Field::try_canonical_extension_type has unexpected behavior from external writers

[limenilbuz]

I created a sample Parquet file with pyarrow that contained Uuid columns.

Here is a test script:

use std::fs::File;
use parquet::arrow::arrow_reader::ParquetRecordBatchReaderBuilder;

fn main() {
    let file = File::open("uuids.parquet").unwrap();
    let reader = ParquetRecordBatchReaderBuilder::try_new(file).unwrap();
    let schema = reader.schema();
    let field = schema.fields.get(0).unwrap();
    println!("{:?}", field);
    match field.try_canonical_extension_type() {
        Ok(extension_type) => println!("I am of extension type: {:?}", extension_type),
        _ => println!("I am NOT an extension type")
    }
}

and the output of cargo run:

$ cargo run
Field { name: "uuids", data_type: FixedSizeBinary(16), metadata: {"ARROW:extension:metadata": "", "ARROW:extension:name": "arrow.uuid"} }
I am NOT an extension type

This is because in the C++ definition of UuidType, the SerDe methods both return and expect the empty string. This causes issues in the Rust SerDe methods for Uuid, which return and expect Option::None.

I found this comment in the original commit for the extension types. The simplest fix is just to accept the empty string as valid metadata.

[limenilbuz]

Ah, maybe not so big an issue after all. If you tell the reader to skip the arrow_metadata with ArrowReaderOptions::new().with_skip_arrow_metadata(true), the Parquet -> Arrow schema conversion step looks at the Parquet logical type instead of the hint from the serialized arrow schema to construct the Uuid extension type with no metadata. This causes Field::try_canonical_extension_type to do the "correct thing".