Because even a "3ms latency" (which was a problem, with respect to "global", that other people have commented out) can absolutely kill the performance for IoT data that may be emitting thousands of updates a second.
Those systems are largely highly localized, and so /strong eventual consistency/ is more important than globally consistent blocking operations.
Also, again with 5B+ people coming online, Master-Slave systems (even distributed ones) still have a huge bottleneck already in the present day. P2P systems (master-master) will scale better in these settings.
I was more curious about the "necessary P2P future of the web" part.
I think there's an assumption here that most of the responsibility for storing the source of truth will move out to things like IoT devices (i.e. fog computing).
And sure, there will probably be a need for that. But regarding the assumption that most web services will go away, I don't there's sufficient evidence to bet on it happening anytime reasonably soon. Data centers and public clouds will probably still be there in the next decade or two.
Twitter is spending over $15M/month on server costs alone to support 333M active monthly users.
Now compare to Pokemon Go's huge explosion of 20M daily users from a while ago.
This problem is only going to get worse with another 5B+ people coming online into the 2020s.
In order to scale, using (what you call) "fog computing" will be absolutely necessary. Cloud services will still be used, of course, but they will be built as P2P systems to take advantage of the "fog".
Cloud infrastructure will always be around, but how apps are built will be a fundamentally different architecture. But when S3 goes out, like it did the other week, we can't suffer worldwide downtime - that will be unacceptable.
Rethink's unfortunate failure to capitalize in this market is a signal that Master-Slave databases (even the best of the best) will have a very small role with respect to the total amount of data flowing through the internet.
Those systems are largely highly localized, and so /strong eventual consistency/ is more important than globally consistent blocking operations.
Also, again with 5B+ people coming online, Master-Slave systems (even distributed ones) still have a huge bottleneck already in the present day. P2P systems (master-master) will scale better in these settings.