We currently send several kHz of device list updates over replication
occisonally, which often causes the replications streams to lag and then
get dropped.
A lot of those updates will actually be duplicates, since we don't send
e.g. device_ids across replication, so let's deduplicate it when we pull
them out of the database.