一个 RabbitMQ 队列上的多个消费者

Question

我正在遵循 RabbitMQ 的本指南： https : //www.rabbitmq.com/tutorials/tutorial-two-java.html 。 我想用一个队列上的多个线程来模拟这个功能。

如果我在启动发送器之前启动接收器，它会按预期工作，如下所示：

[*] Rcvr1 Waiting for messages...
[*] Rcvr2 Waiting for messages...
[x] Rcvr1 Received 'Hello 0'
[x] Rcvr2 Received 'Hello 1'
[x] Rcvr1 Received 'Hello 2'
[x] Rcvr2 Received 'Hello 3'
[x] Rcvr1 Received 'Hello 4'
[x] Rcvr2 Received 'Hello 5'
[x] Rcvr1 Received 'Hello 6'
[x] Rcvr2 Received 'Hello 7'
[x] Rcvr1 Received 'Hello 8'
...

但是，首先启动我的接收器会导致只有一个线程接收消息（要启动的最后一个线程）：

[*] Rcvr2 Waiting for messages...
[*] Rcvr1 Waiting for messages...
[x] Rcvr1 Received 'Hello 9'
[x] Rcvr1 Received 'Hello 10'
[x] Rcvr1 Received 'Hello 11'
[x] Rcvr1 Received 'Hello 12'
[x] Rcvr1 Received 'Hello 13'
[x] Rcvr1 Received 'Hello 14'
[x] Rcvr1 Received 'Hello 15'
...

有趣的是，如果我启动发送器，然后启动接收器，如上所述，然后再次启动发送器（而接收器正在处理第一批）。 发送的第一条消息是串行处理的，而第二批消息是并行处理的，或者至少与剩余的线程一起处理。：

 [*] Rcvr1 Waiting for messages...
 [*] Rcvr2 Waiting for messages...
 [x] Rcvr1 Received '[Batch 1] Hello 0'
 [x] Rcvr1 Received '[Batch 1] Hello 1'
 [x] Rcvr1 Received '[Batch 1] Hello 2'
 [x] Rcvr1 Received '[Batch 1] Hello 3'
 [x] Rcvr1 Received '[Batch 1] Hello 4'
 [x] Rcvr1 Received '[Batch 1] Hello 5'
 [x] Rcvr1 Received '[Batch 1] Hello 6'
 [x] Rcvr1 Received '[Batch 1] Hello 7'
 [x] Rcvr1 Received '[Batch 1] Hello 8'
 [x] Rcvr2 Received '[Batch 2] Hello 1'
 [x] Rcvr1 Received '[Batch 1] Hello 9'
 [x] Rcvr2 Received '[Batch 2] Hello 3'
 [x] Rcvr1 Received '[Batch 1] Hello 10'
 [x] Rcvr2 Received '[Batch 2] Hello 5'
 [x] Rcvr1 Received '[Batch 1] Hello 11'
 [x] Rcvr2 Received '[Batch 2] Hello 7'
 [x] Rcvr1 Received '[Batch 1] Hello 12'
 [x] Rcvr2 Received '[Batch 2] Hello 9'
 [x] Rcvr1 Received '[Batch 1] Hello 13'
 [x] Rcvr2 Received '[Batch 2] Hello 11'

RabbitMQ 显然可以做到这一点，我不确定我做错了什么。 我的简单代码如下：

发件人

public class Send {

    private final static String QUEUE_NAME = "hello";

    public static void main(String[] argv) throws Exception {
        ConnectionFactory factory = new ConnectionFactory();
        factory.setHost("localhost");
        try (Connection connection = factory.newConnection();
             Channel channel = connection.createChannel()) {
            channel.queueDeclare(QUEUE_NAME, true, false, false, null);
            for(int x=0; x<100; x++) {
                String message = "Hello "+x;
                channel.basicPublish("", QUEUE_NAME, null, message.getBytes("UTF-8"));
                System.out.println(" [x] Sent '" + message + "'");
            }
        }
    }
}

接收者

package com.mawv.ingest.rabbitmq;

import com.rabbitmq.client.*;

import java.util.concurrent.Executors;
import java.util.concurrent.ThreadPoolExecutor;

public class Recv {

    private final static String QUEUE_NAME = "hello";

    public static void main(String[] argv) throws Exception {
        ThreadPoolExecutor rcvrPool = (ThreadPoolExecutor) Executors.newFixedThreadPool(10);
        ConnectionFactory factory = new ConnectionFactory();
        factory.setHost("localhost");

