在企業級架構應用中學Netty構建分散式訊息佇列實現原理淺析
首先,在一個企業級的架構應用中,究竟何時需引入訊息佇列呢?本人認為,最經常的情況,無非這幾種:做業務解耦、事件訊息廣播、訊息流控處理。其中,對於業務解耦是作為訊息佇列,要解決的一個首要問題。所謂業務解耦,就是說在一個業務流程處理上,只關注具體的流程,盡到通知的責任即可,不必等待訊息處理的結果。
總得來看,企業級系統模組通訊的方式通常情況下,無非兩種。
同步方式:REST、RPC方式實現;非同步方式:訊息中介軟體(訊息佇列)方式實現。
同步方式的優點:可以基於http協議之上,無需中介軟體代理,系統架構相對而言比較簡單。缺點是:客戶端和服務端緊密耦合,並且要實時線上通訊,否則會導致訊息傳送失敗。
非同步方式的優點:客戶端和服務端互相解耦,雙方可以不產生依賴。缺點是:由於引入了訊息中介軟體,在程式設計的時候會增加難度係數。此外,訊息中介軟體的可靠性、容錯性、健壯性往往成為這類架構的決定性因素。
舉一個本人工作中的例子向大家說明一下:移動業務中的產品訂購中心,每當一個使用者通過某些渠道(營業廳、自助終端等等)開通、訂購了某個套餐之後,如果這些套餐涉及第三方平臺派單的話,產品訂購中心會向第三方平臺發起訂購請求操作。試想一下,如果遇到高峰受理時間段,由於業務受理量的激增,導致一些外圍系統的響應速度降低(比如業務閘道器響應速度不及時、網路延時等等原因),終端使用者開通一個套餐花在主流程的時間會延長很多,這個會造成極不好的使用者體驗,最終可能導致受理失敗。在上述的場景裡面,我們就可以很好的引入一個訊息佇列進行業務的解耦,具體來說,產品訂購中心只要“通知”第三方平臺,我們的套餐開通成功了,並不一定非要同步阻塞地等待其真正的開通處理完成。正因為如此,訊息佇列逐漸成為當下系統模組通訊的主要方式手段。
當今在Java的訊息佇列通訊領域,有很多主流的訊息中介軟體,比如RabbitMQ、ActiveMQ、以及炙手可熱Kafka。其中ActiveMQ是基於JMS的標準之上開發定製的一套訊息佇列系統,效能穩定,訪問介面也非常友好,但是這類的訊息佇列在訪問吞吐量上有所折扣;另外一個方面,比如Kafka這樣,以高效吞吐量著稱的訊息佇列系統,但是在穩定性和可靠性上,能力似乎還不夠,因此更多的是用在服務日誌傳輸、短訊息推送等等對於可靠性不高的業務場景之中。總結起來,不管是ActiveMQ還是Kafka,其框架的背後涉及到很多非同步網路通訊、多執行緒、高併發處理方面的專業技術知識。但本文的重點,也不在於介紹這些訊息中介軟體背後的技術細節,而是想重點闡述一下,如何透過上述訊息佇列的基本原理,在必要的時候,開發定製一套符合自身業務要求的訊息佇列系統時,能夠獲得更加全面的視角去設計、考量這些問題。
因此本人用心開發實現了一個,基於Netty的訊息佇列系統:AvatarMQ。當然,在設計、實現AvatarMQ的時候,我會適當參考這些成熟訊息中介軟體中用到的很多重要的思想理念。
當各位從github上面下載到AvatarMQ的原始碼的時候,可以發現,其中的包結構如下所示:
現在對每個包的主要功能進行一下簡要說明(下面省略字首com.newlandframework.avatarmq)。
broker:訊息中介軟體的伺服器模組,主要負責訊息的路由、負載均衡,對於生產者、消費者進行訊息的應答回覆處理(ACK),AvatarMQ中的中心節點,是連線生產者、消費者的橋樑紐帶。
consumer:訊息中介軟體中的消費者模組,負責接收生產者過來的訊息,在設計的時候,會對消費者進行一個叢集化管理,同一個叢集標識的消費者,會構成一個大的消費者叢集,作為一個整體,接收生產者投遞過來的訊息。此外,還提供消費者接收訊息相關的API給客戶端進行呼叫。
producer:訊息中介軟體中的生產者模組,負責生產特定主題(Topic)的訊息,傳遞給對此主題感興趣的消費者,同時提供生產者生產訊息的API介面,給客戶端使用。
core:AvatarMQ中訊息處理的核心模組,負責訊息的記憶體儲存、應答控制、對訊息進行多執行緒任務分派處理。
model:主要定義了AvatarMQ中的資料模型物件,比如MessageType訊息型別、MessageSource訊息源頭等等模型物件的定義。
msg:主要定義了具體的訊息型別對應的結構模型,比如消費者訂閱訊息SubscribeMessage、消費者取消訂閱訊息UnSubscribeMessage,訊息伺服器應答給生產者的應答訊息ProducerAckMessage、訊息伺服器應答給消費者的應答訊息ConsumerAckMessage。
netty:主要封裝了Netty網路通訊相關的核心模組程式碼,比如訂閱訊息事件的路由分派策略、訊息的編碼、解碼器等等。
serialize:利用Kryo這個優秀高效的物件序列化、反序列框架對訊息物件進行序列化網路傳輸。
spring:Spring的容器管理類,負責把AvatarMQ中的訊息伺服器模組:Broker,進行容器化管理。這個包裡面的AvatarMQServerStartup是整個AvatarMQ訊息伺服器的啟動入口。
test:這個就不用多說了,就是針對AvatarMQ進行訊息路由傳遞的測試demo。
AvatarMQ執行原理示意圖:
首先是訊息生產者客戶端(AvatarMQ Producer)傳送帶有主題的訊息給訊息轉發伺服器(AvatarMQ Broker),訊息轉發伺服器確認收到生產者的訊息,傳送ACK應答給生產者,然後把訊息繼續投遞給消費者(AvatarMQ Consumer)。同時broker伺服器接收來自消費者的訂閱、取消訂閱訊息,併傳送ACK應該給對應的消費者,整個訊息系統就是這樣周而復始的工作。
現在再來看一下,AvatarMQ中的核心模組的組成,如下圖所示:
Producer Manage:訊息的生產者,其主要程式碼在(com.newlandframework.avatarmq.producer)包之下,其主要程式碼模組關鍵部分簡要說明如下:
package com.newlandframework.avatarmq.producer; import com.newlandframework.avatarmq.core.AvatarMQAction; import com.newlandframework.avatarmq.model.MessageSource; import com.newlandframework.avatarmq.model.MessageType; import com.newlandframework.avatarmq.model.RequestMessage; import com.newlandframework.avatarmq.model.ResponseMessage; import com.newlandframework.avatarmq.msg.Message; import com.newlandframework.avatarmq.msg.ProducerAckMessage; import com.newlandframework.avatarmq.netty.MessageProcessor; import java.util.concurrent.atomic.AtomicLong; /** * @filename:AvatarMQProducer.java * @description:AvatarMQProducer功能模組 * @author tangjie<https://github.com/tang-jie> * @blog http://www.cnblogs.com/jietang/ * @since 2016-8-11 */ public class AvatarMQProducer extends MessageProcessor implements AvatarMQAction { private boolean brokerConnect = false; private boolean running = false; private String brokerServerAddress; private String topic; private String defaultClusterId = "AvatarMQProducerClusters"; private String clusterId = ""; private AtomicLong msgId = new AtomicLong(0L); //連線訊息轉發伺服器broker的ip地址,以及生產出來訊息附帶的主題資訊 public AvatarMQProducer(String brokerServerAddress, String topic) { super(brokerServerAddress); this.brokerServerAddress = brokerServerAddress; this.topic = topic; } //沒有連線上訊息轉發伺服器broker就傳送的話,直接應答失敗 private ProducerAckMessage checkMode() { if (!brokerConnect) { ProducerAckMessage ack = new