What happened?
max and min raise on an empty dimension instead of returning the fill value that mean, median and std return. On datetime64 the gap is wider: mean and median raise too, and median disagrees with timedelta64, which returns NaT from the same call.
#11549 covers quantile and describes the rest of the reductions as returning NaN. That is true for mean, median, std and var on numeric data, and not true for these.
What did you expect to happen?
NaN for numeric data and NaT for datetime and timedelta, which is what pandas returns and what the working reductions here already return.
Minimal Complete Verifiable Example
import numpy as np
import pandas as pd
import xarray as xr
empty = xr.DataArray(np.array([], dtype="f8"), dims="x")
empty.mean("x") # nan
empty.max("x") # ValueError: zero-size array to reduction operation fmax which has no identity
pd.Series([], dtype="f8").max() # nan
dates = xr.DataArray(np.array([], dtype="M8[ns]"), dims="x")
dates.mean("x") # ValueError: zero-size array to reduction operation fmin which has no identity
dates.median("x") # UFuncTypeError: ufunc 'add' cannot use operands with types dtype('<M8[ns]')
pd.Series([], dtype="M8[ns]").max() # NaT
xr.DataArray(np.array([], dtype="m8[ns]"), dims="x").median("x") # NaTFull matrix on 8ea7c88:
| empty input |
mean |
median |
max / min |
float64 |
nan |
nan |
ValueError |
int64 |
nan |
nan |
ValueError |
datetime64[ns] |
ValueError |
UFuncTypeError |
ValueError |
timedelta64[ns] |
ValueError |
NaT |
ValueError |
Two things stand out beyond the missing fill value. mean on datetime64 reports a failure in fmin, which is the internal offset subtraction showing through rather than anything the caller did. And median returns NaT for timedelta64 but raises for datetime64, so the two halves of the same path disagree.
sum and prod are not affected. They have identities and correctly return 0 and 1.
Only reducing over the empty dimension fails. xr.DataArray(np.zeros((0, 3)), dims=("x", "y")).max("y") returns shape (0,) as expected.
Environment
xarray 2026.7.1.dev43+ga48f3152c (main at 8ea7c88), numpy 2.2.6, pandas 2.3.3, python 3.12
What happened?
maxandminraise on an empty dimension instead of returning the fill value thatmean,medianandstdreturn. Ondatetime64the gap is wider:meanandmedianraise too, andmediandisagrees withtimedelta64, which returnsNaTfrom the same call.#11549 covers
quantileand describes the rest of the reductions as returningNaN. That is true formean,median,stdandvaron numeric data, and not true for these.What did you expect to happen?
NaNfor numeric data andNaTfor datetime and timedelta, which is what pandas returns and what the working reductions here already return.Minimal Complete Verifiable Example
Full matrix on
8ea7c88:float64nannanint64nannandatetime64[ns]timedelta64[ns]NaTTwo things stand out beyond the missing fill value.
meanondatetime64reports a failure infmin, which is the internal offset subtraction showing through rather than anything the caller did. AndmedianreturnsNaTfortimedelta64but raises fordatetime64, so the two halves of the same path disagree.sumandprodare not affected. They have identities and correctly return 0 and 1.Only reducing over the empty dimension fails.
xr.DataArray(np.zeros((0, 3)), dims=("x", "y")).max("y")returns shape(0,)as expected.Environment
xarray 2026.7.1.dev43+ga48f3152c (main at 8ea7c88), numpy 2.2.6, pandas 2.3.3, python 3.12