Kafka 消费组消费者分配策略

时间 2021-01-01

标签 javascript php css html java python nginx web apache django 栏目 Kafka 繁體版

原文原文链接

微信公众号：苏言论
理论联系实际，畅言技术与生活。javascript

消费组和消费者是kafka中比较重要的概念，理解和掌握原理有利于优化kafka性能和处理消费积压问题。Kafka topic 由多个分区组成，分区分布在集群节点上；php

Topic:topic01  PartitionCount:10       ReplicationFactor:2     Configs:
        Topic: topic01 Partition: 0    Leader: 1       Replicas: 1,4   Isr: 1,4
        Topic: topic01 Partition: 1    Leader: 2       Replicas: 2,5   Isr: 2,5
        Topic: topic01 Partition: 2    Leader: 3       Replicas: 3,1   Isr: 3,1
        Topic: topic01 Partition: 3    Leader: 4       Replicas: 4,2   Isr: 4,2
        Topic: topic01 Partition: 4    Leader: 5       Replicas: 5,3   Isr: 5,3
        Topic: topic01 Partition: 5    Leader: 1       Replicas: 1,3   Isr: 1,3
        Topic: topic01 Partition: 6    Leader: 2       Replicas: 2,4   Isr: 2,4
        Topic: topic01 Partition: 7    Leader: 3       Replicas: 3,5   Isr: 3,5
        Topic: topic01 Partition: 8    Leader: 4       Replicas: 4,1   Isr: 4,1
        Topic: topic01 Partition: 9    Leader: 5       Replicas: 5,2   Isr: 5,2

当外部程序消费topic数据时，kafka将其视为消费组(ConsumerGroup)，每一个消费组包含1个或多个消费者(Consumer)，消费者数量最多能够为分区总数量，并非能够无限量。当消费组中的任意一个消费者终止时，kafka会对消费组进行平衡(Rebalance)，再根据存活消费数和消费者分配策略从新分配消费者。在0.10.x版本中，kafka提供两种分配策略(RangeAssignor、RoundRobinAssignor)，0.11.x 版本新增策略(StickyAssignor)，结构以下；css

1 RangeAssignor 策略

RangeAssignor 以主题为单位，以数据顺序排列可用分区，以字典顺序排列消费者，将topic分区数除以消费者总数，以肯定分配给每一个消费者的分区数；若是没有平均分配，那么前几个消费者将拥有一个额外的分区。实现代码；html

for (String memberId : subscriptions.keySet())
            assignment.put(memberId, new ArrayList<TopicPartition>());

        for (Map.Entry<String, List<String>> topicEntry : consumersPerTopic.entrySet()) {
            String topic = topicEntry.getKey();
            List<String> consumersForTopic = topicEntry.getValue();

            Integer numPartitionsForTopic = partitionsPerTopic.get(topic);
            if (numPartitionsForTopic == null)
                continue;

            Collections.sort(consumersForTopic);

            int numPartitionsPerConsumer = numPartitionsForTopic / consumersForTopic.size();  //topic分区数除以消费者总数
            int consumersWithExtraPartition = numPartitionsForTopic % consumersForTopic.size();  //计算额外分区

            List<TopicPartition> partitions = AbstractPartitionAssignor.partitions(topic, numPartitionsForTopic);
            for (int i = 0, n = consumersForTopic.size(); i < n; i++) {
                int start = numPartitionsPerConsumer * i + Math.min(i, consumersWithExtraPartition);
                int length = numPartitionsPerConsumer + (i + 1 > consumersWithExtraPartition ? 0 : 1);
                assignment.get(consumersForTopic.get(i)).addAll(partitions.subList(start, start + length));
            }
        }

好比有两个topic(topic1 ,topic2) ，每一个topic都有三个分区；java

topic1 ,分区：topic1p0,topic1p1,topic1p2
topic2 ,分区：topic2p0,topic2p1,topic2p2

和一个消费组(consumer_group1)，有(consumer1,consumer2)两个消费者，使用RangeAssignor策略可能会获得以下的分配：python

consumer1: [topic1p0,topic1p1,topic2p0,topic2p1]
consumer2: [topic1p2,topic2p2]

若是此时消费组(consumer_group1)有新的消费者consumer3加入，使用RangeAssignor策略可能会获得以下的分配：nginx

consumer1: [topic1p0,topic2p0]
consumer2: [topic1p2,topic2p2]
consumer3: [topic1p1,topic2p1]

2 RoundRobinAssignor 策略

RoundRobinAssignor 是kafka默认策略，对全部分区和全部消费者循环分配，分区更均衡；实现代码；web

Map<String, List<TopicPartition>> assignment = new HashMap<>();
        for (String memberId : subscriptions.keySet())
            assignment.put(memberId, new ArrayList<TopicPartition>());

        CircularIterator<String> assigner = new CircularIterator<>(Utils.sorted(subscriptions.keySet()));
        for (TopicPartition partition : allPartitionsSorted(partitionsPerTopic, subscriptions)) {
            final String topic = partition.topic();
            while (!subscriptions.get(assigner.peek()).topics().contains(topic))
                assigner.next();
            assignment.get(assigner.next()).add(partition);
        }

继续以上例topic和消费组为例，RoundRobinAssignor 策略可能会获得以下的分配；apache

consumer1: [topic1p0,topic1p1,topic2p2,]
consumer2: [topic2p0,topic2p1,topic1p2]

