2020-09-26 17:54:04 UTC - Addison Higham: Both are valid, it is a tradeoff between resilience and potentially performance. For example, a local nvme is likely to be faster and lower latency than network block device over ceph, but with a local volume, you need to worry a bit more operationally about replacing instances. If you are running k8s on bare metal, that might be much more manageable compared to cloud.
As far as production, it is had to say generally as it is very use case and load dependent, but usually the minimun configuration that is production ready is: 3 bookies, 3 zookeeper nodes, and 2 brokers. Usually minimum effective JVM size is ~2GB of heap ---- 2020-09-27 08:33:47 UTC - Guillaume: I too have the same issue. What would be the reason for the broker to timeout reading from zookeeper? ----
