm, and design callback() method for asynchronous processing when needed.
  3. Reactive programming, simple and efficient implementation.

Apache ShenYu Start Demo

· 2 min read
Kunshuai Zhu
Apache ShenYu Contributor

Environmental preparation#

  • Install JDK1.8+ locally
  • Install Git locally
  • Install Maven locally
  • Choose a development tool, such as IDEA

Pull ShenYu code#

Use Git to clone code

> git clone

Compile code#

Compile with Maven

> cd incubator-shenyu> mvn clean install -Dmaven.javadoc.skip=true -B -Drat.skip=true -Djacoco.skip=true -DskipITs -DskipTests

Start the gateway service#

Use development tools, take IDEA as an example.

Start shenyu-admin (use H2 database by default)


Start shenyu-bootstrap


At this point, shenyu gateway has been activated.

We can open the browser and access the admin console: http://localhost:9095/

Start application service#

Apache ShenYu provides examples for Http, Dubbo, SpringCloud and other applications to access the shenyu gateway, located in the shenyu-example module. Here we take the Http service as an example.

If shenyu-example is not marked as a Maven project by IDEA, you can right-click the pom.xml file in the shenyu-example directory to add it as a Maven project.


Start shenyu-examples-http


At this time, shenyu-examples-http will automatically register the interface method annotated with @ShenyuSpringMvcClient and the related configuration in application.yml to the gateway. When we open the admin console, you can see the relevant configuration in divide and context-path.

Test Http request#

Now use postman to simulate http to request your http service:


Use more plugins#

We can refer to Official Document to use other plugins.

Here is an example of using the param-mapping plugin.

Edit the param-mapping plugin in BasicConfig -> Plugin and set status.


Configure selectors and rules in PluginList -> http process.



Then use postman to make an http request to /http/test/payment.


Http Long Polling Data Synchronization Source Code Analysis

· 30 min read
Apache ShenYu Committer

Apache ShenYu is an asynchronous, high-performance, cross-language, responsive API gateway.

In ShenYu gateway, data synchronization refers to how to synchronize the updated data to the gateway after the data is sent in the background management system. The Apache ShenYu gateway currently supports data synchronization for ZooKeeper, WebSocket, http long poll, Nacos, etcd and Consul. The main content of this article is based on http long poll data synchronization source code analysis.

This paper based on shenyu-2.4.0 version of the source code analysis, the official website of the introduction of please refer to the Data Synchronization Design .

1. Http Long Polling#

Here is a direct quote from the official website with the relevant description.

The mechanism of Zookeeper and WebSocket data synchronization is relatively simple, while Http long polling is more complex. Apache ShenYu borrowed the design ideas of Apollo and Nacos, took their essence, and implemented Http long polling data synchronization function by itself. Note that this is not the traditional ajax long polling!

Http Long Polling mechanism as shown above, Apache ShenYu gateway active request shenyu-admin configuration service, read timeout time is 90s, means that the gateway layer request configuration service will wait at most 90s, so as to facilitate shenyu-admin configuration service timely response to change data, so as to achieve quasi real-time push.

The Http long polling mechanism is initiated by the gateway requesting shenyu-admin, so for this source code analysis, we start from the gateway side.

2. Gateway Data Sync#

2.1 Load Configuration#

The Http long polling data synchronization configuration is loaded through spring boot starter mechanism when we introduce the relevant dependencies and have the following configuration in the configuration file.

Introduce dependencies in the pom file.

        <!--shenyu data sync start use http-->        <dependency>            <groupId>org.apache.shenyu</groupId>            <artifactId>shenyu-spring-boot-starter-sync-data-http</artifactId>            <version>${project.version}</version>        </dependency>

Add the following configuration to the application.yml configuration file.

shenyu:    sync:       http:          url : http://localhost:9095

When the gateway is started, the configuration class HttpSyncDataConfiguration is executed, loading the corresponding Bean.

/** * Http sync data configuration for spring boot. */@Configuration@ConditionalOnClass(HttpSyncDataService.class)@ConditionalOnProperty(prefix = "shenyu.sync.http", name = "url")@Slf4jpublic class HttpSyncDataConfiguration {
    /**     * Http sync data service.     * @param httpConfig             * @param pluginSubscriber        * @param metaSubscribers         * @param authSubscribers         * @return the sync data service     */    @Bean    public SyncDataService httpSyncDataService(final ObjectProvider<HttpConfig> httpConfig, final ObjectProvider<PluginDataSubscriber> pluginSubscriber,                                           final ObjectProvider<List<MetaDataSubscriber>> metaSubscribers, final ObjectProvider<List<AuthDataSubscriber>> authSubscribers) {"you use http long pull sync shenyu data");        return new HttpSyncDataService(Objects.requireNonNull(httpConfig.getIfAvailable()), Objects.requireNonNull(pluginSubscriber.getIfAvailable()),                metaSubscribers.getIfAvailable(Collections::emptyList), authSubscribers.getIfAvailable(Collections::emptyList));    }
    /**     * Http config http config.     * @return the http config     */    @Bean    @ConfigurationProperties(prefix = "shenyu.sync.http")    public HttpConfig httpConfig() {        return new HttpConfig();    }}

HttpSyncDataConfiguration is the configuration class for Http long polling data synchronization, responsible for creating HttpSyncDataService (responsible for the concrete implementation of http data synchronization) and HttpConfig (admin property configuration). It is annotated as follows.

