Hello Antoine, This is an interesting problem, and introducing finer-grained traffic classes certainly makes sense. The three areas that stand out to me are peer declaration, topology inference and system bottlenecks. Peer declaration: Explicit signalling of specialised roles (Example - I only relay hot-blocks) to peers increases the fingerprint/profile. We already see topology inference attacks via relay behaviour; adding public role declarations may expand that surface. Nodes can already drop or deprioritise whatever they wish locally, so explicit signalling may not be necessary. Topology inference: Since topology inference can be drawn from tx-relay timing and relay behaviour, an internal class-based model also allows the node to randomise acceptance, forwarding, and scheduling behaviour per class. Even small amounts of deliberate jitter or probabilistic message handling make it far harder for an observer to infer which peers are responsible for block-relay versus tx-relay traffic. This further reduces the value of explicit capability signalling, since role exposure can be disguised at the behavioural level instead. System bottlenecks: Introducing multiple traffic classes as separate processes and sockets increases resource consumption, as you’ve noted. A single connection can already multiplex all P2P message types. A cleaner approach might be to integrate the class separation internally, without advertising anything to peers. Incoming messages can be classified (Examples - hot blocks, cold blocks, tx, address gossip, etc.) and per-class policies applied locally. Since a node doesn’t need to receive or forward traffic it doesn’t want to handle, it seems unnecessary to declare toggle roles at handshake time. An additional benefit of integrating classes internally is that it gives a natural place for per-class bandwidth accounting and load shedding. Under congestion, hot-block traffic could be prioritised while less critical classes (Example - cold block serving) are throttled, without multiplying sockets or increasing signalling surfaces. Externally the peer sees a normal connection; internally the node routes each message into class-specific queues and rate-limit buckets. This preserves flexibility (isolated CPU paths, independent scheduling) while avoiding multiple processes, multiple VERSION messages, and multiple physical sockets. Conceptually: ``` enum TrafficClass { HOT_BLOCK, COLD_BLOCK, TX_RELAY, ADDR_GOSSIP, META_HEAVY } function classify(msg): if msg.type == BLOCK and is_recent(msg): return HOT_BLOCK if msg.type == BLOCK and is_old(msg): return COLD_BLOCK if msg.type in {TX, INV_TX}: return TX_RELAY if msg.type in {ADDR, ADDRV2}: return ADDR_GOSSIP return META_HEAVY function allow(class): if class == TX_RELAY and !config.tx_relay: return false if class == COLD_BLOCK and !config.archive: return false return true ``` Obviously the exact implementation details would differ, but the idea is to show that a single-process, single-socket design can still support multiple internal relay lanes. Do you see any drawbacks with internalising the traffic classes rather than exposing multiple processes, sockets or service bits? Curious if I’m missing a constraint. Kind regards, Defenwycke On Thursday, December 4, 2025 at 11:16:52 PM UTC Antoine Riard wrote: > Hi list, > > Surfacing an old idea concerning the network-level and the current > meddling of block, > tx and addr messages traffic generally all over one network link. > Historically, for > example, if you consider bitcoin core by default connections are going to > be FULL_RELAY. > Over the last years, there has been few improvements to separate network > links by types > e.g with the introduction of dedicated outbound BLOCK-RELAY connections > [1], without the > segregation at the network-level between the class of traffic really being > pursued, or at > least more flexibility in network mechanisms to signal to a node's peers > what categories > of messages will be processed on a given link. > > Previously it has been shown that leveraging tx-relay's orphan mechanism > can allow to map > a peer's network-topology [2] (sadly, one trick among others). Being able > to infer a peer's > "likely" network topology from tx traffic, one can guess the peers used to > carry block-relay > traffic. From the PoV of an economical node, dissimulating the block-relay > traffic is a very > valuable to minimize the risks of escalation attacks based on > network-topology (e.g for > lightning nodes [3]). > > Segregating more network traffic by class of messages sounds to suppose 1) > being able to signal > among the {ADDR, ADDRV2} service bits if block, addr or tx relay is > supported on a link to be > opened for a pair of a (net_addr, port) or alternatively 2) if network > link are open blindly > with peers, being to signal in the VERSION message or with a dedicated > message what class of > message is supported. There is already a signaling mechanism in the > VERSION message to > disable tx-relay (i.e `fRelay`), however there is no signaling to disable > block-relay over a link. > Alternatively, it has been proposed in the past to add a new early message > among all the other > handshake messages between the VERSION / VERACK flow, but it has never > been implemented [4]. > > For bitcoin backbone, started to natively isolate each class of traffic in > its own process, and > only strictly signaling what is needed in the VERSION message. Though, I'm > starting to reach > the limit of the current network mechanisms, e.g I've an `archive_relayd` > process to service "cold" > blocks, dissociate from the process doing full block-relay traffic, and > this process is emitting versions > messages, with the NODE_NETWORK bit set and the others process would have > NODE_NETWORK_LIMITED. If you're asking the why of dissociating "cold" from > "hot" block relay > servicing, that avoids wasting CPU cycles on a busy code path. > > Anyway, for now I think I can come up with good hacks with the service > field and experimental bit > services. One drawback, it's just one "logical" node might start to occupy > multiple "physical" sockets > of its peers (one for tx-relay, one for block-relay), but network-wide > this might not be the most > ressource-preserving approach, so I'm wondering if better mechanisms are > worthy to muse about. > > Cheers, > Antoine > OTS hash: 22f8cfbd2b1fd093f6bb8737f3ddcdb956f8dadb1b9436dab3c8491e4b5583fd > > [0] > https://github.com/bitcoin/bitcoin/blob/master/src/node/connection_types.h > [1] https://github.com/bitcoin/bitcoin/pull/15759 > [2] https://discovery.ucl.ac.uk/id/eprint/10063352/1/txprobe_final.pdf > [3] https://arxiv.org/pdf/2006.01418 > [4] https://github.com/bitcoin/bips/blob/master/bip-0338.mediawiki > -- You received this message because you are subscribed to the Google Groups "Bitcoin Development Mailing List" group. To unsubscribe from this group and stop receiving emails from it, send an email to bitcoindev+unsubscribe@googlegroups.com. To view this discussion visit https://groups.google.com/d/msgid/bitcoindev/de958fe0-5897-4fb0-9b4c-b41f5c63296bn%40googlegroups.com.