ProducerAckMessage(); ack.setStatus(ProducerAckMessage.FAIL); return ack; } return null; } //啟動訊息生產者 public void start() { super.getMessageConnectFactory().connect(); brokerConnect = true; running = true; } //連線訊息轉發伺服器broker,設定生產者訊息處理鉤子,用於處理broker過來的訊息應答 public void init() { ProducerHookMessageEvent hook = new ProducerHookMessageEvent(); hook.setBrokerConnect(brokerConnect); hook.setRunning(running); super.getMessageConnectFactory().setMessageHandle(new MessageProducerHandler(this, hook)); } //投遞訊息API public ProducerAckMessage delivery(Message message) { if (!running || !brokerConnect) { return checkMode(); } message.setTopic(topic); message.setTimeStamp(System.currentTimeMillis()); RequestMessage request = new RequestMessage(); request.setMsgId(String.valueOf(msgId.incrementAndGet())); request.setMsgParams(message); request.setMsgType(MessageType.AvatarMQMessage); request.setMsgSource(MessageSource.AvatarMQProducer); message.setMsgId(request.getMsgId()); ResponseMessage response = (ResponseMessage) sendAsynMessage(request); if (response == null) { ProducerAckMessage ack = new ProducerAckMessage(); ack.setStatus(ProducerAckMessage.FAIL); return ack; } ProducerAckMessage result = (ProducerAckMessage) response.getMsgParams(); return result; } //關閉訊息生產者 public void shutdown() { if (running) { running = false; super.getMessageConnectFactory().close(); super.closeMessageConnectFactory(); } } public String getTopic() { return topic; } public void setTopic(String topic) { this.topic = topic; } public String getClusterId() { return clusterId; } public void setClusterId(String clusterId) { this.clusterId = clusterId; } }
Consumer Clusters Manage / Message Routing:訊息的消費者叢集管理以及訊息路由模組,其主要模組在包(com.newlandframework.avatarmq.consumer)之中。其中訊息消費者物件,對應的核心程式碼主要功能描述如下:
package com.newlandframework.avatarmq.consumer; import com.google.common.base.Joiner; import com.newlandframework.avatarmq.core.AvatarMQAction; import com.newlandframework.avatarmq.core.MessageIdGenerator; import com.newlandframework.avatarmq.core.MessageSystemConfig; import com.newlandframework.avatarmq.model.MessageType; import com.newlandframework.avatarmq.model.RequestMessage; import com.newlandframework.avatarmq.msg.SubscribeMessage; import com.newlandframework.avatarmq.msg.UnSubscribeMessage; import com.newlandframework.avatarmq.netty.MessageProcessor; /** * @filename:AvatarMQConsumer.java * @description:AvatarMQConsumer功能模組 * @author tangjie<https://github.com/tang-jie> * @blog http://www.cnblogs.com/jietang/ * @since 2016-8-11 */ public class AvatarMQConsumer extends MessageProcessor implements AvatarMQAction { private ProducerMessageHook hook; private String brokerServerAddress; private String topic; private boolean subscribeMessage = false; private boolean running = false; private String defaultClusterId = "AvatarMQConsumerClusters"; private String clusterId = ""; private String consumerId = ""; //連線的訊息伺服器broker的ip地址以及關注的生產過來的訊息鉤子 public AvatarMQConsumer(String brokerServerAddress, String topic, ProducerMessageHook hook) { super(brokerServerAddress); this.hook = hook; this.brokerServerAddress = brokerServerAddress; this.topic = topic; } //向訊息伺服器broker傳送取消訂閱訊息 private void unRegister() { RequestMessage request = new RequestMessage(); request.setMsgType(MessageType.AvatarMQUnsubscribe); request.setMsgId(new MessageIdGenerator().generate()); request.setMsgParams(new UnSubscribeMessage(consumerId)); sendSyncMessage(request); super.getMessageConnectFactory().close(); super.closeMessageConnectFactory(); running = false; } //向訊息伺服器broker傳送訂閱訊息 private void register() { RequestMessage request = new RequestMessage(); request.setMsgType(MessageType.AvatarMQSubscribe); request.setMsgId(new MessageIdGenerator().generate()); SubscribeMessage subscript = new SubscribeMessage(); subscript.setClusterId((clusterId.equals("") ? defaultClusterId : clusterId)); subscript.setTopic(topic); subscript.setConsumerId(consumerId); request.setMsgParams(subscript); sendAsynMessage(request); } public void