There should be a graceful way of decommissioning the broker so that there is absolutely 0 data loss. Decommissioning is not necessarily related to replication (kafka-50).
There should be a way to get the broker out of the cluster only from the produce side. Consumers should be able to continue keep pulling data. When the administrator is sure that all data has been consumed by consumers, broker node can be removed permanently. Same functionality can be used for rolling upgrades without any message loss. If this makes sense, I can go ahead and create JIRA. Sharad
