[mattijn]

Summary

tp.Topology accepts a Polars DataFrame with a column of geometries as WKB or EWKB, such as the GeoDataFrame of polars-st:

import polars_st as st
import topojson as tp

gdf = st.GeoDataFrame({"name": ["abc", "def"], "geometry": ["POLYGON ((0 0, 1 0, 1 1, 0 1, 0 0))", "POLYGON ((1 0, 2 0, 2 1, 1 1, 1 0))"]})
topo = tp.Topology(gdf)

• shapely.from_wkb reads the column geometry (or else the only Binary column) at once; the other columns become the properties, as for a GeoDataFrame. A shared SRID (EWKB) becomes the CRS, also of to_gdf().
• A Series of WKB, a frame of only geometries and a list of DataFrames with object_name work too.
• Polars and polars-st are not imported by topojson. The tests build the WKB with shapely.to_wkb, so they run without polars-st and with Polars 2; polars is added to the dev extras.
• The section Polars DataFrame or Series on Types of input data runs its example while the site is built. The docs extras pin polars<2 next to polars-st, as polars-st 0.10 does not import with Polars 2 (Fix import on Polars 2 (GeoSeries metaclass conflict) oreilles/polars-st#85, not released yet).
• Extract looks up the method of each type of input in a table built once, instead of a chain of elif; a new type is one line.

Africa from a Polars DataFrame gives the same arcs and objects as from a GeoDataFrame. Of the inputs, a Polars DataFrame is the quickest to extract: the geometries are read in one call, and Polars gives its rows as dictionaries quicker than pandas. Most of Extract is still the loop per feature, the same for every input, which a vectorized path (shapely.get_rings, get_coordinates) could avoid.

Not yet: sync() (incremental updates) takes a GeoDataFrame only, and there is no to_polars().

🤖 Generated with Claude Code

[github-actions]

Benchmark result

These PR changes are benchmarked against five files (see tests/benchmark_compute.py for what each one tests), compared with PyPi's latest release and the repository's main branch.

[github-actions]

Benchmark result

These PR changes are benchmarked against five files (see tests/benchmark_compute.py for what each one tests), compared with PyPi's latest release and the repository's main branch.