If you've been following the discussion over Mumps : the IT world's best kept secret? you'll have noticed that one thing that stands out about Caché / MUMPS is its unusually structured data-store of persistent, multi-dimensional ragged-arrays. And Rob Tweed makes a case that this is a better fit than a relational database for internet scale applications. He cites Google's BigTable and Amazon's S3 as examples of a trend away from relational dbs in favour of key value pairs.
Today Steve Yegge has an interesting meditation on a similar theme, about "domain modelling schools" and the need for programmers to have experience of multiple tools and organizing principles. (And while he doesn't note the MUMPS model, it would be another interesting one to add to the list.)
Worth reading, also, this Cringely piece on how relational databases don't work so well in a parallel cloud where disk-access is often a bottleneck so more of the live database is moved into memory. (Killer quote : "Google has THE ENTIRE INTERNET IN MEMORY AT ONCE.") Disk is just used for backup / persistence."