xarray¶

The xarray integration preserves labeled containers; it is not another array provider. With advect[xarray] installed, importing advect.xarray registers DataArray and Dataset as pytrees. Dynamic transforms then operate on their data buffers and rebuild the labels around derivative results.

Contract¶

• Floating and complex data buffers are differentiable. Integer, boolean, string, and object variables are rejected rather than assigned zero gradients.
• Dimensions, coordinates, names, attributes, and dataset variable order are static metadata.
• xarray continues to own alignment, named indexing, and named reductions when those operations lower to supported array primitives.
• DataArray subclasses inherit the same pytree behavior and preserve their concrete container type.

Boundaries¶

This integration is dynamic-only. For a reusable staged kernel, stage the raw array function, call it with field.data, and restore labels outside the program. DataArray.values and .to_numpy() materialize NumPy arrays and remain outside tracing. Data-dependent coordinates, MultiIndex coordinates, Dask execution, and broad groupby, rolling, or interpolation coverage are outside the current contract.

See Arrays and Scientific Python for a complete example.