我有一个主题在Kafka描述如下（通过 /usr/hdp/2.6.5.0-292/kafka/bin/kafka-topics.sh --describe <rest of command> )

Topic:arrival_events    PartitionCount:12   ReplicationFactor:2 Configs:

Topic: arrival_events   Partition: 0    Leader: 1001    Replicas: 1001,1002 Isr: 1001,1002
Topic: arrival_events   Partition: 1    Leader: 1002    Replicas: 1002,1003 Isr: 1002,1003
Topic: arrival_events   Partition: 2    Leader: 1003    Replicas: 1003,1001 Isr: 1003,1001
Topic: arrival_events   Partition: 3    Leader: 1001    Replicas: 1001,1003 Isr: 1003,1001
Topic: arrival_events   Partition: 4    Leader: 1002    Replicas: 1002,1001 Isr: 1002,1001
Topic: arrival_events   Partition: 5    Leader: 1003    Replicas: 1003,1002 Isr: 1003,1002
Topic: arrival_events   Partition: 6    Leader: 1001    Replicas: 1001,1002 Isr: 1001,1002
Topic: arrival_events   Partition: 7    Leader: 1002    Replicas: 1002,1003 Isr: 1002,1003
Topic: arrival_events   Partition: 8    Leader: 1003    Replicas: 1003,1001 Isr: 1003,1001
Topic: arrival_events   Partition: 9    Leader: 1001    Replicas: 1001,1003 Isr: 1003,1001
Topic: arrival_events   Partition: 10   Leader: 1002    Replicas: 1002,1001 Isr: 1002,1001
Topic: arrival_events   Partition: 11   Leader: 1003    Replicas: 1003,1002 Isr: 1003,1002

经纪人大致收到 5-8mil 每天的信息（旅行模式）。
一切都很好，除了几个分区（不超过2-3个）被高延迟阻塞。
随着数据的连续流动，在几天内，这有时也会超过1-2英里。而其他分区舒适地位于0滞后
我已经试着把消费者人数降到12人以下 round robin 也会强制读取其他分区，但没有帮助。
我有什么办法来减少这个滞后的建议吗？消费者通过使用java构建的数据流处理器。