  • @Configuration: indicates that this is a configuration class.
  • @ConditionalOnClass(HttpSyncDataService.class): conditional annotation indicating that the class HttpSyncDataService is to be present.
  • @ConditionalOnProperty(prefix = "shenyu.sync.http", name = "url"): conditional annotation to have the property shenyu.sync.http.url configured.

2.2 Property initialization#

  • HttpSyncDataService

In the constructor of HttpSyncDataService, complete the property initialization.

public class HttpSyncDataService implements SyncDataService, AutoCloseable {
    // omitted attribute field ......

   public HttpSyncDataService(final HttpConfig httpConfig, final PluginDataSubscriber pluginDataSubscriber, final List<MetaDataSubscriber> metaDataSubscribers, final List<AuthDataSubscriber> authDataSubscribers) {        // 1. create data refresh factory        this.factory = new DataRefreshFactory(pluginDataSubscriber, metaDataSubscribers, authDataSubscribers);        // 2. get config of admin         this.httpConfig = httpConfig;        // shenyu-admin url        this.serverList = Lists.newArrayList(Splitter.on(",").split(httpConfig.getUrl()));        // 3. create httpClient, used to initiate requests to admin        this.httpClient = createRestTemplate();        // 4. start a long polling task        this.start();    }

Other functions and related fields are omitted from the above code, and the initialization of the properties is done in the constructor, mainly.

  • creating data processors for subsequent caching of various types of data (plugins, selectors, rules, metadata and authentication data).

  • obtaining the admin property configuration, mainly to obtain the url of the admin, admin with possible clusters, multiple split by a comma (,).

  • creating httpClient, using RestTemplate, for launching requests to admin.

    private RestTemplate createRestTemplate() {        OkHttp3ClientHttpRequestFactory factory = new OkHttp3ClientHttpRequestFactory();
        // connection establishment timeout of 10s        factory.setConnectTimeout((int) this.connectionTimeout.toMillis());
        // The gateway actively requests the configuration service of shenyu-admin, and the read timeout is 90s        factory.setReadTimeout((int) HttpConstants.CLIENT_POLLING_READ_TIMEOUT);        return new RestTemplate(factory);    }
  • Start the long polling task.

2.3 Start the long polling task.#

  • HttpSyncDataService#start()

In the start() method, two things are done, one is to get the full amount of data, that is, to request the admin side to get all the data that needs to be synchronized, and then cache the acquired data into the gateway memory. The other is to open a multi-threaded execution of a long polling task.

private void start() {        // Initialize only once, implemented by atomic classes.         RUNNING = new AtomicBoolean(false);        // It could be initialized multiple times, so you need to control that.        if (RUNNING.compareAndSet(false, true)) {            // fetch all group configs.            // Initial startup, get full data            this.fetchGroupConfig(ConfigGroupEnum.values());
            // A backend service, a thread            int threadSize = serverList.size();            // ThreadPoolExecutor            this.executor = new ThreadPoolExecutor(threadSize, threadSize, 60L, TimeUnit.SECONDS,                    new LinkedBlockingQueue<>(),                    ShenyuThreadFactory.create("http-long-polling", true));                        // start long polling, each server creates a thread to listen for changes.            this.serverList.forEach(server -> this.executor.execute(new HttpLongPollingTask(server)));        } else {  "shenyu http long polling was started, executor=[{}]", executor);        }    }
2.3.1 Fetch Data#
  • HttpSyncDataService#fetchGroupConfig()

ShenYu groups all the data that needs to be synchronized, there are 5 data types, namely plugins, selectors, rules, metadata and authentication data.

public enum ConfigGroupEnum {    APP_AUTH, // app auth data    PLUGIN, // plugin data    RULE, // rule data    SELECTOR, // selector data    META_DATA; // meta data}

The admin may be a cluster, and here a request is made to each admin in a round-robin fashion, and if one succeeds, then the operation to get the full amount of data from the admin and cache it to the gateway is executed successfully. If there is an exception, the request is launched to the next admin.

private void fetchGroupConfig(final ConfigGroupEnum... groups) throws ShenyuException {    // It is possible that admins are clustered, and here requests are made to each admin by means of a loop.        for (int index = 0; index < this.serverList.size(); index++) {            String server = serverList.get(index);            try {                // do execute                this.doFetchGroupConfig(server, groups);                // If you have a success, you are successful and can exit the loop                break;            } catch (ShenyuException e) {                // An exception occurs, try executing the next                // The last one also failed to execute, throwing an exception                // no available server, throw exception.                if (index >= serverList.size() - 1) {                    throw e;                }                log.warn("fetch config fail, try another one: {}", serverList.get(index + 1));            }        }    }
  • HttpSyncDataService#doFetchGroupConfig()

In this method, the request parameters are first assembled, then the request is launched through httpClient to admin to get the data, and finally the obtained data is updated to the gateway memory.

