> Ok, it's probably a pain if one of the black boxes fails, and I do know enough about the underlying architecture behind the services I'm using - I could probably replicate (badly) most if not all of them given time and money, but why would I?
I don't think this is a possibility whose importance can be understated. In most of my apps running on IaaS/PaaS offerings like heroku, key pieces of the infrastructure tend to fail for short times (typically very inconvenient times) fairly frequently, on the order of once or twice a week. The relative cost of this is highly variable and business-dependent, but at best its mildly inconvenient and at worst can be crippling. Its definitely something that needs to be accounted for. That said, I've also been lead on apps that used the same pieces of server infrastructure that had far better uptime when I handle ops myself instead of delegating that to another provider. (Not to pick on the providers—they have a very difficult job to do in a cost-effective way, and I'm quite certain that they do it better at that scale than I would.)
The choice of Erlang for WhatsApp was strategic if for no other reason that that it allows them to dispense with some of those choices. Many of the typical use-cases of Redis, for example, are easily supported by OTP.