init() { super.getMessageConnectFactory().setMessageHandle(new MessageConsumerHandler(this, new ConsumerHookMessageEvent(hook))); Joiner joiner = Joiner.on(MessageSystemConfig.MessageDelimiter).skipNulls(); consumerId = joiner.join((clusterId.equals("") ? defaultClusterId : clusterId), topic, new MessageIdGenerator().generate()); } //連線訊息伺服器broker public void start() { if (isSubscribeMessage()) { super.getMessageConnectFactory().connect(); register(); running = true; } } public void receiveMode() { setSubscribeMessage(true); } public void shutdown() { if (running) { unRegister(); } } public String getBrokerServerAddress() { return brokerServerAddress; } public void setBrokerServerAddress(String brokerServerAddress) { this.brokerServerAddress = brokerServerAddress; } public String getTopic() { return topic; } public void setTopic(String topic) { this.topic = topic; } public boolean isSubscribeMessage() { return subscribeMessage; } public void setSubscribeMessage(boolean subscribeMessage) { this.subscribeMessage = subscribeMessage; } public String getDefaultClusterId() { return defaultClusterId; } public void setDefaultClusterId(String defaultClusterId) { this.defaultClusterId = defaultClusterId; } public String getClusterId() { return clusterId; } public void setClusterId(String clusterId) { this.clusterId = clusterId; } }
訊息的叢集管理模組,主要程式碼是ConsumerContext.java、ConsumerClusters.java。先簡單說一下消費者叢集模組ConsumerClusters,主要負責定義消費者叢集的行為,以及負責訊息的路由。主要的功能描述如下所示:
package com.newlandframework.avatarmq.consumer; import com.newlandframework.avatarmq.model.RemoteChannelData; import com.newlandframework.avatarmq.model.SubscriptionData; import com.newlandframework.avatarmq.netty.NettyUtil; import io.netty.channel.Channel; import java.util.ArrayList; import java.util.Collections; import java.util.List; import java.util.concurrent.ConcurrentHashMap; import org.apache.commons.collections.CollectionUtils; import org.apache.commons.collections.MapUtils; import org.apache.commons.collections.Predicate; /** * @filename:ConsumerClusters.java * @description:ConsumerClusters功能模組 * @author tangjie<https://github.com/tang-jie> * @blog http://www.cnblogs.com/jietang/ * @since 2016-8-11 */ public class ConsumerClusters { //輪詢排程(Round-Robin Scheduling)位置標記 private int next = 0; private final String clustersId; private final ConcurrentHashMap<String/*生產者訊息的主題*/, SubscriptionData/*訊息對應的topic資訊資料結構*/> subMap = new ConcurrentHashMap<String, SubscriptionData>(); private final ConcurrentHashMap<String/*消費者標識編碼*/, RemoteChannelData/*對應的消費者的netty網路通訊管道資訊*/> channelMap = new ConcurrentHashMap<String, RemoteChannelData>(); private final List<RemoteChannelData> channelList = Collections.synchronizedList(new ArrayList<RemoteChannelData>()); public ConsumerClusters(String clustersId) { this.clustersId = clustersId; } public String getClustersId() { return clustersId; } public ConcurrentHashMap<String, SubscriptionData> getSubMap() { return subMap; } public ConcurrentHashMap<String, RemoteChannelData> getChannelMap() { return channelMap; } //新增一個消費者到消費者叢集 public void attachRemoteChannelData(String clientId, RemoteChannelData channelinfo) { if (findRemoteChannelData(channelinfo.getClientId()) == null) { channelMap.put(clientId, channelinfo); subMap.put(channelinfo.getSubcript().getTopic(), channelinfo.getSubcript()); channelList.add(channelinfo); } else { System.out.println("consumer clusters exists! it's clientId:" + clientId); } } //從消費者叢集中刪除一個消費者 public void detachRemoteChannelData(String clientId) { channelMap.remove(clientId); Predicate predicate = new Predicate() { public boolean evaluate(Object object) { String id = ((RemoteChannelData) object).getClientId(); return id.compareTo(clientId) == 0; } }; RemoteChannelData data = (RemoteChannelData) CollectionUtils.find(channelList, predicate); if (data != null) { channelList.remove(data); } } //根據消費者標識編碼,在消費者叢集中查詢定位一個消費者,如果不存在返回null public RemoteChannelData findRemoteChannelData(String clientId) { return (RemoteChannelData) MapUtils.getObject(channelMap, clientId); } //負載均衡,根據連線到broker的順序,依次投遞訊息給消費者。