Fix quantile raising on an empty result

[chiruu12]

Closes #11549.

_wrapper builds the empty result itself instead of moving a q axis that is not there. Both backing functions are unusable on this input: np.nanquantile returns the nan but drops the leading q axis, so moveaxis hands apply_ufunc the wrong rank, and np.quantile raises before returning anything.

The trigger is npa.size == 0, not an empty reduced axis. DataArray(zeros((0, 3))).quantile(0.5, dim="y") reduces over y, which has length 3, and is still empty because x is.

	before	after
empty.quantile(0.5)	AxisError	nan
empty.quantile(0.5, skipna=False)	IndexError	nan
zeros((0,3)).quantile(0.5, dim="y")	ValueError	shape (0,)
empty datetime64 / timedelta64	IndexError	NaT, dtype kept

Datetimes keep their own dtype and NA. That matches the non-empty path, median, and groupby_bins(...).quantile() over an empty bin. Returning float64 nan instead made concat of an empty and a non-empty result fail to promote.

Empty input is validated exactly as non-empty input is, by running the call over one element first so numpy owns the message. Without that, q=1.5 raised on a 5-element array and returned nan on a 0-element one.

skipna=False returns nan too, so both paths agree. That differs from numpy, where np.quantile([], 0.5) raises, but it is what median(skipna=False) and mean(skipna=False) already do on an empty dimension.

Known gap: a pandas extension dtype the non-empty path rejects (string, category, interval, period, boolean) is still accepted when empty, as it already is for object arrays. numpy cannot build a probe array from those dtypes, so only q and method are checked there.

Tests at Variable, DataArray and Dataset level cover the empty reduction, the empty-but-not-reduced dimension, datetime dtype retention, extension dtypes, and the argument validation. Against origin/main, 24 fail and 3 pass; the 3 are guards, two of them against a regression this PR introduced and then fixed.

test_variable, test_dataarray, test_dataset, test_groupby, test_weighted, test_computation, test_duck_array_ops, test_units, test_concat: 2961 passed. ruff check and format clean.

I could not exercise the dask path; dask is not installed in my environment.

[Copilot]

🟡 Changes recommended

The new datetime/timedelta empty-result behavior can conflict with dask gufunc metadata because output_dtypes is still hard-coded to float64, which can break chunked execution.

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

Pull request overview

Fixes quantile on empty results so it returns missing values (rather than raising) and preserves datetime/timedelta dtype semantics, aligning behavior with other reductions and pandas.

Changes:

• Add an empty-result fast path to Variable.quantile() that validates arguments and returns correctly shaped fill values (NaN / NaT).
• Add targeted tests for empty inputs across Variable, DataArray, and Dataset quantile reductions.
• Document the behavior change in the “Bug Fixes” section of the release notes.

File summaries

File	Description
xarray/core/variable.py	Implements the empty-result handling in Variable.quantile() and preserves datetime/timedelta NA semantics.
xarray/tests/test_variable.py	Adds comprehensive tests for empty-quantile behavior, argument validation, dtype retention, and extension dtypes.
xarray/tests/test_dataarray.py	Adds a DataArray.quantile() regression test for empty inputs.
xarray/tests/test_dataset.py	Adds a Dataset.quantile() regression test for empty inputs.
doc/whats-new.rst	Notes the bug fix and the new empty-result behavior in release notes.

Review details

• Files reviewed: 5/5 changed files
• Comments generated: 1
• Review effort level: Lite

---

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