You only needed the full copy of all atproto data for the first version of the relay protocol. Since "relay 1.1" relays became a lot thinner.
Nowadays you can run a relay which maintains the current firehose of events + some amount of backfill (commonly a day, week, or month).
Appviews listen to the relay and can save what they care about and can look to other relays if they need more backfill.
So in practice you have your relays for regular use which handle large amounts of outbound traffic and then you have "archival relays" which store all or large portions of the history.
And in the eventual future "archival relays" will likely end up providing backfill for extremely old history via something closer to IPFS (it's the same underlying data structures so this isn't a major change, just nobody has done it yet).
And of course in the event a particular bit of history is missing, a relay can just ask the PDS for a new copy of the data.
-----
TLDR: The 20usd/month is for a relay with like a month or so worth of backlog attached and you can get by with less/cheaper or with more.
You only needed the full copy of all atproto data for the first version of the relay protocol. Since "relay 1.1" relays became a lot thinner.
Nowadays you can run a relay which maintains the current firehose of events + some amount of backfill (commonly a day, week, or month).
Appviews listen to the relay and can save what they care about and can look to other relays if they need more backfill.
So in practice you have your relays for regular use which handle large amounts of outbound traffic and then you have "archival relays" which store all or large portions of the history.
And in the eventual future "archival relays" will likely end up providing backfill for extremely old history via something closer to IPFS (it's the same underlying data structures so this isn't a major change, just nobody has done it yet).
And of course in the event a particular bit of history is missing, a relay can just ask the PDS for a new copy of the data.
-----
TLDR: The 20usd/month is for a relay with like a month or so worth of backlog attached and you can get by with less/cheaper or with more.