Make the order of encodings in column metadata deterministic

org.apache.parquet.format.converter.ParquetMetadataConverter::toFormatEncodings から開始し、Encoding 値の HashSet が返されるリストになるまでの処理を確認してください。等価な書き込みデータが、決定論的な昇順の ordinal 順序で Encoding メタデータを生成し、プロセス依存の順序を排除できれば、作業は完了です。

Describe the enhancement requested

Background
The list of encodings used for a column is stored in the column metadata in the Parquet file footer.

The elements in this list are enumeration constants (org.apache.parquet.format.Encoding), which are collected in a HashSet when the records and fields are written. Later, when the footer is written, the set of enum constants are converted to a temporary list using the HashSet’s iteration order.

Since Enum::hashCode delegates to Object::hashCode (or, in later JDK versions, to System::identityHashCode), the order of the enum constants in this list can vary between runs in different processes, and thus files with identical encodings can represent this list differently (the elements may appear in different order).

Rationale for changing this behaviour
Two processes running the same version of parquet-java and having identical writer configurations can still produce files that are different at the binary level for the exact same written data.

For redundancy reasons, it is not uncommon to write data to Parquet files on two different machines. To verify that the same data has been written on both machines, it is currently not sufficient to compare the files ate the binary level. Instead, the files must be decoded and their actual data must be compared to ensure they are equal.

If the files can be made identical at the binary level, this verification process would be simplified.

Suggested change
The list of encodings is created in org.apache.parquet.format.converter.ParquetMetadataConverter::toFormatEncodings. A simple solution is to sort this list before it is returned, i.e. always return the Encoding enum constants in ascending ordinal order.

Component(s)

Core