3 StickyAssignor 策略

StickyAssignor 策略是最复杂且是0.11.x 版本出现的新策略，该策略主要做用：django

使topic分区分配尽量均匀的分配给消费者
当某个消费者终止触发从新分配时，尽量保留现有分配，将已经终止的消费者所分配的分区移动到另外一个消费者，避免所有分区从新平衡，节省开销。

这个策略自0.11.x 版本出现后，一直到新版本有不一样bug被发现，低版本慎用。

4 java多线程消费实例

public class KafkaTopicConsumer {   
    private KafkaConsumer<String, String> consumer;
    private int consumerId=0; //消费实例id
    private final long timeOut=10000;
    public KafkaTopicConsumer(int consumerId){
        this.consumerId=consumerId;
        Properties props = new Properties();
        props.put("client.id", "client-" + consumerId);
        props.put("bootstrap.servers","192.168.1.10:9092,192.168.1.11:9092");
        props.put("group.id", "test-group03");
        props.put("zookeeper.session.timeout.ms", "4000");
        props.put("zookeeper.sync.time.ms", "200");
        props.put("enable.auto.commit", "false");
        props.put("auto.offset.reset", "earliest");
        props.put("key.deserializer", "org.apache.kafka.common.serialization.StringDeserializer");
        props.put("value.deserializer", "org.apache.kafka.common.serialization.StringDeserializer");
        //设置分区策略
        props.put("partition.assignment.strategy", "org.apache.kafka.clients.consumer.RangeAssignor");
        consumer = new KafkaConsumer<String, String>(props);
        consumer.subscribe(Arrays.asList("topic1","topic2"));
    }

    public void consume() {
        while (true){
            ConsumerRecords<String, String> records=consumer.poll(timeOut);
            System.out.println("records count:"+records.count());
            for (ConsumerRecord<String, String> record : records) {
              System.out.println(String.format("client-id = %d , topic = %s, partition = %d , offset = %d, key = %s, value = %s", this.consumerId,record.topic(), record.partition(), record.offset(), record.key(), record.value()));
            }
            consumer.commitSync();
        }
    }

    public static void main(String[] args) {
        int threadSize=Integer.parseInt(args[0]);
        for (int i = 0; i < threadSize; i++) {
            int id = i;
            new Thread() {
                @Override
                public void run() {
                    new KafkaTopicConsumer(id).consume();
                }
            }.start();
        }
    }//
}

启动三个多线程实例消费，分区分配到每一个消费者的状况；

GROUP                          TOPIC                          PARTITION  CURRENT-OFFSET  LOG-END-OFFSET  LAG             OWNER
test-group03                   topic2                0          0               3333            3333            client-0_/192.168.1.13
test-group03                   topic1                0          500             3333            2833            client-0_/192.168.1.13
test-group03                   topic2                2          0               3333            3333            client-2_/192.168.1.13
test-group03                   topic1                2          500             3333            2833            client-2_/192.168.1.13
test-group03                   topic2                1          500             3334            2834            client-1_/192.168.1.13
test-group03                   topic1                1          0               3334            3334            client-1_/192.168.1.13

对于大的topic，将topic单独消费以免数据积压和topic各自影响数据处理速度，好比文章开始时提到的10分区的topic(topic01)，根据硬件资源和分区策略设置合理的消费者，数据量大时最优的消费者数量为分区总数。

GROUP                          TOPIC                          PARTITION  CURRENT-OFFSET  LOG-END-OFFSET  LAG             OWNER
test-group02                   topic01                       6          373460          1026328         652868          client-6_/192.168.1.13
test-group02                   topic01                       2          375660          1048756         673096          client-2_/192.168.1.13
test-group02                   topic01                       5          374625          1013157         638532          client-5_/192.168.1.13
test-group02                   topic01                       3          347001          1066967         719966          client-3_/192.168.1.13
test-group02                   topic01                       0          375570          1013261         637691          client-0_/192.168.1.13
test-group02                   topic01                       9          376545          1094088         717543          client-9_/192.168.1.13
test-group02                   topic01                       8          347082          1066948         719866          client-8_/192.168.1.13
test-group02                   topic01                       7          375100          1048827         673727          client-7_/192.168.1.13
test-group02                   topic01                       1          372447          1026467         654020          client-1_/192.168.1.13
test-group02                   topic01                       4          377052          1093926         716874          client-4_/192.168.1.13

5 总结

Kafka提供三种分配策略(RangeAssignor、RoundRobinAssignor、StickyAssignor)，其中StickyAssignor策略是0.11.x 版本新增的，每种策略不尽相同，RangeAssignor策略以主题为单位，以数据顺序排列可用分区，以字典顺序排列消费者计算分配；RoundRobinAssignor 对全部分区和全部消费者循环均匀分配；但这两种分配策略当有消费者终止或加入时均会触发消费组平衡；StickyAssignor 策略当某个消费者终止时，尽量保留现有分配，将已经终止的消费者所分配的分区移动到另外一个消费者，避免所有分区从新平衡，节省开销；对于topic分区数较多、数量较大使用StickyAssignor策略有较大优点。

参考文献

https://kafka.apache.org/0100/javadoc - kafka 0.10.0.1 API
https://kafka.apache.org/0100/documentation.html - kafka DOCUMENTATION