Re: [Din] draft-mazieres-dinrg-scp-04 peer priority algorithm

[David Mazieres <[e-mail]>]

Piers Powlesland <[e-mail]> writes:

> As I understand, this filtering could lead to a situation where the
> neighbours set is empty, since there is no guarantee that any hash
> calculated above will be less than '2^256 * weight(peer_id)'. Is this
> intentional?

There is a guarantee, because a node must belong to all of its own
quorum sets.  So even if a node has no other neighbors, it will still be
its own neighbor... which highlights the fact that maybe "neighbor" is a
misleading term here?  Or maybe at least suggests we should mention
this.

> Also what is the motivation for creating the "neighbours"
> set as opposed to simply selecting a peer from all peers based on
> their priority weighted by their weight?

The goal of this neighbor/priority scheme is twofold:

   1. Ensure that nodes that are not very important (because they appear
      only in a few quorum slices) also do not very often get to dictate
      the consensus value.

   2. Make it likely for nodes to listen to the same nodes for
      nomination values, so as to minimize the number of different
      values nominated.

The neighbor selection achieves #1, while the prioritization achieves
#2.  You seem to be proposing a unified mechanism that would ideally
achieve both goals at once, but don't think actually does.

As an example, suppose that my quorum slices consist of myself plus one
of 4 US nodes, plus one of 100 Chinese nodes.  (Assume the US and
Chinese nodes depends on one another in a robust way so the system is
safe.)  With your scheme, I will usually nominate my own value,
sometimes listen to one of the US nominated values, and almost never
listen to a Chinese node (because my own hash value would have to be
less than 0.01*2^256 and the US nodes would all have to be less then
0.04*2^256 for any Chinese node to have a chance).  In reality, my
quorum slice selection indicates that I consider Chinese nodes, in
aggregate, as important as US nodes--there just happen to be more
Chinese nodes.  So your scheme doesn't properly reflect the importance
of the nodes.

Does that make sense?

David