// Launch a request to the admin backend management system to get all synchronized dataprivate void doFetchGroupConfig(final String server, final ConfigGroupEnum... groups) {    // 1. build request parameters, all grouped enumeration types    StringBuilder params = new StringBuilder();    for (ConfigGroupEnum groupKey : groups) {        params.append("groupKeys").append("=").append("&");    }
    // admin url:  /configs/fetch    String url = server + "/configs/fetch?" + StringUtils.removeEnd(params.toString(), "&");"request configs: [{}]", url);    String json = null;    try {        // 2. get a request for change data        json = this.httpClient.getForObject(url, String.class);    } catch (RestClientException e) {        String message = String.format("fetch config fail from server[%s], %s", url, e.getMessage());        log.warn(message);        throw new ShenyuException(message, e);    }    // update local cache    // 3. Update data in gateway memory    boolean updated = this.updateCacheWithJson(json);    // The update was successful and the method is now complete    if (updated) {"get latest configs: [{}]", json);        return;    }    // not updated. it is likely that the current config server has not been updated yet. wait a moment."The config of the server[{}] has not been updated or is out of date. Wait for 30s to listen for changes again.", server);    // No data update on the server side, just wait 30s    ThreadUtils.sleep(TimeUnit.SECONDS, 30);}

From the code, we can see that the admin side provides the interface to get the full amount of data is /configs/fetch, so we will not go further here and put it in the later analysis.

If you get the result data from admin and update it successfully, then this method is finished. If there is no successful update, then it is possible that there is no data update on the server side, so wait 30s.

Here you need to explain in advance, the gateway in determining whether the update is successful, there is a comparison of the data operation, immediately mentioned.

  • HttpSyncDataService#updateCacheWithJson()

Update the data in the gateway memory. Use GSON for deserialization, take the real data from the property data and give it to DataRefreshFactory to do the update.

    private boolean updateCacheWithJson(final String json) {        // Using GSON for deserialization        JsonObject jsonObject = GSON.fromJson(json, JsonObject.class);        JsonObject data = jsonObject.getAsJsonObject("data");        // if the config cache will be updated?        return factory.executor(data);    }
  • DataRefreshFactory#executor()

Update the data according to different data types and return the updated result. Here, parallelStream() is used for parallel update, and the specific update logic is given to the dataRefresh.refresh() method. In the update result, one of the data types is updated, which means that the operation has been updated.

    public boolean executor(final JsonObject data) {        //updata data in parallelStream        List<Boolean> result = ENUM_MAP.values().parallelStream()                .map(dataRefresh -> dataRefresh.refresh(data))                .collect(Collectors.toList());        //有一个更新就表示此次发生了更新操作        return;    }
  • AbstractDataRefresh#refresh()

The data update logic uses the template method design pattern, where the generic operation is done in the abstract method and the different implementation logic is done by subclasses. 5 data types have some differences in the specific update logic, but there is also a common update logic, and the class diagram relationship is as follows.

In the generic refresh() method, it is responsible for data type conversion, determining whether an update is needed, and the actual data refresh operation.

    @Override    public Boolean refresh(final JsonObject data) {        boolean updated = false;        // convert data        JsonObject jsonObject = convert(data);        if (null != jsonObject) {            // get data            ConfigData<T> result = fromJson(jsonObject);            // does it need to be updated            if (this.updateCacheIfNeed(result)) {                updated = true;                // real update logic, data refresh operation                refresh(result.getData());            }        }        return updated;    }
  • AbstractDataRefresh#updateCacheIfNeed()

The process of data conversion, which is based on different data types, we will not trace further to see if the data needs to be updated logically. The method name is updateCacheIfNeed(), which is implemented by method overloading.

// result is dataprotected abstract boolean updateCacheIfNeed(ConfigData<T> result);
// newVal is the latest value obtained// What kind of data type is groupEnumprotected boolean updateCacheIfNeed(final ConfigData<T> newVal, final ConfigGroupEnum groupEnum) {        // If it is the first time, then it is put directly into the cache and returns true, indicating that the update was made this time        if (GROUP_CACHE.putIfAbsent(groupEnum, newVal) == null) {            return true;        }        ResultHolder holder = new ResultHolder(false);        GROUP_CACHE.merge(groupEnum, newVal, (oldVal, value) -> {            // md5 value is the same, no need to update            if (StringUtils.equals(oldVal.getMd5(), newVal.getMd5())) {      "Get the same config, the [{}] config cache will not be updated, md5:{}", groupEnum, oldVal.getMd5());                return oldVal;            }
            // The current cached data has been modified for a longer period than the new data and does not need to be updated.            // must compare the last update time            if (oldVal.getLastModifyTime() >= newVal.getLastModifyTime()) {      "Last update time earlier than the current configuration, the [{}] config cache will not be updated", groupEnum);                return oldVal;            }  "update {} config: {}", groupEnum, newVal);            holder.result = true;            return newVal;        });        return holder.result;    }

As you can see from the source code above, there are two cases where updates are not required.

  • The md5 values of both data are the same, so no update is needed;
  • The current cached data has been modified longer than the new data, so no update is needed.

