hm. I can reproduce this behaviour & confirm that your answer solves this - as I’m currently tackling the same problem. And somehow it makes sense that it behaves like this, given the possibilities to automatically or manually acknowledge the record.

However, it seems quite uncomfortable to set the max poll records to 1 if I want to have an accurate lag metric - given the fact that the default is 500 for a reason. I assume that this would cause negative performance impact. Therefore I’m asking myself whether there is any simpler solution, anybody came across to measure the lag.

I thought about measuring it on the server side of my Confluent Kafka with JMX but this is also a bit of an overhead. Any other ideas or best-practices how one can solve this issue - and get an accurate consumer-lag? My goal is to raise alerts if consumer-lags build up.