kafka rebalance机制与Consumer多种消费模式案例

rebalance 何时触发？到底干嘛？流程如何？

reblance 何时触发

组订阅发生变更，比如基于正则表达式订阅，当匹配到新的topic创建时，组的订阅就会发生变更。
组的topic分区数发生变更，通过命令行脚本增加了订阅topic的分区数。
组成员发生变更：新加入组以及离开组。

reblance 到底干嘛

一句话：多个Consumer订阅了一个Topic时，根据分区策略进行消费者订阅分区的重分配

Coordinator 到底在那个Broker

找到Coordinator的算法与找到_consumer_offsets目标分区的算法是一致的。

第一步：确定目标分区：Math.abs(groupId.hashCode)%50，假设是12。
第二步：找到_consumer_offsets分区为10的Leader副本所在的Broker，那么该broker即为Group Coordinator。

reblance 流程如何

reblance 流程流程整体如下图所示，值得强调的几点如下：

Coordinator的角色由Broker端担任。
Group Leader 的角色主要有Consumer担任。
加入组请求（JoinGroup）=>作用在于选择Group Leader。
同步组请求（SyncGroup）=>作用在于确定分区分配方案给Coordinator，把方案响应给所有Consumer。

reblance 机制的好处

分区分配权利下放给客户端consumer，因此系统不用重启，既可以实现分区策略的变更。
用户可以自行实现机架感知分配方案。

reblance generation 过滤无用请求

kafka引入 reblance generation ，就是为了防止Consumer group的无效Offset提交。若因为某些原因，consumer延迟提交了Offset，而该consumer被踢出了消费组，那么该Consumer再次提交位移时，携带的就是旧的generation了。

reblance 监听器应用级别实战

reblance 监听器解决用户把位移提交到外部存储的情况，在监听器中实现位移保存和位移的重定向。
onPartitionsRevoked : rebalance开启新一轮的重平衡前会调用，一般用于手动提交位移，及审计功能

onPartitionsAssigned ：rebalance在重平衡结束后会调用，一般用于消费逻辑处理

 Properties props = new Properties();
 props.put("bootstrap.servers", "localhost:9092");
 props.put("group.id", "test");
 props.put("enable.auto.commit", "false");
 props.put("key.deserializer", "org.apache.kafka.common.serialization.StringDeserializer");
 props.put("value.deserializer", "org.apache.kafka.common.serialization.StringDeserializer");
 KafkaConsumer<String, String> consumer = new KafkaConsumer<>(props);
 
 统计rebalance总时长
 final AtomicLong totalRebalanceTimeMs =new AtomicLong(0L)
 
 统计rebalance开始时刻
 final AtomicLong rebalanceStart =new AtomicLong(0L)
 
 
 1 重平衡监听
 consumer.subscribe(Arrays.asList("test-topic"), new ConsumerRebalanceListener(){
     
 @Override
    public void onPartitionsRevoked(Collection<TopicPartition> partitions) {
        for(TopicPartition tp : partitions){
        
            1 保存到外部存储
            saveToExternalStore(consumer.position(tp)) 
            
            2 手动提交位移
            //consumer.commitSync(toCommit);
        }
        
        rebalanceStart.set(System.currentTimeMillis())
    }
}

@Override
public void onPartitionsAssigned(Collection<TopicPartition> partitions) {

    totalRebalanceTimeMs.addAndGet(System.currentTimeMillis()-rebalanceStart.get())
    
    for (TopicPartition tp : partitions) {
           
           consumer.seek(tp，readFromExternalStore(tp))
    }
}
 });

2 消息处理
while (true) {
 ConsumerRecords<String, String> records = consumer.poll(100);
 for (ConsumerRecord<String, String> record : records) {
     buffer.add(record);
 }
 if (buffer.size() >= minBatchSize) {
     insertIntoDb(buffer);
     consumer.commitSync();
     buffer.clear();
 }
}