這裡的均衡演算法直接採用 //輪詢排程(Round-Robin Scheduling),後續可以加入:加權輪詢、隨機輪詢、雜湊輪詢等等策略。 public RemoteChannelData nextRemoteChannelData() { Predicate predicate = new Predicate() { public boolean evaluate(Object object) { RemoteChannelData data = (RemoteChannelData) object; Channel channel = data.getChannel(); return NettyUtil.validateChannel(channel); } }; CollectionUtils.filter(channelList, predicate); return channelList.get(next++ % channelList.size()); } //根據生產者的主題關鍵字,定位於具體的訊息結構 public SubscriptionData findSubscriptionData(String topic) { return this.subMap.get(topic); } }
而ConsumerContext主要的負責管理消費者叢集的,其主要核心程式碼註釋說明如下:
package com.newlandframework.avatarmq.consumer; import com.newlandframework.avatarmq.model.RemoteChannelData; import com.newlandframework.avatarmq.model.SubscriptionData; import java.util.ArrayList; import java.util.Iterator; import java.util.List; import java.util.concurrent.ConcurrentHashMap; import java.util.concurrent.CopyOnWriteArrayList; import org.apache.commons.collections.Predicate; import org.apache.commons.collections.iterators.FilterIterator; /** * @filename:ConsumerContext.java * @description:ConsumerContext功能模組 * @author tangjie<https://github.com/tang-jie> * @blog http://www.cnblogs.com/jietang/ * @since 2016-8-11 */ public class ConsumerContext { //消費者叢集關係定義 private static final CopyOnWriteArrayList<ClustersRelation> relationArray = new CopyOnWriteArrayList<ClustersRelation>(); //消費者叢集狀態 private static final CopyOnWriteArrayList<ClustersState> stateArray = new CopyOnWriteArrayList<ClustersState>(); public static void setClustersStat(String clusters, int stat) { stateArray.add(new ClustersState(clusters, stat)); } //根據消費者叢集編碼cluster_id獲取一個消費者叢集的狀態 public static int getClustersStat(String clusters) { Predicate predicate = new Predicate() { public boolean evaluate(Object object) { String clustersId = ((ClustersState) object).getClusters(); return clustersId.compareTo(clusters) == 0; } }; Iterator iterator = new FilterIterator(stateArray.iterator(), predicate); ClustersState state = null; while (iterator.hasNext()) { state = (ClustersState) iterator.next(); break; } return (state != null) ? state.getState() : 0; } //根據消費者叢集編碼cluster_id查詢一個消費者叢集 public static ConsumerClusters selectByClusters(String clusters) { Predicate predicate = new Predicate() { public boolean evaluate(Object object) { String id = ((ClustersRelation) object).getId(); return id.compareTo(clusters) == 0; } }; Iterator iterator = new FilterIterator(relationArray.iterator(), predicate); ClustersRelation relation = null; while (iterator.hasNext()) { relation = (ClustersRelation) iterator.next(); break; } return (relation != null) ? relation.getClusters() : null; } //查詢一下關注這個主題的消費者叢集集合 public static List<ConsumerClusters> selectByTopic(String topic) { List<ConsumerClusters> clusters = new ArrayList<ConsumerClusters>(); for (int i = 0; i < relationArray.size(); i++) { ConcurrentHashMap<String, SubscriptionData> subscriptionTable = relationArray.get(i).getClusters().getSubMap(); if (subscriptionTable.containsKey(topic)) { clusters.add(relationArray.get(i).getClusters()); } } return clusters; } //新增消費者叢集 public static void addClusters(String clusters, RemoteChannelData channelinfo) { ConsumerClusters manage = selectByClusters(clusters); if (manage == null) { ConsumerClusters newClusters = new ConsumerClusters(clusters); newClusters.attachRemoteChannelData(channelinfo.getClientId(), channelinfo); relationArray.add(new ClustersRelation(clusters, newClusters)); } else if (manage.findRemoteChannelData(channelinfo.getClientId()) != null) { manage.detachRemoteChannelData(channelinfo.getClientId()); manage.attachRemoteChannelData(channelinfo.getClientId(), channelinfo); } else { String topic = channelinfo.getSubcript().getTopic(); boolean touchChannel = manage.getSubMap().containsKey(topic); if (touchChannel) { manage.attachRemoteChannelData(channelinfo.getClientId(), channelinfo); } else { manage.getSubMap().clear(); manage.getChannelMap().clear(); manage.attachRemoteChannelData(channelinfo.getClientId(), channelinfo); } } } //從一個消費者叢集中刪除一個消費者 public static void unLoad(String clientId) { for (int i = 0; i < relationArray.size(); i++) { String id = relationArray.get(i).getId(); ConsumerClusters manage = relationArray.get(i).getClusters(); if (manage.findRemoteChannelData(clientId) != null) { manage.detachRemoteChannelData(clientId); } if (manage.getChannelMap().size() == 0) { ClustersRelation relation = new ClustersRelation(); relation.setId(id); relationArray.remove(id); } } } }
