Or maybe not an actual regression...it's a ~5e-7 difference and one of the things the patch does (at around dask/dataframe/tests/test_rolling.py:270) is _tighten_ the tolerance on that test. I have filed a separate bug (#972516) for the fsspec issues.