Support Python 3.15 and free-threaded 3.15t #11710

Python 3.15.0 is out, and conda-forge has both python-gil and python-freethreading builds. I tried running the test suite on both, as preparation for adding them to CI.

Python 3.15 (conda-forge)

• All test dependencies have 3.15 builds except pyarrow.
• The dask metapackage requires pyarrow, so without a lower bound the solver silently falls back to dask/distributed 2023.3.0. Using dask-core >=2026 + distributed instead gives dask 2026.8.0, and otherwise the same versions as test-py314.
• Test results: 1 failed, 22016 passed. TestDataArray::test_to_dask_dataframe needs pyarrow (via dask.dataframe) but is only guarded by requires_dask.

Free-threaded Python 3.15t (PyPI wheels)

conda-forge has no free-threaded 3.15 builds of numpy, pandas, scipy etc. yet, but PyPI has cp315t wheels for numpy, pandas, scipy, pyarrow, matplotlib, cartopy, rasterio and lxml. There are no cp315t wheels for cftime, netCDF4, h5py, numcodecs, bottleneck, numexpr and numba.

With a conda python-freethreading interpreter and PyPI wheels: 150 failed, 2 errors, 14646 passed. With PYTHON_GIL=1, 148 of these still fail, so most aren't free-threading issues:

• 137: h5netcdf is installed without h5py (no cp315t wheel), so requires_h5netcdf doesn't skip. Environment issue.
• 5: test_ensure_warnings_not_elevated. ensure_warnings in xarray/testing/assertions.py reassigns warnings.filters, which has no effect with context-aware warnings (sys.flags.context_aware_warnings, on by default in free-threaded builds). So the public xr.testing.assert_* functions elevate warnings to errors there.
• 3: test_rechunk_multi_dimensional_da_cftime is missing requires_cftime.
• 3: datatree tests assume netCDF4 is installed (NameError: name 'nc4' is not defined in a fixture, and test_default_write_engine).

Only 4 failures depend on the GIL: test_open_mfdataset_manyfiles[scipy-20-*-*-5] (file_cache_maxsize=5, parallel dask reads). CachingFileManager evicts and closes a scipy file opened with mmap while another thread still uses its data. The resulting RuntimeWarning leaves the LRU cache inconsistent and ends in a KeyError. These pass with PYTHON_GIL=1.

Upstream status

Python 3.15 (checked when there are builds on conda-forge and wheels on PyPI):

• numpy: 2.5.4 on conda-forge and PyPI
• numba: 0.68.0 on conda-forge and PyPI (Python 3.15 numba/numba#10581)
• dask: pure Python (on conda-forge use dask-core; the dask metapackage needs pyarrow)
• netCDF4: abi3 builds of 1.7.4 on conda-forge and PyPI
• h5netcdf: pure Python
• bottleneck: 1.6.0 on conda-forge; PyPI wheels need a release (support merged in TST/WHL: add support for CPython 3.15 bottleneck#589, release requested in REL: new release needed bottleneck#591)
  Check when: a bottleneck release after 1.6.0 ships cp315 wheels on PyPI.
• cftime: 1.6.5 on conda-forge, abi3 wheels of 1.6.6 on PyPI
• numcodecs: 0.17.0 on conda-forge; no PyPI wheels: Wheels for Python 3.15 and 3.15t zarr-developers/numcodecs#874 (Start testing python 3.15 zarr-developers/numcodecs#842 was closed without merging)
  Check when: a numcodecs release ships cp315 wheels on PyPI (0.17.0 doesn't).
• pyarrow: wheels of 26.0.0 on PyPI; py315 builds of 25.0.0 on conda-forge (Rebuild for python 3.15 conda-forge/pyarrow-feedstock#184)
• h5py: 3.16.0 on conda-forge; no PyPI wheels: Python 3.15 compatible wheels (limited ABI?) h5py/h5py#2972 (closed; abi3 wheels from REL: publish trailing abi3 wheels in addition to version-specific ones h5py/h5py#2885 are not released yet)
  Check when: an h5py release ships cp315 or abi3 wheels on PyPI (3.16.0 has neither).
• numexpr: 2.14.2 on conda-forge; no PyPI wheels: Python 3.15 wheels numexpr#579
  Check when: a numexpr release ships cp315 wheels on PyPI (2.14.2 doesn't).
• cf-units (needed by iris): abi3 builds of 3.3.1 on conda-forge and PyPI
• pyyaml (needed by dask): 6.0.3 on conda-forge; no PyPI wheels: Python 3.15 support yaml/pyyaml#951
  Check when: a pyyaml release ships cp315 wheels on PyPI (6.0.3 doesn't).

