Hmily：高性能異步分布式事務TCC框架

Hmily框架特性

• 無縫集成Spring,Spring boot start。

• 無縫集成Dubbo,SpringCloud,Motan等rpc框架。

• 多種事務日志的存儲方式（redis，mongdb,mysql等）。

• 多種不同日志序列化方式（Kryo,protostuff,hession）。

• 事務自動恢復。

• 支持內嵌事務的依賴傳遞。

• 代碼零侵入,配置簡單靈活。

Hmily為什么這么高性能？

1.采用disruptor進行事務日志的異步讀寫（disruptor是一個無鎖，無GC的並發編程框架）

package com.hmily.tcc.core.disruptor.publisher;
import com.hmily.tcc.common.bean.entity.TccTransaction;
import com.hmily.tcc.common.enums.EventTypeEnum;
import com.hmily.tcc.core.concurrent.threadpool.HmilyThreadFactory;
import com.hmily.tcc.core.coordinator.CoordinatorService;
import com.hmily.tcc.core.disruptor.event.HmilyTransactionEvent;
import com.hmily.tcc.core.disruptor.factory.HmilyTransactionEventFactory;
import com.hmily.tcc.core.disruptor.handler.HmilyConsumerDataHandler;
import com.hmily.tcc.core.disruptor.translator.HmilyTransactionEventTranslator;
import com.lmax.disruptor.BlockingWaitStrategy;
import com.lmax.disruptor.IgnoreExceptionHandler;
import com.lmax.disruptor.RingBuffer;
import com.lmax.disruptor.dsl.Disruptor;
import com.lmax.disruptor.dsl.ProducerType;
import org.springframework.beans.factory.DisposableBean;
import org.springframework.beans.factory.annotation.Autowired;
import org.springframework.stereotype.Component;

import java.util.concurrent.Executor;
import java.util.concurrent.LinkedBlockingQueue;
import java.util.concurrent.ThreadPoolExecutor;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.atomic.AtomicInteger;

/**
 * event publisher.
 *
 * @author xiaoyu(Myth)
 */
@Component
public class HmilyTransactionEventPublisher implements DisposableBean {

    private Disruptor<HmilyTransactionEvent> disruptor;

    private final CoordinatorService coordinatorService;

    @Autowired
    public HmilyTransactionEventPublisher(final CoordinatorService coordinatorService) {
        this.coordinatorService = coordinatorService;
    }

    /**
     * disruptor start.
     *
     * @param bufferSize this is disruptor buffer size.
     * @param threadSize this is disruptor consumer thread size.
     */
    public void start(final int bufferSize, final int threadSize) {
        disruptor = new Disruptor<>(new HmilyTransactionEventFactory(), bufferSize, r -> {
            AtomicInteger index = new AtomicInteger(1);
            return new Thread(null, r, "disruptor-thread-" + index.getAndIncrement());
        }, ProducerType.MULTI, new BlockingWaitStrategy());

        final Executor executor = new ThreadPoolExecutor(threadSize, threadSize, 0, TimeUnit.MILLISECONDS,
                new LinkedBlockingQueue<>(),
                HmilyThreadFactory.create("hmily-log-disruptor", false),
                new ThreadPoolExecutor.AbortPolicy());

        HmilyConsumerDataHandler[] consumers = new HmilyConsumerDataHandler[threadSize];
        for (int i = 0; i < threadSize; i++) {
            consumers[i] = new HmilyConsumerDataHandler(executor, coordinatorService);
        }
        disruptor.handleEventsWithWorkerPool(consumers);
        disruptor.setDefaultExceptionHandler(new IgnoreExceptionHandler());
        disruptor.start();
    }

    /**
     * publish disruptor event.
     *
     * @param tccTransaction {@linkplain com.hmily.tcc.common.bean.entity.TccTransaction }
     * @param type           {@linkplain EventTypeEnum}
     */
    public void publishEvent(final TccTransaction tccTransaction, final int type) {
        final RingBuffer<HmilyTransactionEvent> ringBuffer = disruptor.getRingBuffer();
        ringBuffer.publishEvent(new HmilyTransactionEventTranslator(type), tccTransaction);
    }

