Nexus-H3: A Resource-Frugal Urban Mobility Observatory for Continuous Telemetry Preservation and Reproducible Benchmarking

Although metropolitan transportation agencies increasingly provide real-time operational APIs, these data streams are inherently ephemeral and rarely distributed as accessible historical archives. Consequently, researchers studying long-term urban mobility dynamics are forced to operate persistent collection infrastructure, creating a substantial barrier to reproducible spatiotemporal research. To address this, we introduce Nexus-H3, an open urban mobility observatory that continuously collects, preserves, harmonizes, and publishes Grand Paris transportation telemetry as a reproducible research benchmark. We publicly release the resulting continuously updated spatiotemporal dataset covering N = 8,277 Uber H3 Resolution-9 hexagons across 37 compiled columns (6 spatiotemporal keys, 7 dynamic mobility and meteorological streams, and 24 static urban morphology, transit, and 5G infrastructure features), governed by a strict temporal regime separation policy designed to prevent source-induced covariate shift. By decoupling normalized dynamic telemetry from static topology prior to on-demand Star-Schema compilation, Nexus-H3 reduces long-term storage requirements to approximately 55–60 GB per year of continuous 3-minute city-wide observations (~2.08 billion raw multi-modal records), avoiding over 400 GB of denormalized Star-Schema storage. This is achieved through a duplicate-safe lease state machine in serverless PostgreSQL, constant-memory streaming extraction, 24-slice daily Parquet compaction, and a three-tier DuckDB compilation framework that offloads heavy analytical joins to in-browser WebAssembly and client-side Google Colab runtimes. Systems evaluations demonstrate a 535.7x reduction in traced Python heap allocations, a 64x critical-path ingestion latency speedup, and a 40.9% compiled file-size reduction via client-side row-group consolidation, while baseline forecasting and zero-shot spatial transfer workflows on our v1.0 release snapshot validate dataset fidelity and out-of-the-box usability for spatiotemporal machine learning.