Lest anyone think I am just taking potshots at MSFT when I suggest that Google running on Win2003 might be less reliable than the home-grown architecture they use now, an anecdote might help.
When I worked at CNN.com, we ran almost exclusively Sun Solaris on Sun hardware and when I left, we had more than 100 machines in services. Having just one machine out of service, though it might be one of twelve identical systems, was a Big Deal and it didn’t take a day to get it resolved. Not that it happened very often: I’m trying to think of a second instance.
We had a fellow join us from MSFT and he took a tour of one of our data centers/machine rooms, remarking that it was nowhere near as impressive as Microsoft’s Internet Data Center (you could hear the initial caps in his phrasing). Why, we have so many machines, fully a quarter of them are down for maintenance or being rebooted at any given time, he explained.
A quarter. One in four. Twenty five percent. Offline, out of service . . .
So take the 10,000 to 15,000 servers mentioned in a comment to the previous post, and add 25% to the costs of the licenses you would need. It would get pricy very quickly.