Based on the goroutine profile, I see around 200 clients connected, but nearly 12,000 revisions stuck waiting to get a sequence number to be written to the database.
This seems like it might be a client trying to push something, timing out, reconnecting and retrying, pushing more and more stuff into the queue. We fixed an issue that may be causing this in an upcoming version of Sync Gateway (https://github.com/couchbase/sync_gateway/pull/4222). We expect this release to ship early 2020.
I’d expect adding more SG nodes to help this situation. Did you see exactly the same behaviour? Or just more spread out over the nodes. I presume you’re routing traffic to SG thorugh a load balancer so you can distribute requests from clients?