Seeing it run

Four runnable programs under demos/. Each takes a real extract or feed; data/ is ignored by git, so bring your own — country extracts from Geofabrik, city ones from BBBike, and any GTFS zip.

A route, and the search behind it

python demos/route_city.py liechtenstein.osm.pbf \
    --from 47.2260,9.5230 --to 47.1300,9.5210 --profile walking

writes a map of the route with every settled node behind it — blue where both searches went, orange for the nodes A* never had to look at. It is the quickest way to see what a heuristic buys: the same 11.2-minute Seattle route settles 41,161 branches under Dijkstra, reaching across Lake Washington to Bellevue, and 12,172 under A*, which never leaves the corridor.

python demos/route_by_clock.py seattle.osm.pbf     # the same trip at every hour
python demos/route_transit.py kcm.zip --date 2026-08-17   # every timetable technique, one itinerary each

route_transit.py --walk 0 drops the footpaths, which is the quickest way to see what crossing the street buys.

The node board

python demos/serve.py data/Seattle.osm.pbf     # then open http://localhost:8000
python demos/serve.py data/Seattle.osm.pbf --gtfs data/kcm.zip --date 2026-08-17

Click an origin, click a destination, and the route comes back with the search drawn underneath. Under the map sits a node board: the three steps drawn as a graph. Layers feed an Environment, a technique binds to one and becomes a planner, a Query asks it something. Drag the nodes, pull the wires out, plug them in somewhere else. There is no second representation of what the controls mean — the board is the query, so unplugging a wire really does remove an argument, and the page says which one.

The map is a node too, and the graph runs out through the map and back: the map gives the query an origin and a destination and takes a route and a space in return. Cross the two points and the trip reverses. Unplug space and no search space gets built — ten megabytes of GeoJSON nothing was listening for. Unplug route and the query still runs and still reports what it cost, with nothing drawing it.

A node whose configuration the board has never answered greys out and spins until it does, and because a node's identity includes everything upstream of it, the spinners spread exactly as far as the rebuild does and no further. A change that costs nothing shows nothing. Where the work can count itself, the node shows how far along it is: contraction reports nodes retired, then arcs assembled; a landmark table reports searches run. Where it cannot — a file parser that yields no counts — the node says so rather than inventing a bar.

Watch that bar on a big network and it teaches you something the timings table hides. Contracting Seattle's walking graph settles 99.5% of its 554,393 nodes in the first minute and spends two more on the rest: the last nodes left are the most connected, and that is where the shortcut search does its real work.

That is what makes the refusals worth causing on purpose. Wire a GTFS layer into Dijkstra and the node turns red with the library's own sentence:

Dijkstra cannot route over timetable layers; it accepts scalar

A dropdown in the board's toolbar loads a starting point — road with landmarks, road with a contraction hierarchy, plain Dijkstra as the control, walking that reads the clock, and either timetable model when the demo was given a feed. These are places to begin, not modes: load one and pull it apart. The pins stay put across a change, which is the point of having more than one.

Rewiring is cheap because everything is cached by a canonical spelling of the node and everything upstream of it, so returning to a shape you had before is free. That matters because the expensive things are exactly the ones worth comparing. Seattle is 258,029 nodes and 590,671 edges from a 65 MB extract, read in about five seconds; swapping the technique node re-runs the same query the other way:

Same two pins throughout; the first four return the same route.

See also

• The shelf — every paper implemented here.
• What preprocessing buys — the same comparison, measured properly.
• How it is built — the contract, the layout, and the decisions underneath.