In other cases, the data needs to be updated.

At this point, we have finished analyzing the logic of the start() method to get the full amount of data for the first time, followed by the long polling operation. For convenience, I will paste the start() method once more.

    private void start() {        // It could be initialized multiple times, so you need to control that.        if (RUNNING.compareAndSet(false, true)) {            // fetch all group configs.            // Initial startup, get full data            this.fetchGroupConfig(ConfigGroupEnum.values());
            // one background service, one thread            int threadSize = serverList.size();            // custom thread pool            this.executor = new ThreadPoolExecutor(threadSize, threadSize, 60L, TimeUnit.SECONDS,                    new LinkedBlockingQueue<>(),                    ShenyuThreadFactory.create("http-long-polling", true));            // start long polling, each server creates a thread to listen for changes.            this.serverList.forEach(server -> this.executor.execute(new HttpLongPollingTask(server)));        } else {  "shenyu http long polling was started, executor=[{}]", executor);        }    }
2.3.2 Execute Long Polling Task#
  • HttpLongPollingTask#run()

The long polling task is HttpLongPollingTask, which implements the Runnable interface and the task logic is in the run() method. The task is executed continuously through a while() loop, i.e., long polling. There are three retries in each polling logic, one polling task fails, wait 5s and continue, 3 times all fail, wait 5 minutes and try again.

Start long polling, an admin service, and create a thread for data synchronization.

class HttpLongPollingTask implements Runnable {
        private String server;
        // Default retry 3 times        private final int retryTimes = 3;
        HttpLongPollingTask(final String server) {            this.server = server;        }
        @Override        public void run() {            // long polling            while (RUNNING.get()) {                for (int time = 1; time <= retryTimes; time++) {                    try {                        doLongPolling(server);                    } catch (Exception e) {                        // print warnning log.                        if (time < retryTimes) {                            log.warn("Long polling failed, tried {} times, {} times left, will be suspended for a while! {}",                                    time, retryTimes - time, e.getMessage());                            // long polling failed, wait 5s and continue                            ThreadUtils.sleep(TimeUnit.SECONDS, 5);                            continue;                        }                        // print error, then suspended for a while.                        log.error("Long polling failed, try again after 5 minutes!", e);                        // failed all 3 times, wait 5 minutes and try again                        ThreadUtils.sleep(TimeUnit.MINUTES, 5);                    }                }            }            log.warn("Stop http long polling.");        }    }
  • HttpSyncDataService#doLongPolling()

Core logic for performing long polling tasks.

  • Assembling request parameters based on data types: md5 and lastModifyTime.
  • Assembling the request header and request body.
  • Launching a request to admin to determine if the group data has changed.
  • Based on the group that has changed, go back and get the data.
private void doLongPolling(final String server) {        // build request params: md5 and lastModifyTime        MultiValueMap<String, String> params = new LinkedMultiValueMap<>(8);        for (ConfigGroupEnum group : ConfigGroupEnum.values()) {            ConfigData<?> cacheConfig = factory.cacheConfigData(group);            if (cacheConfig != null) {                String value = String.join(",", cacheConfig.getMd5(), String.valueOf(cacheConfig.getLastModifyTime()));                params.put(, Lists.newArrayList(value));            }        }        // build request heaad and body        HttpHeaders headers = new HttpHeaders();        headers.setContentType(MediaType.APPLICATION_FORM_URLENCODED);        HttpEntity httpEntity = new HttpEntity(params, headers);        String listenerUrl = server + "/configs/listener";        log.debug("request listener configs: [{}]", listenerUrl);        JsonArray groupJson = null;        //Initiate a request to admin to determine if the group data has changed        //Here it just determines whether a group has changed or not        try {            String json = this.httpClient.postForEntity(listenerUrl, httpEntity, String.class).getBody();            log.debug("listener result: [{}]", json);            groupJson = GSON.fromJson(json, JsonObject.class).getAsJsonArray("data");        } catch (RestClientException e) {            String message = String.format("listener configs fail, server:[%s], %s", server, e.getMessage());            throw new ShenyuException(message, e);        }        // Depending on the group where the change occurred, go back and get the data        /**         * The official website explains here.         * After the gateway receives the response message, it only knows which Group has made the configuration change, and it still needs to request the configuration data of that Group again.         * There may be a question here: why not write out the changed data directly?         * We also discussed this issue in depth during development, because the http long polling mechanism can only guarantee quasi-real time, if the processing at the gateway layer is not timely, * or the administrator frequently updates the configuration, it is very difficult to get the information from the gateway layer.         * If it is not processed in time at the gateway level, or if the administrator updates the configuration frequently, it is very likely to miss the push of a configuration change, so for security reasons, we only inform a group that the information has changed.         *For security reasons, we only notify a group of changes.         * Personal understanding.         * If the change data is written out directly, when the administrator frequently updates the configuration, the first update will remove the client from the blocking queue and return the response information to the gateway.         * If a second update is made at this time, the current client is not in the blocking queue, so this time the change is missed.         * The same is true for untimely processing by the gateway layer.         * This is a long polling, one gateway one synchronization thread, there may be time consuming process.         * If the admin has data changes, the current gateway client is not in the blocking queue and will not get the data.         */        if (groupJson != null) {            // fetch group configuration async.            ConfigGroupEnum[] changedGroups = GSON.fromJson(groupJson, ConfigGroupEnum[].class);            if (ArrayUtils.isNotEmpty(changedGroups)) {      "Group config changed: {}", Arrays.toString(changedGroups));                // Proactively get the changed data from admin, depending on the grouping, and take the data in full                this.doFetchGroupConfig(server, changedGroups);            }        }    }

One special point needs to be explained here: In the long polling task, why don't you get the changed data directly? Instead, we determine which group data has been changed, and then request admin again to get the changed data?