ACK Queue Dispatch:主要是broker分別向對應的訊息生產者、消費者傳送ACK訊息應答,其主要核心模組是在:com.newlandframework.avatarmq.broker包下面的AckPullMessageController和AckPushMessageController模組,主要職責是在broker中收集生產者的訊息,確認成功收到之後,把其放到訊息佇列容器中,然後專門安排一個工作執行緒池把ACK應答傳送給生產者。
Message Queue Dispatch:生產者訊息的分派,主要是由com.newlandframework.avatarmq.broker包下面的SendMessageController派發模組進行任務的分派,其中訊息分派支援兩種策略,一種是記憶體緩衝訊息區裡面只要一有訊息就通知消費者;還有一種是對訊息進行緩衝處理,累計到一定的數量之後進行派發,這個是根據:MessageSystemConfig類中的核心引數:SystemPropertySendMessageControllerTaskCommitValue(com.newlandframework.avatarmq.system.send.taskcommit)決定的,預設是1。即一有訊息就派發,如果改成大於1的數值,表示訊息緩衝的數量。現在給出SendMessageController的核心實現程式碼:
package com.newlandframework.avatarmq.broker; import com.newlandframework.avatarmq.core.SemaphoreCache; import com.newlandframework.avatarmq.core.MessageSystemConfig; import com.newlandframework.avatarmq.core.MessageTaskQueue; import com.newlandframework.avatarmq.core.SendMessageCache; import com.newlandframework.avatarmq.model.MessageDispatchTask; import java.util.Timer; import java.util.TimerTask; import java.util.concurrent.Callable; import java.util.concurrent.ConcurrentLinkedQueue; import java.util.concurrent.atomic.AtomicBoolean; import java.util.logging.Level; import java.util.logging.Logger; /** * @filename:SendMessageController.java * @description:SendMessageController功能模組 * @author tangjie<https://github.com/tang-jie> * @blog http://www.cnblogs.com/jietang/ * @since 2016-8-11 */ public class SendMessageController implements Callable<Void> { private volatile boolean stoped = false; private AtomicBoolean flushTask = new AtomicBoolean(false); private ThreadLocal<ConcurrentLinkedQueue<MessageDispatchTask>> requestCacheList = new ThreadLocal<ConcurrentLinkedQueue<MessageDispatchTask>>() { protected ConcurrentLinkedQueue<MessageDispatchTask> initialValue() { return new ConcurrentLinkedQueue<MessageDispatchTask>(); } }; private final Timer timer = new Timer("SendMessageTaskMonitor", true); public void stop() { stoped = true; } public boolean isStoped() { return stoped; } public Void call() { int period = MessageSystemConfig.SendMessageControllerPeriodTimeValue; int commitNumber = MessageSystemConfig.SendMessageControllerTaskCommitValue; int sleepTime = MessageSystemConfig.SendMessageControllerTaskSleepTimeValue; ConcurrentLinkedQueue<MessageDispatchTask> queue = requestCacheList.get(); SendMessageCache ref = SendMessageCache.getInstance(); while (!stoped) { SemaphoreCache.acquire(MessageSystemConfig.NotifyTaskSemaphoreValue); MessageDispatchTask task = MessageTaskQueue.getInstance().getTask(); queue.add(task); if (queue.size() == 0) { try { Thread.sleep(sleepTime); continue; } catch (InterruptedException ex) { Logger.getLogger(SendMessageController.class.getName()).log(Level.SEVERE, null, ex); } } if (queue.size() > 0 && (queue.size() % commitNumber == 0 || flushTask.get() == true)) { ref.commit(queue); queue.clear(); flushTask.compareAndSet(true, false); } timer.scheduleAtFixedRate(new TimerTask() { public void run() { try { flushTask.compareAndSet(false, true); } catch (Exception e) { System.out.println("SendMessageTaskMonitor happen exception"); } } }, 1000 * 1, period); } return null; } }
訊息分派採用多執行緒並行派發,其內部通過柵欄機制,為訊息派發設定一個屏障點,後續可以暴露給JMX介面,進行對整個訊息系統,訊息派發情況的動態監控。比如發現訊息積壓太多,可以加大執行緒並行度。訊息無堆積的話,降低執行緒並行度,減輕系統負荷。現在給出訊息派發任務模組SendMessageTask的核心程式碼:
