具有手動提交偏移量的kafka多線程消費者：KafkaConsumer對於多線程訪問是不安全的

Question

我使用ArrayBlockingQueue將 Kafka 消費者與接收器分離：

1. Kafka的多線程消費，每個線程一個kafka消費者；
2. Kafka 消費者手動管理偏移量；
3. Kafka 消費者將消息內容和包含 OFFSET 的回調 function 包裝到 Record object 中，並將其發送到ArrayBlockingQueue ；
4. Sink 從ArrayBlockingQueue獲取記錄並處理它。 Sink只有在處理成功后，才會調用Record object的回調function（通知Kafka消費者commitSync）

操作過程中遇到了一個錯誤，困擾了我好幾天。 我不明白問題的哪一部分是錯誤的：

11:44:10.794 [pool-2-thread-1] ERROR com.alibaba.kafka.source.KafkaConsumerRunner - [pool-2-thread-1] ConcurrentModificationException: KafkaConsumer is not safe for multi-threaded access
java.util.ConcurrentModificationException: KafkaConsumer is not safe for multi-threaded access
    at org.apache.kafka.clients.consumer.KafkaConsumer.acquire(KafkaConsumer.java:1824)
    at org.apache.kafka.clients.consumer.KafkaConsumer.acquireAndEnsureOpen(KafkaConsumer.java:1808)
    at org.apache.kafka.clients.consumer.KafkaConsumer.commitSync(KafkaConsumer.java:1255)
    at com.alibaba.kafka.source.KafkaConsumerRunner$1.call(KafkaConsumerRunner.java:75)
    at com.alibaba.kafka.source.KafkaConsumerRunner$1.call(KafkaConsumerRunner.java:71)
    at com.alibaba.kafka.sink.Sink.run(Sink.java:25)
    at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1149)
    at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:624)
    at java.lang.Thread.run(Thread.java:748)

源代碼：

隊列.java

public class Queues {
    public static volatile BlockingQueue[] queues;

    /**
     * Create Multiple Queues.
     * @param count The number of queues created.
     * @param capacity The Capacity of each queue.
     */
    public static void createQueues(final int count, final int capacity) {
        Queues.queues = new BlockingQueue[count];
        for (int i=0; i<count; ++i) {
            Queues.queues[i] = new ArrayBlockingQueue(capacity, true);
        }
    }
}

記錄

@Builder
@Getter
public class Record {
    private final String value;
    private final Callable<Boolean> ackCallback;
}

水槽.java

public class Sink implements Runnable {
    private final int queueId;

    public Sink(int queueId) {
        this.queueId = queueId;
    }

    @Override
    public void run() {
        while (true) {
            try {
                Record record = (Record) Queues.queues[this.queueId].take();
                // (1) Handler: Write to database
                Thread.sleep(10);
                // (2) ACK: notify kafka consumer to commit offset manually
                record.getAckCallback().call();
            } catch (Exception e) {
                e.printStackTrace();
                System.exit(1);
            }
        }
    }
}

KafkaConsumerRunner

@Slf4j
public class KafkaConsumerRunner implements Runnable {
    private final String topic;
    private final KafkaConsumer<String, String> consumer;

    public KafkaConsumerRunner(String topic, Properties properties) {
        this.topic = topic;
        this.consumer = new KafkaConsumer<>(properties);
    }

    @Override
    public void run() {
        // offsets to commit
        Map<TopicPartition, OffsetAndMetadata> offsetsToCommit = new HashMap<>();
        // Subscribe topic
        this.consumer.subscribe(Collections.singletonList(this.topic));
        // Consume Kafka Message
        while (true) {
            try {
                ConsumerRecords<String, String> consumerRecords = this.consumer.poll(10000L);
                for (TopicPartition topicPartition : consumerRecords.partitions()) {
                    for (ConsumerRecord<String, String> consumerRecord : consumerRecords.records(topicPartition)) {
                        // (1) Restore [partition -> offset] Map
                        offsetsToCommit.put(topicPartition, new OffsetAndMetadata(consumerRecord.offset()));
                        // (2) Put into queue
                        int queueId = topicPartition.partition() % Queues.queues.length;
                        Queues.queues[queueId].put(Record.builder()
                                .value(consumerRecord.value())
                                .ackCallback(this.getAckCallback(offsetsToCommit))
                                .build());
                    }
                }
            } catch (ConcurrentModificationException | InterruptedException e) {
                log.error("[{}] {}", Thread.currentThread().getName(), ExceptionUtils.getMessage(e), e);
                System.exit(1);
            }
        }
    }

    private Callable<Boolean> getAckCallback(Map<TopicPartition, OffsetAndMetadata> offsets) {
        return new AckCallback<Boolean>(this.consumer, new HashMap<>(offsets)) {
            @Override
            public Boolean call() throws Exception {
                try {
                    this.getConsumer().commitSync(this.getOffsets());
                    return true;
                } catch (Exception e) {
                    log.error(String.format("[%s] %s", Thread.currentThread().getName(), ExceptionUtils.getMessage(e)), e);
                    return false;
                }
            }
        };
    }