The official explanation here is.

After the gateway receives the response information, it only knows which Group has changed its configuration, and it needs to request the configuration data of that Group again. There may be a question here: Why not write out the changed data directly? We have discussed this issue in depth during development, because the http long polling mechanism can only guarantee quasi-real time, and if it is not processed in time at the gateway layer, it will be very difficult to update the configuration data. If the gateway layer is not processed in time, > or the administrator updates the configuration frequently, it is likely to miss the push of a configuration change, so for security reasons, we only inform a group that the information has changed.

My personal understanding is that.

If the change data is written out directly, when the administrator updates the configuration frequently, the first update will client remove the blocking queue and return the response information to the gateway. If a second update is made at this time, then the current client is not in the blocking queue, so this time the change is missed. The same is true for the gateway layer's untimely processing. This is a long polling, one gateway one synchronization thread, there may be a time-consuming process. If admin has data changes, the current gateway client is not in the blocking queue and will not get the data.

We have not yet analyzed the processing logic of the admin side, so let's talk about it roughly. At the admin end, the gateway client will be put into the blocking queue, and when there is a data change, the gateway client will come out of the queue and send the change data. So, if the gateway client is not in the blocking queue when there is a data change, then the current changed data is not available.

When we know which grouping data has changed, we actively get the changed data from admin again, and get the data in full depending on the grouping. The call method is doFetchGroupConfig(), which has been analyzed in the previous section.

At this point of analysis, the data synchronization operation on the gateway side is complete. The long polling task is to keep making requests to admin to see if the data has changed, and if any group data has changed, then initiate another request to admin to get the changed data, and then update the data in the gateway's memory.

Long polling task flow at the gateway side.

3. Admin Data Sync#

From the previous analysis, it can be seen that the gateway side mainly calls two interfaces of admin.

  • /configs/listener: determine whether the group data has changed.
  • /configs/fetch: get the changed group data.

If we analyze directly from these two interfaces, some parts may not be well understood, so let's start analyzing the data synchronization process from the admin startup process.

3.1 Load Configuration#

If the following configuration is done in the configuration file application.yml, it means that the data synchronization is done by http long polling.

shenyu:  sync:      http:        enabled: true

When the program starts, the configuration of the data synchronization class is loaded through springboot conditional assembly. In this process, HttpLongPollingDataChangedListener is created to handle the implementation logic related to long polling.

/** * Data synchronization configuration class * Conditional assembly via springboot * The type Data sync configuration. */@Configurationpublic class DataSyncConfiguration {
    /**     * http long polling.     */    @Configuration    @ConditionalOnProperty(name = "shenyu.sync.http.enabled", havingValue = "true")    @EnableConfigurationProperties(HttpSyncProperties.class)    static class HttpLongPollingListener {
        @Bean        @ConditionalOnMissingBean(HttpLongPollingDataChangedListener.class)        public HttpLongPollingDataChangedListener httpLongPollingDataChangedListener(final HttpSyncProperties httpSyncProperties) {            return new HttpLongPollingDataChangedListener(httpSyncProperties);        }    }}

3.2 Data change listener instantiation#

  • HttpLongPollingDataChangedListener

The data change listener is instantiated and initialized by means of a constructor. In the constructor, a blocking queue is created to hold clients, a thread pool is created to execute deferred tasks and periodic tasks, and information about the properties of long polling is stored.

    public HttpLongPollingDataChangedListener(final HttpSyncProperties httpSyncProperties) {        // default client (here is the gateway) 1024        this.clients = new ArrayBlockingQueue<>(1024);        // create thread pool        // ScheduledThreadPoolExecutor can perform delayed tasks, periodic tasks, and normal tasks        this.scheduler = new ScheduledThreadPoolExecutor(1,                ShenyuThreadFactory.create("long-polling", true));        // http sync properties        this.httpSyncProperties = httpSyncProperties;    }

In addition, it has the following class diagram relationships.

The InitializingBean interface is implemented, so the afterInitialize() method is executed during the initialization of the bean. Execute periodic tasks via thread pool: updating the data in memory (CACHE) is executed every 5 minutes and starts after 5 minutes. Refreshing the local cache is reading data from the database to the local cache (in this case the memory), done by refreshLocalCache().

