(I’m assuming that Google maps users to servers in some way, presumably with a primary cluster, a hot backup, and an emergency data backup in the Bahamas. That would seem to be a reasonable strategy from a response-time perspective, allowing the big clients to be grouped with small ones for load balancing, or for servers to be geographically optimized. This is not unlike the way Facebook used to cluster servers by university network. Also note that Microsoft does the same thing with Hotmail, they just roll changes less frequently.)

Under the above assumption, Google can offline a primary cluster (swapping in the hot backup), run the update, online the primary, and repeat without adversely affecting overall availability. Rolling out in waves also allows allows them to catch scalability/functionality bugs. So, on the one hand you have the “can you switch me to the new stuff” email, and on the other you get either the “hey, why can’t I access the system” or the “hey, the software’s buggy” email. I suspect the Suits would rather send the former type of email than the latter, assuming it is made clear that they can’t say “but I always want it new and right”.

As to annotation, I wonder if it would be more beneficial to integrate automatic annotation into Analytics, or to provide a convenient way to automagically extract raw data from Analytics ala the Twitter API. The latter seems like a more general solution to the problem of doing site-specific data processing / rendering, and would be rather more “Web 2.0″ or “Social Web” or whatever the right buzzword is. Being able to do easy raw data extraction would also provide the opportunity to do fun things like displaying stats in the site dashboard or on the pages themselves when viewed by an admin without needing to further mod Analytics itself to support those features.