Just pushed an empty commit to test against the main branch of fastparquet (given dask/fastparquet#738 has been merged). Hope that's okay

Yep. I thought that restarting the tests would have the same effect, but maybe not.

This does not seem to be picking up fastparquet main in every run. Do I need to make a release? The (parquet) tests do all pass for me locally.

Only the upstream test will pick up fastparquet from main. So yeah I think a release would be necessary to get the regular tests passing.

OK; well upstream seems to be failing on loads of dataframe tests (from warnings?). They include some parquet, but I cannot tell if these are actually parquet failures or due to whatever broke elsewhere.

Comparing the failing test summary between the linux Python 3.9 build (which is using pandas=1.4.0) and the upstream build, the changes in this PR + the fastparquet fixes result in dask/dataframe/io/tests/test_parquet.py::test_roundtrip[fastparquet*] and dask/dataframe/io/tests/test_parquet.py::test_timestamp96 passing. #8623 should address the other failures on pandas=1.4.0

OK, so we merge and I'll release fastparquet? There are a couple of PRs waiting there that would be nice to get in first.

A fastparquet release would be great. Do you have a rough estimate on how long that might take? It would be good to have the Python 3.9 builds pass here, so if it will still be a bit before those other PRs can get in, then we might consider using the dev version of fastparquet in our 3.9 builds

Coming... dask/fastparquet#739

Woo 🎉

Ok I am going to merge this and then we should be all green???? :)

Thank you @martindurant for the quick turn around on this!

Hope! Note that conda-forge did it's job, which is what matters here, but some pypi wheels didn't build. I'll track those down later.

Thanks @martindurant @jsignell!