i think that it needs to be a big fast cache and many small archive repositories. renting a small one that only answers to you and to the caches you make arrangements with doesn't have to be 99% uptime, it can be down for a few minutes and only the very newest events probably won't already be cached.