Sorry, we had more downtime earlier. Lemmy seems to guzzle up RAM, and when it ran out everything started to collapse.
I've decided to schedule nightly restarts of the lemmy backend, which seems to be all that's necessary to free the RAM. Now we are on 0.19 there is a persistent federation queue, so a restart no longer causes all in-progress federation to be lost.
I've scheduled it for 3am NZ time, when the least amount of people will be online. Hopefully this solves it!