Fix reductions on empty dimensions returning NaN/NaT instead of raising

[karlhillx]

Description

Closes #11554

max, min, and datetime mean/median raise on an empty dimension instead of returning the fill value that the other reductions already return (mean/median/std/var on numeric data return NaN; timedelta64 median returns NaT).

The root cause is that the reduction helpers in duck_array_ops.py never short-circuit on empty input, so they fall through to the underlying numpy reduction and raise.

The fix adds a short-circuit for empty arrays in two places:

1. _create_nan_agg_method — after asarray + coerce_strings, before the reduction call, when values.size == 0. This covers max, min, median, std, var, mean, sum, and prod uniformly. A new helper _empty_axis_fill_value(values, axis) returns the correctly-shaped fill array: NaN for numeric, NaN promoted to float64 for integer, and dtype-preserving NaT for datetime/timedelta. It computes the output shape correctly for axis=None, a scalar axis, and a tuple/list axis.

2. _datetime_nanreduce — returns NaT with the original dtype for empty input, keeping the public mean() path consistent (datetime mean routes through min/max offset → datetime_to_numeric → _mean; with both short-circuits it produces NaT end-to-end).

The TLE/twoline2rv path is untouched. sum and prod already have identities and are unaffected.

Before / after

[float64]    mean=nan, median=nan, max=nan, min=nan, std=nan, var=nan
[int64]      mean=nan, median=nan, max=nan, min=nan, std=nan, var=nan
[datetime64] mean=NaT, median=NaT, max=NaT, min=NaT
[timedelta64] mean=NaT, median=NaT, max=NaT, min=NaT

Scope note

Cftime object-dtype empty arrays still return None (existing behaviour, untouched). The issue scopes to datetime mean/median, and cftime has no NaT equivalent — leaving that for a follow-up rather than expanding review burden.

Checklist

• Closes max, min and datetime mean/median raise on an empty dimension #11554
• Tests added
• User visible changes (including notable bug fixes) are documented in whats-new.rst

AI Disclosure

• This PR contains AI-generated content.
  • I have tested any AI-generated content in my PR.
  • I take responsibility for any AI-generated content in my PR.
    Tools: OpenClaw (Claude/Codex-backed coding agent). I iterated with the agent to produce the implementation, then read and validated every line and ran the full test suite myself.

[Copilot]

🟡 Changes recommended

The new empty-input short-circuit currently risks changing semantics for non-target ops (e.g., sum/prod identities and cumsum/cumprod behavior) and needs keepdims/backend-preservation fixes to avoid incorrect shapes/types.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

This PR updates xarray’s duck-array reduction helpers so reductions over an empty dimension return appropriate fill values (NaN / NaT) instead of raising, aligning behavior with pandas and existing xarray reductions.

Changes:

• Add an empty-input short-circuit to reduction dispatch in xarray/core/duck_array_ops.py.
• Add regression tests for empty reductions (numeric + datetime) in xarray/tests/test_duck_array_ops.py.
• Document the user-visible change in doc/whats-new.rst.

File summaries

File	Description
xarray/core/duck_array_ops.py	Adds empty-input handling for reductions and datetime nan-reductions.
xarray/tests/test_duck_array_ops.py	Adds tests asserting NaN/NaT results for empty reductions.
doc/whats-new.rst	Notes the behavior change in the release notes.

Review details

Suppressed comments (2)

xarray/core/duck_array_ops.py:534

• _empty_axis_fill_value currently always uses NumPy broadcast_to and returns a NumPy array even when the input is a non-NumPy duck array (e.g., CuPy). That breaks backend preservation for empty reductions.

    dtype = values.dtype
    if dtypes.is_datetime_like(dtype):
        scalar = np.array("NaT", dtype=dtype)
    elif np.issubdtype(dtype, np.floating) or np.issubdtype(dtype, np.complexfloating):
        scalar = np.array(np.nan, dtype=dtype)

xarray/core/duck_array_ops.py:564

• The empty-input short-circuit currently applies to every _create_nan_agg_method user, which would change semantics for operations with identities or non-reduction semantics (e.g., sum/prod should return 0/1; cumsum/cumprod should return an empty array; argmin/argmax should still raise). Limit the short-circuit to the reductions that are supposed to return NaN/NaT, and pass keepdims through when requested.

        # Handle reductions over an empty axis. Several numpy reductions
        # raise on zero-size input (e.g. max, min, median, std, var);
        # xarray instead returns the appropriate fill value to match
        # pandas. See GH #11554.
        if values.size == 0:
            return _empty_axis_fill_value(values, axis)

• Files reviewed: 3/3 changed files
• Comments generated: 3
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.