I've researched the Kafka monitoring systems and found out that there are two types of them. First is consuming __consumer_offsets topic and Second is using kafka api.
It seems using kafka broker api is much more accurate, if cluster has many consumers - which means producing lots of __consumer_offsets, creating lag for consuming them and shows totally inaccurate lag monitoring.
The only disadvantage seems sending api intervally can be burdensome to broker, but it looks very slight.
I have no idea why the most famous lag monitoring systems like burrow, kafka manager use this inaccurate method.