The latency limit is caused by the activity queue that was introduced in v19.

Servers can only talk as fast as round time allows, because Lemmy instances now keep track that each event actually does get federated, and in the right order.

That last point means each event only gets sent once acknowledgement of the last one is received, creating a hard limit for how many events can be communicated, depending on ping. A mere two per second with a latency of 500ms.

This serial process will obviously need to be parallelized. But that’s difficult.