#dusk $DUSK
Everyone argues about consensus. Almost nobody looks one layer below it.
@Dusk doesn't use random gossip to move blocks between nodes. It uses Kadcast — a structured overlay, where each node's position determines who it forwards to. The docs give the reason in one line: less bandwidth, and more predictable latency.
Predictable is the word that matters here.
Random gossip is robust but noisy. A message might reach you in 200ms or 900ms depending on luck. For most chains that's fine. For a chain selling ~10 second deterministic finality to institutions, propagation variance is not a cosmetic detail — it's part of the settlement promise. You can't guarantee finality timing on top of a transport layer that shrugs.
The second half is the audit. Blaize reviewed the Rust implementation and scored it 9.8 out of 10, but the findings are the interesting part: deviations from the original Kadcast specification, missed edge cases in idle-node processing, and ambiguous handling of reserved fields in message headers. All resolved or verified except two informational items.
Reserved fields and idle nodes. Not glamorous. Exactly the kind of thing that turns into a weird production incident three years later.
The open question I keep sitting with: a structured overlay means the topology is derivable rather than random. That's what buys the predictability. Does it also make traffic patterns easier to observe for a chain whose entire value proposition is confidentiality? I genuinely don't know, and I haven't found a public analysis that answers it.
For a privacy chain — would you trade some propagation predictability for a messier, harder-to-map network?
Everyone argues about consensus. Almost nobody looks one layer below it.
@Dusk doesn't use random gossip to move blocks between nodes. It uses Kadcast — a structured overlay, where each node's position determines who it forwards to. The docs give the reason in one line: less bandwidth, and more predictable latency.
Predictable is the word that matters here.
Random gossip is robust but noisy. A message might reach you in 200ms or 900ms depending on luck. For most chains that's fine. For a chain selling ~10 second deterministic finality to institutions, propagation variance is not a cosmetic detail — it's part of the settlement promise. You can't guarantee finality timing on top of a transport layer that shrugs.
The second half is the audit. Blaize reviewed the Rust implementation and scored it 9.8 out of 10, but the findings are the interesting part: deviations from the original Kadcast specification, missed edge cases in idle-node processing, and ambiguous handling of reserved fields in message headers. All resolved or verified except two informational items.
Reserved fields and idle nodes. Not glamorous. Exactly the kind of thing that turns into a weird production incident three years later.
The open question I keep sitting with: a structured overlay means the topology is derivable rather than random. That's what buys the predictability. Does it also make traffic patterns easier to observe for a chain whose entire value proposition is confidentiality? I genuinely don't know, and I haven't found a public analysis that answers it.
For a privacy chain — would you trade some propagation predictability for a messier, harder-to-map network?