package com.newlandframework.avatarmq.core; import com.newlandframework.avatarmq.msg.ConsumerAckMessage; import com.newlandframework.avatarmq.msg.Message; import com.newlandframework.avatarmq.broker.SendMessageLauncher; import com.newlandframework.avatarmq.consumer.ClustersState; import com.newlandframework.avatarmq.consumer.ConsumerContext; import com.newlandframework.avatarmq.model.MessageType; import com.newlandframework.avatarmq.model.RequestMessage; import com.newlandframework.avatarmq.model.ResponseMessage; import com.newlandframework.avatarmq.model.RemoteChannelData; import com.newlandframework.avatarmq.model.MessageSource; import com.newlandframework.avatarmq.model.MessageDispatchTask; import com.newlandframework.avatarmq.netty.NettyUtil; import java.util.concurrent.Callable; import java.util.concurrent.Phaser; /** * @filename:SendMessageTask.java * @description:SendMessageTask功能模組 * @author tangjie<https://github.com/tang-jie> * @blog http://www.cnblogs.com/jietang/ * @since 2016-8-11 */ public class SendMessageTask implements Callable<Void> { private MessageDispatchTask[] tasks; //訊息柵欄器,為後續進行訊息JMX實時監控預留介面 private Phaser phaser = null; private SendMessageLauncher launcher = SendMessageLauncher.getInstance(); public SendMessageTask(Phaser phaser, MessageDispatchTask[] tasks) { this.phaser = phaser; this.tasks = tasks; } public Void call() throws Exception { for (MessageDispatchTask task : tasks) { Message msg = task.getMessage(); if (ConsumerContext.selectByClusters(task.getClusters()) != null) { RemoteChannelData channel = ConsumerContext.selectByClusters(task.getClusters()).nextRemoteChannelData(); ResponseMessage response = new ResponseMessage(); response.setMsgSource(MessageSource.AvatarMQBroker); response.setMsgType(MessageType.AvatarMQMessage); response.setMsgParams(msg); response.setMsgId(new MessageIdGenerator().generate()); try { //訊息派發的時候,發現管道不可達,跳過 if (!NettyUtil.validateChannel(channel.getChannel())) { ConsumerContext.setClustersStat(task.getClusters(), ClustersState.NETWORKERR); continue; } RequestMessage request = (RequestMessage) launcher.launcher(channel.getChannel(), response); ConsumerAckMessage result = (ConsumerAckMessage) request.getMsgParams(); if (result.getStatus() == ConsumerAckMessage.SUCCESS) { ConsumerContext.setClustersStat(task.getClusters(), ClustersState.SUCCESS); } } catch (Exception e) { ConsumerContext.setClustersStat(task.getClusters(), ClustersState.ERROR); } } } //若干個並行的執行緒共同到達統一的屏障點之後,再進行訊息統計,把資料最終彙總給JMX phaser.arriveAndAwaitAdvance(); return null; } }
Message Serialize:訊息的序列化模組,主要基於Kryo。其主要的核心程式碼為:com.newlandframework.avatarmq.serialize包下面的KryoCodecUtil、KryoSerialize完成訊息的序列化和反序列化工作。其對應的主要核心程式碼模組是:
package com.newlandframework.avatarmq.serialize; import com.esotericsoftware.kryo.pool.KryoPool; import io.netty.buffer.ByteBuf; import java.io.ByteArrayInputStream; import java.io.ByteArrayOutputStream; import java.io.IOException; /** * @filename:KryoCodecUtil.java * @description:KryoCodecUtil功能模組 * @author tangjie<https://github.com/tang-jie> * @blog http://www.cnblogs.com/jietang/ * @since 2016-8-11 */ public class KryoCodecUtil implements MessageCodecUtil { private KryoPool pool; public KryoCodecUtil(KryoPool pool) { this.pool = pool; } public void encode(final ByteBuf out, final Object message) throws IOException { ByteArrayOutputStream byteArrayOutputStream = null; try { byteArrayOutputStream = new ByteArrayOutputStream(); KryoSerialize kryoSerialization = new KryoSerialize(pool); kryoSerialization.serialize(byteArrayOutputStream, message); byte[] body = byteArrayOutputStream.toByteArray(); int dataLength = body.length; out.writeInt(dataLength); out.writeBytes(body); } finally { byteArrayOutputStream.close(); } } public Object decode(byte[] body) throws IOException { ByteArrayInputStream byteArrayInputStream = null; try { byteArrayInputStream = new ByteArrayInputStream(body); KryoSerialize kryoSerialization = new KryoSerialize(pool); Object obj = kryoSerialization.deserialize(byteArrayInputStream); return obj; } finally { byteArrayInputStream.close(); } } }
