Avoid DataArray construction when scanning for cell_measures

[dcherian]

Summary

CFAccessor.cell_measures scans every variable's cell_measures attribute by iterating obj.coords.values() and obj.data_vars.values():

all_attrs = [
    ChainMap(da.attrs, da.encoding).get("cell_measures", "")
    for da in obj.coords.values()
]
...
all_attrs += [
    ChainMap(da.attrs, da.encoding).get("cell_measures", "")
    for da in obj.data_vars.values()
]

Iterating those mappings constructs a full DataArray (resolving its coordinates) per variable just to read one attribute. Because cell_measures is reached from __getitem__ → _get_all_cell_measures, anything that does ds.cf[[var]] pays an O(nvars) DataArray-construction cost, and code that loops over variables calling ds.cf[[var]] (e.g. grid/style probing) ends up O(nvars²).

A downstream profile of a metadata endpoint over a many-variable dataset showed this cell_measures attr-scan as a top contributor to Dataset._construct_dataarray time:

ds.cf[[var]] → _getitem → _get_all_cell_measures → cell_measures
  → for da in obj.data_vars.values()  →  _construct_dataarray  (per variable)

Change

Read the cell_measures attribute off the bare Variable objects (obj.variables for a Dataset; obj.coords.variables + obj.variable for a DataArray), which carry .attrs/.encoding without constructing a DataArray. This mirrors how standard_names already reads from self._obj._variables.

Behavior is unchanged (verified against airds for both the Dataset and DataArray paths). Full test_accessor.py passes (208 passed, 1 skipped).

🤖 Generated with Claude Code