osmviews

Python client for OSMViews, a world-wide ranking of geographic locations by how much they are looked at on OpenStreetMap-based maps. See the main project for background.

OSMViews aggregates a year of OpenStreetMap map-tile access logs into a single raster covering the whole planet. This package reads a copy of that raster from local disk and answers point queries.

Usage

# pip install osmviews
import osmviews

with osmviews.open("osmviews.tiff") as o:
    # rank() is 0.0 (nobody looks here) to 1.0 (one of the most-viewed places on
    # Earth). Coordinates are x, y — longitude then latitude, as in GeoJSON;
    # values drift weekly.
    shibuya = o.rank(139.7013, 35.6586)  # Tokyo, Shibuya     ~0.69
    altstetten = o.rank(8.4889, 47.3915)  # Zürich, Altstetten ~0.66
    ushuaia = o.rank(-68.3030, -54.8019)  # Ushuaia            ~0.56
    sahara = o.rank(13.0000, 23.0000)  # Sahara             ~0.00
    assert shibuya > altstetten > ushuaia > sahara

The package does not download anything. Fetch the dataset (~594 MB, regenerated weekly) from osmviews.DOWNLOAD_URL however you like, then pass the path to osmviews.open.

An OSMViews instance is safe to share across threads: every query takes a lock only briefly, and tile decoding happens outside it. Decoded tiles are kept in a small LRU cache (osmviews.open(path, cache_tiles=...), 0 disables it), so queries clustered in one region stay fast. o.metrics() returns a snapshot of counters (cache hit rate, decode time, …) worth logging at the end of a long run.

The file is memory-mapped, so it must not be modified or truncated while an OSMViews is open.

Performance

Rough numbers on an Apple M5, CPython 3.13 (from benchmarks/bench.py): rank() returns in ~0.5 µs when the tile is already cached and ~57 µs on a miss that has to read and inflate one. Each decoded tile is 256 KiB; the default LRU holds 64 of them (~16 MiB), and the GeoTIFF is memory-mapped rather than read onto the heap. For bulk lookups, submit points in roughly spatial order (e.g. sorted by tile or by S2 cell ID) so neighbouring queries reuse cached tiles.

This is pure Python; it is not trying to be fast. If throughput matters, the Rust client answers the same query in tens of nanoseconds.

No dependencies

Pure Python 3.11+, standard library only (mmap, zlib, struct, array). The TIFF header parsing and the map projection are done in-package.

Contributing

Contributions are welcome — see CONTRIBUTING.md. The design and its rationale are written up in TECHNICAL_DESIGN.md.

Sponsoring

This package and the OSMViews pipeline behind it are maintained by Sascha Brawer as a volunteer effort. If your project relies on them, please consider sponsoring continued maintenance and future development via GitHub Sponsors.

License

MIT — see LICENSE.