Consumer组内消息均衡实战

Consumer 单线程封装，实现多个消费者来消费（浪费资源）

实例主题：

ConsumerGroup 实现组封装
ConsumerRunnable 每个线程维护私有的KafkaConsumer实例

public class  Main {

    public static void main(String[] args) {
        String brokerList = "localhost:9092";
        String groupId = "testGroup1";
        String topic = "test-topic";
        int consumerNum = 3;

        核心对外封装
        ConsumerGroup consumerGroup = new ConsumerGroup(consumerNum, groupId, topic, brokerList);
        consumerGroup.execute();
    }
}

import java.util.ArrayList;
import java.util.List;
public class ConsumerGroup {    

    private List<ConsumerRunnable> consumers;
    
    public ConsumerGroup(int consumerNum, String groupId, String topic, String brokerList) {
        consumers = new ArrayList<>(consumerNum);
        for (int i = 0; i < consumerNum; ++i) {
            ConsumerRunnable consumerThread = new ConsumerRunnable(brokerList, groupId, topic);
            consumers.add(consumerThread);
        }
    }
    public void execute() {
        for (ConsumerRunnable task : consumers) {
            new Thread(task).start();
        }
    }
}

import org.apache.kafka.clients.consumer.ConsumerRecord;
import org.apache.kafka.clients.consumer.ConsumerRecords;
import org.apache.kafka.clients.consumer.KafkaConsumer;
import java.util.Arrays;
import java.util.Properties;

public class ConsumerRunnable implements Runnable {

    private final KafkaConsumer<String, String> consumer;

    public ConsumerRunnable(String brokerList, String groupId, String topic) {
        Properties props = new Properties();
        props.put("bootstrap.servers", brokerList);
        props.put("group.id", groupId);
        props.put("enable.auto.commit", "true");        //本例使用自动提交位移
        props.put("auto.commit.interval.ms", "1000");
        props.put("session.timeout.ms", "30000");
        props.put("key.deserializer", "org.apache.kafka.common.serialization.StringDeserializer");
        props.put("value.deserializer", "org.apache.kafka.common.serialization.StringDeserializer");
        this.consumer = new KafkaConsumer<>(props);
        consumer.subscribe(Arrays.asList(topic));   // 本例使用分区副本自动分配策略
    }

    @Override
    public void run() {
        while (true) {
            ConsumerRecords<String, String> records = consumer.poll(200);   
            for (ConsumerRecord<String, String> record : records) {
                System.out.println(Thread.currentThread().getName() + " consumed " + record.partition() +
                        "th message with offset: " + record.offset());
            }
        }
    }
}

一个Consumer，内部实现多线程消费（consumer压力过大）

实例主题：

ConsumerHandler 单一的Consumer实例，poll后里面会跑一个线程池，执行多个Processor线程来处理
Processor 业务逻辑处理方法

进一步优化建议；

ConsumerHandler 设置手动提交位移，负责最终位移提交consumer.commitSync();。
ConsumerHandler设置一个全局的Map<TopicPartion,OffsetAndMetadata> offsets，来管理Processor消费的位移。
Processor 负责批处理完消息后，得到消息的最大位移，并更新offsets数组
ConsumerHandler 根据 offsets，位移提交后会清空offsets集合。
ConsumerHandler设置重平衡监听

public class Main {
    public static void main(String[] args) {
    
        String brokerList = "localhost:9092,localhost:9093,localhost:9094";
        String groupId = "group2";
        String topic = "test-topic";
        int workerNum = 5;

