Gossip propagation

Exact transmission evidence

Follow the actual senders

One rumor starts at node zero. At the beginning of each round, every node that already knows the rumor selects distinct peers, excluding itself. Each sender attempts the configured number of contacts, capped by the number of other nodes. Delivered messages teach their recipients at the end of that round. Newly informed nodes cannot forward the rumor until the next round. The lines in the diagram come from these actual transmission records. Choose a sender to inspect only its traffic, or choose all senders to see the whole round. The message ledger repeats every source, destination and delivery outcome.

Change the network and workload

Cluster nodes controls membership from two through forty. Contacts per sender controls fanout, not guaranteed new recipients: a contacted peer may already know the rumor. Scenario seed reproduces peer choices and loss decisions. Packet loss percent applies a deterministic seeded drop decision to otherwise reachable transmissions; one hundred percent prevents any new learning. Split network separates nodes below floor of N divided by two from the remaining nodes. Cross-group attempts are blocked until Heal at round. A healing value of twenty-one keeps the split throughout the twenty-round experiment; one means there are no blocked transmission rounds. Every model-control edit rebuilds the trace and pauses inspection at round zero.

Try a complete and an incomplete spread

Restore defaults loads twenty nodes, fanout two, seed forty-two, zero loss and no partition. Inspect round five: this particular trace reaches all twenty nodes. That observed result is not a universal round-count guarantee. Now set packet loss to one hundred percent and inspect the final round: only node zero knows the rumor, while attempted messages accumulate. Restore defaults, enable Split network and leave healing at twenty-one. Nodes in the second group cannot learn the rumor because no successful path crosses the split. Increase fanout and observe that additional attempts do not overcome an unavailable connection.

Reproduce the healing boundary

Set five nodes, fanout eight, zero packet loss, Split network enabled and healing at round eight. The effective fanout is four because a sender cannot select itself or duplicate another target in the same round. Round one informs the other node in the first group. At round seven, the second group is still isolated. At round eight, cross-group messages become deliverable and all five nodes learn the rumor. Compare the attempted, blocked and delivered totals before and after healing. Contacts that were blocked in earlier rounds are not queued or retransmitted individually; future rounds make new contact attempts.

Use the sixty-second inspector

Play reveals twenty synchronous rounds over sixty inspection seconds, three seconds per round. Pause freezes the selected state. Next round advances exactly one round. Restart returns to the beginning of the same scenario, while Restore defaults also resets model inputs. Speed changes viewing pace between half, normal and double speed. Inspect time scrubs backward or forward without recomputing randomness. Show final round jumps to round twenty. Sender filter changes only which transmission paths are drawn; it does not change any model result or the complete ledger. Node numbers and knowledge state are repeated in readable text, so color is not the only way to interpret the scene.

Account for work and limitations

Attempts equal delivered plus lost plus partition-blocked messages. Delivered messages equal first-time learning events plus redundant deliveries, where redundant includes a second message reaching the same new recipient in one round. The counters continue after full spread to make that ongoing work visible. The model uses synchronous rounds, fixed membership, one persistent rumor, seeded hash-based peer selection and immediate within-round delivery. It has no push-pull exchange, suspicion timers, membership incarnations, bounded rumor lifetime, bandwidth queue, durable state or real packet capture. Nodes are positioned for readability rather than geographic distance. A successful run is evidence about this fixture, not a guarantee of logarithmic completion time or a reproduction of a particular production gossip protocol.