Graph execution under attack
The extensive LSQB-derived and adversarial query benchmark: upstream provenance, backend matrices, historical cohorts, and performance corrections.
Read the query benchmark →Updated · September 19Graph algorithms, full paths
Eleven execution classes over identical graphs, across three source pins. Explicit output contracts, a completed 2.147-billion-entry path run, and two findings where measuring which class paid what overturned the obvious explanation.
Read the algorithm benchmark →Kernels · September 22Graph kernels, five participants
PageRank, connected components, BFS and triangle counting across five in-memory participants, one of them the neo4j-labs/graph library, checked against an independent reference before timing. Two measurement artifacts corrected, one kernel change timed with bit identity established first, one commit timed against a regression under an unchanged protocol, every cell that got worse listed beside what the previous campaign said.
Graph stores under strain
Fifteen backends, fourteen real graphs, nine hard gates. Embedded Rust now matches Neo4j's reach at 117 million edges and beats it on speed in every same-machine pair; Neo4j's memory lead is down to 10 to 40%. 365 runs in 34 verified bundles, every failure retained.
Read the strain ledger →Graph Nuts · September 27Graph algorithms on Sail
PageRank and WCC through Pecan, Nutmeg Banda and Grenada. Reference, delta/frontier and fused contraction methods, with complete-call time, sampled memory and a build-and-run tutorial.
Read the Sail measurements →