Parquet metadata written can be corrupt due to Buffer comparison for DECIMAL type

[MatthewJin-at]

Steps to reproduce

This program reproduces it on 1.7.0:

import {ParquetSchema, ParquetWriter} from '@dsnp/parquetjs';

export async function writeCorruptParquetFile() {
    const parquetWriter = await ParquetWriter.openFile(
        new ParquetSchema({
            num: {type: 'DECIMAL' as const, precision: 20},
        }),
        'corrupt.parquet',
    );

    // These are the numbers for 10^15 + 127 through 10^15 + 129
    const buffer127 = Buffer.from([0x00, 0x00, 0x03, 0x8d, 0x7e, 0xa4, 0xc6, 0x80, 0x7f]);
    const buffer128 = Buffer.from([0x00, 0x00, 0x03, 0x8d, 0x7e, 0xa4, 0xc6, 0x80, 0x80]);
    const buffer129 = Buffer.from([0x00, 0x00, 0x03, 0x8d, 0x7e, 0xa4, 0xc6, 0x80, 0x81]);

    await parquetWriter.appendRow({
        num: buffer127,
    });
    await parquetWriter.appendRow({
        num: buffer128,
    });
    await parquetWriter.appendRow({
        num: buffer129,
    });

    await parquetWriter.close();
}

Expected behaviour

The file should have 3 numbers with accurate rowgroup metadata.

Actual behaviour

This writes a file with 3 numbers but corrupt metadata - the max_value of the row group is the 128 value instead of the 129 value.

➜  parquet meta corrupt.parquet                                                   [12:39:23]

File path:  corrupt.parquet
Created by: @dsnp/parquetjs
Properties: (none)
Schema:
message root {
  required binary num (DECIMAL(20,0));
}


Row group 0:  count: 3  37.00 B records  start: 4  total(compressed): 111 B total(uncompressed):186 B
--------------------------------------------------------------------------------
     type      encodings count     avg size   nulls   min / max
num  BINARY    _ BB_     3         37.00 B    0       "1000000000000127" / "1000000000000128"

➜  parquet check-stats corrupt.parquet                                            [12:39:30]
corrupt.parquet has corrupt stats: Max should be >= all values.

Any logs, error output, etc?

n/a

Any other comments?

I think this is because the library uses < and > instead of Buffer.compare() to determine the statistics min_value/max_values.