This repo exists mostly to make releases of parquet-mr more often. The only difference to upstream is solution for PARQUET-686.
The change that we had made that upstream only made in statistics v2 is to change binary comparison to be unsigned and declare all statistics priori to that change as corrupted. This lets us more quickly take advantage of binary statistics and removes burden on user to know whether they should account for signed binary comparison in their values.
As a result statistics on binary columns produced by this fork and mainline will be different. This fork will follow lexicographical ordering while upstream won't. It's important when reading those with upstream version to ignore the statistics as it can lead to wrong results.
Most of the diff is contained in parquet-hadoop/src/main/java/org/apache/parquet/format/converter/ParquetMetadataConverter.java which should be carefully reconciled with upstream on merges. We ignore parquet.strings.signed-min-max.enabled option and always return statistics if parquet-column/src/main/java/org/apache/parquet/CorruptStatistics.java reports them as not corrupted.