memory bug with Dataset starting in version 2025.3.0 in combination with dask

What happened?

Hi,
I think I found a memory bug that happens when using xarray from version 2025.3.0 when also dask in any version is present. The memory of the very fist Dataset created is never released. For all later created Datasets it works and a workaround for me is in fact to initialize a small Dataset at the beginning of the code.

What did you expect to happen?

a deleted Dataset should release the memory as in xarray version 2025.1.2 or older.

Minimal Complete Verifiable Example

import xarray as xr
import numpy as np

# starting with defining a tiny Dataset would mitigate the problem, as only the very first Dataset is never released from memory
# xr.Dataset({}, coords={ "a": [1]})

def dummy(n):
    ds = xr.Dataset(        {
            "A": (["x", "y"], np.random.randn(n, n))
        },
        coords={
            "x": range(n),
            "y": range(n),
        },
        )


dummy(25000)
input("Check your memory usage now... ~4GB is not released")

# Dockerfile to reproduce
# FROM python:3.11.5-slim-bullseye
# RUN pip install xarray==2025.1.2 dask==2025.5.1
# Using xarray==2025.1.2 or lower shows correct behavior

MVCE confirmation

• Minimal example — the example is as focused as reasonably possible to demonstrate the underlying issue in xarray.
• Complete example — the example is self-contained, including all data and the text of any traceback.
• Verifiable example — the example copy & pastes into an IPython prompt or Binder notebook, returning the result.
• New issue — a search of GitHub Issues suggests this is not a duplicate.
• Recent environment — the issue occurs with the latest version of xarray and its dependencies.

Relevant log output

using docker stats to monitor memory usage (while waiting for user prompt):
MEM USAGE 4.728GiB

When using xarray version 2025.1.2:
MEM USAGE: 82.77MiB

Anything else we need to know?

initializing a tiny Datasets at the top of the code mitigates the problem
xr.Dataset({}, coords={ "a": [1]})

funnily, even calling xr.show_versions() does....

Feels like the very first call to Dataset leaves a reference somewhere, so it is not picked up by the garbage collector.

Might be related to: https://github.com/pydata/xarray/issues/9807
But here we have a much simpler minimal example.

Environment

Start by running the supplied MVCE with xarray 2025.4.0 and dask 2025.5.1, then compare it with xarray 2025.1.2 while monitoring memory. Investigate the first xr.Dataset creation and verify that deleting it releases memory without requiring a preliminary Dataset or xr.show_versions() call.