    /**     * is called in the afterPropertiesSet() method of the InitializingBean interface, which is executed during the initialization of the bean     */    @Override    protected void afterInitialize() {        long syncInterval = httpSyncProperties.getRefreshInterval().toMillis();        // Periodically check the data for changes and update the cache
        // Execution cycle task: Update data in memory (CACHE) is executed every 5 minutes and starts after 5 minutes        // Prevent the admin from starting up first for a while and then generating data; then the gateway doesn't get the full amount of data when it first connects        scheduler.scheduleWithFixedDelay(() -> {  "http sync strategy refresh config start.");            try {                // Read data from database to local cache (in this case, memory)                this.refreshLocalCache();      "http sync strategy refresh config success.");            } catch (Exception e) {                log.error("http sync strategy refresh config error!", e);            }        }, syncInterval, syncInterval, TimeUnit.MILLISECONDS);"http sync strategy refresh interval: {}ms", syncInterval);    }
  • refreshLocalCache()

Update for each of the 5 data types.

    // Read data from database to local cache (in this case, memory)    private void refreshLocalCache() {        //update app auth data        this.updateAppAuthCache();        //update plugin data        this.updatePluginCache();        //update rule data        this.updateRuleCache();        //update selector data        this.updateSelectorCache();        //update meta data        this.updateMetaDataCache();    }

The logic of the 5 update methods is similar, call the service method to get the data and put it into the memory CACHE. Take the updateRuleData method updateRuleCache() for example, pass in the rule enumeration type and call ruleService.listAll() to get all the rule data from the database.

    /**     * Update rule cache.     */    protected void updateRuleCache() {        this.updateCache(ConfigGroupEnum.RULE, ruleService.listAll());    }
  • updateCache()

Update the data in memory using the data in the database.

// cache Mapprotected static final ConcurrentMap<String, ConfigDataCache> CACHE = new ConcurrentHashMap<>();
/**     * if md5 is not the same as the original, then update lcoal cache.     * @param group ConfigGroupEnum     * @param <T> the type of class     * @param data the new config data     */    protected <T> void updateCache(final ConfigGroupEnum group, final List<T> data) {        // data serialization        String json = GsonUtils.getInstance().toJson(data);        // pass in md5 value and modification time        ConfigDataCache newVal = new ConfigDataCache(, json, Md5Utils.md5(json), System.currentTimeMillis());        // update group data        ConfigDataCache oldVal = CACHE.put(newVal.getGroup(), newVal);"update config cache[{}], old: {}, updated: {}", group, oldVal, newVal);    }

The initialization process is to start periodic tasks to update the memory data by fetching data from the database at regular intervals.

Next, we start the analysis of two interfaces.

  • /configs/listener: determines if the group data has changed.
  • /configs/fetch: fetching the changed group data.

3.3 Data change polling interface#

  • /configs/listener: determines if the group data has changed.

The interface class is ConfigController, which only takes effect when using http long polling for data synchronization. The interface method listener() has no other logic and calls the doLongPolling() method directly.

   /** * This Controller only when HttpLongPollingDataChangedListener exist, will take effect. */@ConditionalOnBean(HttpLongPollingDataChangedListener.class)@RestController@RequestMapping("/configs")@Slf4jpublic class ConfigController {
    @Resource    private HttpLongPollingDataChangedListener longPollingListener;        // Omit other logic
    /**     * Listener.     * Listen for data changes and perform long polling     * @param request  the request     * @param response the response     */    @PostMapping(value = "/listener")    public void listener(final HttpServletRequest request, final HttpServletResponse response) {        longPollingListener.doLongPolling(request, response);    }
  • HttpLongPollingDataChangedListener#doLongPolling()

Perform long polling tasks: If there are data changes, they will be responded to the client (in this case, the gateway side) immediately. Otherwise, the client will be blocked until there is a data change or a timeout.

/**     * Execute long polling: If there is a data change, it will be responded to the client (here is the gateway side) immediately.     * Otherwise, the client will otherwise remain blocked until there is a data change or a timeout.     * @param request     * @param response     */    public void doLongPolling(final HttpServletRequest request, final HttpServletResponse response) {        // compare group md5        // Compare the md5, determine whether the data of the gateway and the data of the admin side are consistent, and get the data group that has changed        List<ConfigGroupEnum> changedGroup = compareChangedGroup(request);        String clientIp = getRemoteIp(request);        // response immediately.        // Immediate response to the gateway if there is changed data        if (CollectionUtils.isNotEmpty(changedGroup)) {            this.generateResponse(response, changedGroup);  "send response with the changed group, ip={}, group={}", clientIp, changedGroup);            return;        }
         // No change, then the client (in this case the gateway) is put into the blocking queue        // listen for configuration changed.        final AsyncContext asyncContext = request.startAsync();        // AsyncContext.settimeout() does not timeout properly, so you have to control it yourself        asyncContext.setTimeout(0L);        // block client's thread.        scheduler.execute(new LongPollingClient(asyncContext, clientIp, HttpConstants.SERVER_MAX_HOLD_TIMEOUT));    }
  • HttpLongPollingDataChangedListener#compareChangedGroup()

To determine whether the group data has changed, the judgment logic is to compare the md5 value and lastModifyTime at the gateway side and the admin side.

