You pick up the phone, you get a dial tone, not a fail whale. You put in a plug, you get electricity. When a bank's core systems go down, it makes the evening news. It's a whole 'nother world from OMG sharding RoR lol that web kids think is "scalability".
It also really sucks when you run into a problem and can't open the damn thing up. Case in point: I was working with some code that was instrumented with HP OpenView OVPA to collect performance analytics. When the new version of OVPA came out, we were planning a huge rollout to bring all 4k+ servers up to date. Problem is, the new major release had moved to a pthreads library and changed a bunch of the threading code, causing a 2 second lockup on start due to some sloppy code. For the vast majority of server software, this is almost undetectable, but these were CGIs that were invoked millions of times per hour. It took me about a day to locate the problem with truss (Solaris' strace). I had to spend a week in 5am meetings with HP's offshore development team to convince them that it was their bug. Of course, it took another week to get a patch.
Unless you need many thousands of servers the majority of funded web application companies could very well afford the best of what the OTP world has to offer them. But since that isn't 'sexy tech' they want to re-invent those wheels.
Twitter has since switched to Erlang, I think they've realised that customer satisfaction beats new technology any day.
http://www.theregister.co.uk/2009/04/01/twitter_on_scala/
Memory error, I should have my dimms checked ;)
> You could perform the same services by a rack of servers running OTHER software.
But with the software that is there today the cost of re-writing it would be prohibitive and that's why the mainframe market is still as large as it is.
It's nothing to do with them being physically large.
Converting a couple of billions of lines of COBOL to something more modern, re-test the whole thing and add a bunch of magic to get the same kind of reliability is not a task any sane CTO at a bank or insurance company is going to sign off on.
Just because one x86 box doesn't have the required level of reliability doesn't mean you can't build a reliable system out of them.
You can't define mainframes as a separate market just because they're physically large.