Luxas on Nostr: It is using read replication with the D1 database. Each of the worker invocations ...

It is using read replication with the D1 database. Each of the worker invocations gets its own sqlite connection to a local replica closest to the end-user, so it is effectively running thousands of independent sqlite instances in parallel. I agree on the write performance, as that could still be a bottleneck to the single D1 instance. But as you're well aware, Nostr relays are usually a read-heavy workload, so this at scale should outperform a local postgres database on a single machine.