  • If the md5 value is different, then it needs to be updated.
  • If the lastModifyTime on the admin side is greater than the lastModifyTime on the gateway side, then it needs to be updated.
 /**     * Determine if the group data has changed     * @param request     * @return     */    private List<ConfigGroupEnum> compareChangedGroup(final HttpServletRequest request) {        List<ConfigGroupEnum> changedGroup = new ArrayList<>(ConfigGroupEnum.values().length);        for (ConfigGroupEnum group : ConfigGroupEnum.values()) {            // The md5 value and lastModifyTime of the data on the gateway side            String[] params = StringUtils.split(request.getParameter(, ',');            if (params == null || params.length != 2) {                throw new ShenyuException("group param invalid:" + request.getParameter(;            }            String clientMd5 = params[0];            long clientModifyTime = NumberUtils.toLong(params[1]);            ConfigDataCache serverCache = CACHE.get(;            // do check. determine if the group data has changed            if (this.checkCacheDelayAndUpdate(serverCache, clientMd5, clientModifyTime)) {                changedGroup.add(group);            }        }        return changedGroup;    }
  • LongPollingClient

No change data, then the client (in this case the gateway) is put into the blocking queue. The blocking time is 60 seconds, i.e. after 60 seconds remove and respond to the client.

class LongPollingClient implements Runnable {      // omitted other logic            @Override        public void run() {            try {                // Removal after 60 seconds and response to the client                this.asyncTimeoutFuture = scheduler.schedule(() -> {                    clients.remove(LongPollingClient.this);                    List<ConfigGroupEnum> changedGroups = compareChangedGroup((HttpServletRequest) asyncContext.getRequest());                    sendResponse(changedGroups);                }, timeoutTime, TimeUnit.MILLISECONDS);
                // Add to blocking queue                clients.add(this);
            } catch (Exception ex) {                log.error("add long polling client error", ex);            }        }
        /**         * Send response.         *         * @param changedGroups the changed groups         */        void sendResponse(final List<ConfigGroupEnum> changedGroups) {            // cancel scheduler            if (null != asyncTimeoutFuture) {                asyncTimeoutFuture.cancel(false);            }            // Groups responding to changes            generateResponse((HttpServletResponse) asyncContext.getResponse(), changedGroups);            asyncContext.complete();        }    }

3.4 Get Change Data Interface#

  • /configs/fetch: get change data;

Get the grouped data and return the result according to the parameters passed in by the gateway. The main implementation method is longPollingListener.fetchConfig().

@ConditionalOnBean(HttpLongPollingDataChangedListener.class)@RestController@RequestMapping("/configs")@Slf4jpublic class ConfigController {
    @Resource    private HttpLongPollingDataChangedListener longPollingListener;
    /**     * Fetch configs shenyu result.     * @param groupKeys the group keys     * @return the shenyu result     */    @GetMapping("/fetch")    public ShenyuAdminResult fetchConfigs(@NotNull final String[] groupKeys) {        Map<String, ConfigData<?>> result = Maps.newHashMap();        for (String groupKey : groupKeys) {            ConfigData<?> data = longPollingListener.fetchConfig(ConfigGroupEnum.valueOf(groupKey));            result.put(groupKey, data);        }        return ShenyuAdminResult.success(ShenyuResultMessage.SUCCESS, result);    }      // Other interfaces are omitted
  • AbstractDataChangedListener#fetchConfig()

Data fetching is taken directly from CACHE, and then matched and encapsulated according to different grouping types.

    /**     * fetch configuration from cache.     * @param groupKey the group key     * @return the configuration data     */    public ConfigData<?> fetchConfig(final ConfigGroupEnum groupKey) {        // get data from CACHE        ConfigDataCache config = CACHE.get(;         switch (groupKey) {            case APP_AUTH: // app auth data                List<AppAuthData> appAuthList = GsonUtils.getGson().fromJson(config.getJson(), new TypeToken<List<AppAuthData>>() {                }.getType());                return new ConfigData<>(config.getMd5(), config.getLastModifyTime(), appAuthList);            case PLUGIN: // plugin data                List<PluginData> pluginList = GsonUtils.getGson().fromJson(config.getJson(), new TypeToken<List<PluginData>>() {                }.getType());                return new ConfigData<>(config.getMd5(), config.getLastModifyTime(), pluginList);            case RULE:   // rule data                List<RuleData> ruleList = GsonUtils.getGson().fromJson(config.getJson(), new TypeToken<List<RuleData>>() {                }.getType());                return new ConfigData<>(config.getMd5(), config.getLastModifyTime(), ruleList);            case SELECTOR:  // selector data                List<SelectorData> selectorList = GsonUtils.getGson().fromJson(config.getJson(), new TypeToken<List<SelectorData>>() {                }.getType());                return new ConfigData<>(config.getMd5(), config.getLastModifyTime(), selectorList);            case META_DATA: // meta data                List<MetaData> metaList = GsonUtils.getGson().fromJson(config.getJson(), new TypeToken<List<MetaData>>() {                }.getType());                return new ConfigData<>(config.getMd5(), config.getLastModifyTime(), metaList);            default:  // other data type, throw exception                throw new IllegalStateException("Unexpected groupKey: " + groupKey);        }

3.5 Data Change#

In the previous websocket data synchronization and zookeeper data synchronization source code analysis article, we know that the admin side data synchronization design structure is as follows.

