Scott Kidder created KAFKA-12231:
------------------------------------
Summary: Consumer Lag increases linearly until a Consumer-Group
Rebalance is initiated
Key: KAFKA-12231
URL: https://issues.apache.org/jira/browse/KAFKA-12231
Project: Kafka
Issue Type: Bug
Components: core
Affects Versions: 2.6.0
Environment: Kubernetes 1.12
Reporter: Scott Kidder
Attachments: Consumer Lag by Partition.png, Consumer Lag on a Single
Partition.png, Lag drop on rebalance.png, max-consumer-lag.png
I observed a linear increase in consumer lag reading from a single topic (480
partitions) across multiple consumers for multiple hours. The increase in lag
was stopped by initiating a consumer-group rebalance by replacing one of the
consumers (this was in Kubernetes, so deleting a consumer pod and seeing its
replacement pod join) at 07:46UTC on the chart below.
!max-consumer-lag.png!
The lag was observed across all topic partitions, but only briefly on each:
!Consumer Lag by Partition.png!
!Consumer Lag on a Single Partition.png!
For additional context, this was a Golang consumer using v1.27.2 of the Shopify
Sarama Kafka client. Consumers used the Sticky Partition Assignor to plan
assignments. So, even after the consumer-group rebalance, the majority of
consumers kept their original assignments. Nothing about the data being
consumed & processed from Kafka could explain these punctuated spikes in
consumer lag. There were no errors or significant messages in the Kafka broker
logs before or after the rebalance.
The lag dropped within 2 minutes of the consumer-group rebalance (initiated at
07:46, lag fell at 07:48):
!Lag drop on rebalance.png!
--
This message was sent by Atlassian Jira
(v8.3.4#803005)