It looks like the lack of persistent storage for the federated activity queue is leading to instances running out of memory in a matter of hours. See my comment for more details.
Furthermore, this leads to data loss, since there is no other consistency mechanism. I think it might be a high priority issue, taking into account the current momentum behind growth of Lemmy…
Normally with XDCR you can specify which documents to replicate out of a bucket. It doesn’t have to be the entire bucket. So if you had certain types (comments, upvotes, etc) then only those would sync when the target comes online.
I did check into Apache CouchDB, the open source upstream, and replication is there. We use Enterprise Couchbase at work and it’s a dream but there are some tools that I use that only use Apache CouchDB (Inkdrop for example). It’s worth looking into.