Kafka 0.9.0 New Java Consumer API fetching duplicate records

问题

I am new to kafka and i am trying to prototype a simple consumer-producer message queue (traditional queue) model using Apache kafka 0.9.0 Java clients.

From the producer process, i am pushing 100 random messages to a topic configured with 3 partitions. This looks fine.

I created 3 consumer threads with same group id, subscribed to the same topic. auto commit enabled. Since all 3 consumer threads are subscribed to same topic i assume that each consumer will get a partition to consume and will commit the offset logs per partition.

But i am facing weird problem here. all my messages are duplicated. i get x time more records at consumer side from each of my thread. Since each of my consumer thread does infinite loop to poll from topic i have to kill the process.

I even tried with single thread and still i get duplicate records x times and still continues.

Could any please help me identify what mistake i am doing here.

I am posting my consumer code for your reference.

public class ConsumerDemo {

public static void main(String[] args) {

    ThreadFactory threadFactory = new ThreadFactoryBuilder().setNameFormat("Consumer-%d").build();
    ExecutorService executor = Executors.newFixedThreadPool(3, threadFactory);

    executor.submit(new ConsumerThread("topic1", "myThread-1"));
    executor.submit(new ConsumerThread("topic1", "myThread-2"));
    executor.submit(new ConsumerThread("topic1", "myThread-3"));

    //executor shutdown logic is skipped
}
}

Consumer Thread:

public class ConsumerThread  implements Runnable {

private static final String KAFKA_BROKER = "<<IP:port>>";

private final KafkaConsumer<String, String> consumer;

    public ConsumerThread(String topic, String name) {
        Properties props = new Properties();
        props.put("bootstrap.servers", ConsumerThread.KAFKA_BROKER);
        props.put("group.id", "DemoConsumer");
        props.put("enable.auto.commit", "true");
        props.put("auto.commit.interval.ms", "6000");
        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);
        this.consumer.subscribe(Collections.singletonList(topic));
    }


    public void run() {
        try {
            boolean isRunning = true;
            while (isRunning) {
                ConsumerRecords<String,String> records= consumer.poll(10L);
                System.out.println("Partition Assignment to this Consumer: "+consumer.assignment());
                Iterator it = records.iterator();
                while(it.hasNext()) {
                    ConsumerRecord record = (ConsumerRecord)it.next();
                    System.out.println("Received message from thread : "+Thread.currentThread().getName()+"(" + record.key() + ", " + (String)record.value() + ") at offset " + record.offset());
                }
            }
            consumer.close();
        }
        catch (Exception e) {
            e.printStackTrace();
        }
    }
}

Also very importantly, i am aiming for exactly once semantic. I know am 1000 mile away for that. Any help is really appreciated.

Observation: Debug sysout prints all 3 tpoics. Does this mean that partitions are not assigned to each consumer?

Partition Assignment to this Consumer: [topic1-1, topic1-0, topic1-2]

Kafka experts, apart from above problem i am looking 2 other inputs.

Please help me understand what is wrong in above code.
In general, how exactly once schematic can be implemented. Example if possible.
Exception scenarios like consumer down. how to handle without loss of message.

Thanks in advance.

回答1:

Well I figured out whats wrong with my code/undertanding.

I should have read Kafka documentation completely before i jump start to prototyping.

Here is what i found.

By default Kafka guarantees at least once schematic. This means consumer gets the message at least once (possibly multiple times. I assumed that if i have 3 partitions and create 3 consumer, Kafka API will take care of assigning a partition to only one consumer randomly. This is wrong.

So i manually assigned a partition to each consumer to make sure one that my consumer owns the partition and controls the offset like below

consumer = new KafkaConsumer(props)    
TopicPartition partition = new TopicPartition(topic, partitionNum);
consumer.assign(Collections.singletonList(partition));

Exactly once scenario: To make sure we consume a message exaclty once, we need to have control on the offset. Although i have not tried so far but based on what i learnt from lot of googling is that its better way to save the offset along with the data. Preferably same transaction. Both data and offset are definitely saved or rolled back for retry.

Any other solutions are appreciated.

回答2:

Are you consuming your messages slower than the session timeout? In that case, you risk rebalancing occur which might lead to double consumption.

回答3:

Rule out duplication due to producer retries or retries within your application layer.

On the producer side, if "request.timeout.ms" is not appropriately configured for your network/cluster, then it is possible that due to a slow start (initializing the producer, connection establishment etc), initial requests will time out on the producer, but would have been actually processed by the broker/server. This will lead to duplication on retry.

来源：https://stackoverflow.com/questions/34405124/kafka-0-9-0-new-java-consumer-api-fetching-duplicate-records

标签

apache-kafka

kafka-consumer-api