Parquet-java sometimes produces 0-size compressed data in data page v2

[mapleFU]

Describe the bug

See:

1. GH-31992: [C++][Parquet] Handling the special case when DataPageV2 values buffer is empty arrow#45252 (comment)
2. File: https://github.com/user-attachments/files/18292070/snappy_bug.parquet.gz

V2 data pages do not compress def/rep levels, only the values. In the (not uncommon) case where all data is null, Parquet-java may write 0 compressed bytes for a 0-size decompressed data. However, a 0-size compressed buffer is not a valid input for compressors. The data page then fails decompressing in the C++ and Rust implementations.

Component(s)

Core

[pitrou]

It's not an enhancement, it's a bug in Parquet-java.

[ConeyLiu]

Let me take a look at this. What the compressed bytes size is written out on the C++ side for this case?

[pitrou]

I could try directly with Arrow C++, but Python tells me it should be a single 0 byte for Snappy:

>>> snappy.compress(b"")
b'\x00'

And, of course, that depends on the compression algorithm. See for example Zstd:

>>> zstd.compress(b"")
b'(\xb5/\xfd \x00\x01\x00\x00'
>>> zstd.compress(b"").hex()
'28b52ffd2000010000'

[pitrou]

But I think that's irrelevant anyway. What you should do if there are 0 uncompressed bytes is simply to set is_compressed to false in the DataPageV2 header, and write 0 uncompressed bytes:
https://github.com/apache/parquet-format/blob/a498aa9a377edcdbc5da802cf9f1763a2e409411/src/main/thrift/parquet.thrift#L664-L669

[ConeyLiu]

@pitrou, thanks for the background information and suggestions you provided. I will fix it after the Chinese New Year holiday.