package com.newlandframework.avatarmq.serialize; import com.esotericsoftware.kryo.Kryo; import com.esotericsoftware.kryo.io.Input; import com.esotericsoftware.kryo.io.Output; import com.esotericsoftware.kryo.pool.KryoPool; import com.google.common.io.Closer; import java.io.IOException; import java.io.InputStream; import java.io.OutputStream; /** * @filename:KryoSerialize.java * @description:KryoSerialize功能模組 * @author tangjie<https://github.com/tang-jie> * @blog http://www.cnblogs.com/jietang/ * @since 2016-8-11 */ public class KryoSerialize { private KryoPool pool = null; private Closer closer = Closer.create(); public KryoSerialize(final KryoPool pool) { this.pool = pool; } public void serialize(OutputStream output, Object object) throws IOException { try { Kryo kryo = pool.borrow(); Output out = new Output(output); closer.register(out); closer.register(output); kryo.writeClassAndObject(out, object); pool.release(kryo); } finally { closer.close(); } } public Object deserialize(InputStream input) throws IOException { try { Kryo kryo = pool.borrow(); Input in = new Input(input); closer.register(in); closer.register(input); Object result = kryo.readClassAndObject(in); pool.release(kryo); return result; } finally { closer.close(); } } }
Netty Core:基於Netty對producer、consumer、broker的網路事件處理器(Handler)進行封裝處理,核心模組在:com.newlandframework.avatarmq.netty包之下。其中broker的Netty網路事件處理器為ShareMessageEventWrapper、producer的Netty網路事件處理器為MessageProducerHandler、consumer的Netty網路事件處理器為MessageConsumerHandler。其對應的類圖為:
可以看到,他們共同的父類是:MessageEventWrapper。該類的程式碼簡要說明如下:
package com.newlandframework.avatarmq.netty; import com.newlandframework.avatarmq.core.HookMessageEvent; import io.netty.channel.ChannelHandlerContext; import io.netty.channel.ChannelInboundHandlerAdapter; import org.springframework.aop.framework.ProxyFactory; import org.springframework.aop.support.NameMatchMethodPointcutAdvisor; /** * @filename:MessageEventWrapper.java * @description:MessageEventWrapper功能模組 * @author tangjie<https://github.com/tang-jie> * @blog http://www.cnblogs.com/jietang/ * @since 2016-8-11 */ public class MessageEventWrapper<T> extends ChannelInboundHandlerAdapter implements MessageEventHandler, MessageEventProxy { final public static String proxyMappedName = "handleMessage"; protected MessageProcessor processor; protected Throwable cause; protected HookMessageEvent<T> hook; protected MessageConnectFactory factory; private MessageEventWrapper<T> wrapper; public MessageEventWrapper() { } public MessageEventWrapper(MessageProcessor processor) { this(processor, null); } public MessageEventWrapper(MessageProcessor processor, HookMessageEvent<T> hook) { this.processor = processor; this.hook = hook; this.factory = processor.getMessageConnectFactory(); } public void handleMessage(ChannelHandlerContext ctx, Object msg) { return; } public void beforeMessage(Object msg) { } public void afterMessage(Object msg) { } //管道鏈路啟用 public void channelActive(ChannelHandlerContext ctx) throws Exception { super.channelActive(ctx); } //讀管道資料 public void channelRead(ChannelHandlerContext ctx, Object msg) throws Exception { super.channelRead(ctx, msg); ProxyFactory weaver = new ProxyFactory(wrapper); NameMatchMethodPointcutAdvisor advisor = new NameMatchMethodPointcutAdvisor(); advisor.setMappedName(MessageEventWrapper.proxyMappedName); advisor.setAdvice(new MessageEventAdvisor(wrapper, msg)); weaver.addAdvisor(advisor); //具體的如何處理管道中的資料,直接由producer、consumer、broker自行決定 MessageEventHandler proxyObject = (MessageEventHandler) weaver.getProxy(); proxyObject.handleMessage(ctx, msg); } //管道鏈路失效,可能網路連線斷開了,後續如果重連broker,可以在這裡做文章 public void channelInactive(ChannelHandlerContext ctx) throws Exception { super.channelInactive(ctx); } public void setWrapper(MessageEventWrapper<T> wrapper) { this.wrapper = wrapper; } }
整個AvatarMQ訊息佇列系統的執行情況,可以參考:Netty構建分散式訊息佇列(AvatarMQ)設計指南之架構篇,裡面說的很詳細了,本文就不具體演示了。
下圖是VisualVM監控AvatarMQ中broker伺服器的CPU使用率曲線。
可以發現,隨著訊息的堆積,broker進行訊息投遞、ACK應答的壓力增大,CPU的使用率明細提高。現在具體看下broker的CPU使用率增高的原因是呼叫哪個熱點方法呢?
