Kafka - 消息零丢失实战

Kafka消息0丢失实战

当你用Kafka处理业务时，是否担心过消息神秘失踪？下面将从SpringBoot整合实战出发，穿透生产者→Broker→消费者全链路

1、 消息丢失的三大场景

场景1：生产者自信发送

// 致命陷阱代码示例

@Bean
public ProducerFactory<String, String> producerFactory() {
    Map<String, Object> configs = new HashMap<>();
    configs.put(ProducerConfig.ACKS_CONFIG, "0"); // 发后即忘
    configs.put(ProducerConfig.RETRIES_CONFIG, 0); // 不重试
    return new DefaultKafkaProducerFactory<>(configs);
}

// 异步发送忘记回调

kafkaTemplate.send("order-topic", orderId, json).addCallback(
    result -> logger.info("发送成功"),  // 成功日志
    ex -> logger.error("发送失败")     // 错误吞没
);

场景2：Broker的死亡错觉

# 危险配置示范
auto.create.topics.enable=true     # 自动创建主题埋雷
unclean.leader.election.enable=true # 允许脏选举
min.insync.replicas=1             # 单副本存活即工作

场景3：消费者的自信提交

// 问题配置

@KafkaListener(topics = "order-topic")
public void handle(Order order) {
    try {
        paymentService.process(order);  // 处理耗时操作
    } finally {
        // 没有手动提交偏移量！
    }
}

// 错误配置：自动提交间隔过长

spring.kafka.consumer.enable-auto-commit=true
spring.kafka.consumer.auto-commit-interval=5000

2、生产者端的钢铁防线

1. 同步发送+重试策略（金融级防护）

@Bean
public KafkaTemplate<String, String> reliableKafkaTemplate() {
    Map<String, Object> configs = new HashMap<>();
    configs.put(ProducerConfig.ACKS_CONFIG, "all"); // 必须所有副本确认
    configs.put(ProducerConfig.RETRIES_CONFIG, Integer.MAX_VALUE);
    configs.put(ProducerConfig.MAX_IN_FLIGHT_REQUESTS_PER_CONNECTION, 1); // 防止乱序
    configs.put(ProducerConfig.ENABLE_IDEMPOTENCE_CONFIG, true); // 开启幂等
    return new KafkaTemplate<>(new DefaultKafkaProducerFactory<>(configs));
}

// 发送模板

ListenableFuture<SendResult<String, String>> future = kafkaTemplate.send("order-topic", key, value);
future.get(10, TimeUnit.SECONDS); // 同步等待确认

2. 事务消息（分布式事务防护）

@Bean
public KafkaTransactionManager<String, String> transactionManager() {
    return new KafkaTransactionManager<>(producerFactory());
}
@Transactional
public void processOrder(Order order) {
    paymentService.charge(order);
    kafkaTemplate.send("payment-topic", order.getId(), order.toJson());
    inventoryService.reduceStock(order); 
}

3. 监控指标看护

// 注册监控指标

metrics.addMetric("producer-error-rate", (tags) -> {
    return producer.metrics().get("record-error-rate").value();
});

3、Broker集群的堡垒计划

1. 存活确认矩阵

# broker关键配置
unclean.leader.election.enable=false    # 禁止脏选举
min.insync.replicas=2                  # 至少2个副本确认
default.replication.factor=3           # 默认3副本
log.flush.interval.messages=10000      # 刷盘策略
log.flush.interval.ms=1000

2. ISR机制源码解析

// Kafka源码片段（Partition.scala）

def inSyncReplicas = {
  val leaderLogEndOffset = localLogOrException.logEndOffset
  remoteReplicas.filter { replica =>
    replica.logEndOffset >= leaderLogEndOffset &&
    (time.milliseconds - replica.lastCaughtUpTimeMs) < replicaLagTimeMaxMs
  }
}

3. 磁盘防护策略

# 使用JBOD而不是RAID（Kafka最佳实践）
log.dirs=/data/kafka/log1,/data/kafka/log2,/data/kafka/log3
# 监控脚本示例
df -h | grep /data/kafka | awk '{if ($5 > 85) print "ALERT: "$6" usage over 85%"}'

4、消费者端的终极防御

1. 手动提交+死信队列

@KafkaListener(topics = "order-topic", groupId = "payment-group")
public void listen(
    ConsumerRecord<String, String> record,
    Acknowledgment ack,
    Consumer<String, String> consumer) {
    try {
        paymentService.process(record.value());
        ack.acknowledge(); // 手动提交
    } catch (Exception ex) {
        // 记录原始消息到死信队列
        kafkaTemplate.send("order-dlq", record.key(), record.value());
        // 重置偏移量到5秒前
        consumer.seek(record.topic(), record.partition(), record.offset() - 1);
    }
}

2. 消费者组反脆弱模式

# 关键配置
spring.kafka.consumer.enable-auto-commit=false
spring.kafka.listener.ack-mode=MANUAL_IMMEDIATE
spring.kafka.consumer.isolation.level=read_committed

3. 消费延迟监控

// 计算消费延迟

long lag = record.timestamp() - System.currentTimeMillis();
metrics.recordGauge("consumer.lag", lag);

5、全链路防护方案

1. 端到端校验设计

// 消息指纹校验

public class MessageWrapper {
    private String payload;
    private String checksum; // SHA256(payload + salt)
}

// 生产者端

String salt = "kafka-secure-2023";
String checksum = DigestUtils.sha256Hex(payload + salt);
template.send("topic", new MessageWrapper(payload, checksum));

// 消费者端

if (!DigestUtils.sha256Hex(message.getPayload() + salt).equals(message.getChecksum())) {
    throw new InvalidMessageException();
}

2. 混沌工程测试用例

@SpringBootTest
public class KafkaChaosTest {
    @Autowired
    private KafkaChaosRunner chaosRunner;
    @Test
    public void testNetworkPartition() {
        chaosRunner.networkPartition("kafka-broker1", Duration.ofMinutes(5));
        // 验证消息不丢失
    }
}

3. 消息轨迹追踪

// 使用OpenTelemetry实现

Span sendSpan = tracer.spanBuilder("kafka.send")
                     .setAttribute("message.key", key)
                     .startSpan();
try (Scope scope = sendSpan.makeCurrent()) {
    kafkaTemplate.send("topic", key, value);
} finally {
    sendSpan.end();
}

配置核查清单

✅ 生产者acks=all且开启幂等

✅ broker禁用unclean leader选举

✅ 消费者关闭自动提交

✅ 事务消息开启read_committed

✅ 监控Producer/Consumer/Broker关键指标