Gossip swallows unicast errors

[vistra]

#1504 changed the behaviour of gossip multicast so it will attempt transmitting to the entire list instead of stopping on the first failure. As a side effect, the multicast call does not return an error even if some of the transmissions failed.
Since unicasts are implemented as a multicast to a list of length 1, the resulting behaviour hides unicast errors.

Some possible solutions are:

• Create a separate unicast API that can return an error
• Change the multicast behaviour to return an error in case of a list of length 1
• Change the multicast to return a list of errors, one for each transmission failure

[electricmonk]

@OdedWx do you have a preference here? I'm in favor of a separate unicast API, but I could live with a list of errors as well.

[netoneko]

Do you guys think it would make sense to move to something more established, like libp2p?

I am asking this because we will inevitably solve the same problems that authors of libp2p faced in the past, and IPFS uses it successfully in production.

Do we really have to build our own network stack? Maybe we should research the alternatives.

[electricmonk]

it's actually not a bad idea, instead of developing real p2p (relay) capabilities

[netoneko]

This error did come back to bite us in relation to #1526: the halving of the chunks never happened because we never returned an error that would trigger it.

[netoneko]

Relevant conversation: #1514