Right now, every single comment and post vote is federated and is a single row in PostgreSQL tables.
Each person and community has a home instance. They only accurate counts of total post, comments, and votes is on that home instance. And even then it is theoretically possible that to save disk space and/or improve performance, the home of a community could purge older data (or have loss of data).
For lemmy to lemmy, I think instead of federating the votes independent of the posts and comments, there could be sharing of aggregate data directly.
The model for this is how profiles of a person are federated. When a person revises their profile on their home instance, every other instance has to get the updated change. Such as a new image or revised bio. Same with the profile of a community is revised, such as changing image or sidebar of a community.
The code redesign could start out by making person and community aggregate count sharing part of those revisions. Those numbers are not that time-sensitive, the statistics of the number of users, posts, comments in a community could be behind by many hours without any real impact on the end-user experience of reading posts and comments on Lemmy.
With votes, posts it is more tricky. But some experiments could be done such as only sending post aggregates when a comment on that post is created, deleted, or edited.... and a more back-fill-oriented bulk operation take care of correcting and discovering out of sync information.