ffsdb

benchmark posture

FFS evals for graph, vector, and write-back engine paths.

The eval posture is measured first and caveated in public. Faster rows, slower rows, comparator stack overhead, machine class, and raw commands all matter more than a single heroic number.

Evals

cargo run -p ffs-evals --release

Comparator Measured path Result
Postgres + pgvector HNSW query, same M and ef_construction 3.01x faster
LanceDB HNSW query vs indexed IVF-PQ path 3.03x faster
LanceDB HNSW query vs brute-force path 4.78x faster
sled KV range scan over 100-key windows 3.94x faster
sled Random u64 insert path 2.89x faster
instant-distance HNSW query, dim 128, k=10 2.64x faster
petgraph 1-hop graph traversal 9.28x faster
FFS integrated Filter, traverse, write-back, reopen tested
Run command cargo run -p ffs-evals --release
Machine class Apple M-series local run, as documented in the README
Rule publish faster and slower paths; name the comparator stack being measured

Clean vector comparator: Postgres + pgvector uses the same HNSW algorithm, M=16, and ef_construction=200. In the README run, ffs::Hnsw queried at 126 us/query vs 380 us/query for indexed pgvector.

Loud graph-server numbers are included only with caveats: they measure FFS native CSR calls against Kuzu Cypher or Neo4j Bolt+Cypher parse/bind/plan/execute paths, so read them as stack overhead evidence, not final database parity.

Caveated path Result in README Read it as
FFS native CSR vs Kuzu Cypher ~29,000x 1-hop traversal Native slice iteration vs embedded Cypher parse/bind/plan/execute/FFI stack overhead.
FFS native CSR vs Neo4j Bolt+Cypher ~1,200x bulk load, ~195,000x 1-hop traversal Native in-process CSR vs server round trips plus Cypher stack; useful as stack-cost evidence.
Predicate-aware graph + vector 100% recall at 10% and 1% selectivity Filtered HNSW pays more latency than post-filtering, but avoids returning the wrong top-k under sparse predicates.

context for numbers

Comparisons How FFS should be read against the comparator products. Engine The mechanics behind the measured paths. Roadmap Where the public proof ladder goes next. Use Cases The production workloads the evals should eventually model. Home The short positioning page.