Remove granular role registration

[ravenac95]

Re: Core Platform Meeting (2019/09/17)

The registry currently allows granular registration of roles within the compute committee but this should be changed so that the per-runtime compute execution roles (txnscheduler/compute/merge nodes) are now a single "compute" role from the viewpoint of the registry.

The scheduler should still schedule individual roles as-is and would pick from all valid nodes registered with the new "compute" role.

[bennetyee]

@lostnfound incentive Qs: how worried are we that we would get too many servers of one type and not enough of another? should we take the max of the hw requirements of non-validator compute and validators and merge the two categories together?

[kostko]

should we take the max of the hw requirements of non-validator compute and validators and merge the two categories together?

In light of recent architecture discussions, we should probably not combine validator and other compute node roles as we can imagine there being validator-only nodes and there being nodes that only run specific runtimes?

So it should be at least four classes:

• Global: Consensus nodes.
• Per-runtime: Compute execution nodes (txnscheduler/compute/merge).
• Per-runtime: Storage nodes.
• Per-runtime: Key manager nodes.

[kostko]

Updated the issue description to be more clear what needs to be done.

[pro-wh]

txnscheduler was not a per-runtime role

edit: hm neither was merge

edit: hm, other parts of our code consider txnscheduler to be a per-runtime role

edit: oh for some roles suitability is per-runtime but registration is not per-runtime. but it's not consistent for the merge workers right now

[kostko]

Oh just noticed this while implementing the runtime maintenance fee payments -- it seems that the committee scheduler currently does not take runtimes into account for merge and storage nodes.

This is incorrect and should be fixed (all these things should be per-runtime instead of global). Can you do it as part of this issue?