    @Override
    public void destroy() {
        disruptor.shutdown();
    }
}

在這里bufferSize 的默認值是4094 * 4,用戶可以根據自行的情況進行配置。

HmilyConsumerDataHandler[] consumers = new HmilyConsumerDataHandler[threadSize];
for (int i = 0; i < threadSize; i++) {
    consumers[i] = new HmilyConsumerDataHandler(executor, coordinatorService);
}
disruptor.handleEventsWithWorkerPool(consumers);

這里是采用多個消費者去處理隊列里面的任務。

2.異步執行confrim,cancel方法。

package com.hmily.tcc.core.service.handler;

import com.hmily.tcc.common.bean.context.TccTransactionContext;
import com.hmily.tcc.common.bean.entity.TccTransaction;
import com.hmily.tcc.common.enums.TccActionEnum;
import com.hmily.tcc.core.concurrent.threadpool.HmilyThreadFactory;
import com.hmily.tcc.core.service.HmilyTransactionHandler;
import com.hmily.tcc.core.service.executor.HmilyTransactionExecutor;
import org.aspectj.lang.ProceedingJoinPoint;
import org.springframework.beans.factory.annotation.Autowired;
import org.springframework.stereotype.Component;

import java.util.concurrent.Executor;
import java.util.concurrent.LinkedBlockingQueue;
import java.util.concurrent.ThreadPoolExecutor;
import java.util.concurrent.TimeUnit;

/**
 * this is transaction starter.
 *
 * @author xiaoyu
 */
@Component
public class StarterHmilyTransactionHandler implements HmilyTransactionHandler {

    private static final int MAX_THREAD = Runtime.getRuntime().availableProcessors() << 1;

    private final HmilyTransactionExecutor hmilyTransactionExecutor;

    private final Executor executor = new ThreadPoolExecutor(MAX_THREAD, MAX_THREAD, 0, TimeUnit.MILLISECONDS,
            new LinkedBlockingQueue<>(),
            HmilyThreadFactory.create("hmily-execute", false),
            new ThreadPoolExecutor.AbortPolicy());

    @Autowired
    public StarterHmilyTransactionHandler(final HmilyTransactionExecutor hmilyTransactionExecutor) {
        this.hmilyTransactionExecutor = hmilyTransactionExecutor;
    }

    @Override
    public Object handler(final ProceedingJoinPoint point, final TccTransactionContext context)
    throws Throwable {
        Object returnValue;
        try {
            TccTransaction tccTransaction = hmilyTransactionExecutor.begin(point);
            try {
                //execute try
                returnValue = point.proceed();
                tccTransaction.setStatus(TccActionEnum.TRYING.getCode());
                hmilyTransactionExecutor.updateStatus(tccTransaction);
            } catch (Throwable throwable) {
                //if exception ,execute cancel
                final TccTransaction currentTransaction = hmilyTransactionExecutor.getCurrentTransaction();
                executor.execute(() -> hmilyTransactionExecutor
                                 .cancel(currentTransaction));
                throw throwable;
            }
            //execute confirm
            final TccTransaction currentTransaction = hmilyTransactionExecutor.getCurrentTransaction();
            executor.execute(() -> hmilyTransactionExecutor.confirm(currentTransaction));
        } finally {
            hmilyTransactionExecutor.remove();
        }
        return returnValue;
    }
}

當try方法的AOP切面有異常的時候，采用線程池異步去執行cancel，無異常的時候去執行confrim方法。

這里有人可能會問：那么cancel方法異常，或者confrim方法異常怎么辦呢？
答：首先這種情況是非常罕見的，因為你上一面才剛剛執行完try。其次如果出現這種情況，在try階段會保存好日志，Hmily有內置的調度線程池來進行恢復，不用擔心。

有人又會問：這里如果日志保存異常了怎么辦？
答：首先這又是一個牛角尖問題，首先日志配置的參數，在框架啟動的時候，會要求你配置的。其次，就算在運行過程中日志保存異常，這時候框架會取緩存中的，並不會影響程序正確執行。最后，萬一日志保存異常了，系統又在很極端的情況下down機了，恭喜你，你可以去買彩票了，最好的解決辦法就是不去解決它。

