We've got a few code paths and tests that use fastparquet instead of pyarrow for serialization. This is a quite a bit of overhead and we end up directing people to pyarrow anyway because it's faster to download dataframes that way (even without the BQ Storage API, since pyarrow is faster at concatenating record buffers than pandas is at concatenating dataframes).

When we go to 2.0, I think that's probably a good time to drop fastparquet.