Gossip: more reliable publishing #387
Conversation
|
Can't find the definition of "direct peer" in the spec, but per my mental model two direct peers mutually configured as "direct peer" always exchange messages among each other. Then I also wonder why direct peers are blindly considered for outbound even if they don't subscribe to the specific topic (but this is another question :-), and maybe it's fine) |
| if (topicMeshPeers != null) { | ||
| // we are subscribed to the topic | ||
| val addFromNonMeshCount = max(0, params.D - topicMeshPeers.size) | ||
| val nonMeshTopicPeers = (getTopicPeers(topic) - topicMeshPeers) |
There was a problem hiding this comment.
Do we care about the score of the topic peers in this case?
There was a problem hiding this comment.
Do we care about the score of the topic peers in this case?
I think the scoring is already "captured" by the mesh. Outside that, seems reasonable to me to pick random peers.
On Friday I had a similar comment suggesting "why don't we select by sorting them by score" and Anton had a good reply to it.
There was a problem hiding this comment.
Yeah, good question. Here is the quotation from the spec:
PublishThreshold: when a peer's score drops below this threshold, self-published messages are not propagated towards this peer when (flood) publishing. This threshold should be negative, and less than or equal to the gossip threshold.
I believe this mean that both 'flood' and 'regular' publishing should consider this threshold.
It's probably counter eclipse attack measure.
So I believe we still need to truncate those with the score < PublishThreshold but still consider them when there are few peers to publish
Does it make sense?
There was a problem hiding this comment.
Yes that seems like a convincing interpretation to me.
@tbenr you are right. There is a mess with direct peers. Just created a separate issue for this: #389 |
|
Addressed the peer score question when publishing. |
|
|
||
| test.mockRouters.forEach { | ||
| it.subscribe(topic) | ||
| // it.sendToSingle(createGraftMessage(topic)) |
The quote from the Gossip spec:
According to spec we should forward the message we are publishing to the mesh peers only.
However the mesh could be not that stable (for example a peer may be evicted from mesh due to slight gossip downscorring (any score
<0)). Thus it could lead to the situation, when the mesh is empty but there are still peers subscribed to the topic outside of the mesh. According to the spec we are not able to publish.The situation could not be that dramatic due to
IHAVE/IWANTmechanism which should finally make the message published, but that could induce significant publish latency.To overcome this it does make sense to take extra peers outside of the mesh when the mesh size is
< D.This is also the subject for a PR to the Gossip spec