3.ThreadLocal緩存的使用。

/**
 * transaction begin.
 *
 * @param point cut point.
 * @return TccTransaction
 */
public TccTransaction begin(final ProceedingJoinPoint point) {
    LogUtil.debug(LOGGER, () -> "......hmily transaction！start....");
    //build tccTransaction
    final TccTransaction tccTransaction = buildTccTransaction(point, TccRoleEnum.START.getCode(), null);
    //save tccTransaction in threadLocal
    CURRENT.set(tccTransaction);
    //publishEvent
    hmilyTransactionEventPublisher.publishEvent(tccTransaction, EventTypeEnum.SAVE.getCode());
    //set TccTransactionContext this context transfer remote
    TccTransactionContext context = new TccTransactionContext();
    //set action is try
    context.setAction(TccActionEnum.TRYING.getCode());
    context.setTransId(tccTransaction.getTransId());
    context.setRole(TccRoleEnum.START.getCode());
    TransactionContextLocal.getInstance().set(context);
    return tccTransaction;
}

首先要理解，threadLocal保存的發起者一方法的事務信息。這個很重要，不要會有點懵逼。rpc的調用，會形成調用鏈，進行保存。

/**
 * add participant.
 *
 * @param participant {@linkplain Participant}
 */
public void enlistParticipant(final Participant participant) {
    if (Objects.isNull(participant)) {
        return;
    }
    Optional.ofNullable(getCurrentTransaction())
    .ifPresent(c -> {
        c.registerParticipant(participant);
        updateParticipant(c);
    });
}

4.GuavaCache的使用

package com.hmily.tcc.core.cache;

import com.google.common.cache.CacheBuilder;
import com.google.common.cache.CacheLoader;
import com.google.common.cache.LoadingCache;
import com.google.common.cache.Weigher;
import com.hmily.tcc.common.bean.entity.TccTransaction;
import com.hmily.tcc.core.coordinator.CoordinatorService;
import com.hmily.tcc.core.helper.SpringBeanUtils;
import org.apache.commons.lang3.StringUtils;

import java.util.Optional;
import java.util.concurrent.ExecutionException;

/**
 * use google guava cache.
 * @author xiaoyu
 */
public final class TccTransactionCacheManager {

    private static final int MAX_COUNT = 10000;

    private static final LoadingCache<String, TccTransaction> LOADING_CACHE =
        CacheBuilder.newBuilder().maximumWeight(MAX_COUNT)
        .weigher((Weigher<String, TccTransaction>) (string, tccTransaction) -> getSize())
    .build(new CacheLoader<String, TccTransaction>() {
        @Override
        public TccTransaction load(final String key) {
            return cacheTccTransaction(key);
        }
    });

    private static CoordinatorService coordinatorService = SpringBeanUtils.getInstance().getBean(CoordinatorService.class);

    private static final TccTransactionCacheManager TCC_TRANSACTION_CACHE_MANAGER = new TccTransactionCacheManager();

    private TccTransactionCacheManager() {

    }

    /**
     * TccTransactionCacheManager.
     *
     * @return TccTransactionCacheManager
     */
    public static TccTransactionCacheManager getInstance() {
        return TCC_TRANSACTION_CACHE_MANAGER;
    }

    private static int getSize() {
        return (int) LOADING_CACHE.size();
    }

    private static TccTransaction cacheTccTransaction(final String key) {
        return Optional.ofNullable(coordinatorService.findByTransId(key)).orElse(new TccTransaction());
    }

    /**
     * cache tccTransaction.
     *
     * @param tccTransaction {@linkplain TccTransaction}
     */
    public void cacheTccTransaction(final TccTransaction tccTransaction) {
        LOADING_CACHE.put(tccTransaction.getTransId(), tccTransaction);
    }