        Runnable rcvr1 = () -> {
            try {
                Connection connection = factory.newConnection();
                Channel channel = connection.createChannel();
                channel.queueDeclare(QUEUE_NAME, true, false, false, null);

                System.out.println(" [*] Rcvr1 Waiting for messages...");
                DeliverCallback deliverCallback = (consumerTag, delivery) -> {
                    Envelope envelope = delivery.getEnvelope();
                    String message = new String(delivery.getBody(), "UTF-8");
                    System.out.println(" [x] Rcvr1 Received '" + message + "'");
                    long deliveryTag = envelope.getDeliveryTag();
                    channel.basicAck(deliveryTag, true);
                    try {
                        Thread.sleep(1000);
                    } catch (Exception ex) { }

                };
                channel.basicConsume(QUEUE_NAME, false, deliverCallback, consumerTag -> {  });

            } catch(Exception ex){
                ex.printStackTrace();
            }
        };
        Runnable rcvr2 = () -> {
            try {
                Connection connection = factory.newConnection();
                Channel channel = connection.createChannel();
                channel.queueDeclare(QUEUE_NAME, true, false, false, null);

                System.out.println(" [*] Rcvr2 Waiting for messages...");
                DeliverCallback deliverCallback = (consumerTag, delivery) -> {
                    Envelope envelope = delivery.getEnvelope();
                    String message = new String(delivery.getBody(), "UTF-8");
                    System.out.println(" [x] Rcvr2 Received '" + message + "'");
                    long deliveryTag = envelope.getDeliveryTag();
                    channel.basicAck(deliveryTag, true);
                    try {
                        Thread.sleep(1000);
                    } catch (Exception ex) {
                    }
                };
                channel.basicConsume(QUEUE_NAME, false, deliverCallback, consumerTag -> {
                });
            } catch(Exception ex){
                ex.printStackTrace();
            }
        };
        rcvrPool.execute(rcvr1);
        rcvrPool.execute(rcvr2);

    }
}

我也绑定了这个例子，正如他们描述的那样，看到了相同的结果。 https://self-learning-java-tutorial.blogspot.com/2015/09/rabbitmq-one-producer-and-multiple.html

我假设我的设置有问题。

Answer 1

根据RabbitMQ api：

“虽然一个 Channel 可以被多个线程使用，但确保一次只有一个线程执行一个命令很重要。同时执行命令可能会导致抛出 UnexpectedFrameError”

首先，我认为您应该为不同的线程使用不同的通道。

最后，我认为第一个线程因为空闲而终止，所以只有第二个线程处于活动状态并完成整个工作。 在这种情况下，一根线程就足够了。

看看 java 8 的 ThreadPoolExecutor api：

https://docs.oracle.com/javase/8/docs/api/java/util/concurrent/ThreadPoolExecutor.html

例如，您可以找到：

“默认情况下，即使是核心线程也只有在新任务到达时才最初创建和启动，但这可以使用方法 prestartCoreThread() 或 prestartAllCoreThreads() 动态覆盖。如果您使用非空构造池，您可能想要预启动线程队列”

和

“如果池中当前有超过 corePoolSize 的线程，如果空闲时间超过 keepAliveTime（请参阅 getKeepAliveTime(TimeUnit)），多余的线程将被终止。”

您应该使用 prestartAllCoreThreads() 或 prestartCoreThreads() 使核心线程即使在空闲时也能启动，或者使用 getKeepAliveTime(TimeUnit) 使它们即使空闲也能保持活动状态。

Answer 2

看起来我缺少一个关键的通道配置。 这解决了我的问题：

channel.basicQos(1);

这就是 RabbitMQ 所说的。

公平调度

您可能已经注意到调度仍然不能完全按照我们想要的方式工作。 例如，在有两个 worker 的情况下，当所有奇数消息都很重，偶数消息很轻时，一个 worker 会一直很忙，而另一个 worker 几乎不做任何工作。 好吧，RabbitMQ 对此一无所知，仍然会均匀地发送消息。

发生这种情况是因为 RabbitMQ 只是在消息进入队列时分派消息。 它不考虑消费者未确认消息的数量。 它只是盲目地将每条第 n 条消息分派给第 n 条消费者。

为了解决这个问题，我们可以使用带有 prefetchCount = 1 设置的 basicQos 方法。 这告诉 RabbitMQ 一次不要给一个工人多个消息。 或者，换句话说，在处理并确认前一条消息之前，不要向工作人员发送新消息。 相反，它会将它分派给下一个不忙的工人。