背景
kafka没有重试机制不支持消息重试，也没有死信队列，因此使用kafka做消息队列时，如果遇到了消息在业务处理时出现异常，就会很难进行下一步处理。应对这种场景，需要自己实现消息重试的功能。
如果不想自己实现消息重试机制，建议使用RocketMQ作为消息队列，RocketMQ的消息重试机制相当完善，对于开发者使用也非常友好，详见https://help.aliyun.com/document_detail/43490.html。
方案
申请一个新的kafka topic作为重试队列，步骤如下：
- 创建一个topic作为重试topic用于接受等待重试的消息
- 普通topic消费者给待重试的消息设置下一次的消费事件后发送到重试topic
- 从重试topic获取待重试消息储存到redis的zset中，并以下一次消费时间排序
- 定时任务从redis获取到达消费事件的消息，并把消息发送到对应的topic
- 同一个消息重试次数过多则不再重试
代码实现
重试消息的javaBean
public class KafkaRetryRecord {
public static final String KEY_RETRY_TIMES = "retryTimes";
private String key;
private String value;
private Integer retryTimes;
private String topic;
private Long nextTime;
public KafkaRetryRecord(){
}
public String getKey() {
return key;
}
public void setKey(String key) {
this.key = key;
}
public String getValue() {
return value;
}
public void setValue(String value) {
this.value = value;
}
public Integer getRetryTimes() {
return retryTimes;
}
public void setRetryTimes(Integer retryTimes) {
this.retryTimes = retryTimes;
}
public String getTopic() {
return topic;
}
public void setTopic(String topic) {
this.topic = topic;
}
public Long getNextTime() {
return nextTime;
}
public void setNextTime(Long nextTime) {
this.nextTime = nextTime;
}
public ProducerRecord parse(){
Integer partition = null;
Long timestamp = System.currentTimeMillis();
List<Header> headers = new ArrayList<>();
ByteBuffer retryTimesBuffer = ByteBuffer.allocate(4);
retryTimesBuffer.putInt(retryTimes);
retryTimesBuffer.flip();
headers.add(new RecordHeader(KafkaRetryRecord.KEY_RETRY_TIMES, retryTimesBuffer));
ProducerRecord sendRecord = new ProducerRecord(
topic, partition, timestamp, key, value, headers);
return sendRecord;
}
}
消费端的消息发送到重试队列
public class KafkaRetryService {
private static final Logger log = LoggerFactory.getLogger(KafkaRetryService.class);
/**
* 消息消费失败后下一次消费的延迟时间(秒)
* 第一次重试延迟10秒;第二次延迟30秒,第三次延迟1分钟...
*/
private static final int[] RETRY_INTERVAL_SECONDS = {10, 30, 1*60, 2*60, 5*60, 10*60, 30*60, 1*60*60, 2*60*60};
/**
* 重试topic
*/
@Value("${spring.kafka.topics.retry}")
private String retryTopic;
@Autowired
private KafkaTemplate<String, String> template;
public void consumerLater(ConsumerRecord<String, String> record){
// 获取消息的已重试次数
int retryTimes = getRetryTimes(record);
Date nextConsumerTime = getNextConsumerTime(retryTimes);
if(nextConsumerTime == null) {
return;
}
KafkaRetryRecord retryRecord = new KafkaRetryRecord();
retryRecord.setNextTime(nextConsumerTime.getTime());
retryRecord.setTopic(record.topic());
retryRecord.setRetryTimes(retryTimes);
retryRecord.setKey(record.key());
retryRecord.setValue(record.value());
String value = JSON.toJSONString(retryRecord);
template.send(retryTopic, null, value);
}
/**
* 获取消息的已重试次数
*/
private int getRetryTimes(ConsumerRecord record){
int retryTimes = -1;
for(Header header : record.headers()){
if(KafkaRetryRecord.KEY_RETRY_TIMES.equals(header.key())){
ByteBuffer buffer = ByteBuffer.wrap(header.value());
retryTimes = buffer.getInt();
}
}
retryTimes++;
return retryTimes;
}
/**
* 获取待重试消息的下一次消费时间
*/
private Date getNextConsumerTime(int retryTimes){
// 重试次数超过上限,不再重试
if(RETRY_INTERVAL_SECONDS.length < retryTimes) {
return null;
}
Calendar calendar = Calendar.getInstance();
calendar.add(Calendar.SECOND, RETRY_INTERVAL_SECONDS[retryTimes]);
return calendar.getTime();
}
}
处理待消费的消息
public class RetryListener {
private static final Logger log = LoggerFactory.getLogger(RetryListener.class);
private static final String RETRY_KEY_ZSET = "_retry_key";
private static final String RETRY_VALUE_MAP = "_retry_value";
@Autowired
private RedisTemplate<String,Object> redisTemplate;
@Autowired
private KafkaTemplate<String, String> kafkaTemplate;
@KafkaListener(topics = "${spring.kafka.topics.retry}")
public void consume(List<ConsumerRecord<String, String>> list) {
for(ConsumerRecord<String, String> record : list){
KafkaRetryRecord retryRecord = JSON.parseObject(record.value(), KafkaRetryRecord.class);
/**
* TODO 防止待重试消息太多撑爆redis,可以将待重试消息按下一次重试时间分开存储放到不同介质
* 例如下一次重试时间在半小时以后的消息储存到mysql,并定时从mysql读取即将重试的消息储储存到redis
*/
// 通过redis的zset进行时间排序
String key = UUID.randomUUID().toString();
redisTemplate.opsForHash().put(RETRY_VALUE_MAP, key, record.value());
redisTemplate.opsForZSet().add(RETRY_KEY_ZSET, key, retryRecord.getNextTime());
}
}
/**
* 定时任务从redis读取到达重试时间的消息,发送到对应的topic
*/
@Scheduled(cron="0/2 * * * * *")
public void retryFormRedis() {
long currentTime = System.currentTimeMillis();
Set<ZSetOperations.TypedTuple<Object>> typedTuples =
redisTemplate.opsForZSet().reverseRangeByScoreWithScores(RETRY_KEY_ZSET, 0, currentTime);
redisTemplate.opsForZSet().removeRangeByScore(RETRY_KEY_ZSET, 0, currentTime);
for(ZSetOperations.TypedTuple<Object> tuple : typedTuples){
String key = tuple.getValue().toString();
String value = redisTemplate.opsForHash().get(RETRY_VALUE_MAP, key).toString();
redisTemplate.opsForHash().delete(RETRY_VALUE_MAP, key);
KafkaRetryRecord retryRecord = JSON.parseObject(value, KafkaRetryRecord.class);
ProducerRecord record = retryRecord.parse();
kafkaTemplate.send(record);
}
// TODO 发生异常将发送失败的消息重新扔回redis
}
}
消息重试
public class ConsumeListener {
private static final Logger log = LoggerFactory.getLogger(ConsumeListener.class);
@Autowired
private KafkaRetryService kafkaRetryService;
@KafkaListener(topics = "${spring.kafka.topics.test}")
public void consume(List<ConsumerRecord<String, String>> list) {
for(ConsumerRecord record : list){
try {
// 业务处理
} catch (Exception e){
log.error(e.getMessage());
// 消息重试
kafkaRetryService.consumerLater(record);
}
}
}
}
