Juicy one. https://issues.apache.org/jira/browse/KAFKA-2530 I hope it's related to http://search-hadoop.com/m/uyzND1XVyK12UNtd32/kafka+orphaned&subj=Consumer+lag+lies+orphaned+offsets ! :)
Otis -- Monitoring * Alerting * Anomaly Detection * Centralized Log Management Solr & Elasticsearch Support * http://sematext.com/ On Fri, Sep 4, 2015 at 3:59 AM, John Watson <jo...@dropbox.com> wrote: > It would seem (if the metrics registry is accurate) that replica > fetcher threads can persist after a leadership election. Even when the > broker itself is elected leader. > This also seems to occur after a reassignment too (as evident by the 5 > different thread entries for the same partition in the registry) > > This leads to inaccurate MaxLag and ConsumerLag metrics to be > reported. Since it's whatever the last reported value is. > > Restarting the offending broker clears these at least. >