        ConsumerHandler consumers = new ConsumerHandler(brokerList, groupId, topic);
        consumers.execute(workerNum);
        try {
            Thread.sleep(1000000);
        } catch (InterruptedException ignored) {}
        consumers.shutdown();
    }
}

import java.util.Arrays;
import java.util.Properties;
import java.util.concurrent.ArrayBlockingQueue;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.ThreadPoolExecutor;
import java.util.concurrent.TimeUnit;

public class ConsumerHandler {

    private final KafkaConsumer<String, String> consumer;
    private ExecutorService executors;

    public ConsumerHandler(String brokerList, String groupId, String topic) {
        Properties props = new Properties();
        props.put("bootstrap.servers", brokerList);
        props.put("group.id", groupId);
        props.put("enable.auto.commit", "true");
        props.put("auto.commit.interval.ms", "1000");
        props.put("session.timeout.ms", "30000");
        props.put("key.deserializer", "org.apache.kafka.common.serialization.StringDeserializer");
        props.put("value.deserializer", "org.apache.kafka.common.serialization.StringDeserializer");
        consumer = new KafkaConsumer<>(props);
        consumer.subscribe(Arrays.asList(topic));
    }
    public void execute(int workerNum) {
    
        executors = new ThreadPoolExecutor(workerNum, workerNum, 0L, TimeUnit.MILLISECONDS,
                new ArrayBlockingQueue<>(1000), new ThreadPoolExecutor.CallerRunsPolicy());
                
        while (true) {
            ConsumerRecords<String, String> records = consumer.poll(200);
            
            for (final ConsumerRecord record : records) {
                executors.submit(new Processor(record));
            }
            
            
        }
    }
    public void shutdown() {
        if (consumer != null) {
            consumer.close();
        }
        if (executors != null) {
            executors.shutdown();
        }
        try {
            if (!executors.awaitTermination(10, TimeUnit.SECONDS)) {
                System.out.println("Timeout.... Ignore for this case");
            }
        } catch (InterruptedException ignored) {
            System.out.println("Other thread interrupted this shutdown, ignore for this case.");
            Thread.currentThread().interrupt();
        }
    }

}

import org.apache.kafka.clients.consumer.ConsumerRecord;
public class Processor implements Runnable {

    private ConsumerRecord<String, String> consumerRecord;

    public Processor(ConsumerRecord record) {
        this.consumerRecord = record;
    }

    @Override
    public void run() {
        System.out.println(Thread.currentThread().getName() + " consumed " + consumerRecord.partition()
            + "th message with offset: " + consumerRecord.offset());
    }
}

方案对比

第一种方案：建议采用 Consumer 单线程封装，实现多个消费者来消费（浪费资源），这样能很好地保证分区内消费的顺序，同时也没有线程切换的开销。
第二种方案：实现复杂，问题在于可能无法维护分区内的消息顺序，注意消息处理和消息接收解耦了。

Consumer指定分区消费案例实战（Standalone Consumer）

Standalone Consumer assign 用于接收指定分区列表的消息和Subscribe是矛盾的。只能二选一。
多个 Consumer 实例消费一个 Topic 借助于 group reblance可谓是天作之合。

若要精准控制，assign逃不了。

poperties props = new Properties();
props.put("bootstrap.servers", brokerList);
props.put("group.id", groupId);
props.put("enable.auto.commit", "false");
props.put("auto.commit.interval.ms", "1000");
props.put("session.timeout.ms", "30000");
props.put("key.deserializer", "org.apache.kafka.common.serialization.StringDeserializer");
props.put("value.deserializer", "org.apache.kafka.common.serialization.StringDeserializer");
consumer = new KafkaConsumer<>(props);

List<TopicPartion> partitions = new ArrayList<>();

List<PartitionInfo>  allPartitions = consumer.partitionsFor("kaiXinTopic")

if(allPartitions != null && !allPartitions.isEmpty){
    for(PartitionInfo partitionInfo : allPartitions){
    
        partitions.add(new TopicPartition(partitionInfo.topic(),partitionInfo.partition()))
    }
    
    consumer.assign(partitions)
}

while (true) {
   ConsumerRecords<String, String> records = consumer.poll(100);
   for (ConsumerRecord<String, String> record : records) {
       buffer.add(record);
   }
   if (buffer.size() >= minBatchSize) {
       insertIntoDb(buffer);
       
       consumer.commitSync();
       
       buffer.clear();
}