I encourage thinking about this as an indexing problem rather than a Sync Gateway problem. The query for user docs will be dwarfed by the query when doing a replication, presuming there are more actual documents than user documents.
My guess would be that upgrading the Server nodes causes some of this index to drop out of memory and fall back to quite slow disks.
This index only contains the keys of documents and 98k keys should be on a the small megabytes level of data and quick to query.