    /**
     * acquire TccTransaction.
     *
     * @param key this guava key.
     * @return {@linkplain TccTransaction}
     */
    public TccTransaction getTccTransaction(final String key) {
        try {
            return LOADING_CACHE.get(key);
        } catch (ExecutionException e) {
            return new TccTransaction();
        }
    }

    /**
     * remove guava cache by key.
     * @param key guava cache key.
     */
    public void removeByKey(final String key) {
        if (StringUtils.isNotEmpty(key)) {
            LOADING_CACHE.invalidate(key);
        }
    }

}

在參與者中，我們使用了ThreadLocal，而在參與者中，我們為什么不使用呢？

其實原因有二點：首先.因為try，和confrim 會不在一個線程里，會造成ThreadLocal失效。當考慮到RPC集群的時候，可能會負載到不同的機器上。

這里有一個細節就是：

private static TccTransaction cacheTccTransaction(final String key) {
    return Optional.ofNullable(coordinatorService.findByTransId(key)).orElse(new TccTransaction());
}

當GuavaCache里面沒有的時候，會去查詢日志返回，這樣就保證了對集群環境的支持。以上4點造就了Hmily是一個異步的高性能分布式事務TCC框架的原因。

Hmily如何使用？
（https://github.com/yu199195/hmily/tree/master/hmily-tcc-demo）

首先因為之前的包命名問題，框架包並沒有上傳到maven中心倉庫，固需要使用者自己拉取代碼，編譯deploy到自己的私服。

1.dubbo用戶

• 在你的Api接口項目引入

  <dependency>
      <groupId>com.hmily.tcc</groupId>
      <artifactId>hmily-tcc-annotation</artifactId>
      <version>{you version}</version>
  </dependency>

• 在你的服務提供者項目引入

  <dependency>
      <groupId>com.hmily.tcc</groupId>
      <artifactId>hmily-tcc-dubbo</artifactId>
      <version>{you version}</version>
  </dependency>

• 配置啟動bean

  <!-- Aspect 切面配置，是否開啟AOP切面-->
  <aop:aspectj-autoproxy expose-proxy="true"/>
  <!--掃描框架的包-->
  <context:component-scan base-package="com.hmily.tcc.*"/>
  <!--啟動類屬性配置-->
  <bean id="hmilyTransactionBootstrap" class="com.hmily.tcc.core.bootstrap.HmilyTransactionBootstrap">
      <property name="serializer" value="kryo"/>
      <property name="recoverDelayTime" value="120"/>
      <property name="retryMax" value="3"/>
      <property name="scheduledDelay" value="120"/>
      <property name="scheduledThreadMax" value="4"/>
      <property name="repositorySupport" value="db"/>
      <property name="tccDbConfig">
          <bean class="com.hmily.tcc.common.config.TccDbConfig">
              <property name="url"
                        value="jdbc:mysql://192.168.1.98:3306/tcc?useUnicode=true&amp;characterEncoding=utf8"/>
              <property name="driverClassName" value="com.mysql.jdbc.Driver"/>
              <property name="username" value="root"/>
              <property name="password" value="123456"/>
          </bean>
      </property>
  </bean>

當然配置屬性很多，這里只給出了demo，具體可以參考這個類：

package com.hmily.tcc.common.config;

import com.hmily.tcc.common.enums.RepositorySupportEnum;
import lombok.Data;

/**
 * hmily config.
 *
 * @author xiaoyu
 */
@Data
public class TccConfig {


    /**
     * Resource suffix this parameter please fill in about is the transaction store path.
     * If it's a table store this is a table suffix, it's stored the same way.
     * If this parameter is not filled in, the applicationName of the application is retrieved by default
     */
    private String repositorySuffix;

    /**
     * log serializer.
     * {@linkplain com.hmily.tcc.common.enums.SerializeEnum}
     */
    private String serializer = "kryo";

    /**
     * scheduledPool Thread size.
     */
    private int scheduledThreadMax = Runtime.getRuntime().availableProcessors() << 1;

    /**
     * scheduledPool scheduledDelay unit SECONDS.
     */
    private int scheduledDelay = 60;

    /**
     * retry max.
     */
    private int retryMax = 3;

