RE: Attachment, I don't understand this at all.
For the twin peaks of the users, I could understand that if perhaps there was an 18-hour or so timeout on the user logins, so that they start dropping off faster than new users are logging in. The response time on the other hand is baffling; the drop in users online is only minor, but the effect on the response time is enormous. I find it hard to believe that there's genuinely far less people online at that moment.
All I can think of is the way the thumbnails are now handled* is causing problems, and it's interfering with serving full images as well (I'm assuming the thumbnail response time will match up to the full response time like the small part currently does). Maybe once all the latest submissions (especially popular ones) are cached it initially takes the strain off later sessions (when people check their notifications and see the same work), then the cache starts overwriting things as people browse around and performance degenerates again. I don't even know if that's possible or how it works, but that's the only thing I could come up with to explain such a massive difference in response time in comparison to such a small user discrepancy.
Unless of course there genuinely is a remarkable dropoff in users. The west coast goes to bed while the east coast haven't logged on yet, and everyone in between just doesn't exist. I find that unlikely, otherwise such a pattern would surely be visible on other sites too, e.g.
Steam.
*<yak[away]> in this case, completely removing the entire filesystem storeage and upload related part of Alkora's code and replacing it with something that ... uses a caching thumbnail service to make previews of the original upload in whatever resolutions the site requiers
<yak[away]> restrictions being arbitrary of course. to save on disk space and cache-miss thumbnail generation times
