I will abstain on this wrangle of "when",
Instead I'd like to address some of the network topology health issues that's been brought up in this debate.
Due to how blocks are being broadcast by miners at the moment, it is not difficult to find the origin node of these blocks. These more influential origin nodes are a minority, about <100 out of ~6000, <2%. These data points are important to certain attack vectors. It is highly recommended that pools adopt broadcast logic that rotates broadcasting nodes and increase their node count.. Eloipool has this implanted for those seeking to adopt/see it in action in the wild.
China is a particular worse-case due to the sporadic nature of their internet infrastructure, especially connecting from/to outside of gfw, on a average node-walk I can get up to a 10% difference while I know for a fact some of the nodes shown to be down are up.
In F2Pool's case, I see 6 replay nodes, I don't know if that's enough or that's all the nodes F2Pool runs, but it may be beneficial to set up multi-homing with shadowsocks over mptcp to increase the stability. also see if you can get a CERNET connection to be part of your rotations since their backbone is quite good.
comments, question and grievances welcome.