Why not provide the data in Parquet format?
#11
by
albertvillanova
HF staff
- opened
Currently, your data is offered in JSON-Lines format, compressed in ZIP archives.
However, adopting the Parquet format could offer several advantages, such as:
- Efficient storage: Parquet is a columnar storage format that applies compression automatically, leading to smaller file sizes compared to JSON-Lines.
- Embedded schema: Parquet files include metadata with the data schema, simplifying schema management during processing.
- Interoperability: Parquet is natively supported by a wide range of data processing frameworks and analytics tools, such as Apache Spark, Presto, DuckDB, which makes it more versatile for downstream analysis and processing.
- Faster Query Performance: Since Parquet is a columnar format, it allows for efficient reading of specific columns without scanning the entire file. This improves query performance, especially in large datasets.
- Better Compression Ratios: Parquet supports multiple compression algorithms (e.g., Snappy, Gzip, Brotli), and its columnar format often yields better compression ratios than row-based formats like JSON-Lines, reducing storage costs.
- Efficient Data Skipping: Parquet files store statistics for each column, such as min/max values, enabling data skipping during queries, which further improves performance by reducing the amount of data that needs to be read.
- Schema Evolution: Parquet supports schema evolution, allowing fields to be added, removed, or changed without breaking backward compatibility, making it easier to adapt to evolving data models.
We could use this thread to get more feedback from the community.
We’ve kept the format of this beta dataset similar to what we offer on our platform (tailored towards more use cases), while we’re gathering initial feedback, like this one 😀 . Thanks for outlining the advantages of parquet and opening this thread for comments!