I’d say that in an ideal world (no chance of attack, nodes with enough resources and correctly structured mesh) the gossip part would probably not be needed. But given that’s not the case, it’s helping a lot. The original post didn’t imply removing it at all. We just hinted on optimisation potential.
That’s the next metric we’ve been investigating and we’ll publish a similar post in a few days. As a spoiler, I can say that about a quarter of outgoing bandwidth is spent on IHAVE messages, which is quite substantial. More to follow on this - I’ll link from here.
Glad to hear the bloom filter approach sounds interesting. There’s lots to be investigated before a final design and implementation, but yeah, I agree there are edge cases of false positives that should be avoided.