I think you have a good thought on a small scale, but it doesn't work on a larger scale when you're serving billions of searches a day, because there would be too little accuracy in the results, plus they have the resources to pull a data center off line and continue to function, so what they do is create a live working version and then take one (or more) data center 'offline' or 'out of the loop' and then update that data center while still serving the old dataset on the rest, then they repeat the process until they've 'rolled through' all of the data centers... This is why Caffeine's been visible off and on, but is not available all the time yet. (AFAIK)
Basically, right now they have two full working versions of data they are serving results from and they are moving through data centers, removing the old dataset (Big Daddy) and replacing it with Caffeine while the data center is unavailable for public view, but when the switch over is complete the old datasets will be gone from all data centers and Caffeine will be 'live' all the time for all searches and searchers.
So, right now Caffeine 'jumps in and out' because people hit different data centers on different searches and some have Caffeine installed and others are still running Big Daddy.