Free-threaded Python 3.15t (checked when there are cp315t wheels on PyPI):

• conda-forge builds: the py315 migration started in Start py315 migration conda-forge/conda-forge-pinning-feedstock#8911, but the free-threading migration only covers 3.14t so far
  Check when: conda-forge starts a python315t migration and numpy/pandas/scipy have cp315t builds.
• numpy: cp315t wheels of 2.5.4
• numba (also needed by numbagg): Free-threading use cases and test candidates numba/numba#10473 (free-threading in general); 3.15t support is announced for the next release in Python 3.15 numba/numba#10581
  Check when: a numba release ships cp315t wheels on PyPI (0.68.0 only has cp315).
• dask: pure Python
• netCDF4: cp315t wheels of 1.7.5 (Free-threaded wheels for Python 3.15t Unidata/netcdf4-python#1509)
• h5netcdf: pure Python (but needs h5py, see below)
• bottleneck: release requested in REL: new release needed bottleneck#591 (see also Py_MOD_GIL_NOT_USED declared but global state unprotected for free-threading bottleneck#532 on free-threading safety)
  Check when: a bottleneck release ships cp315t wheels on PyPI.
• cftime (also needed by nc-time-axis and iris): cp315t wheels of 1.6.6.1 (1.6.6 release did not include free-threaded wheels Unidata/cftime#421, WHL: include all default build targets (including free-threaded ones) Unidata/cftime#424)
• numcodecs (also needed by zarr): Wheels for Python 3.15 and 3.15t zarr-developers/numcodecs#874 (see also ENH: Free-Threading (No-GIL) Support for Python 3.13+ zarr-developers/numcodecs#749 for free-threading in general; 3.14t support in Add free-threaded CPython (3.14t) support zarr-developers/numcodecs#858)
  Check when: a numcodecs release ships cp315t wheels on PyPI (0.17.0 doesn't).
• pyarrow: cp315t wheels of 26.0.0
• h5py: Support for free-threading CPython h5py/h5py#2475 (free-threading support; abi3t wheels planned, see Python 3.15 compatible wheels (limited ABI?) h5py/h5py#2972)
  Check when: an h5py release ships cp315t or abi3t wheels on PyPI.
• numexpr: Python 3.15 wheels numexpr#579 (2.14.2 already has cp314t wheels, so the next release will likely include cp315t)
  Check when: a numexpr release ships cp315t wheels on PyPI.
• cf-units (needed by iris): no issue yet; 3.3.1 only has abi3 wheels
  Check when: a cf-units release ships cp315t or abi3t wheels on PyPI.
• pyyaml (needed by dask): Python 3.15 support yaml/pyyaml#951 (6.0.3 already has cp314t wheels; see also Free-threaded support blockers yaml/pyyaml#870 for free-threading in general)
  Check when: a pyyaml release ships cp315t wheels on PyPI.

Other test dependencies without 3.15t wheels on PyPI (checked 2026-10-13): psutil and tornado (distributed), google-crc32c (zarr) and lz4 (dask extras). psutil, tornado and lz4 build from source on 3.15t; google-crc32c builds but falls back to a pure Python implementation with a RuntimeWarning on import. cf-units doesn't build without udunits2, so iris has to be left out of the 3.15t env.

Proposal for CI

• Add test-py315 alongside the 3.14 envs, and only replace 3.14 once pyarrow has 3.15 builds on conda-forge. Until then, move pyarrow from the accel feature into its own feature and use dask-core in the 3.15 env.
• Add test-py315t (ubuntu only) with a conda python-freethreading interpreter and PyPI wheels, without h5netcdf until h5py has cp315t wheels.
• Fix the test guards and ensure_warnings above, and look into the file manager race.

[This is Claude Code on behalf of Michael Niklas]