Loading
How FalkorDB decides
Loading the hero recording…
Scene 1 of 9
Alerts, deploys, warehouse lineage, docs and people stream into one FalkorDB graph per tenant.
The real query
Alerts, deploys, warehouse lineage, docs and people stream into one FalkorDB graph per tenant.
FalkorDB stores each relationship type as a sparse adjacency matrix and traverses by matrix multiplication. PROPAGATES_TO has – edges among – nodes: – of its matrix is empty.
Checkout pages. algo.BFS spreads from it over the failure-flow edges, one hop at a time: that frontier is the blast radius (Q-BLAST-1, – ms engine p50).
Three deploys landed in the window. From D-rec-12, algo.SPpaths finds no path to checkout: 0 rows, so it cannot be blamed. – lights up (G2, – ms engine p50).
A path is not enough. Edges become pipes sized by capacity: the email deploy's max flow is –, the outage needs –. Rejected by G5; the payment pipe runs full.
The smallest set of edges that contains the failure is the min-cut: FalkorDB computes the max-flow (– ms) and, in the same query, the edges it saturates. It is tried first on a GRAPH.COPY branch (– ms for – nodes): the protected services are no longer reached.
People close to the payment code and its past fixes rank by harmonic centrality; the top responder steps forward, and a maximum spanning forest draws the escalation chain.
The model can only point. Every read is GRAPH.RO_QUERY with a timeout and a LIMIT; after approval one atomic query writes the decision back, and a separate read-only connection re-reads it: –.
Switch FalkorDB off and it refuses to confirm anything: no graph, no claim. Remove algo.SPpaths and algo.maxFlow together and the right cause falls from – to – of – cases. Faultline calls every algo.* procedure the pinned engine ships, – of them, and each changes a decision. No path, no claim.
Drawn from the hero recording (public/replay/s1-hero, exported from FalkorDB) · numbers from evals/results/latest.json · per-type edge counts from db.meta.stats on golden_acme · the matrix scene shows how FalkorDB stores a graph (its design docs: sparse adjacency matrices, GraphBLAS). Layout is fixed and illustrative.