    @Getter
    @AllArgsConstructor
    abstract class AckCallback<T> implements Callable<T> {
        private final KafkaConsumer<String, String> consumer;
        private final Map<TopicPartition, OffsetAndMetadata> offsets;
    }
}

應用.java

public class Application {
    private static final String TOPIC = "YEWEI_TOPIC";
    private static final int QUEUE_COUNT = 1;
    private static final int QUEUE_CAPACITY = 4;
    
    private static void createQueues() {
        Queues.createQueues(QUEUE_COUNT, QUEUE_CAPACITY);
    }

    private static void startupSource() {
        if (null == System.getProperty("java.security.auth.login.config")) {
            System.setProperty("java.security.auth.login.config", "jaas.conf");
        }

        Properties properties = new Properties();
        properties.put(ConsumerConfig.GROUP_ID_CONFIG, "ConsumerGroup1");
        properties.put(ConsumerConfig.BOOTSTRAP_SERVERS_CONFIG, "cdh1:9092,cdh2:9092,cdh3:9092");
        properties.put(ConsumerConfig.KEY_DESERIALIZER_CLASS_CONFIG, org.apache.kafka.common.serialization.StringDeserializer.class);
        properties.put(ConsumerConfig.VALUE_DESERIALIZER_CLASS_CONFIG, org.apache.kafka.common.serialization.StringDeserializer.class);
        properties.put(ConsumerConfig.MAX_POLL_RECORDS_CONFIG, 2);
        properties.put(ConsumerConfig.ENABLE_AUTO_COMMIT_CONFIG, false);
        properties.put(ConsumerConfig.AUTO_OFFSET_RESET_CONFIG, "earliest");
        properties.put(CommonClientConfigs.SECURITY_PROTOCOL_CONFIG, "SASL_PLAINTEXT");
        properties.put(SaslConfigs.SASL_MECHANISM, "PLAIN");

        ExecutorService executorService = Executors.newFixedThreadPool(QUEUE_COUNT);
        for (int queueId = 0; queueId < QUEUE_COUNT; ++queueId) {
            executorService.execute(new KafkaConsumerRunner(TOPIC, properties));
        }
    }

    private static void startupSinks() {
        ExecutorService executorService = Executors.newFixedThreadPool(QUEUE_COUNT);
        for (int queueId = 0; queueId < QUEUE_COUNT; ++queueId) {
            executorService.execute(new Sink(queueId));
        }
    }

    public static void main(String[] args) {
        Application.createQueues();
        Application.startupSource();
        Application.startupSinks();
    }
}

Answer 1

我想通了這個問題。 Kafka 消費者在自己的線程中運行，也被 Sink 線程回調。 KafkaConsumer 的poll和commitSync方法只能應用於一個線程。 請參閱org.apache.kafka.clients.consumer.KafkaConsumer#acquireAndEnsureOpen 。

改為： Sink回調不直接使用consumer object，而是向LinkedTransferQueue發送ACK消息。 KafkaConsumerRunner 每次都會輪詢 LinkedTransferQueue 並批量發送 ACK

@Slf4j
public class KafkaConsumerRunner implements Runnable {
    private final String topic;
    private final BlockingQueue ackQueue;
    private final KafkaConsumer<String, String> consumer;

    public KafkaConsumerRunner(String topic, Properties properties) {
        this.topic = topic;
        this.ackQueue = new LinkedTransferQueue<Map<TopicPartition, OffsetAndMetadata>>();
        this.consumer = new KafkaConsumer<>(properties);
    }

    @Override
    public void run() {
        // Subscribe topic
        this.consumer.subscribe(Collections.singletonList(this.topic));
        // Consume Kafka Message
        while (true) {
            while (!this.ackQueue.isEmpty()) {
                try {
                    Map<TopicPartition, OffsetAndMetadata> offsets = (Map<TopicPartition, OffsetAndMetadata>) this.ackQueue.take();
                    this.consumer.commitSync(offsets);
                } catch (InterruptedException e) {
                    e.printStackTrace();
                }
            }

            ...
        }
    }

    private Callable<Boolean> getAckCallback(Map<TopicPartition, OffsetAndMetadata> offsets) {
        return new AckCallback<Boolean>(new HashMap<>(offsets)) {
            @Override
            public Boolean call() throws Exception {
                try {
                    ackQueue.put(offsets);
                    return true;
                } catch (Exception e) {
                    log.error(String.format("[%s] %s", Thread.currentThread().getName(), ExceptionUtils.getMessage(e)), e);
                    System.exit(1);
                    return false;
                }
            }
        };
    }

    ...
}