the discovery should probably just happen once a day on the backend job, then each user is just served a cache based map on their existing list. basically your backend should know 99% of all npub's outbox relays without checking for each user.

then a user comes with a follow list, you should be able to use that to return a cluster map nearly instantly

then req all latest notes limit 1 per author to every relay at once, a browser can handle easily 50 websockets