Ah I have been down this path. It is the zookeeper client. It resolves and caches the ip addresses: https://github.com/apache/zookeeper/blob/bd9a1448f9b29859092e6bdca93da121ec166b7a/src/java/main/org/apache/zookeeper/client/StaticHostProvider.java#L108
I believe they are cached forever. We have had success with HAProxy to manage connections. We have also written and patched Kafka to use our service discovery system to always have an ip to resolve too. On Fri, Mar 18, 2016 at 10:49 AM Ralph Rodkey <ralph.rod...@angieslist.com> wrote: > Is there any way to make Kafka 0.8.2.2 honor the DNS TTLs of its Zookeeper > ensemble? I'm specifying the Zookeepers by hostname, but Kafka doesn't seem > to ever re-resolve those hostnames after startup. I've tried setting the > relevant JVM system properties via "-Dnetworkaddress.cache.ttl=0 > -Dnetworkaddress.cache.negative.ttl=0", but it doesn't seem like Kafka's > behavior was affected by those options. > > I don't see any reference to this issue in the bug tracker. Is there a way > to deal with Zookeeper server replacement short of restarting Kafka? Are > there other approaches I can take that won't result in downtime? The only > option I can think of is to do a rolling restart of the Kafka servers after > each Zookeeper server is replaced, which seems distasteful. >
