Because those who ignore the pub-sub research IBM did in the 80s are doomed to re-implement it badly.
The problem in large scale information networks, more in the Facebook way than the Twitter way, is that you you potentially trigger a cascading effect in information updates if you go pub-sub. Specifically, applications that do interesting things with social graphs have to go beyond basically doing message passing.
I naturally look at things from a recommendations angle, but if you've got a new edge that enters the graph that may affect other edges that are connected to the end points. Those may in turn affect the edges that are connected to those nodes and so on. You want to avoid something that effectively becomes a breadth first traversal of the graph doing updates since that's well, slow, to put it mildly.
This is why large-scale graph algorithms like PageRank work on constantly regenerating static matrices rather than doing regeneration of the ranks dynamically, but that naturally is problematic when you're working on data sets where the most recent data is the most important and is being generated at very high rates.