This is a concept that doesn't even exist in most other systems, so there isn't an agreed name for it (we call it a "slice" in Spyre). It's similar to a Kubernetes "pod" but not quite.
Exactly - we really struggled with this. A node (for example, when doing a Docker info when having $SWARM hosts) implies to the user physical isolation. As a distributed systems nerd, I pushed for "anything that doesn't make a contract it's on a different machine". Segment, slice, pod, chunk, block - something other than implying the isolation and therefore fault tolerance of the overall system.
This level abstraction seems unnecessary. We need to just let-go of the concept of physical or virtual hosts. A machine should boot and join a cluster, advertising its capabilities (ssd drives, gpu's, enhanced networking, it's availability zone or region, etc...) and you should never need to think about that stuff, period.
I don't care how many devices/segments/nodes/slices/dynos/widgets are in my cluster. I care that I have X GB of total memory and Y cpu's. I want to check a box to make an app or service highly available (on more than one node, and in more than one zone) and Ronco™ set it and forget it.
Everything else is just noise.
I love the ideas behind Kubernetes and Fleet/CoreOS (and this) but everyone is SO excited about these low level technologies that have yet to be composed into beautiful experiences.
A tweet from Kelsey Hightower sums it up perfectly:
It's going to be nearly impossible for people
to evaluate and chose a container management platform
during the gold rush.
Absolutely agree with that. We released yesterday a project called Hypernetes[1], which has been designed to solve the "I care that I have X GB of total memory and Y cpu's. I want to check a box to make an app or service highly available" problem.