Skip to content

ParquetSink: restore ability to provide additional user metadata into the encoded parquet file. #10223

@wiedld

Description

@wiedld

Describe the bug

We add our own metadata to the parquet file. Currently, we do so using the WriterProperties' kv_metadata and the ArrowWriter. We want to start performing parquet writes with datafusion's ParquetSink, however a recent change has removed this ability to add our own metadata.

There was a change to unify the different writer options across sink types, specifically to make COPY TO and create external table have a uniform configuration. Users can now specify the configuration at the SQL level API (e.g. COPY <src> TO <sink> (<config_options>)). This was a good high level change; however, a side effect of the implementation was the removal of the ability to add our own metadata.

The current implementation (after the above change) now derives the writer properties from the TableParquetOptions. This conversion always sets the sorting_columns and user-defined kv_metadata as None, as demonstrated in the first commit of the fix.

To Reproduce

The hardcoded setting of the user metadata to None is demonstrated in this commit.

Expected behavior

The expected behavior is to be able to set our own metadata. Ideally, to have user-inserted metadata as an option at the SQL level API.

COPY source_table TO 'sink' STORED AS PARQUET OPTIONS ('format.metadata' 'key:value')

The expected outcome is demonstrated in this commit.

Additional context

No response

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions