VIGO 1.0.0-rc.1 · build b901da90bb49 · Apple M2 · 8 cores · 16 GiB RAM. 5,000 pairs per city: 20% random, 40% near-stop and 40% stop-offset. Serial measurements with retained service dates.
Latency compares the same pairs accepted by all nine interfaces, excluding process-first responses from both batches. Caller time includes adapter, transport and decoding; City loading and rendering are separate. Speedup is the ratio of the selected latency quantiles, not the median of per-pair speedups. VIGO uses the Node CLI to the native engine; VIGO-py and R5py are Python interfaces.
Memory is the maximum RSS reported by macOS /usr/bin/time for each invocation. Routing takes the larger peak across the two query batches; building uses the original one-time preparation. These are measured process peaks, not minimum RAM requirements or summed process-tree totals. Child processes and external services may fall outside the reported scope, including VIGO-py’s native runtime.
Valhalla flags assumed schedules in all 16,558 eligible journeys. Three New York GraphHopper errors are excluded from the matched cohort. Eligibility is adapter acceptance; routes and interface semantics can differ. Blocked results include rejected candidates.
Benchmark code ↗ · Data & methodology ↗
