Merrill is buying top of the line Cisco gear. Google is making its own switches and routers because of costs. Cisco is expensive because the median time to fail

Merrill is buying top of the line Cisco gear. Google is making its own switches and routers because of costs. Cisco is expensive because the median time to failure is so high. Google and Amazon own their datacenters, but they're playing the bargain basement game. You won't see racks upon racks of Nexus 7k gear in Google, at least not to my knowledge (please correct me if I'm wrong). You're right that live loads are different from test loads (entropy) but having a separate set of test servers still holds a lot of value. It's just like Aphyr said, when your tolerances are higher, but your costs of delivering said tolerances as an aggregate are lower, you can afford to over-provision. A lot of these failures come from blocked I/O or network capacity failures that just wouldn't happen in a production finance environment because of over-provisioning. I hope this lends some light onto your queries :).