pyarrow is an optional dependency of google-cloud-bigquery, but it's made mandatory by python-bigquery-sqlalchemy.

As pyarrow is quite large on disk — 100 MB on x86_64 Linux — I don't want to install it when it's unused in my application (AFAICT I don't need google-cloud-bigquery-storage either but that's not huge).

I suggest:

1. Removing the direct dependencies on pyarrow and google-cloud-bigquery-storage
2. Add a bqstorage extra that depends on google-cloud-bigquery[bqstorage]. That'll respect upstream's version bounds without introducing local bounds that could cause conflicts for users.
3. Document that users wanting improved performance with large result sets should install bigquery-sqlalchemy[bqstorage].

There's an existing PR at #470 but it looks like it has stalled out, so I'm filing this issue to provide a blueprint for someone who's able to do this work.