Various data change listeners are subclasses of DataChangedListener.

When the data is modified on the admin side, event notifications are sent through the Spring event handling mechanism. The sending logic is as follows.

/** * Event forwarders, which forward the changed events to each ConfigEventListener. * Data change event distributor: synchronize the change data to ShenYu gateway when there is a data change in admin side * Data changes rely on Spring's event-listening mechanism: ApplicationEventPublisher --> ApplicationEvent --> ApplicationListener * */@Componentpublic class DataChangedEventDispatcher implements ApplicationListener<DataChangedEvent>, InitializingBean {
  // other logic omitted
    /**     * Call this method when there are data changes     * @param event     */    @Override    @SuppressWarnings("unchecked")    public void onApplicationEvent(final DataChangedEvent event) {        // Iterate through the data change listeners (it's generally good to use a kind of data synchronization)        for (DataChangedListener listener : listeners) {            // What kind of data has changed            switch (event.getGroupKey()) {                case APP_AUTH: // app auth data                    listener.onAppAuthChanged((List<AppAuthData>) event.getSource(), event.getEventType());                    break;                case PLUGIN:  // plugin data                    listener.onPluginChanged((List<PluginData>) event.getSource(), event.getEventType());                    break;                case RULE:    // rule data                    listener.onRuleChanged((List<RuleData>) event.getSource(), event.getEventType());                    break;                case SELECTOR:   // selector data                    listener.onSelectorChanged((List<SelectorData>) event.getSource(), event.getEventType());                    break;                case META_DATA:  // meta data                    listener.onMetaDataChanged((List<MetaData>) event.getSource(), event.getEventType());                    break;                default:  // other data type, throw exception                    throw new IllegalStateException("Unexpected value: " + event.getGroupKey());            }        }    }}

Suppose, the plugin information is modified and the data is synchronized by http long polling, then the actual call to listener.onPluginChanged() is org.apache.shenyu.admin.listener. AbstractDataChangedListener#onPluginChanged.

    /**     * In the operation of the admin, there is an update of the plugin occurred     * @param changed   the changed     * @param eventType the event type     */    @Override    public void onPluginChanged(final List<PluginData> changed, final DataEventTypeEnum eventType) {        if (CollectionUtils.isEmpty(changed)) {            return;        }        // update CACHE        this.updatePluginCache();        // execute change task        this.afterPluginChanged(changed, eventType);    }

There are two processing operations, one is to update the memory CACHE, which was analyzed earlier, and the other is to execute the change task, which is executed in the thread pool.

  • HttpLongPollingDataChangedListener#afterPluginChanged()
    @Override    protected void afterPluginChanged(final List<PluginData> changed, final DataEventTypeEnum eventType) {        // execute by thread pool        scheduler.execute(new DataChangeTask(ConfigGroupEnum.PLUGIN));    }
  • DataChangeTask

Data change task: remove the clients in the blocking queue in turn and send a response to notify the gateway that a group of data has changed.

class DataChangeTask implements Runnable {        //other logic omitted          @Override        public void run() {            // If the client in the blocking queue exceeds the given value of 100, it is executed in batches            if (clients.size() > httpSyncProperties.getNotifyBatchSize()) {                List<LongPollingClient> targetClients = new ArrayList<>(clients.size());                clients.drainTo(targetClients);                List<List<LongPollingClient>> partitionClients = Lists.partition(targetClients, httpSyncProperties.getNotifyBatchSize());               // batch execution                partitionClients.forEach(item -> scheduler.execute(() -> doRun(item)));            } else {                // execute task                doRun(clients);            }        }
        private void doRun(final Collection<LongPollingClient> clients) {            // Notify all clients that a data change has occurred            for (Iterator<LongPollingClient> iter = clients.iterator(); iter.hasNext();) {                LongPollingClient client =;                iter.remove();                // send response to client                client.sendResponse(Collections.singletonList(groupKey));      "send response with the changed group,ip={}, group={}, changeTime={}", client.ip, groupKey, changeTime);            }        }    }

At this point, the data synchronization logic on the admin side is analyzed. In the http long polling based data synchronization is, it has three main functions.

  • providing a data change listening interface.
  • providing the interface to get the changed data.
  • When there is a data change, remove the client in the blocking queue and respond to the result.

Finally, three diagrams describe the long polling task flow on the admin side.

  • /configs/listener data change listener interface.

  • /configs/fetch fetch change data interface.

  • Update data in the admin backend management system for data synchronization.

4. Summary#

This article focuses on the source code analysis of http long polling data synchronization in the ShenYu gateway. The main knowledge points involved are as follows.

  • http long polling is initiated by the gateway side, which constantly requests the admin side.
  • change data at group granularity (authentication information, plugins, selectors, rules, metadata).
  • http long polling results in getting only the change group, and another request needs to be initiated to get the group data.
  • Whether the data is updated or not is determined by the md5 value and the modification time lastModifyTime.