更新时间:2022-12-14 GMT+08:00

Consumer消费数据存在重复消费现象

问题背景与现象

当数据量较大时会频繁的发生rebalance导致出现重复消费的情况,关键日志如下:

2018-05-12 10:58:42,561 | INFO | [kafka-request-handler-3] | [GroupCoordinator 2]: Preparing to restabilize group DemoConsumer with old generation 118 | kafka.coordinator.GroupCoordinator (Logging.scala:68)
2018-05-12 10:58:43,245 | INFO | [kafka-request-handler-5] | [GroupCoordinator 2]: Stabilized group DemoConsumer generation 119 | kafka.coordinator.GroupCoordinator (Logging.scala:68)
2018-05-12 10:58:43,560 | INFO | [kafka-request-handler-7] | [GroupCoordinator 2]: Assignment received from leader for group DemoConsumer for generation 119 | kafka.coordinator.GroupCoordinator (Logging.scala:68)
2018-05-12 10:59:13,562 | INFO | [executor-Heartbeat] | [GroupCoordinator 2]: Preparing to restabilize group DemoConsumer with old generation 119 | kafka.coordinator.GroupCoordinator (Logging.scala:68)
2018-05-12 10:59:13,790 | INFO | [kafka-request-handler-3] | [GroupCoordinator 2]: Stabilized group DemoConsumer generation 120 | kafka.coordinator.GroupCoordinator (Logging.scala:68)
2018-05-12 10:59:13,791 | INFO | [kafka-request-handler-0] | [GroupCoordinator 2]: Assignment received from leader for group DemoConsumer for generation 120 | kafka.coordinator.GroupCoordinator (Logging.scala:68)
2018-05-12 10:59:43,802 | INFO | [kafka-request-handler-2] | Rolled new log segment for '__consumer_offsets-17' in 2 ms. | kafka.log.Log (Logging.scala:68)
2018-05-12 10:59:52,456 | INFO | [group-metadata-manager-0] | [Group Metadata Manager on Broker 2]: Removed 0 expired offsets in 0 milliseconds. | kafka.coordinator.GroupMetadataManager (Logging.scala:68)
2018-05-12 11:00:49,772 | INFO | [kafka-scheduler-6] | Deleting segment 0 from log __consumer_offsets-17. | kafka.log.Log (Logging.scala:68)
2018-05-12 11:00:49,773 | INFO | [kafka-scheduler-6] | Deleting index /srv/BigData/kafka/data4/kafka-logs/__consumer_offsets-17/00000000000000000000.index.deleted | kafka.log.OffsetIndex (Logging.scala:68)
2018-05-12 11:00:49,773 | INFO | [kafka-scheduler-2] | Deleting segment 2147948547 from log __consumer_offsets-17. | kafka.log.Log (Logging.scala:68)
2018-05-12 11:00:49,773 | INFO | [kafka-scheduler-4] | Deleting segment 4282404355 from log __consumer_offsets-17. | kafka.log.Log (Logging.scala:68)
2018-05-12 11:00:49,775 | INFO | [kafka-scheduler-2] | Deleting index /srv/BigData/kafka/data4/kafka-logs/__consumer_offsets-17/00000000002147948547.index.deleted | kafka.log.OffsetIndex (Logging.scala:68)
2018-05-12 11:00:49,775 | INFO | [kafka-scheduler-4] | Deleting index /srv/BigData/kafka/data4/kafka-logs/__consumer_offsets-17/00000000004282404355.index.deleted | kafka.log.OffsetIndex (Logging.scala:68)
2018-05-12 11:00:50,533 | INFO | [kafka-scheduler-6] | Deleting segment 4283544095 from log __consumer_offsets-17. | kafka.log.Log (Logging.scala:68)
2018-05-12 11:00:50,569 | INFO | [kafka-scheduler-6] | Deleting index /srv/BigData/kafka/data4/kafka-logs/__consumer_offsets-17/00000000004283544095.index.deleted | kafka.log.OffsetIndex (Logging.scala:68)
2018-05-12 11:02:21,178 | INFO | [kafka-request-handler-2] | [GroupCoordinator 2]: Preparing to restabilize group DemoConsumer with old generation 120 | kafka.coordinator.GroupCoordinator (Logging.scala:68)
2018-05-12 11:02:22,839 | INFO | [kafka-request-handler-4] | [GroupCoordinator 2]: Stabilized group DemoConsumer generation 121 | kafka.coordinator.GroupCoordinator (Logging.scala:68)
2018-05-12 11:02:23,169 | INFO | [kafka-request-handler-1] | [GroupCoordinator 2]: Assignment received from leader for group DemoConsumer for generation 121 | kafka.coordinator.GroupCoordinator (Logging.scala:68)
2018-05-12 11:02:49,913 | INFO | [kafka-request-handler-6] | Rolled new log segment for '__consumer_offsets-17' in 2 ms. | kafka.log.Log (Logging.scala:68)

其中Preparing to restabilize group DemoConsumer with old generation表示正在发生rebalance。

可能原因

参数设置不合理。

原因分析

原因:由于参数设置不当,数据量大时数据处理时间过长,导致频繁发生balance,此时offset无法正常提交,导致重复消费数据。

原理:每次poll的数据处理完后才提交offset,如果poll数据后的处理时长超出了session.timeout.ms的设置时长,此时发生rebalance导致本次消费失败,已经消费数据的offset无法正常提交,所以下次重新消费时还是在旧的offset消费数据,从而导致消费数据重复。

解决办法

建议用户在Manager页面调整以下服务参数:

request.timeout.ms=100000

session.timeout.ms=90000

max.poll.records=50

heartbeat.interval.ms=3000

其中:

request.timeout.ms要比session.timeout.ms大10s。

session.timeout.ms的大小设置要在服务端参数group.min.session.timeout.ms和group.max.session.timeout.ms之间。

以上参数可以根据实际情况进行适当的调整,特别是max.poll.records,这个参数是为了控制每次poll数据的records量,保证每次的处理时长尽量保持稳定。目的是为了保证poll数据以后的处理时间不要超过session.timeout.ms的时间。

参考信息

  • poll之后的数据处理效率要高,不要阻塞下一次poll
  • poll方法和数据处理建议异步处理