從下圖可以看出,熱點方法是:SemaphoreCache的acquire。
這個是因為broker接收來自生產者訊息的同時,會先把訊息快取起來,然後利用多執行緒機制進行訊息的分派,這個時候會對訊號量維護的許可集合進行獲取操作,獲取成功之後,才能進行任務的派發,主要防止臨界區的共享資源競爭。這裡的Semaphore是用來控制多執行緒訪問共享資源(生產者過來的訊息),類似作業系統中的PV原語,P原語相當於acquire(),V原語相當於release()。
寫在最後
本文通過一個基於Netty構建分散式訊息佇列系統(AvatarMQ),簡單地闡述了一個極簡訊息中介軟體的內部結構、以及如何利用Netty,構建生產者、消費者訊息路由的通訊模組。一切都是從零開始,開發、實現出精簡版的訊息中介軟體!本系列文章的主要靈感源自,自己業餘時間,閱讀到的一些訊息佇列原理闡述文章以及相關開源訊息中介軟體的原始碼,其中也結合了自己的一些理解和體會。由於自身技術水平、理解能力方面的限制,不能可能擁有大師一樣高屋建瓴的視角,本文有說得不對、寫的不好的地方,懇請廣大同行批評指正。現在,文章寫畢,算是對自己平時學習的一些經驗總結,在這之前,對於訊息中介軟體都是很簡單的使用別人造好的輪子,沒有更多的深入瞭解背後的技術細節,只是單純的覺得別人寫的很強大、很高效。其實有的時候提升自己能力,要更多的深究其背後的技術原理,舉一反三,而不是簡單的蜻蜓點水,一味地點到為止,長此以往、日復一日,自身的技術積累就很難有質的飛躍。
AvatarMQ一定還有許多不足、瓶頸甚至是bug,確實它不是一個完美的訊息中介軟體,真因為如此,還需要不斷地進行重構優化。後續本人還會持續更新、維護這個開源專案,希望有興趣的朋友,共同關注!
文章略長,謝謝大家的觀賞,如果覺得不錯,還請多多推薦!
相關文章
- 大型網站架構系列:分散式訊息佇列(一)網站架構分散式佇列
- 分散式訊息佇列分散式佇列
- 架構文摘:訊息佇列設計精要架構佇列
- 訊息佇列系列一:訊息佇列應用佇列
- Netty 實現原理淺析Netty
- 打造企業級微服務平臺架構,分散式應用場景管理微服務架構分散式
- LAMMP架構的企業級應用架構
- 大型網站架構系列:訊息佇列(二)網站架構佇列
- MicrosoftNet企業級應用架構設計(中)ROS應用架構
- 訊息佇列在大型分散式系統中的實戰要點分析!佇列分散式
- 淺析Vite本地構建原理Vite
- 高併發架構訊息佇列面試題解析架構佇列面試題
- 淺談訊息佇列佇列
- 大型分散式網站架構:快取在分散式系統中的應用分散式網站架構快取
- 架構設計之NodeJS操作訊息佇列RabbitMQ架構NodeJS佇列MQ
- 分散式訊息佇列知識圖譜分散式佇列
- 快速理解Kafka分散式訊息佇列框架Kafka分散式佇列框架
- Redis實現訊息佇列Redis佇列
- 進階高階IoT架構-教你如何簡單實現一個訊息佇列架構佇列
- rabbitmq訊息佇列原理MQ佇列
- 分散式網站架構後續:zookeeper技術淺析分散式網站架構
- 分散式訊息佇列RocketMQ--事務訊息--解決分散式事務的最佳實踐分散式佇列MQ
- 企業實施分散式架構的挑戰以及應對建議 | 上海 ATEC 大會實錄分散式架構
- 企業實施分散式架構的挑戰以及應對建議|上海ATEC大會實錄分散式架構
- 淺析雲原生應用安全組織架構架構
- 基於訊息佇列 RocketMQ 的大型分散式應用上雲最佳實踐佇列MQ分散式
- .NET應用架構設計—重新認識分層架構(現代企業級應用分層架構核心設計要素)應用架構
- 訊息佇列之Kafka——從架構技術重新理解Kafka佇列Kafka架構
- JavaWeb專案架構之Redis分散式日誌佇列JavaWeb架構Redis分散式佇列
- 分散式訊息佇列:如何保證訊息的順序性分散式佇列
- 分散式之訊息佇列複習精講分散式佇列
- Zookeeper和Curator-Framework實踐之:分散式訊息佇列Framework分散式佇列
- Java開發微服務實現分散式架構應用總結Java微服務分散式架構
- 企業應用平臺架構架構
- H5-APP在企業系統中的架構應用H5APP架構
- 電商架構淺析架構
- Spring Cloud雲服務架構 - 企業分散式微服務雲架構構建SpringCloud架構分散式微服務
- redis訊息佇列簡單應用Redis佇列