I can't say too much but I work for a company that supplies software to businesses that need to take thousands of transactions a minute on data that's got to be current with real-time and frequent price changes streaming data to 10s of thousands of people and third parties that are themselves big companies.

Downtime is not tolerated - I've been surgically torn apart by boards of major listed corporations over small amounts of downtime (it was something of a pleasure to be so professionally abused, strangely).

For all this we have found the best solution to use a monolithic DB with SAN-based replication, C libraries to connect to the SQL db, and a process-based custom-built application server with an embedded (unfashionable) scripting language for ease.

None of these things are fashionable, but no-one's beaten our performance.

The trouble (from my perspective) with all these social sites is that downtime is tolerated (Facebook is always screwing up; I consider them a partial joke technically) and the data is easily distributed (think Google) because the demands on consistency are low.

I'm sure our time will be up one day, but you go trendy at your peril. I'm an advocate of functional programming languages, but it's not realistic to get hundreds of these guys to switch seamlessly and reliably to a new paradigm without serious pain. Ditto NoSQL.