Community detection, iterative resident passes, no self-join per round
You are looking at Bitcoin value-flow graph1.78M nodes · 7.37B edges in one Semurg substrate. Ask it to detect communities by label propagation over 7.37B edges and it runs the
real query live on this box, on data you can
download and re-run yourself ▸.
Click an example in the agent rail, or ask to propagate more rounds, e.g. propagate 8 rounds. Every run is measured end to end, on
this box, on the real dataset. Nothing is pre-computed.
the live Bitcoin value-flow graph, transactions and addresses linked by value-flows (prev-tx to tx to address), parsed straight off the P2P network · Bitcoin mainnet P2P (our own block-parser node, no bitcoind, no REST). Label-propagation community detection: every node adopts the majority label of its neighbors, swept round after round until the clusters stabilize. Each round is a single co-located image pass instead of a self-join re-materialized per round. Ask the agent (right); the number is measured here, now, on data you can download.
the graph-analytics leader
Copy out before it can iterate. A dedicated graph-analytics engine copies the whole graph out into a separate cluster before the first round, and on a relational store every propagation round is a self-join re-materialized from scratch.
semurg, one engine
Every round runs in place. Each iterative round is one resident pass over the in-memory adjacency, no per-round self-join and no copy to a separate cluster, because the compute runs on the data.
The ETL hop and the per-round self-join both disappear, the iteration happens where the graph lives.
Bitcoin value-flow graph: 1.78M nodes, 7.37B edges. Measured live on this box. Download it and re-run it yourself.
ask the agent to detect communities, the live throughput renders here
This is a RAW win, and we say so plainly: label-propagation community detection is iterative by nature, several sweeps of the whole graph until the labels settle. On a relational store each round is a self-join re-materialized from scratch; a dedicated graph-analytics engine first copies the whole graph out into a separate cluster before it can iterate at all. Here every round is a single resident pass over the in-memory adjacency, because the compute runs on the data. Every number is measured live on the AVX-512 box, on a public dataset you can download and re-run yourself, the same engine that also serves graph traversal, full-text and vectors.