使用 2 个消费者配置时消费者吞吐量缓慢 [英] Slow consumer throuput when using 2 consumer-configuration

问题描述

使用 spring-integration-kafka 扩展和以下配置:

Using the spring-integration-kafka extention and the following configuration:

<int-kafka:zookeeper-connect id="zookeeperConnect"
    zk-connect="#{kafkaConfig['zooKeeperUrl']}" zk-connection-timeout="10000"
    zk-session-timeout="10000" zk-sync-time="2000" />

<int-kafka:consumer-context id="consumerContext" consumer-timeout="5000" zookeeper-connect="zookeeperConnect">
    <int-kafka:consumer-configurations>
        <int-kafka:consumer-configuration
                group-id="realtime-services-consumer-grp" 
                value-decoder="purchaseDecoder" 
                key-decoder="kafkaReflectionDecoder"
                max-messages="5" >
            <int-kafka:topic id="purchase" streams="1" />
        </int-kafka:consumer-configuration>
        <int-kafka:consumer-configuration 
                group-id="realtime-services-consumer-gw"
                value-decoder="eventDecoder" 
                key-decoder="kafkaReflectionDecoder" 
                max-messages="10" >
            <int-kafka:topic id="event" streams="1" />
        </int-kafka:consumer-configuration>
    </int-kafka:consumer-configurations>
</int-kafka:consumer-context>

<int-kafka:inbound-channel-adapter
    id="kafkaInboundChannelAdapter" kafka-consumer-context-ref="consumerContext"
    auto-startup="true" channel="inputFromKafka">
    <int:poller fixed-delay="20" time-unit="MILLISECONDS" />
</int-kafka:inbound-channel-adapter>

例如，当我评论第一个 consumer-configuration 时，我每分钟可以有 300 个事件而没有问题.但是当两者都被激活时.我的吞吐量非常低.来自这两个主题的总吞吐量低于每分钟 50.

When for example I comment the first consumer-configuration I can have 300 event per minute with no issue. But when both are activated. I have a very low throughput. Total of throughput coming from both topic is less than a 50 per minutes.

有谁知道为什么我在阅读 2 个主题时表现如此糟糕?我在配置中做错了什么?

Anyone know why I have such poor performance when reading from 2 topic? What did I do wrong in the configuration?

推荐答案

感谢您指出这一点！

在与我当地的 Kafka claster 进行一番激烈斗争后，我已经能够重现您的问题，并且我为您提供了一些解决方法:-).

After some big fight with my local Kafka claster I have been able to reproduce your issue and I am with some workaround for you :-).

首先不是round-robin，而是一个接一个:

First of all it isn't round-robin, but one by one:

for (final ConsumerConfiguration<K, V> consumerConfiguration : getConsumerConfigurations().values()) {
    Map<String, Map<Integer, List<Object>>> messages = consumerConfiguration.receive();

如果 KafkaStream 中没有消息，在 consumer-timeout="5000" 期间，每个 consumerConfiguration 在后台被阻止代码>现在.因此，来自 <int-kafka:inbound-channel-adapter> 的整个 poll 任务被阻塞，直到超时或更糟:如果每个主题没有消息整个等待超时是超时的总和！

Where each of those consumerConfiguration are blocked on the background during that consumer-timeout="5000", if there is no messages in the KafkaStream right now. Hence entire poll task from the <int-kafka:inbound-channel-adapter> is blocked until that timeout or even worse: if each topic doesn't have message the whole waiting time out is sum of timeouts!

要解决这个问题，您可以减少 consumer-timeout="5000" 或提供几个 和 用于每个主题.

To overcome the issue, you can decrease consumer-timeout="5000" or provide several <int-kafka:consumer-context> and therefore <int-kafka:inbound-channel-adapter> for each topic.

是的，这看起来很奇怪，而且在发布之前我们还没有找到时间来看看这真的很糟糕，但无论如何请随时提出 JIRA 问题来修复它.

Yes, it looks odd and it is really bad that we haven't found a time to take a look to this before release, but anyway feel free to raise JIRA issue to fix it.

谢谢！

这篇关于使用 2 个消费者配置时消费者吞吐量缓慢的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！