Good point, but that would make me a little nervous. What happens if the compaction takes a while and the replica gets far behind? And wouldn't the compaction time just keep getting longer and longer as data grows?

Can you repartition the data online to keep the cleanup work bounded?