What about Shingle recording ? There are prototype drives already built that can near 100TB on a single device. It has limitations, of course, like 2GB+ block sizes, but it'll get interesting. Give it 3-5 years (as usual, for mythical tech that might ship someday).

  So, this is bullshit. You can keep secrets as long as the people involved think secrecy is warranted.

  Google have an astonishing track record of not leaking projects to the press. They've worked on some incredible stuff, and the vast majority don't get leaked at all, or get leaked accidentally. Huge numbers of internal/infrastructure projects never get told about outside the company. Sure, some projects are pre-announced because by working with outside companies they assume there will be leaks (ChromeOS, Android).

  Internally people get told "Please don't leak unannounced projects. A leak could cause your co-workers to have to launch an unfinished or unpolished project ahead of time, reducing the impact of months or years of their time".

  The problem with Apple is that they work with a lot of outside agents, all of whom can leak without thinking of the personal consequences to friends, just financial/legal ones (which can be avoided). Their own engineers have a pretty good track record of keeping quiet about 'important' things.

"That's nearly twice what Google spend on a datacenter" - yeah. They'll likely stock it with Xserves, rather than decent priced hardware!

Yeah, we had problems in Google with these too; we have large networks of machines that used to use multiple different NTP servers (for resilience). Turns out not all NTP servers implemented leap seconds the same way, and many cluster based applications get upset when they aren't synchronized to within 100ms.

  Now, we run a dry-run of a fake leap-second with all software a few weeks before the leap-second failover. It's the only way to be 100% sure that applications changed since the last leap second won't have problems. Though, most unittest frameworks now have the ability to implement second skewing, since the suffering caused by the 2005 leap second.

  The main problem is that the POSIX description of how to do a leap second is retarded; you basically go from 00:00:00 to 00:00:59, some apps also get upset when they see the same time twice.

John

  Um...not as such. There are many parts of Google Apps that are better off running on a cluster than a 'server farm'.

  Especially searching. If you need to search through your 100GB of indexed documents, you want to be able to transparently break up that search query over multiple machines.

  Obviously, a cluster is just a fabric running on top of a server farm. But having a fabric that spots slow/loaded machines and repairs or unloads them, and gives a rich API to things like GFS/Bigtable/Chubby/Mapreduce etc. means a big difference to Apps developers.