render_shapes(method="datashader") fails with KeyError: 'global' when the element has no "global" coordinate system #782

AI-generated--manually reviewed

Description

render_shapes(..., method="datashader") raises KeyError: 'global' when the shapes element has no "global" coordinate system. This breaks the Visium HD example notebooks (technology_visium_hd, technology_visium_hd_mouse_4.0.1): data from spatialdata-io has no global CS, and datashader is used either explicitly or automatically (more than 10k shapes).

MRE (PEP 723, run with uv run mre.py)

# /// script
# requires-python = ">=3.12"
# dependencies = [
#     "spatialdata @ git+https://github.com/scverse/spatialdata.git@main",
#     "spatialdata-plot @ git+https://github.com/scverse/spatialdata-plot.git@main",
# ]
# ///
import matplotlib

matplotlib.use("Agg")

import spatialdata as sd
import spatialdata_plot  # noqa: F401
from spatialdata.transformations import get_transformation, set_transformation

sdata = sd.datasets.cells()

# Move every element from "global" to a custom CS, as spatialdata-io does for Visium HD.
cs = "my_sample"
for _, name, element in sdata.gen_spatial_elements():
    set_transformation(element, {cs: get_transformation(element, "global")}, set_all=True)
assert "global" not in sdata.coordinate_systems

sdata.pl.render_shapes("cell_boundaries", method="matplotlib").pl.show(coordinate_systems=cs)  # OK
sdata.pl.render_shapes("cell_boundaries", method="datashader").pl.show(coordinate_systems=cs)  # KeyError: 'global'

(spatialdata is installed from main because datasets.cells() is not in a release yet.)

Traceback (abridged)

File .../spatialdata_plot/pl/render.py, in _render_shapes
    plot_width, plot_height, x_ext, y_ext, factor = _get_extent_and_range_for_datashader_canvas(
        transformed_element, "global", fig_params
    )
File .../spatialdata_plot/pl/_datashader.py, in _get_extent_and_range_for_datashader_canvas
    extent = _fast_extent(spatial_element, coordinate_system)
File .../spatialdata_plot/pl/utils.py, in _fast_extent
    return _element_extent_fast(element, coordinate_system) or get_extent(element, coordinate_system=coordinate_system)
File .../spatialdata_plot/pl/utils.py, in _element_extent_fast
    matrix = transformations[coordinate_system].to_affine_matrix(("x", "y"), ("x", "y"))
KeyError: 'global'

Root cause

In the datashader branch of _render_shapes, the already-transformed geometry is wrapped with ShapesModel.parse(gpd.GeoDataFrame(data=shapes.drop("geometry", axis=1), ...)), and the canvas extent is then computed in a hardcoded "global" CS. Because shapes.drop(...) keeps .attrs, the new element inherits the original transformations (here only my_sample) and never gets a "global" entry, so the lookup in _element_extent_fast fails.

The lookup came in with #703 (fast axis-aligned extent), so that's the likely regression point; not confirmed by bisecting.

Possible fix

The geometry is already in the target CS, so pass transformations={"global": Identity()} to that ShapesModel.parse call (or clear attrs before parsing). A regression test rendering shapes with datashader in a non-global CS would cover this.