• which are "obvious" and perhaps even necessary for productive coding of distributed systems software; and
• which have huge economies of scale (one shared cluster for all customers beats the pants off the performance+availability of your company's puny little three-node private cluster), and yet...
• which other major clouds don't support at all.
Effectively, all the "clouds" currently only offer between 30% and 90% of what you'd want in something that called itself "a cloud." Nobody has a "whole cloud" (AWS is closest, but still not there.)
Designing for portability between these clouds would be like writing assembly intended to be portable between processor architectures, when only one architecture had an ALU, only one had registers, and only one could conditionally branch. It would be madness.
---
Personally, I feel like, to be able to sensibly design for portability between cloud providers, they'd need a lot more features in common than they have now.
Maybe we could invent a minimum common standard to hold the cloud providers' stacks to—maybe a small one at first, with a growing list of expectations over time; or maybe a "core" spec, and then a number of "levels" of support atop it. Then you could say you've targeted "IaaS Level 3", and clouds could claim to support that, and cloud-abstraction libraries like Fog could actually do something useful.
