Pyarrow is sensitive to the order of the columns upon load of partitioned Files.
With the function pa.Table.from_pandas(dataframe, schema=my_schema) we can apply a schema to a dataframe. I noticed that the returned pa.Table object does use the ordering of pandas columns rather than the schema columns. Furthermore it is possible to have columns in the schema but not in the DataFrame (and hence in the resulting pa.Table).
This behaviour requires a lot of fiddling with the pandas Frame in the first place if we like to write compatible partitioned files. Hence I argue that for pa.Table.from_pandas, and any other comparable function, the schema should be the principal source for the Table structure and not the columns and the ordering in the pandas DataFrame. If I specify a schema I simply expect that the resulting Table actually has this schema.
Here is a little example. If you remove the reordering of df2 everything works fine: