Benchmarks
How fast is TripleTrail?
We measure TripleTrail against other open-source engines on public benchmark suites and on spatial and routing work. The harness and the raw results are in the source repository, so you can run them yourself.
Where TripleTrail leads
General SPARQL on the BSBM explore and update workloads, spatial joins on real OpenStreetMap data, routing, combined route and spatial queries, and bulk loading.
Where others lead
PostGIS answers simple box and nearest-neighbour lookups in under a millisecond. maplib is faster on BSBM business-intelligence queries and on the RINF Sparqloscope set. QLever stores data in less disk space.
How we measured
One Hetzner ccx43 server (16 dedicated vCPU, 64 GB RAM), November 2026. Each engine ran alone, with its result cache off, using its own fastest correct form of each query.
Berlin SPARQL Benchmark (BSBM)
The classic benchmark for SPARQL stores: an e-commerce dataset with three workloads. Explore is a mix of typical lookups, BI is heavy analytical queries, update mixes writes in. Query mixes per hour, higher is better.
| Workload | TripleTrail | Oxigraph | maplib | QLever | Jena TDB2 |
|---|---|---|---|---|---|
| Explore, 1 M triples | 69,042 | 33,191 | 28,981* | 2,904* | 5,389 |
| Explore, 10 M triples | 112,307 | 14,766 | 16,482* | 5,379* | 1,630 |
| BI, 1 M triples | 17,768 | 350 | 35,228† | 8,506 | 1,716 |
| BI, 10 M triples | 4,757 | 47 | 8,571† | 3,477 | 190 |
| Update, 1 M triples | 40,804 | 26,021 | 2,517* | n/a | 4,001 |
| Update, 10 M triples | 60,869 | 15,093 | 206* | n/a | 1,640 |
* Skipped the DESCRIBE and CONSTRUCT queries (2 of 12), because the BSBM test driver asks for RDF/XML, which these endpoints do not serve. † maplib 0.20.26 skipped BI query 5 and returned wrong result counts on queries 7 and 8. QLever has no update workload result. 1 M and 10 M triples are BSBM scales 2,785 and 28,480 products.
Sparqloscope: one query per SPARQL feature
Sparqloscope, from the QLever team at the University of Freiburg, generates about 100 queries that each test one SPARQL feature: joins, OPTIONAL, MINUS, property paths, regular expressions and more. Geometric mean per query, lower is better.
| Dataset | TripleTrail | maplib | QLever | Oxigraph | Jena TDB2 |
|---|---|---|---|---|---|
| BSBM, 1 M triples | 3.2 ms | 5.1 ms | 5.8 ms | 60.1 ms | 81.7 ms |
| RINF, European rail infrastructure register | 13.7 ms | 7.1 ms | 30.7 ms | 182.9 ms | 206.2 ms |
TripleTrail, maplib, QLever and Jena answered all 102 queries; Oxigraph timed out on one.
Spatial and routing
Our own benchmark for spatial networks: a synthetic rail network of 10.5 million triples modelled on RINF, and real OpenStreetMap data for Norway. PostGIS with GiST indexes and pgRouting is the reference. Warm median time, lower is better.
| Question | TripleTrail | PostGIS / pgRouting | QLever | Jena GeoSPARQL |
|---|---|---|---|---|
| Everything in a box / polygon | 3.5 / 2.1 ms | 0.3 / 0.2 ms | 729 / 552 ms | 32 / 50 ms |
| Everything within a distance | 2.0 ms | 1.7 ms | 135 ms | 29 ms |
| 10 nearest signals | 5.3 ms | 0.3 ms | 135 ms | 3.01 s |
| Join: signals within 20 m of track elements | 1.12 s | 851 ms | 2.77 s | timeout (300 s) |
| Join, OSM Norway: stations × admin units | 218 ms | 312 ms | 5.86 s | 6.86 s |
| Join, OSM Norway: tracks × land use | 519 ms | 1.30 s | 4.23 s | timeout (600 s) |
| Join, OSM Norway: tracks within 5 m of tracks | 309 ms | 5.08 s | 1.46 s | 155 s |
| Track length per region | 1.93 s | 4.94 s | n/a | n/a |
| Kilometre post → point on the track | 4.5 ms | 9.5 ms | n/a | n/a |
| Shortest route (Dijkstra) | 40 ms | 1.04 s | n/a | n/a |
| Shortest route, contraction hierarchies | 5.1 ms | n/a | n/a | n/a |
| Route with constraints | 31 ms | 922 ms | n/a | n/a |
| Five shortest routes | 93 ms | 21.3 s | n/a | n/a |
| 100 × 100 travel cost matrix | 647 ms | 24.7 s | n/a | n/a |
| Route, then signals within 50 m of it | 766 ms | 1.07 s | n/a | n/a |
| Load 10.5 M triples and build indexes | 6.5 s | 29.8 s | 9.9 s | 150.5 s |
| Size on disk | 746 MiB | 893 MiB | 204 MiB | 1.94 GiB |
n/a: the engine has no such function. pgRouting rebuilds its graph from SQL on every call, which is most of its time on point-to-point routes; TripleTrail keeps named networks in memory. The first spatial query after a restart builds TripleTrail's spatial index (125 ms for the box selection), and the first route builds the network (about 5 s); the full results list every cold time.
Method and fairness
- All engines ran on the same server, one at a time, with the same data. PostGIS reads the same geometries as the RDF engines.
- Every engine uses its own fastest correct query form; results are checked against each other, and wrong answers are reported, not counted as fast.
- Versions: PostgreSQL 16, PostGIS 3.4, pgRouting 3.6, Apache Jena Fuseki GeoSPARQL 6.2 (TDB2), QLever (November 2026 image), maplib 0.20.26, Oxigraph (latest release at the time of the run).
- Every suite, query and raw result is in the
bench/folder of the source repository, with a one-command run on your own machine or on a rented server. - Benchmarks are a guide, not a promise. Your data and queries decide. Send us a sample and we will measure it.