I very rarely see events older than a month in my timeline having been reacted to, commented on, etc.
Most relays could store a rolling 1-3 months of data and be very useful. Keep hot the metadata kinds 0/3/10002 and then a month of data.
The concern however is a similar problem to broken http links on legacy internet.
I think the archive relay is certainly the direction headed. We don’t need to keep 2 year old data indexed as well as recent. Or we can purge older revisions or replaced events (if stored).
An alternative to the archive relay is paid relays that persist a pubkey/members data for longer periods or forever. I think this will be common too.
All relays all data.. that will die the next couple Nostr 10X jumps.