> Would that keep compatibility with existing files produces by Parquet C++?
Changing the lz4 implementation to be compatible with parquet-mr/hadoop would break compatibility with any existing files that were written by Parquet C++ using lz4 compression. I believe that it is not possible to reliably detect, from inspection of the first few bytes, which implementation variant was used by the writer. But I could be misinformed, as I do not have expert knowledge of LZ4 compression.