[matschaer]

Summary

• No copy of single-variable datasets (cd24e93): interp2poi and interp2grid call ds_utils.ds_to_da(..., copy=False). When there is one data variable and it spans all dimensions, they get a view via expand_dims instead of the np.stack copy made by to_array. The default stays copy=True.
• NaN-skipping coarsening (01e36a0): both coarsening sites (_coarsen_interp and slice_coarsen) use coarsen(...).reduce(_window_nanmean). This is a 2-line helper that moves the window axes last and calls bottleneck.nanmean. A coarsened cell with some NaN gets the mean of its valid values; only all-NaN cells stay NaN. This prepares for regridding the snow limit, which is NaN below the orography.

Performance (operational grids, later calls with cached neighbors)

Case	main	branch
ICON-CH2 → POIs, nearest	0.18 s	0.12 s (0.1–0.2 GB less memory)
IFS → seamless 2 km, nearest_alt	0.39 s	0.24 s
ICON-CH1 → seamless 2 km, nearest_alt	0.32 s	0.22 s (+0.05 GB peak)
gridefix-read (1.1 M cells → 252 × 376, 100 fields)	0.13 s	0.11 s

Known regression: ICON-CH2 in LV95 → seamless 2 km gets slower, from 0.23 to 0.29 s with nearest_alt and from 0.34 to 0.39 s with temp_fixgrad. The source there is a sliced, non-contiguous view, and the reshape now copies it with a gather. Compacting the data as main did is slower in 9 of the 12 operational grid cases.

Changes of results

• Coarsened cells with partial NaN are no longer NaN (see above).
• Coarsened float32 values differ by float32 rounding, which grows with the window size. For 2 m temperature that is up to 3e-5 K for a 2 × 2 window and 2.4e-4 K for 10 × 10 (global 1° benchmark); the relative error is about 2e-6 for a 40 × 40 window. ERA5-Land → RhiresY is identical.

Tests

• New test_ds_to_da_no_copy: checks the view, the default copy, multiple variables, and a dimension that only a coordinate has.
• New test_window_nanmean and test_window_nanmean_large_window: check partial and all-NaN blocks, float32, integers equal to np.mean, and a 10 × 10 window within rtol=1e-6 of the float64 nanmean.
• All 262 tests pass; ruff and mypy are clean; pylint gives 10.00.

🤖 Generated with Claude Code