I'm assuming you mean you shut down the follower and the leader did not reflect this in isr shrink-rate. I think the reason for this is that the controller would have sent a leaderAndIsr request to the (existing leader) with the shrunk ISR.
On the (redundant) leader transition the leader updates the ISR (i.e., shrinks) but does not mark the shrink-rate meter. We currently only mark the meter if the follower is slow/stuck. If it is down, then the controller may detect that and send the leaderAndIsr request before the ISR shrink-thread gets a chance to detect and shrink the ISR. We should mark the ISR shrink when receiving a leaderAndIsr request from the controller with a shrunk ISR - I filed https://issues.apache.org/jira/browse/KAFKA-2309 for this. Thanks, Joel On Thu, Jul 02, 2015 at 12:07:46AM +0800, taosheng53 wrote: > hello, > when i use jmx to monitor kafka, i find the value of the isrshrinkrate > always 0,never change when one broker has been shutdown.(i have a cluster > containing three nodes),i want to know if there is some problem?
