Never invent the wheel if you haven't done your research yet.
Colleagues of mine successfully employed CouchDB in a situation where devices were offline for an x amount of time (running on tablets in airplanes), once they had internet again they would reliably start syncing data with the main database again. This was a number of years ago though, I haven't heard anything about CouchDB since.
Who says you need to feed it directly to the replica. Logical replication provides infrastructure for decoding changes, and you may fetch it any way you want / how often you want / feed it wherever you want (file, another database, ...).
It would be trivial to write something that connects to a bunch of machines regularly, fetches the decoded increment and feed it somewhere (say, to a single database over a shared connection).
Colleagues of mine successfully employed CouchDB in a situation where devices were offline for an x amount of time (running on tablets in airplanes), once they had internet again they would reliably start syncing data with the main database again. This was a number of years ago though, I haven't heard anything about CouchDB since.