    /**
     * recoverDelayTime Unit seconds
     * (note that this time represents how many seconds after the local transaction was created before execution).
     */
    private int recoverDelayTime = 60;

    /**
     * Parameters when participants perform their own recovery.
     * 1.such as RPC calls time out
     * 2.such as the starter down machine
     */
    private int loadFactor = 2;

    /**
     * repositorySupport.
     * {@linkplain RepositorySupportEnum}
     */
    private String repositorySupport = "db";

    /**
     * disruptor bufferSize.
     */
    private int bufferSize = 4096 * 2 * 2;

    /**
     * this is disruptor consumerThreads.
     */
    private int consumerThreads = Runtime.getRuntime().availableProcessors() << 1;

    /**
     * db config.
     */
    private TccDbConfig tccDbConfig;

    /**
     * mongo config.
     */
    private TccMongoConfig tccMongoConfig;

    /**
     * redis config.
     */
    private TccRedisConfig tccRedisConfig;

    /**
     * zookeeper config.
     */
    private TccZookeeperConfig tccZookeeperConfig;

    /**
     * file config.
     */
    private TccFileConfig tccFileConfig;

}

2.SpringCloud用戶

• 需要引入

   <dependency>
         <groupId>com.hmily.tcc</groupId>
         <artifactId>hmily-tcc-springcloud</artifactId>
         <version>{you version}</version>
  </dependency>

• 配置啟動bean 如上。

3.Motan用戶

• 需要引入

  <dependency>
      <groupId>com.hmily.tcc</groupId>
      <artifactId>hmily-tcc-motan</artifactId>
      <version>{you version}</version>
   </dependency>

• 配置啟動bean 如上。

hmily-spring-boot-start

• 那這個就更容易了，只需要根據你的RPC框架去引入不同的jar包。
• 如果你是dubbo用戶，那么引入

  <dependency>
       <groupId>com.hmily.tcc</groupId>
       <artifactId>hmily-tcc-spring-boot-starter-dubbo</artifactId>
       <version>${your version}</version>
   </dependency>

• 如果你是SpringCloud用戶，那么引入

  <dependency>
       <groupId>com.hmily.tcc</groupId>
       <artifactId>hmily-tcc-spring-boot-starter-springcloud</artifactId>
       <version>${your version}</version>
   </dependency>

• 如果你是Motan用戶，那么引入

  <dependency>
       <groupId>com.hmily.tcc</groupId>
       <artifactId>hmily-tcc-spring-boot-starter-motan</artifactId>
       <version>${your version}</version>
   </dependency>

• 然后在你的yml里面進行如下配置：

  hmily:
      tcc :
          serializer : kryo
          recoverDelayTime : 128
          retryMax : 3
          scheduledDelay : 128
          scheduledThreadMax :  10
          repositorySupport : db
          tccDbConfig :
                   driverClassName  : com.mysql.jdbc.Driver
                   url :  jdbc:mysql://192.168.1.98:3306/tcc?useUnicode=true&characterEncoding=utf8
                   username : root
                   password : 123456
  
          #repositorySupport : redis
          #tccRedisConfig:
                   #masterName: mymaster
                   #sentinel : true
                   #sentinelUrl : 192.168.1.91:26379;192.168.1.92:26379;192.168.1.93:26379
                   #password  : foobaredbbexONE123
  
  
         # repositorySupport : zookeeper
         #         host      : 92.168.1.73:2181
         #         sessionTimeOut      :  100000
         #         rootPath  : /tcc
  
         # repositorySupport : mongodb
         #       mongoDbUrl  : 192.168.1.68:27017
         #       mongoDbName  :  happylife
         #       mongoUserName  : xiaoyu
         #       mongoUserPwd   : 123456
  
         # repositorySupport : file
         #         path      : /account
         #         prefix    :  account

就這么簡單，然后就可以在接口方法上加上@Tcc注解，進行愉快的使用了。當然因為篇幅問題，很多東西只是簡單的描述，尤其是邏輯方面的。

下面是github地址：https://github.com/yu199195/hmily