Unlike some relational replication schemes, XDCR does not maintain some queue that fills up. The destination can be unavailable for long periods of time (months) and XDCR will continue to retry replication but won’t impact the mutation on the source cluster. This is also true for DCP which is the underlying replication protocol for intra cluster replication as well as XDCR
replication is done with timestamps and there is a way for replication to retrieve updates after a given timestamp without managing a queue that is significant in size. So when you fall behind with DCP or XDCR, if we cannot find the updates for a given timestamp in memory, we can go to disk and retrieve these items. So there is no large replication queue per say. The “queues” for replication are in memory.