The maximum concurrent logins per hour is calculated by running through the logs and incrementing the counter for every login and decrementing for every logout, and finding the maximum for each hour.
The method of doing this was changed on 30 Oct 2002. Before that, only logins that had matching logout records were counted. This underestimated the number of concurrent logins, but only slightly. I believe that most logins that don't have a matching logout don't last very long.
However, this seriously misrepresented when the user last logged in. Both the general stats and the last-login date are based on login and logout records, the same program was used to generate them. We must be more cautious in determining the last-login date because the deletion policy depends on it.
So, on 30 Oct 2002, we decided to forgo the matching of the logins to logouts. It sped up the stats program considerably. But the result is an inaccurate jump in the maximum concurrent login stats -- there was no jump in the load on the servers at that time.
The log entries show that there were many cases with the pattern: