Reframed after profiling (the original to_rgba framing was wrong — it's end-to-end noise).
Root cause
render_shapes (matplotlib) is dominated by per-shape Patch construction, identical across no-color / categorical / continuous color (cProfile, 200k circles, ~23 s):
_build_shape_patches (_geometry.py): ~13 s — one mpatches.Circle/Polygon per shape; each Patch.__init__ resolves default colors via to_rgba (~4×/patch) + recomputes a per-patch transform.PatchCollection.set_paths: ~6 s — converting all N patches to Paths.
Fix: PathCollection of pre-built Paths
Build Path objects (not Patch objects) and return a matplotlib.collections.PathCollection. These are exactly the Paths PatchCollection.set_paths produces internally, so it's byte-identical (same backend), preserves holes (compound paths) and the existing per-path affine step, keeps all color/alpha/outline logic, is one code path, and needs no signature change.
Measured (200k circles, build+draw): ~2.6× faster, ~5.8× less memory (11.7 s/149 MB vs 30.1 s/867 MB), color-mode-independent.
Rejected: EllipseCollection/PolyCollection — incompatible with the coordinate-system affine applied per-path (EllipseCollection has one shared read-only path), PolyCollection can't represent holes, and EllipseCollection tessellates circles coarsely. Not byte-identical.
Plan: plans/issue-733.md. Datashader + as_points paths already optimized — untouched.
Reframed after profiling (the original
to_rgbaframing was wrong — it's end-to-end noise).Root cause
render_shapes(matplotlib) is dominated by per-shapePatchconstruction, identical across no-color / categorical / continuous color (cProfile, 200k circles, ~23 s):_build_shape_patches(_geometry.py): ~13 s — onempatches.Circle/Polygonper shape; eachPatch.__init__resolves default colors viato_rgba(~4×/patch) + recomputes a per-patch transform.PatchCollection.set_paths: ~6 s — converting all N patches toPaths.Fix:
PathCollectionof pre-builtPathsBuild
Pathobjects (notPatchobjects) and return amatplotlib.collections.PathCollection. These are exactly thePathsPatchCollection.set_pathsproduces internally, so it's byte-identical (same backend), preserves holes (compound paths) and the existing per-path affine step, keeps all color/alpha/outline logic, is one code path, and needs no signature change.Measured (200k circles, build+draw): ~2.6× faster, ~5.8× less memory (11.7 s/149 MB vs 30.1 s/867 MB), color-mode-independent.
Rejected:
EllipseCollection/PolyCollection— incompatible with the coordinate-system affine applied per-path (EllipseCollectionhas one shared read-only path),PolyCollectioncan't represent holes, andEllipseCollectiontessellates circles coarsely. Not byte-identical.Plan:
plans/issue-733.md. Datashader +as_pointspaths already optimized — untouched.