Fediverse servers can quickly get more expensive if you have a few thousand users, or even a few dozen but somebody has a post go viral. That’s because every retrieval of a post always goes to the original user’s server, every like does too, etc, and this generates a flood of events which quickly gets expensive to process.

Just ask the maintainers of the botsin.space Mastodon server who couldn’t afford to keep it running, and now put the server in archival mode and not allowing new posts.

A PDS only publishes static data and don’t have to process incoming events, making it very easy to run one behind a caching server very cheaply.

There is another problem: these other relays are all copies of the Bluesky relay, where the official app publishes the messages of its users, so they are not independent from each other; if I publish my posts on a relay other than Bluesky’s I will not be able to communicate with them.

Not entirely correct.

Every individual users’ account host (PDS) publishes directly locally, the relay then collects published posts from known PDS servers (including both bluesky’s own and others’ self hosted servers) and display everything. A PDS server can sync to multiple relays. Relays can even sync to each other, which is practical because PDS servers publish through content addressing for posts in user repositories so it’s easy to verify completeness.

So sure if somebody uses an app connected to a filtering / partial / out of sync relay they might not see everything. This is not an architectural limit in the protocol, however.