geors

Rust-powered drop-in replacement for geopy.

Replaces every compute-bound function in geopy with a Rust implementation using Karney's geodesic algorithm via geographiclib-rs. I/O-bound operations (geocoding, rate limiting) forward transparently to geopy.

Performance

Batch throughput: 42 million point-pairs/sec (rayon-parallel).

Verification

Verified to sub-nanometer precision against geographiclib and geopy:

• 10,000 Karney GeodTest cases -- max distance error: 7.5e-9 m (~7 nanometers), 0 cases >1mm
• 15 real-world city pairs -- max error: 9.3e-10 m (<1 nanometer)
• 13 edge cases -- identical points, poles, antipodal, antimeridian, near-antipodal (Vincenty failure cases) -- all exact match
• 1,000 random direct/inverse roundtrips -- max error: 5.6e-9 m
• 1,000 random cross-checks vs geopy -- max error: 5.6e-9 m
• Batch vs scalar consistency -- exact match (0.0 m difference)

pytest tests/test_python_compat.py -v     # 57 tests
python tests/verify_accuracy.py           # 37/37 comprehensive checks
python tests/bench_distance.py            # full benchmark suite

Installation

pip install geors

Usage

Drop-in replacement

# Before
from geopy.distance import geodesic, great_circle

# After (same API, 125x faster)
from geors.distance import geodesic, great_circle

Distance calculations

from geors.distance import geodesic, great_circle

# Geodesic (ellipsoidal, Karney's algorithm)
d = geodesic((40.7128, -74.0060), (51.5074, -0.1278))
print(d.km)        # 5570.248...
print(d.miles)     # 3461.358...
print(d.meters)    # 5570248.4...
print(d.feet)      # 18274764.0...
print(d.nautical)  # 3007.6...

# Great circle (spherical, Haversine)
d = great_circle((40.7128, -74.0060), (51.5074, -0.1278))
print(d.km)        # 5564.847...

Low-level functions

from geors._geors import distance, geodesic

# Scalar
meters = distance.geodesic_distance(40.7128, -74.0060, 51.5074, -0.1278)

# Full inverse (returns dict with s12, azi1, azi2, a12, ...)
result = distance.geodesic_inverse(40.7128, -74.0060, 51.5074, -0.1278)

# Direct problem (start + bearing + distance -> endpoint)
result = geodesic.direct(40.7128, -74.0060, 51.37, 5_554_000.0)

# Destination
lat2, lon2 = distance.geodesic_destination(40.7128, -74.0060, 51.37, 5_554_000.0)

Batch operations (numpy arrays, rayon-parallel)

import numpy as np
from geors._geors import distance

lat1 = np.random.uniform(-90, 90, 1_000_000)
lon1 = np.random.uniform(-180, 180, 1_000_000)
lat2 = np.random.uniform(-90, 90, 1_000_000)
lon2 = np.random.uniform(-180, 180, 1_000_000)

# Returns numpy array, computed in parallel across all cores
distances = distance.geodesic_distance_batch(lat1, lon1, lat2, lon2)
distances_gc = distance.great_circle_distance_batch(lat1, lon1, lat2, lon2)

Geocoding (forwards to geopy)

# Geocoding is I/O-bound, so we forward to geopy transparently
from geors.geocoders import Nominatim

geolocator = Nominatim(user_agent="my-app")
location = geolocator.geocode("New York City")
print(location.latitude, location.longitude)

What's Rust, what's forwarded

Architecture

geors/
  crates/
    geo-math/     # High-precision math (trig in degrees, error-free sums, Accumulator)
    geors/        # Core library (geodesic, great-circle, polygon area, constants)
  src/            # PyO3 bindings (py_distance, py_geodesic, py_units)
  python/geors/   # Drop-in Python wrapper (Distance class, geocoder forwarding)

Built with PyO3 + maturin. Wheels for Linux, macOS, and Windows.

License

MIT