There is no database. All GTFS data is loaded from CSV files at startup and held in memory. This gives extremely fast query times at the cost of startup time (10-30 seconds for index building).

The RAPTOR index is wrapped in ArcSwap, which allows atomic pointer swaps. When new GTFS data is loaded via POST /api/gtfs/reload, the entire index is rebuilt in a background thread and swapped in atomically. No request is ever blocked or sees partial data.

Trips with identical stop sequences are grouped into patterns. This dramatically reduces memory usage and speeds up the RAPTOR scan phase, because the algorithm only needs to evaluate one entry per pattern instead of one per trip.

Valhalla supports indoor maneuvers such as elevators, stairs, escalators, and building enter/exit transitions. Transfers are classified by parent_station: outdoor transfers (different parent_station) always get a Valhalla walking route for the map polyline, while indoor transfers (same parent_station) are only enriched when indoor maneuvers exist in OSM. Transfer polylines use the Valhalla shape (actual walking route) when available, falling back to a straight line otherwise. This enrichment only runs when routing.maneuvers is enabled in config.yaml.