Joyent's Triton avoids this completely: Their whole data centre is a Docker "host", and you never have to care about it. The way it should be.
(I'm not paid by Joyent, nor am I a customer -- I just like what they've done to push the model forward.)
My big question is around a service like AWS Lambda. Is that not already the logical conclusion of container based computation? If magic units of computation can run instantly on demand, what more do you need?
No hosts, but also no OS images, and no specific containerization tools!
> If magic units of computation can run instantly on demand, what more do you need?
Some control over flow, scaling and batching wrap everything up for me. Startup times for my code are non-zero even if the environment is, and adding on queues with a "batch grab" means I can scale things far more sensibly (I can cram a lot of stuff into a single matrix mul if I can pull 100 items at a time from a queue).
I really, really miss picloud :(
To that extent, I'm not sure that tossing out the idea of hosts is great yet.
Can you expand on this please? I was not aware of the problem.
It's not any worse than you'd have if you were starting from scratch with your own hardware, but Bryan Cantrill is firmly against magic block storage. His suggestion is that you rewrite your persistence layer to know about SDC or "deal with it".
It's not necessarily that bad since a lot of containers don't need state, but my personal axe is that I work with databases almost exclusively.
But I agree... At this level of abstraction ECS is far more attractive because of the rest of the AWS services and ecosystem.
Joyent deserves a lot of credit for removing the extra layer. However Triton is missing key functionality like a good load balancer.
Abstractions are important, but pure functionality is more important. How can anyone beat AWS?
(Disclaimer: I am building open source tooling in top of ECS, and have published a head to head review of ECS vs Triton)