To run a witness, especially, on a brand new chain, would only require about 2gb of memory, even with a slow spinning disk, at least for backup. But if it was getting a block every minute you might find you want it to have an SSD, and the connection has to be super reliable.

I take a bit of a different angle on reliabilities of witness servers - so long as, I figure, there is 22 witnesses, it doesn't matter if half of them miss blocks a few times a day. I may look into it down the track to have the scheduling system account for this and deprioritise failing nodes automatically and slowly re-add their slots in the schedule.

There is no real issue if the witnesses are recalculating the schedule every few minutes, in reality, it's not that complex a process, I don't believe it impacts on block production. The way it works here, and with the culture, it makes the job very rigid. If 8 out of 10 blocks make it on, then the total capacity of the network is only reduced by 20%, and until it reaches over 3000tx per block do you start to see a ceiling on this. or so. I forget how fast it goes, but the network has never been saturated yet, at 95% or so uptime on nodes.

As a rule, nodes only have something like 6 seconds to prepare to make a block. The schedule is not mapped out very far ahead, for obvious reasons that this would make attacking by schedule, a specific node difficult.

I personally don't think the scheduling system is the optimal solution, at all, but it seems to work so far.

By the way, running a witness, you would not want to rent by the hour, as they have to be running every 3 seconds to grab new blocks. I can show you VPS services that will cost you under 10 euros a month and be ridiculously ample.