Parquet scan returns unvalidated Utf8 for top-level binary columns read as strings

Start with schema_coercion.rs and apply_file_schema_type_coercions, then compare the top-level path with the nested-field handling in #27067. Reproduce the invalid UTF-8 case from the issue and trace how the schema adapter casts the column. Done means reading binary data as a string returns an error for invalid UTF-8 and never yields an unvalidated string array.

Describe the bug

Reading a Parquet file with a top-level BINARY column through a table schema that declares the column as a string (Utf8, LargeUtf8 or Utf8View) produces a string array that has not been checked for valid UTF-8. That covers the binary_as_string option, and any explicit schema. When a value is not valid UTF-8:

• in a debug build, the Parquet reader panics in OffsetBuffer::into_array (parquet/src/arrow/buffer/offset_buffer.rs:139)
• in a release build, the scan returns a StringArray holding invalid UTF-8 (validate_full() fails), which is unsound for any kernel that reads the values as &str

apply_file_schema_type_coercions rewrites a binary file field to the table's string type before the Parquet reader runs. The reader validates UTF-8 only for columns annotated UTF8, which a BINARY column is not, so nothing checks the bytes. The module docs on schema_coercion.rs already note that this "hands back an unvalidated string array".

This is long-standing: 55.2.0 behaves the same. It is the top-level form of #27065. That issue covers nested fields, which regressed in 56 and are fixed in #27067 by keeping nested binary fields out of the coercion.

To Reproduce

With arrow/parquet 60.1.0 and tokio:

use std::sync::Arc;

use arrow::array::{Array, ArrayRef, BinaryArray};
use arrow::datatypes::{DataType, Field, Schema};
use arrow::record_batch::RecordBatch;
use datafusion::prelude::*;
use parquet::arrow::ArrowWriter;

#[tokio::main]
async fn main() -> datafusion::error::Result<()> {
    // A top-level binary column whose second value is not valid UTF-8.
    let values: ArrayRef = Arc::new(BinaryArray::from(vec![b"one".as_slice(), &[0xff]]));
    let batch = RecordBatch::try_from_iter([("s", values)])?;

    let path = std::env::temp_dir().join("top_level_binary.parquet");
    let mut writer = ArrowWriter::try_new(std::fs::File::create(&path)?, batch.schema(), None)?;
    writer.write(&batch)?;
    writer.close()?;

    // Read it with a table schema that declares the column as a string.
    let table_schema = Schema::new(vec![Field::new("s", DataType::Utf8, true)]);
    let ctx = SessionContext::new();
    ctx.register_parquet(
        "t",
        path.to_str().unwrap(),
        ParquetReadOptions::default().schema(&table_schema),
    )
    .await?;

    let batches = ctx.sql("SELECT s FROM t").await?.collect().await?;
    let column = batches[0].column(0);
    println!("type: {}", column.data_type());
    println!("validate_full: {:?}", column.to_data().validate_full());
    Ok(())
}

On branch-56 and on 55.2.0, a debug build panics:

called `Result::unwrap()` on an `Err` value: InvalidArgumentError("Invalid UTF8 sequence at string index 1 (3..4): invalid utf-8 sequence of 1 bytes from index 0")

A release build returns an invalid array:

type: Utf8
validate_full: Err(InvalidArgumentError("Invalid UTF8 sequence at string index 1 (3..4): invalid utf-8 sequence of 1 bytes from index 0"))

Expected behavior

Reading BINARY data as a string never produces an unvalidated string array. Invalid UTF-8 returns an error, as it already does for nested fields, which go through the validating cast.

Additional context

There are two places this could be fixed:

• In DataFusion, by not rewriting binary file fields to string types at all. The column is read as binary, and the schema adapter's cast to the table's string type validates it. Pruning should be unaffected, since the opener already moves such casts to the literal side for statistics. For Utf8View tables the file field can be read as BinaryView, so the cast only validates and does not copy.
• In the Parquet reader, by validating whenever the target type is a string type. That is being discussed in apache/arrow-rs#12047, together with an opt-in InvalidUtf8::Replace policy that would let binary_as_string users replace invalid sequences instead of getting an error.

Either way, a query over binary data that is not valid UTF-8, read as a string, would start returning an error where it currently returns an invalid array.

Willingness to contribute

I plan to contribute a fix for this bug independently