Some of you may have noticed that federated actions are slow to synchronize between Lemmy instances. This is most likely because of the setting "Federation worker count" under /admin. It determines how many federation activities can be sent out at once. The default value is 64 which is enough for small or medium sized instances. But for large instances it needs to be increased.
Grep the server logs for "Maximum number of activitypub workers reached" and "Activity queue stats" to confirm that this affects you. For lemmy.ml I just changed the value to 512, you have to experiment what is sufficient. The new value is only applied after restarting Lemmy. In my case changing the value through the website didnt work (maybe because its overloaded). Instead I had to update local_site.federation_worker_count directly in the database.
Edit: I had to increase the value to 160k for lemmy.ml. Now the stats arent getting logged anymore, so Im not sure if the pending queue is still building up or not.
Going from 512 to 160,000 is a massive parameter change.
Network replication like this presents a ton of issues with servers going up and down, database insert locking the tables, desire for backfill and integrity checks, etc.
Today things are going poorly, this posting has an example: https://lemmy.ml/post/1239920 -- comments are not showing up on other instances after hours.
From a denial-of-service perspective, intentional or accidental, I think we need to start discussing the protocol for federation. When servers connect to each others, how frequently, how they behave when getting errors from a remote host.
Is lemmy_server doing all the federation replication in-code? I would propose moving it to a independent service - perhaps a shell application - and start looking at replication (and associated queues) as a core thing to manage, optimize, and operate. It isn't working seamlessly and having hundreds of servers creates a huge amount of complexity.