Context

Python 3.14t introduces two substantial changes that impact Dask:

• contextvars are propagated from the parent thread to new threads that are being spawned. When it comes to thread pools, however, it means that a worker thread may be already running when the user calls submit, and thus won't get the change in context.
• warnings.catch_warnings now is backed by contextvars, whereas previously was process-global. This also impacts @pytest.mark.filterwarnings.

This change in CPython was necessary to allow catching warnings in threads without interfering with other threads.
It's worth noting that, for Dask, this behaviour is relevant also in GIL-enabled builds (think about divisions by zero in numpy). On Python 3.14t, this behaviour is on by default. On 3.14, it is opt-in with

python -X thread_inherit_context=1 -X context_aware_warnings=1

or with equivalent environment variables.
See upstream documentation:

• https://docs.python.org/3/using/cmdline.html#envvar-PYTHON_THREAD_INHERIT_CONTEXT
• https://docs.python.org/3/using/cmdline.html#envvar-PYTHON_CONTEXT_AWARE_WARNINGS

Impact

The direct consequence for most Dask users is that catching warnings on the threaded scheduler becomes erratic on 3.14t, because dask.threaded keeps thread pools warm across calls to get/compute/persist.

Consider this:

import warnings
import dask.array as da

a = da.zeros(1)
a.compute()

b = da.zeros(1) / da.zeros(1)
with warnings.catch_warnings(category=RuntimeWarning, action='ignore'):
    b.compute()

the above fails to catch the warning on 3.14t. If you comment out a.compute(), however, it will work again.

Same will happen here:

import dask.array as da
import pytest


def test_1():
    a = da.zeros(1)
    a.compute()


@pytest.mark.filterwarnings("ignore:invalid value encountered in divide:RuntimeWarning")
def test_2():
    b = da.zeros(1) / da.zeros(1)
    b.compute()

pytest will fail with RuntimeWarning. pytest -k test_2 will pass.
In-depth discussion here: pytest-dev/pytest#14077

This PR

• scheduler="threads" (which is the default for dask.array and dask.dataframe when there is no distributed.Client) will now capture the contextvars on the thread calling get/compute/persist and replicate them in the worker threads. This is on all Python versions and regardless of thread pool warmup.
• Fix regression on Python 3.14t where warnings.catch_warnings and @pytest.mark.filterwarnings would not catch anything if the thread pool was already warm (e.g. if another test already called get/compute/persist).