Relays should store observed bandwidth in state file
arma suggested I write a proposal for this, but I believe it is a flat out bug. When a relay goes down, it looses its 24 hour observed bandwidth window. For fast relays, this can be catastropic, as it takes weeks for them to regain their bandwidth due to our slow descriptor refresh interval. The slow descriptor refresh interval can be shortened, but that requires careful study to make sure we do not increase the descriptor load on clients.
If we just instead stored this info in the state file to survive crashes, high bandwidth relays would still spend a loooong time gathering bandwidth (and maybe this is a really good thing?), but will at least not have to restart when they die.