Scenario: Servers has to be restarted (or crash) during rush hour with 100M active users. Even being generous and allowing for 10000 logins per second, it will

Scenario: Servers has to be restarted (or crash) during rush hour with 100M active users. Even being generous and allowing for 10000 logins per second, it will take almost 3 hours for everyone to relogin. Oops. Ok, so that 100M is clearly not viable for services that actually expect a significant amount of their users to use the service at roughly the same time (that is - it would clearly not handle Facebook amount of load). A much more relevant metric would be how many simultaneous sessions you can handle while being able to claim good service (say, the 99th percentile being able to login without getting time-outs). Also, generally speaking, most authentication systems will involve at least a few server-side writes for logins, meaning the difference to new registrations shouldn't be all that big in a well designed and compartmentalized system. Disclaimer: Yeah, I used to work for a company that did authentication servers for ISPs.