Currently pyarrow reads empty parts as float64, even if the underlying columns have other dtypes. This can cause problems for pandas downstream, as certain operations are only valid on certain dtypes, even if the columns are empty.
Copying the comment Uwe over:
This is the expected behaviour as an empty string column in Pandas is simply an empty column of type object. Sadly object does not tell us much about the type of the column at all. We return numpy.float64 in this case as it's the most efficient type to store nulls in Pandas.
This seems unintuitive at best to me. An empty object column in pandas is treated differently in many operations than an empty float64 column (str accessor is available, excluded from numeric operations, etc..). Having an empty file read in as a different dtype than was written could lead to errors in processing code downstream. Would arrow be willing to change this behavior?
We should probably use another method than `field.type.to_pandas_dtype()` in this case. The column saved in Parquet should be saved with `NA` as type which sadly does not provide enough information.
We also store the original dtype in the Pandas metadata that is used for the actual DataFrame reconstruction later on. If we would also pick up the metadata when it was written, we should be able to correctly reconstruct the dtype.