can write empty parquet but not read about polars

[corwinjoy]

@adamreeve @wukan1986
I have put in a PR to fix this issue. The fix was straightforward, but I am new to this project and am hoping to get help with the associated unit tests.

[taki-mekhalfa]

Is not it sensible to keep the behavior as it? Parquet files without a schema? I think write_parquet should error when trying to write schema-less parquets.

Duckdb fails to read the file too: Error: Invalid Error: Failed to read Parquet file "test.parquet": Need at least one non-root column in the file

[wukan1986]

Hi, @taki-mekhalfa

This problem easily occurs when reading a folder, because the folder contains files generated by other software using pandas and saved daily. There is no data on the new day, but the files are generated.
However, if you read the folder directly here, an error will be reported.

[adamreeve]

Is not it sensible to keep the behavior as it? Parquet files without a schema? I think write_parquet should error when trying to write schema-less parquets.

I disagree, it is completely valid to create an empty Polars DataFrame, so why should it not be possible to round-trip this via Parquet too? I would also not describe such a file as schema-less, there is a schema and it has zero columns.

As a counter-example to DuckDB, PyArrow can read and write Parquet files with zero columns without any problem, as shown in the example in the description.

[corwinjoy]

I have started working on this. This has also been reported as part of (#9016) where @gregates helpfully gave some Rust code for reproducing this directly.