This makes me wonder what the testing method is for this sort of thing. Presumably you'd do as much as possible in VM's to get repeatable states with updates from every conceivable starting point.
But what about physical machines? Exactly how many repeatable update tests on physical machines are done? You get a combinatorial explosion if you need to test on 100 different configurations for hardware and you have 100 different starting points in your pre-update OS image. But something like that (10k tests on physical machines, millions for VM's) is what I would _expect_ they are doing.
I expect this too, but at such insane scale, it won't catch everything, and no one should expect this. The only solution (as far as I know) is very slow releases (starting with something like 0.01% of users) and watching what happens.
But what about physical machines? Exactly how many repeatable update tests on physical machines are done? You get a combinatorial explosion if you need to test on 100 different configurations for hardware and you have 100 different starting points in your pre-update OS image. But something like that (10k tests on physical machines, millions for VM's) is what I would _expect_ they are doing.